Formulation Simplified

Formulation Simplified
Formulation Simplified
Finding the Sweet Spot through Design and
Analysis of Experiments with Mixtures
Mark J. Anderson
Patrick J. Whitcomb
Martin A. Bezener
A PRODUCTIVITY PRESS BOOK

Routledge
Taylor & Francis Group
711 Third Avenue, New York, NY 10017
© 2018 by Mark J. Anderson, Patrick J. Whitcomb, Martin A. Bezener
Productivity Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Printed on acid-free paper
International Standard Book Number-13: 978-1-1380-5604-6 (Hardback)

International Standard Book Number-13: 978-1-1380-5601-5 (Paperback)
International Standard Book Number-13: 978-1-3151-6557-8 (eBook)
This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made
to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all
materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all
material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been
obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future
reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in
any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, micro-
filming, and recording, or in any information storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.
copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-
8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that
have been granted a photocopy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identi-
fication and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the Productivity Press site at

www.productivitypress.com
Contents
Preface........................................................................................................ix
Acknowledgments .................................................................................... xi
Authors .....................................................................................................xiii
Introduction ............................................................................................. xv
1 Getting Your Toe into Mixtures ...................................................1

All That Glitters Is Not Gold.......................................................................3
Generating a Beautiful Response Surface—Like a String of Rubies
on a Gold Strand! ........................................................................................7
Details on Modeling the Performance of a Two-Component Mixture........11
Practice Problems ......................................................................................14
Problem 1.1 ............................................................................................ 15
Problem 1.2............................................................................................17
Appendix 1A: Cubic Equations for Mixture Modeling (and Beyond) .....21
Appendix 1B: Statistical Details on How to Order Up Just the Right
Model .........................................................................................................25
Testing for Lack of Fit ...........................................................................26
Model Summary Statistics .....................................................................29
2 Triangulating Your Region of Formulation ...............................31
The Simplex Centroid Design ...................................................................33
The Black and Blue Moon Beer Cocktail ................................................35
Diving Under the Response Surface to Detail the Underlying
Predictive Model .......................................................................................41
Taking Cost into Account .........................................................................44
Do Not Put a Square Peg into a Triangular Hole ....................................46
v
vi ◾ Contents

Problem 2.1............................................................................................48
Problem 2.2 ...........................................................................................49
Appendix 2A: The Special Cubic (and Advice on Interpreting
Coefficients) ...............................................................................................50
3 Simplex Lattice Designs to Any Degree You Like .....................55
Working with Four Components in Tetrahedral Space ...........................55
Building a Simplex Lattice Design ............................................................57
Augmented Simplex Lattice: When in Doubt, Build Them Stout ............60
Using Augmented Simplex Lattice Mixture Design to Optimize
an Extra-Virgin Olive Oil ..........................................................................62
Problem 3.1............................................................................................71
Problem 3.2 ...........................................................................................71
4 Mixture Constraints That Keep Recipes Reasonable .................73
Setting Minimum Constraints ...................................................................73
Expanding the Constrained Space via Mathematical Coding..................75
Why It Was Worth Reading This Chapter and What’s in It for You
as a Formulator .........................................................................................79
Practice Problem .......................................................................................80
Problem 4.1............................................................................................80
Appendix 4A: Upper (“U”) Pseudo Coding to Invert Mixture Space......82
5 Optimal Design to Customize Your Experiment .......................87
Extreme Vertices Design: Shampoo Experiment .....................................89
Optimal Designs Customize Your Experiment as You See Fit for
Any Feasible Region..................................................................................93
Take This Tableting Case-Study Now and Call the Statistician
in the Morning ..........................................................................................94
Optimal Design Simplified .......................................................................95
Take-Home Advice on Deploying Optimal Design .................................97
Practice Problem .......................................................................................98
Problem 5.1 ............................................................................................98
Appendix 5A: An Algorithm for Finding Vertices ....................................99
6 Getting Crafty with Multicomponent Constraints ................... 101
How Multicomponent Constraints Differ from Simple Ones ................ 101
MCCs Made as Easy as Making a Pound Cake ......................................104
Contents ◾ vii
Ratio Constraints ..................................................................................... 107

Appendix 6A: Combining Components .................................................108
7 Multiple Response Optimization Hits the Spot .......................113
Desirability Simplified ............................................................................. 114
Framing the Sweet Spot and Narrowing It Down to Achieve
Quality by Design (QbD) ........................................................................ 119
Practice Problem .....................................................................................124
Problem 7.1 ..........................................................................................124
8 Screening for Vital Components ..............................................127
High-Octane Simplex Screening Designs ..............................................128
Measuring the Effect of a Component ...................................................133
Extreme Vertices Design for Non-Simplex Screening ............................134
Appendix 8A: Trace Plots—Cox versus Piepel Direction ......................138
9 Working Amounts, Categorical and Process Factors
into the Mix .............................................................................143
Mixture-Amount Experiments—Not Just the Composition but
How Much of It ....................................................................................... 143
Contending with Categorical Variables .................................................. 149
Practice Problem ..................................................................................... 152
Problem 9.1 .......................................................................................... 152
Appendix 9A: Alternatives for Modeling Results from
Combined Designs .................................................................................. 152
10 Blocking and Splitting Designs for Ease of Experimentation.....155
Blocking to Remove Known Sources of Variation ................................. 155
Split Plots to Handle Hard-to-Change Factors or Components
in Combined Designs .............................................................................160
A Case Where the Process, Rather Than the Mixture, Is Hard
to Change ................................................................................................164
11 Practical Magic for Making the Most of a Mixture ..................165
The Failure of Fillers and Perils of Parts ................................................ 165
Strategy for Formulation Experimentation ............................................. 171
References ...................................................................................... 175

About the Software ........................................................................ 179
Index .............................................................................................. 181
Preface
All that is gold does not glitter; Not all those who wander are lost.
—J. R. R. Tolkien (The Fellowship of the Ring)
This book rounds out our series of “Simplified” books (Anderson and
Whitcomb, 2015, 2016) into a trilogy on the design of experiments (DOE).
It may not achieve the stature of Tolkien’s towering trio—The Lord of the
Rings, but the detailing of mixture design completes our quest to provide
the statistical tools needed by modern-day industrial experimenters. The
beneficiaries of this third “Simplified” book will be formulators of alloys,
beverages, chemicals, cosmetics, construction materials (such as concrete),
food, flavors, pharmaceuticals, paints, plastics, pulp, paper rubber, textiles,
and so forth, that is, any product made from stuff.
Formulation Simplified is derived from a popular workshop on mixture
design that my coauthor, Pat, developed over twenty years ago. He’s worked
unstintingly to continuously incorporate new statistical methods that prove
to be of practical use. More recently, statistician, Martin Bezener, joined our
team at Stat-Ease and took to mixtures like a barista to coffee. However, it’s
one thing to be trained intensively by expert instructors like Pat or Martin, but
another thing to learn on your own from a book. That’s where I come in by
making these powerful statistical tools of experimental design and analysis as
unintimidating as possible in a self-study, written format. Luckily, I can rely on
Pat and Martin to bolster any inadequate mathematical details, thus helping us
maintain statistical rigor throughout. If we feel that this may create too much
information (TMI) for some readers, the in-depth explanations go into sidebars
or appendices that can be glossed over (at least on the first go-through!).
What differentiates Formulation Simplified from the standard statisti-
cal texts on mixture design by Cornell (2002) and Smith (2005) is that we
make things relatively easy and fun to read. To convey my experience that
ix
x ◾ Preface
“experimental design can lead to feelings of pure joy, especially when it

points in the direction of improvement” (Lynne Hare), I provide a whole
new collection of delightful original studies that illustrate the essentials of
mixture design and analysis. Solid industrial examples are offered as prob-
lems at the end of many chapters for those who are serious about trying
new tools on their own. Statistical software to do the computations can be
freely accessed via a website developed in support of this book. There you
will also find answers posted. See About the Software for instructions.
Early in my career as a chemical engineer working in process develop-
ment, I discovered that many of the chemists in our research and devel-
opment center—especially those most brilliant in their field—failed to
appreciate the power of planned experimentation. Furthermore, they
disliked the mathematical aspects of statistical analysis. To top off their
DOE phobia, these otherwise competent chemists also dismissed predictive
models based only on empirical data. Ironically, in the hands of subject
matter experts like these elite chemists, the statistical methods of mixture
design and analysis provide the means for rapidly converging optimal
compositions—the “sweet spot,” as I like to call it. That is the true gold
that awaits those who are willing to take a new path.
Mark J. Anderson
Acknowledgments
Pat, Martin, and I are indebted to several contributors to the development

of mixture-design methods, particularly, Wendell Smith and Greg Piepel.
However, we must especially acknowledge John Cornell. His landmark
book, Experiments with Mixtures (Wiley, 1981) served as our template. With
John’s passing in 2016, our community of industrial statisticians lost a bright
light—a fellow with unbounded enthusiasm for the design of experiments
(DOE) geared to formulators. We dedicate this book to John with the hope
that it carries his torch for mixture design forward.
Mark J. Anderson
xi
Authors
Mark J. Anderson, PE, CQE, MBA, is a principal and general manager

of Stat-Ease, Inc. (Minneapolis, Minnesota). He is a chemical engineer by
profession, who also has a diverse array of experience in process development
(earning a patent), quality assurance, marketing, purchasing, and general
management. Before joining Stat-Ease, he spearheaded an award-winning
quality improvement program (generating millions of dollars in profit for an
international manufacturer) and served as general manager for a medical
device manufacturer. His other achievements include an extensive portfolio
of published articles on the design of experiments (DOE). Anderson authored
(with Whitcomb) DOE Simplified: Practical Tools for Effective Experimentation,
3rd Edition (Productivity Press, 2015) and RSM Simplified: Optimizing
Processes Using Response Surface Methods for Design of Experiments, 2nd
Edition (Productivity Press, 2016).
Patrick J. Whitcomb, MS, is the founding principal of Stat-Ease, Inc.

Before starting his own business, he worked as a chemical engineer, quality
assurance manager, and plant manager. Whitcomb developed Design-Ease®
software, an easy-to-use program for design of two-level and general facto-
rial experiments, and Design-Expert® software, an advanced user’s program
for response surface, mixture, and combined designs. He has provided
consulting and training on the application of design of experiments (DOE)
and other statistical methods for decades. In 2013, the Minnesota Federation
of Engineering, Science, and Technology Societies (MFESTS) awarded
Whitcomb the Charles W. Britzius Distinguished Engineer Award for his
lifetime achievements.
xiii
xiv ◾ Authors
Martin A. Bezener, PhD, is a principal and statistician with Stat-Ease,

Inc. He did his graduate studies at the University of Minnesota, Twin Cities.
There he spent a year at the Statistical Consulting Center of the School of
Statistics working on a wide variety of projects with university researchers.
He also taught undergraduate-level statistics for several years. In addition to
his role as a consultant on DOE, Martin takes point at Stat-Ease for researching
new methodology and developing algorithms for coding into the publisher’s
software.
Introduction
There are many paths to enlightenment. Be sure to take one with

a heart.
—Lao Tzu
To avoid disheartenment, this book swoops down from above the forest to
treetop level, where it remains for the most part. Go ahead and enjoy the
ride for the first pass through Formulation Simplified—it will be a far easier
read than any other statistical textbook you are likely to see, especially, if
you skip the formula-laden appendices and the serious sidebars (read the
trivial ones just for fun).
A MODEL FOR PREDICTING WHAT FORMULAS

DO TO GENERAL READERSHIP
Mathematics Professor Roger Penrose prefaces his book The Emperor’s
New Mind (Oxford University Press, 1989) with this note to readers
(excerpted): “At a number of places I have resorted to the use of math-
ematical formulae,…unheeding of warnings…that each such formula
will cut the general readership by half…I recommend [you] ignore that
line completely and skip over to the next actual line of text!…If armed
with new confidence, one may return to that neglected formula and try
to pick out some salient features. The text itself may be helpful in letting
one know what is important and what can be safely ignored about it.”
Great advice!
(Kudos to colleague Neal Vaughn for alerting us about this rule of
thumb!)
xv
xvi ◾ Introduction
However, if you genuinely hope to master the tools for design and analysis of
experiments with mixtures go back and do your homework via the practice
problems. Be sure to pursue the links to web-based content that provides
many of the details you will need to interpret the statistical analyses and
graphics from the software we make available or others that offer the same
features (there are several good alternatives that can be easily searched out
if not already at your fingertips from your enterprise’s server).
Finally, to leave no leaf unturned, consider going back to the first two
books in this trilogy—DOE Simplified and RSM Simplified. Even if you have
already read these two books, it will be useful to leaf through them (pun
intended) and review the detailed statistical tools presented that remain
useful for mixture design and analysis, for example—diagnostics for model
validation.
See the flowchart for specific parts of the prior Simplified books that will
be very relevant to what’s covered in Formulation Simplified.
DOE simplified
(Chapters 1–4)
RSM simplified
(Chapters 1, 2, 6–9, 11
and glossary of terms)
Formulation
simplified
Flowchart of recommended readings
Ultimately this all becomes just an academic exercise (like reading a text-
book about how to ride a bicycle) unless you actually take these tools for
a spin on your own. It will not be hard to find a proper application for
mixture design and analysis—just consider your favorite food or beverage
and search out the sweet spot in their formulation. You will see plenty of
other ideas throughout the book on experiments you can do at home, but
better yet, dive in on something that will benefit your sponsor or employer.
Introduction ◾ xvii
The secret weapon you will develop by reading this book and putting its
tools into practice is the ability to handle many ingredients simultaneously—
not just one at a time, as dictated from time immemorial by “the scientific
method.” As you will see from example after example, one needn’t hold
all else constant while changing only one thing. Instead, take advantage
of modern-day parallel processing schemes with mixture designs that
provide multicomponent testing. This is the forest we hope you will see
from the highest level, which then will provide the necessary motivation
for sharpening up your ax before going back to hacking at the trees of
formulation development.
Chapter 1
Getting Your Toe

into Mixtures
Simplicity is the ultimate sophistication.

—Leonardo da Vinci
Come on in—the water’s fine! Ok, maybe you’d do best by first sampling
the temperature of the pool with just your finger or toe. That’s what we
will try to do in this chapter—start with the simple stuff before getting too
mathematical and statistical about mixture design and modeling for optimal
formulation. Our two previous books, DOE Simplified and RSM Simplified,
both featured chapters on mixture design that differentiate this tool from
factorials and response surface methods (RSM), respectively. However, if
you did not read these books, that’s OK. We will start with an empty pool
and fill it up for you!
It’s natural to think of mixtures as liquids, such as the composition of
chemicals a pool owner must monitor carefully to keep it sanitary. However,
mixtures can be solids too, such as cement or pharmaceutical excipients—
substances that bind active ingredients into pills. The following two defini-
tions of mixtures leave the form of matter open:
◾ “Mixtures are combinations of ingredients (components) that together

produce an end product having one or more properties of interest”—
Cornell and Piepel (2008).
1
2 ◾ Formulation Simplified
◾ “What makes a mixture?

1. The factors are ingredients of a mixture.
2. The response is a function of proportions, not amounts.
– Given these two conditions, fixing the total (an equality constraint)
facilitates modeling of the response as a function of component
proportions”—Whitcomb (2009).
The first definition by Cornell and Piepel provides a practical focus on

products and the interest that formulators will naturally develop for certain
properties of their mixture (as demanded by their clients!). However, the
second specification of a mixture from Pat presented more concise conditions
that provide a better operational definition. He suggests that formulators ask
themselves an easy question: “If I double everything, will I get a different
result?” If the answer is no, such as it would be for a sip of sangria from the
glass versus the carafe, for example (strictly for the purpose of tasting!), then
mixture design will be the best approach to experimentation.
THE QUINTESSENTIALS OF MIXTURES

Mixture experiments date back to ancient times when it was thought that
all matter came from four elements: water, earth, fire, and air. For centu-
ries, alchemists sought the magical fifth element, called the “quintessence,”
which would convert base metals to gold. Petrochemicals made from
“black gold” (oil) became the focus of Henri Scheffé’s pioneering article
in the field of statistical design and analysis of mixtures—Experiments
with Mixtures (1958). Perhaps this interest was sparked by his work in
World War II, which remains obscure (not revealed at the time) but can
be described in general terms as having to do with “the effects of impact
and explosion” (Daniel and Lehmann, 1979). Thus one assumes that
Scheffé studied highly exothermic mixtures in his formative years as an
industrial statistician.
Later, Sahrmann et al. (1987) created a stir in the field for their mix-
ture experiments on Harvey Wallbanger’s popular cocktail in the 1970s.
A potential problem with mixture experiments on alcoholic drinks is
that, unless the tasters are professional enough to refrain from drinking
(Continued )
Getting Your Toe into Mixtures ◾ 3
the little they sip, after several samples, the amount of alcohol ingested
could matter—not just the proportions. Therefore, one should never
permit sensory evaluators to consume alcoholic beverages—only sip,
spit and rinse mouth afterward with water. Keep that in mind if you
wish to apply the methods of Cornell’s landmark book (2002) to such
purpose (e.g., if you become inspired by our case study in Chapter 2
on blending beers).
PS: Of course, some mixtures are better liberally applied—for example,
primer paint—the more, the better for hiding power. This would be a good
candidate for a “mixture-amount” design of formulation experiment. They
require more complicated approaches and modeling so let’s set this aside
for now. We will devote our full attention to “mixture-amount” experi-
ments towards the end of the book.
With mixtures, the property studied depends on the proportions of

the components present, but not on the amount.
—Henri Scheffé
All That Glitters Is Not Gold

Let’s now dive in on the shallow end of design and analysis of experiments
with mixtures. The easiest approach will be to lead by example via a case
study.
Some years ago, Mark enjoyed a wonderful exhibit on ancient gold at the
Dallas Art Museum. It explained how goldsmiths adulterated gold with small
amounts of copper to create a lower melt-point solder that allowed them
to connect intricately designed filigrees to the backbone of bracelets and
necklaces. This seemed very mysterious, given that copper actually melts at
a higher temperature than gold! However, when mixed, these two metals
melt at a lower temperature than either one alone. This is a very compelling
example of synergism—a surprisingly beneficial combination of ingredients
that one could never predict until they are actually mixed for experimental
purposes (Figure 1.1).
Figure 1.1 This exquisite necklace, now in London’s British Museum, came from
a necropolis (burial site) on Rhodes. It features Artemis, the Greek goddess of hunting.
(Courtesy of Bridgeman Art Library, London/New York.)
A EUREKA MOMENT!
You may recall from studying Archimedes’ principle of buoyancy in which
this Greek mathematician, physicist, and inventor who lived from 287–212
BC was asked by his King (Hiero of Syracuse) to determine whether
a crown was pure gold or was alloyed with a cheaper, lighter metal.
Archimedes was confused as to how to prove this, until one day, when
he started observing the overflow of water from his bathtub, he suddenly
realized that, since gold is denser, a given weight of gold represents
a smaller volume than an equal weight of the cheap alloy. Therefore, a
given weight of gold would displace less water. Delighted at his discovery,
Archimedes ran home without his clothes, shouting “Eureka! Eureka!”
which means “I have found it! I have found it!” When you make your
discovery with the aid of mixture design for an optimal formulation, feel
free to yell “Eureka!” as well, but wait until you get dressed.
PS: If you have a copy of DOE Simplified, 3rd Edition, see the Chapter 9
sidebar “Worth its weight in gold?” It provides information on a linear blend-
ing model we derived based on the individual densities of copper versus
the much heavier (nearly double) gold.
The ancient Greek and Roman goldsmiths mixed their solder by a simple
recipe of 2 parts gold and 1 part copper (Humphrey et al., 1998). The use
of “parts,” while extremely convenient for formulators as a unit of measure
is very unwieldy for doing mathematical modeling of product performance.
The reason is obvious, the more parts of one material that you add, the
more diluted the other ingredients become, but there is no quantitative
accounting for this. For example, some goldsmiths added 1 part of silver to
the original recipe. That now brings the total to 4 parts, and thus the gold
becomes diluted further (2 parts out of 4% or 50%, versus the original con-
centration of 2/3 or about 67%). Therefore, one of the first things we must
do is wean formulators wanting to use modern tools of mixture design off
the old-fashioned use of parts. In this case, it will be convenient to spec-
ify the metal mixture by weight fraction—scaled from zero (0) to one (1).
However, all that matters is that the total is fixed, such as one for the weight
fraction. Alternatively, if our goldsmith used a 50-milliliter crucible, then
the ingredients could be specified by volume—provided that when added
together, they will always be 50 mL. You will see various units of measure-
ments used in mixture designs throughout this book, although the most
common may be by weight. Regardless, the first thing we will always spec-
ify is the total.
Getting back to the task at hand, let’s see the results for the temperature
at which various mixtures of gold and copper begin to melt. Assume this
was done in ancient times when measurements were not very accurate.
(This is a pretend experiment!) We’ve covered the entire range from zero to
one of each metal (Table 1.1).
Notice that the table sorts the blends by their purity of gold. The
actual order for experimentation can be assumed to be random. As
emphasized in both our previous books on statistical design, random-
ization provides insurance against lurking variables such as warm-up
effects from the furnace, cross-contamination in the crucible, learning
curves of operators and so forth. As the inventor of modern-day
industrial statistics, R. A. Fisher said, “Designing an experiment is like
gambling with the devil: only a random strategy can defeat all his
betting systems.”
Table 1.1 Melt Points of Copper versus Gold and Mixtures of the Two
Gold Copper Melt Point
Blend # Point Type Blend Type (wt fraction) (wt fraction) (Deg C)
1 Vertex Pure 0.00 1.00 1073
2 “ “ 0.00 1.00 1063
3 “ “ 0.00 1.00 1083
4 Axial check Quarter 0.25 0.75 955
blend
5 Third edge Third 0.33 0.67 951
6 Centroid Binary 0.50 0.50 926
7 Third edge Third 0.67 0.33 929
8 Axial check Quarter 0.75 0.25 952
blend
9 Vertex Pure 1.00 0.00 1049
10 “ “ 1.00 0.00 1036
PURE BLENDS REPLICATED TO PROVIDE PURE ERROR

To provide an estimate of pure error, the “pure blends” (yes, that is an
oxymoron) are replicated several times, with one more for copper (blend
#s 1–3) than for gold (#s 9–10), thus saving on the precious metal. These
replicates provide three degrees of freedom (df), two from the copper
(df = 3 − 1) and one from the gold (df = 2 − 1). As a benchmark for
testing lack of fit (see Appendix 1B for details on this statistic), this is a
marginal number of degrees of freedom (df) for pure error. If gold were
not so costly, one more replicate would be better. That increases the df
for pure error to four, which we recommend at a minimum.
Another important element of this experimental design is the replication

designated in the descriptor columns (point type and blend type) by ditto
marks (“). We advise that at least three blends be replicated in the random-
ized plan, preferably four or more. These provide a measure of pure error,
desirable for statistical purposes, but as a practical matter, the replicates offer
an easy way for formulators to get a feel for their inevitable variations in
blending the materials and measuring the response(s)—simply look at the
results from run-to-run made by the same recipe.
Generating a Beautiful Response Surface—Like

a String of Rubies on a Gold Strand!
Ok, perhaps we are getting carried away in our enthusiasm for using data
from a well-designed mixture experiment to produce a very useful plot
of predicted response at any given composition. Here is our equation, fit-
ted from the experimental data by least squares regression, for modeling
the melt point (m.p.) as a function of the two ingredients: gold and copper,
symbolized by x1 and x2; respectively. These input values are expressed on a
coded scale of zero to one, which statisticians prefer for modeling mixtures.
Melt point = 1043 x1 + 1072 x 2 − 536 x1x 2
This mixture model, developed by Henri Scheffé (1958), is derived from the
conventional second-order polynomial for process RSM, called a quadratic
equation. The mathematical details are spelled out by Cornell (2002). Two
things distinguish Scheffé’s polynomial from that used for RSM. First, there
is no intercept. Normally this term represents the response when factors are
set to zero—set by standard coding to their midpoints for process modeling.
However, a mixture would disappear entirely if all the components went to
zero—we can’t have that! The second aspect of this second-order mixture-
model that differs from those used for RSM is that it lacks the squared
terms. Again, refer to Cornell’s book for the mathematical explanation, but
suffice it to say for our purposes that the x1 x2 terms capture the nonlinear
blending behavior—in this case, one that is synergistic, that is—a desirable
combination of two components.
THE JARGON OF DESIGN OF EXPERIMENTS

ON MIXTURES VERSUS PROCESS
Cornell and other experts are very particular on how one describes the ele-
ments of design and analysis for mixture experiments. For example, always
refer to the manipulated variables as “components”—not factors. Those of
you who are familiar with factorial DOE and RSM will see other aspects
that closely parallel in this book on mixture design but which are named
differently. One of the traps you may fall into is referring to the second-
order mixture-term xixj as an interaction. If you say this in the presence of
the real experts, you’d best duck and cover as school children were advised
(Continued )
in the Cuban Missile Crisis (dating ourselves here!)—the proper descriptor

is “nonlinear blending.” Although it seems picky, there is a good reason
for this: Curvature and interaction terms that appear in process models
become partially confounded due to the mixture constraint that all com-
ponents sum to a fixed total. Do not fight this—just don’t say these words!
To develop a high level of expertise in any technical field, one must learn
the technical terms and express them with great care to maintain precision
in communication. This strategy can be a pain, but it provides great gain.
Observe that although this experiment requires the control of two inputs—
gold versus copper, only one X-axis is needed on the response surface plot
shown in Figure 1.2. That is because of the complete inverse correlation of
one component with the other—as one goes up the other goes down and
vice versa. In statistical terms, this can be expressed as r = −1, where r sym-
bolizes correlation, and the minus sign indicates the inverse relationship.
Let’s see how that model for m.p. connects to the graph. First, the
coefficient of 1043 for x1 estimates that temperature in degrees Celsius at
which pure gold melts. On the other hand, pure copper melts at a higher
temperature—estimated from this experiment to be 1072°C. Always keep in
mind that results will vary from any given experiment, which represents only
a sampling of the true population of all possible results from your process—
an unknown and unknowable value. The predicted values represented by
Two component mix

1100
1050
Melting point (deg C)
1000
950
900
A: Gold (wt %) 0.0 25.0 50.0 75.0 100.0

B: Copper (wt %) 100.0 75.0 50.0 25.0 0.0
Figure 1.2 Response surface for melt point of copper versus gold and their mixtures.
the solid line in Figure 1.2 are simply an estimate. This is accentuated by the
addition of 95% confidence bands (dashed) to the plot. What really counts is
that, as a practical matter, the predictions serve the purpose of the goldsmith
for using copper to formulate an optimal jewelry-solder.
The most intriguing feature of this mixture model is the large nega-
tive coefficient of 536 on the x1x2 terms. The analysis of variance (ANOVA)
shows the term to be significant at p < 0.0001—a less than 1:10,000 chance
of it being this large if the true effect were null. (For a primer on ANOVA
and p-values, refer to DOE Simplified.) So together gold and copper melt at a
lower temperature than either one alone—isn’t that amazing!
OTHER DEPRESSING COMBINATIONS OF MATERIALS

Depressed m.p. from mixtures of one material with another, such as gold
with copper, are not that uncommon. The point at which a mixture of
two such substances reaches the minimum melting temperature is called
the “eutectic.” For example, an ideal solder for electronic circuitry is made
from 63% tin (m.p. 450°F) and 37% lead (m.p. 621°F)—together these met-
als melt at a lower temperature (361°F) than either one in pure form. The
constituents crystallize simultaneously at this temperature from molten
liquid solution by what chemists call a eutectic reaction. The term comes
from the Greek eutektos, meaning “easily melted.”
The most prevalent eutectic reaction that we encounter in Minnesota
occurs when our highway workers spread salt on roads to aid snow
removal. The eutectic point for sodium chloride occurs at 23.3 wt% in
water at a freezing point of minus six degrees Fahrenheit (−6°F). As salt
is added to a mixture of water and ice on winter roads, some of the ice
melts due to the depression of the m.p. That causes heat to be absorbed
from the asphalt or concrete surface, which is no big deal—it’s got lots to
give. However, in a well-insulated environment like the jacket of an old-
fashioned ice-cream maker this effect becomes very chilling (and useful!).
Mathematically, due to the coding on a zero to one scale for each compo-
nent, the maximum impact of this second-order effect (x1x2) occurs at the
0.5–0.5 (“50/50”) blend. Some quick figuring will help you see that this must
be so. First, multiply 0 by 1 and 1 by 0 to get the products at either end of
the scale. If you do not compute zero in both cases, then perhaps you pos-
sess the street smarts to be a vendor like the one we quote in the sidebar
below. Now things get a lot harder because fractions are involved. Multiply
¼ by ¾ and ¾ by ¼ to work out the result for the two axial check blends
that this design specifies the centroid and the vertices. If you got the first
calculation, we trust you know that either way this product comes to three-
sixteenths. This is a little less than the 1/4 of the result you get from multi-
plying 0.5 by 0.5 for the “50/50” blend at the centroid.
A FISHY WAY TO BLEND 50/50

A New Orleans street vendor was asked how he could sell Gulf shrimp-cakes
so cheap. “Well,” he explained, “I have to mix in some big old Mississippi
catfish that the trawler dredges off the river bottom when it makes a shrimp
run. Nevertheless, I mix them up 50:50—one shrimp, one catfish.”
If you look closely at the curve in Figure 1.2, you may notice that the mini-
mum actually occurs just a little to the right of the 0.5–0.5 point. This is due
to the gold having a lower m.p. than the copper, thus favoring a bit more of
this noble metal. A computerized search for the minimum using a hill-climbing
algorithm finds the minimum at 0.55 weight fraction gold, and thus 0.45% copper
is required to make the two components total to 1.
Now for a major disclaimer: A mixture experiment like this one on
gold and copper will only produce an approximation of the true response
surface—it may not be accurate, particularly for the fine points such as
the eutectic temperature. In the end, you must ask yourself as a formulator
whether the results can be useful for improving your recipe. In this case, the
next step would be to select a composition that meets the needs of solder
for goldsmithing fine jewelry. Determine the predicted m.p. from the graph
or more precisely via the mathematical model. Then run a confirmation test
to see how close you get. As a practical matter, this might be off by some
degrees and yet still be useful for improving your process.
TRIAL OF THE Pyx

In Anglo Saxon times the debasing of gold coin was punished by the
loss of the hand. In later years, the adulteration of precious metals was
prohibited by the Goldsmiths’ Company of London (founded 1180). The
composition of gold sovereigns was fixed eventually at eleven-twelfths
(Continued )
fine gold, and one-twelfth alloy (copper). So accurate became the com-
position and weight of the coin issued from the mint that at the 1871 trial
of the “Pyx” the jury reported that every piece they separately examined,
representing many millions of pounds sterling, was found to be accurate
for both weight and fineness. The term “Pyx,” Greek in origin, refers to
the wooden chest in which newly-minted coins are placed for presenta-
tion to the expert jury of assayers assembled once a year at the Hall of
the Worshipful Company of Goldsmiths in the United Kingdom. This
ceremony dates to 1282.
Source: Encyclopaedia Britannica, 10th Edition (1902).
Details on Modeling the Performance

of a Two-Component Mixture
Our example of blending glittery metals for jewelry provides a specific
application of mixture design and modeling. Now that we’ve enticed
you this far, it’s time to consider some general guidelines for setting
up a formulation experiment and analyzing the results. Let’s start
with the Scheffé equations for predicting the response from two
components.
First order ( linear ): y = β1x1 + β2 x 2
Second order (quadratic): y = β1x1 + β1x1 + β12 x1x 2
The hat (^), properly known as a circumflex, over the letter y symbolizes
that we want to predict this response value. The β (beta) symbols represent
coefficients, fitted via regression.
We detail the third order (cubic), which you may never need, in the
Appendix 1A. There, for added measure, we also spell out the fourth-
order (quartic) Scheffé equation. By this stage, very complex behavior can
be modeled for all practical purposes. However, this process of model-
building could continue to infinite orders of the inputs x to approximate
any true surface in what mathematicians refer to as a Taylor polynomial
series.
HOW TO ORDER UP JUST THE RIGHT

MODEL FOR YOUR MIXTURE DATA
Refer to Appendix 1B for clever statistical procedures that sequentially test
polynomial orders, layer by layer, for significance and lack of fit. This pro-
cess produces a model that provides an adequate explanation of the results
with as few predictor terms as possible, which is, “parsimonious.” We spell
it out for the gold jewelry case. Do not be put off by mathematically-intense
statistical procedures—software handles all the calculations. You just need
to get comfortable with the outputs and competent in concluding which
model they favor.
The mentally disturbed do not employ the Principle of Scientific

Parsimony: the simplest theory to explain a given set of facts. They
shoot for the baroque.
—Philip K. Dick (“Valis,” p. 16, Mariner Books, 2011)
The second-order equation not only may suffice for your needs to character-
ize the two primary components in your formulation, but it also could reveal a
surprising nonlinear blending effect. The possibilities are illustrated graphically
in Figure 1.3, which presumes that the higher the response, the better.
Response—higher the better =>
Nonlinear: synergism
1/4β12
ing
lend
ng ar b
l endi Line
ar b
Line
1/4β12
β1
Nonlinear: antagonism
β2
0 0.25 0.5 0.75 1 X1

X2 1 0.75 0.5 0.25 0
Figure 1.3 Graphical depiction of second-order mixture-model.

Notice that we tilted the linear blending line upwards, in other words, β1
exceeds β2. So this response surface predicts better performance for pure
x1 than for pure x2. If together these two ingredients produce at the same
rates as when working alone, then at the 0.5–0.5 midpoint the response will
fall on the linear blending line. However, you hope that they really hit it
off and produce more than either one alone. Then the response will curve
upwards—producing the maximum deflection at the midpoint. This syner-
gistic (positive) nonlinear blending effect equals one-fourth (0.5 * 0.5) of the
second-order coefficient. Unfortunately, some components just do not work
very well together, and things get antagonistic. Then the response curves
downward and the β12 coefficient becomes negative.
ISOBOLOGRAMS
In 1871 T.R. Fraser introduced a graphical tool called the “isobologram.” It
characterized departures from additivity between combinations of drugs.
Although it differs a bit in shape from our graph in Figure 1.2, the isobo-
logram is essentially equivalent—it plots the dose-response surface associ-
ated with the combination superimposed on a plot of the same contour
under the assumption of additivity, that is—linear blending. The observed
results are called the “isobol,” generally produced for the combinations
of individual drug dosages that produce a 50% response by the subjects.
If the isobol falls below the line of additivity, a synergism is claimed,
because less of the drugs will be needed. On the other hand, if the
isobol rises above the line, then the drugs are presumed to be antagonis-
tic. However, there are two major shortcomings associated with the use
of isobolograms. They do not account for data variability, and they are
restricted to only a few components.
Source: Meadows, S.L. et al., Environ. Health Perspect., 110, 979, 2002.
In this example, we made the response one where higher is better. Thus,
a positive β12 coefficient is desirable for this nonlinear blending effect.
However, in the first example—blending of copper into gold—the negative
nonlinear coefficient was what the jewelry maker hoped to see. Thus, a
synergistic deflection off the linear blending slope on the response surface
could be positive or negative, depending on the goal being maximization or
minimization.
When experimenting with mixtures, it pays to design an experiment that

provides enough unique blends to fit the second-order Scheffé polynomial
model. Then you can detect possible nonlinear blending effects. Hopefully,
these will prove as advantageous as the synergistic combination of copper
with gold for soldering fine jewelry. Meanwhile, it’s just as well to know
about antagonistic ingredients so you can keep these separated!
AN ANALOGY FOR NONLINEAR BLENDING

THAT WILL WORK FOR YOUR MANAGER
In the go–go years of the computer industry of the late 1990s (before
the dot-com bust) a movement developed for extreme programming
(XP)—a form of agile development. One of its premises, which many
software executives found counter-intuitive is that two people working
at a single computer are just as productive as they would be if kept in
separate cubicles, but (this is the payoff!) this pair programming increases
software quality. Thus, better code emerges without impacting time to
deliver. Wouldn’t that be a wonderful case of synergistic (nonlinear)
blending?
The idea of creating teams of two is nothing new. It’s easy to imagine
cavemen pairing up to hunt down mastodons more effectively. However,
an age-old problem for managers of such tasks, for example, a police
chief setting up patrol cars, is choosing the right people to team together.
Unfortunately, some combinations turn out to be antagonistic, that is—
they produce less as a partnership than either one alone. That creates
several headaches all around.
How does one know which elements will prove not only to be com-
patible but also—create a synergy? Experiment!
Practice Problems
To practice using the statistical techniques you learned in Chapter 1, work
through the following problems. Statistical software used for such compu-
tations can be accessed freely via a website developed in support of this
book. There you will also find answers posted. See “About the Software”
for instructions.
A LIFE PRESERVER FOR THOSE WHO BELIEVE

STATISTICS IS OVER THEIR HEADS
Our first two books, DOE Simplified and RSM Simplified, offer all the
statistical tools that one needs to choose predictive models and validate
them statistically. However, leaving nothing to chance, we offer you read-
ers a stimulating web-based primer on the fundamental statistics used in
the design of experiments. This dynamic, interactive presentation, called
“PreDOE” will fill in gaps on analysis of variance (ANOVA), p-values, and
so on. It normally requires a fee, but Stat-Ease will waive this for readers
of the Simplified series by Anderson and Whitcomb. If you are not sure
whether you need to invest time in PreDOE, which may engage your
attention for a number of hours, start with the online self-assessment on
statistics for the design of experiments. To get the link and validation,
email workshops@statease.com with reference to the page number of this
offer for free access to PreDOE web-based training.
Errors, like straws, upon the surface flow; He who would search for
pearls must dive below.
—John Dryden (1678)
When p is low, null must go. When p is high, null will fly.
–Author unknown
P-values are the calculated probability, used to evaluate statistical
significance in a hypothesis test.
Problem 1.1
To reinforce the basics of mixture modeling presented in this chapter, we
will start you off with some obvious questions that stem from this imaginary,
but commonplace, situation in our heartland of the United States.
The old truck on your hobby farm gets very poor gas mileage. Luckily
you can purchase fuel from a wholesaler who serves the agricultural
market—A low-grade gasoline that produces 10 miles to the gallon (mpg)
then it’s alright to drive the old truck all the way back into the city where
you usually dwell. It’s cheap; only 3 dollars a gallon. Another possibility is
to purchase the highly refined premium gasoline that increases the engine
efficiency to 14 mpg. However, it costs 4 dollars a gallon.
Consider these questions:

1. Assuming you drive 1,000 miles per year going back and forth from
your hobby farm, which grade of gasoline should you buy to minimize
your annual fuel cost?
2. Now suppose the wholesaler offers to blend these two fuels 50/50 at
$3.55 per gallon: How does this differ from the linear blend of prices?
3. Furthermore, you discover that your old truck gets 13 mpg with this
blend of gasoline: Is this a synergism for fuel economy?
4. Should you buy the 50/50 blend of the two grades of gasoline? (Do
not assume this will be so. Even if synergism is evident, the beneficial
deflection off the linear blending point may not achieve the level of the
best pure component. However, in this case, the solution requires an
economic analysis—look for the best bottom line on costs per year.)
INVERSE TRANSFORMATION PUTS

MILEAGE COMPARISONS ON TRACK
When the price of gas went over 4 dollars a gallon, I started paying atten-
tion to which of my three cars went where. For example, my wife and
her sister traveled 100 miles the other day to do some work at the home
of their elderly parents. They had our old minivan loaded up, but, after
thinking about it getting only about 15 miles per gallon (mpg), I moved
all the stuff over to my newer Mazda 6 Sports Wagon, which gets 25 mpg.
That meant no zoom-zoom for me that day going to work, but it was
worth enduring the looks of scorn from the other road warriors.
National Public Radio’s (NPR) All Things Considered show on June
19, 2008, led off with this quiz: “Which saves more gas: trading in a
16-mile-a-gallon gas guzzler for a slightly more efficient car that gets
20 mpg? Alternatively, going from a gas-sipping sedan of 34-mpg to a
hybrid that gets 50 mpg?” Of course, the counter-intuitive answer is the
one that’s correct—the first choice.
This is referred to as a “math illusion” studied by Richard Larrick, a man-
agement professor at Duke University. He found it easy to fool college stu-
dents into making the wrong choice in puzzlers like that posed by NPR.
Larrick suggests that it makes far more sense to report fuel efficiency in terms
of gallons per 10,000 miles (gpm)—an average distance driven per year by
(Continued )
the typical USA car owner. Professor Larrick was inspired to promote “gpm”
(vs. mpg) after realizing in the end that he’d be better off trading in the fam-
ily minivan and only gaining 10 miles per gallon with a station wagon, rather
than swapping his second car, a small sedan, for a highly efficient hybrid.
Are you still not sure about the NPR puzzler? Imagine you and your
spouse work at separate locations that require an annual commute of pre-
cisely 10,000 miles per year for both of you driving separately (two auto-
mobiles). Then your 16-mpg guzzler consumes 625 gallons (10,000/16).
By trading that for a 20-mpg car, you will need only 500 gallons the next
year—a savings of 125 gallons. On the other hand, your spouse drives
the far more efficient 34 mpg sedan—it requires only 294 gallons of
gas per year (10,000/34). Upgrading this to the 50-mpg hybrid saves just
94 gallons! We will let you do the math on this last bit.
It is surprising how something as simple as an inverse transformation
makes things so much clearer.
PS: For more details on transformations, refer to Chapter 4, “Dealing
with Nonnormality via Response Transformations” in the 3rd edition of
DOE Simplified.
Problem 1.2
This exercise stems from an experiment done by Mark with help from his
daughter Katie. To demonstrate an experiment on mixtures, they blew up a
plastic film canister—not just once, but over a dozen times. The explosive
power came from Alka Seltzer®—an amalgam of citric acid, sodium bicar-
bonate (baking soda) and aspirin (Figure 1.4).
You can see the experimental apparatus pictured: launching tube, a
container with water, the tablets, plastic film canister (Fuji’s works best), a
scale and stop-watch. Research via the Internet produced many write-ups
on making Alka Seltzer “rockets.” These are generally recommended when
using only a quarter of one tablet, and they advocate experimentation on
the amount of water, starting by filling the canister halfway. Mark quickly
discovered that the tablets break apart very easily, so he found it most con-
venient and least variable to simply put in a whole tablet every time (a con-
stant). It then took a steady hand to quickly snap on the top of the canister,
over which Katie placed the launching tube and Mark prepared to press his
stopwatch. (Subsequent research on this experiment indicated it would have
Figure 1.4 Apparatus for film-canister rocketry.
been far less nerve-wracking to stick the tablet on the lid with chewing gum,
put water in the container, put the lid on, and then tip it over—shooting the
canister into the air.) After some seconds the explosion occurred—propelling
the lid from the back porch to nearly the roof of his two-story home.
HEADS UP—DO NOT PICK PRANKSTERS AS

YOUR ASSISTANT ON ROCKET SCIENCE
Those of you who are fans of Gary Larsen’s Far Side series of cartoons
may recall a classic on depicting a white-coated scientist putting the
last nail on the nosecone of a big rocket. In the background you see
his assistant sneaking up with an inflated paper bag—poised to pop it!
Mark’s rocketry assistant, Katie, discovered that enough fizz remained
in the canister to precipitate the second blow up. On randomly chosen
runs she would sneak up on her father while he recorded the first shot’s
results and blast away. The only saving grace for Mark was the ready
availability of Alka Seltzer for an ensuing headache.
Before designing this experiment, Mark did some range finding to discover
that only 4 cubic centimeters (cm3) of water in the 34 cm3 canister would
produce a very satisfactory explosion. However, it would not do to fill the
container completely because the Alka Seltzer effervesced too quickly and
prevented placement of the lid. After some further fiddling, Mark found
that a reasonable maximum of water would be 20 cm3—more than half full.
He then set up a two-component mixture design that provided the extreme

vertices (4–20 cm3 of water), the centroid (12 cm3) and axial check blends at
8 and 16 cm3. Mark replicated the vertices and centroid to provide measures
of pure error for testing lack of fit.
Just for fun, Mark asked several masters-level engineers, albeit not rocket
scientists, but plenty smart, what they predicted—the majority guessed it
would make no difference how much water given a minimum to wet the
tablet and not so full it would prevent the top going on. This becomes the
null hypothesis for statistical testing—assuming no effect due to changing
the mix of air and water in the film canister.
Table 1.2 shows the results of flight time in seconds for various blends of
water versus air. Looking at the data, sorted by the amount of water, do you
agree with these engineers that this component makes no difference? You
may be somewhat uncertain with only an “intraocular test”—statistical analysis
would be far more definitive to assess the significance of the spread in flight
times relative to the variation due to blending errors and the perilous process
of launching the rockets. Now would be a good time to fire up your favor-
ite statistical software, assuming it provides the capability for mixture design,
modeling, analysis, response surface graphics, and multiple-response optimi-
zation. In case you have no such program readily available, we offer one via
the Internet—see the “About the Software” section for the website location and
instructions for downloading. To get started with the software, try reproducing
the outputs embedded in the answer to this problem posted at the same site
(in portable document format—.pdf). In the following chapters, we will lead
you to more detailed tutorials on using this particular DOE program.
Table 1.2 Results from Film-Canister Rocket Experiment

Blend # Run Type A: Water (cm3) B: Air (cm3) Flight Time (s)
1 2 Vertex 4 30 1.88
2 6 Vertex 4 30 1.87
3 4 AxialCB 8 26 1.75
4 3 Center 12 22 1.60
5 8 Center 12 22 1.72
6 5 AxialCB 16 18 1.75
7 1 Vertex 20 14 1.47
8 7 Vertex 20 14 1.53
BLASTING OFF FROM TUCSON, ARIZONA

After touring the Titan Missile Museum south of Tucson, Arizona, Mark
found the toy pictured in their gift shop. This product, made by a local
inventor (CSC Toys LLC), improves the aerodynamics of the seltzer-powered
rocket by the addition of a nose cone and fins (Figure 1.5).
Like these film canister rockets, the thrust of the Titan missile depended
on two components, albeit many orders of magnitude more powerful—
a precisely controlled combination of nitrogen tetroxide (oxidizer) and
hydrazine (fuel) that spontaneously ignited upon contact. This extreme
exothermic chemical behavior is characterized as “hypergolic.” The fuels
were stable only at 58°F–62°F, which meant that temperature control was
critical. In 1980 a worker dropped a 9-pound socket from his wrench
down a silo and punctured the fuel tank. Fortunately, the 8,000-pound
nuclear warhead, more destructive than all the bombs exploded in all of
World War II, landed harmlessly several hundred feet away. Some years
later the Titans were replaced with MX “Peacekeeper” rockets that used
solid fuel.
Figure 1.5 The MIGHTY Seltzer Rocket pictured from a launch pad in Tucson.
Appendix 1A: Cubic Equations for Mixture

Modeling (and Beyond)
The full cubic (third order) equation for modeling a two-component mixture
is shown below:
y = β1x1 + β2 x 2 + β12 x1x 2 + δ12 x1x 2 ( x1 − x 2 )
Notice that the coefficient on the highest order, non-linear blending term
is distinguished by the Greek letter delta. Think of the letter “d” (delta) as
a symbol for the differences (“d” for difference) pictured in Figure 1A.1. It
depicts a very unusual response surface for two components with only first
and third order behavior—the second-order coefficient was zeroed out to
provide a clearer view of how the new term superimposes a wave around
the linear blending line. Also, to add another wrinkle (pun intended) to this
surface, the coefficient is negative.
y = β1x1 + β2 x 2 − δ12 x1x 2 ( x1 − x 2 )
See if you can bend your brain around this complex mixture model: It’s
challenging!
At the 50–50 blend point the components are equal, so the offset is
zeroed (x1 – x2 = 0). When x1 exceeds x2 to the right of the midpoint,
g
ndin
ar ble
Line
3/32δ12
3/32δ12
g
ndin
a r ble
Line
β1
β2
0 0.25 0.5 0.75 1 X1

X2 1 0.75 0.5 0.25 0
Figure 1A.1 Graphical depiction of a cubic mixture model.

the difference δ (delta) is positive; thus (due to the negative coefficient)

the curve deflects downward. To the left, the difference δ becomes nega-
tive causing the response to go up (a negative times a negative makes
a positive!). The maximum wave height from the linear blending of
3/32 delta occurs at one-fourth (1/4) and three-fourths (3/4) blend points
(¾ * ¼ * (¾ – ¼) = 3/32).
FOR THOSE OF YOU MORE FAMILIAR WITH RSM MODELS

You may wonder why the third-order equation for mixtures is more com-
plex than that used to model similar behavior in a process. Remember
that the coding for process models goes from –1 to +1. When you cube
these quantities, the positives stay positive (+1 * +1 * +1 = +1), and the
negative stays negative (−1 * −1 * −1 = −1). Thus, you get wavy, up-and-
down, behavior in the surface. But in mixtures, the coded units are 0 to
1, which will always be positive, so you model waviness via a difference
of components.
PS: Since this chapter has kept things simple by focusing only on two
components, the cubic mixture model is missing one general term that
involves three components: xixjxk. You will see this term highlighted in
future chapters that delve into a “special” cubic model, which shortcuts
some unnecessary complexities and thus makes things a lot easier for
formulators.
For illustrative purposes, only, we dramatized the impact of the cubic term
in Figure 1A.1. Usually, it creates a far subtler “shaping” of the surface such
you see illustrated in Figure 1A.2, which shows a cubic model (solid line) fit-
ting noticeably better than the simpler quadratic (dotted).
Think of polynomial terms as shape parameters, becoming subtler in
their effect as they increase by order. Linear (first order) terms define the
slope. As shown in the blending case of gold and copper we went through
earlier in this first chapter of Formulation Simplified, the second order
(quadratic) fits curvature. The cubic order that we’ve just introduced in the
Appendix accommodates asymmetry in the response surface.
It’s good at this stage to simply consider mixture design as a special form
of RSM, which relies on empirical, not mechanistic, model building. In other
words, it’s best that you don’t try relating specific model parameters to the
underlying chemistry and physics of your formulation behavior. However,
0 0.25 0.5 0.75 1 X1

X2 1 0.75 0.5 0.25 0
Figure 1A.2 A subtler surface fitted best with a cubic model.
asymmetry is more prevalent in mixture experiments than it is in process

experiments. Thus, if the data suffices to fit a surface very precisely to the
third-order model, it might capture subtle non-linearity that the quadratic
would not; that is, this lower equation would exhibit a significant lack of fit.
AN IDIOM FOR NONLINEAR BLENDING: “CHEMISTRY”

Chemistry is the study of matter and its interactions. It amazes us by
unexpected reactions between particular substances. The word “chem-
istry” is often used to expressively describe a potently positive pairing,
such as an irresistible attraction between two lovers. Less often one may
hear of a “bad chemistry” building up in a group that includes antagonis-
tic elements. Thus this term “chemistry” has become a word that gener-
ally describes non-linear blending effects.
“We have very good chemistry team this year”. Phil Housley, the
Minnesota hockey great, assessing his 2009 Stillwater Area (Minnesota)
High School team.
In any case, to fit this cubic equation, one must design an experiment with
at least four unique blends, whereas three suffices (at the bare minimum)
to fit the quadratic. The more complex the behavior you want to model, the
more work you must do as a formulator. You get what you pay for.
If you have plenty of materials and time to mix them together—not to

mention the capability for making many response measurements, you could
design an experiment that fits a quartic Scheffé polynomial model. Here it is in
general terms for however many components (“q”) you care to experiment on:
q q −1 q q −1 q q −1 q
∑β x + ∑ ∑β x x + ∑ ∑ δ x x ( x − x ) + ∑ ∑ γ x x ( x − x )
2
y = i i ij i j ij i j i j ij i j i j
i =1 i <j j i <j j i <j j
q −2 q −1 q q −2 q −1 q q −2 q −1 q
+ ∑∑∑
i <j j<k k
βiijk x 2i x jx k + ∑∑∑
i <j j<k k
βijjk x i x j2 x k + ∑ ∑ ∑β
i <j j<k k
xxx 2
ijkk i j k
q −3 q −2 q −1 q
+ ∑ ∑ ∑ ∑β
i <j j<k k <l l
xxx x
ijkl i j k l
Notice that squared terms now appear. Although statistical software (such
as the one we provide to you readers) will handle the design and analysis
of a mixture experiment geared to this fourth order, it is very unlikely that
this will provide any practical gain over the fit you get from cubic or qua-
dratic models. For response surface modeling, it’s good to keep in mind
the principle of parsimony, which advises that when confronted with many
equally accurate explanations of a scientific phenomenon it’s best to choose
the simplest one (Anderson and Whitcomb, 2005, Chapter 1, sidebar “How
Statisticians Keep Things Simple”).
OUT OF ORDER?
Back in the days when computer-aided mixture modeling was limited to
cubic, an industrial statistician cornered Mark at a conference and com-
plained that he needed quartic to fit a formulation over the entire experi-
mental region. Quadratic fit fine for most of the results but fell short where
the performance fell off very rapidly. Mark tried a trick that his doctor told
him after he injured his shoulder playing softball. “When does it hurt,” the
medico asked. “Only when I throw a softball,” said Mark. “Just don’t do
that,” the doctor advised. In similar fashion, Mark—being ever practical—
suggested that one could simply not look at the response surface where
it drops off and gets fit inaccurately because no one cares at that point.
(Continued )
This flippant advice is more helpful than you might think. If you can
apply your subject matter knowledge and do some pre-experimentation
to restrict the focus of the mixture design to a desired region, the degree
of Scheffé polynomial required to approximate the response surface will
likely be less, thus reducing the number of blends required by simplify-
ing the modeling needed for adequate prediction power. For example,
why model all the Rocky Mountains when you are really interested only
in exploring one of the peaks?
Some might say that this question is academic, but that’s OK because
I am an academic.
—Kevin Dorfman, a Professor at the University

of Minnesota, speaking on esoteric research
in his specialty—chemical engineering on the
macromolecular scale
Appendix 1B: Statistical Details on How

to Order Up Just the Right Model
As noted in the sidebar “How to Order Up Just the Right Model for Your
Mixture Data,” it’s best that predictive models be kept as parsimonious
as possible. We refer to this as the KISS Principle, that is, Keep It Simple,
Statistically. It can be accomplished via a statistical table called “sequential
model sum of squares” (SMSS). Refer to Chapter 4 (pp. 76–77 and 96–98)
of RSM Simplified for details on this process for fitting polynomial process
models, which works equally well for mixture data. For example, Table 1A.1
shows the SMSS for the gold–copper blending case.
The SMSS provides an accounting of variation and associated p-values
(Prob > F). To KISS, you should add a higher-level source of terms (i.e.,
order) only if it explains a significant amount of variation (i.e., p < 0.05)
beyond what’s already removed by lower orders. In this case, it appears at
first that the model may not rise above the mean, there being no signifi-
cance to the linear terms due to the p of 0.385 being far above the 0.05
threshold. However, the next level of quadratic comes out highly significant,
so much so that the exact p-value is not worth noting, it being less than 1 in
10,000 (0.0001) probability that the 498 F value could occur by chance.
Table 1A.1 Sequential Model Sums of Squares

Sum of Mean p-Value
Source Squares Df Square F Value Prob > F
Mean versus 1.003 * 107 1 1.003 * 107
total
Linear versus 3542 1 3542 0.84 0.3850
mean
Quadratic 33094 1 33094 498.0 <0.0001
versus linear
Cubic versus 1.996 1 1.996 0.026 0.8775
quadratic
Quartic versus 30.71 1 30.71 0.36 0.5772
cubic
Residual 432.4 5 86.48
Total 1.007 * 107 10 1.007 * 106
Furthermore, because the quadratic term is significant, all the linear terms
come back into the model to maintain hierarchy, which we explain in Chapter 5
of DOE Simplified in our sidebar (p. 103) on “Preserving Family Unity.” To put it
simply, parents must always be included with their children. Therefore, in this
case, where the quadratic term AB merits inclusion in the model, both A and B
must come back in for support, even though these linear terms, on their own
are insignificant. It would be especially problematic in the case of mixture not
to include the main ingredients in the model. That would not make any sense.
The computational details on the SMSS displayed in Table 1A.1, are
shown diagrammatically in Figure 1A.3.
Testing for Lack of Fit

The lack-of-fit (LOF) test compares the deviation of actual points from the
fitted surface, relative to pure error. If a model has a significant lack of fit,
it should be investigated before being used for prediction (refer to Chapter 3,
pp. 64–65, of RSM Simplified for details). You can see in Figure 1A.4 that
the linear model does not fit nearly as well as the quadratic for the melting
point of gold–copper blends.
The deviations of the points from the fitted line, indicated by the arrowed
lines, obviously far exceed what one would expect from the variations
within the replicated points for pure gold (left) and pure copper (right).
Total
SS = 10,071,131
df = 10
Mean Residual
SS = 10,034,029 SS = 37,102
df = 1 df = 9
Linear Residual 3542

F= 1 = 0.84
SS = 3,542 SS = 33,559
33559
df = 1 df = 8 8
Quadratic Residual 33094

F= 1 = 498
SS = 33,094 SS = 465
465
df = 1 df = 7 7
Cubic Residual 2
F= 1 = 0.026
SS = 2 SS = 463 463
df = 1 df = 6 6
Quartic Residual 32.7

SS = 30.7 SS = 432 F= 1 = 0.36
432
df = 1 df = 5 6
Figure 1A.3 Calculations for sequential model sums of squares in Table 1A.1.
1100
1050
Melting point (deg C)
1000
950
900
A: Gold (wt %) 0.0 25.0 50.0 75.0 100.0

B: Copper (wt %) 100.0 75.0 50.0 25.0 0.0
Figure 1A.4 Response surface for the linear model of gold–copper melting points.
Table 1A.2 shows the lack of fit statistics for models ranging from linear to
quartic for the gold-copper blending case. Not surprisingly, the linear model gets
rejected by the p-value being far below the 0.05 benchmark for significance.
The quadratic model wins out. Going to the next level of cubic or beyond, an
order which does no good—thus creates an overcomplicated model.
Figure 1A.5 diagrams the derivation by source for the LOF statistics.
Table 1A.2 Lack-of-Fit Tests

Sum of p-Value
Source Squares Df Mean Square F Value Prob > F
Linear 33274.76 5 6654.95 70.18 0.0026
Quadratic 180.64 4 45.16 0.48 0.7570
Cubic 178.64 3 59.55 0.63 0.6442
Quartic 147.93 2 73.97 0.78 0.5336
Pure error 284.50 3 94.83
Total
SS = 10,071,131
df = 10
Mean Residual
SS = 10,034,029 SS = 37,102
df = 1 df = 9
Linear Lack of fit Pure error 33275

F= 5 = 70.18
SS=3,542 SS = 33,275 SS = 284
284
df = 1 df = 5 df = 3 3
Quadratic Lack of fit Pure error 181

F= 4 = 0.48
SS = 33,094 SS = 181 SS = 284
df = 1 df = 4 df = 3 284
3
Cubic Lack of fit Pure error 179

F= 3 = 0.63
SS = 2 SS = 179 SS = 284
df = 1 df = 3 df = 3 284
3
Quartic Lack of fit Pure error 148

F= 2 = 0.78
SS = 30.7 SS = 148 SS = 284
df = 1 df = 2 df = 3 284
3
Figure 1A.5 Calculations for sequential model sums of squares in Table 1A.2.
Model Summary Statistics

Last, but not least, we turn our attention to the R-squared (R2) statistic—
a measure of fit that is very familiar to most experimenters. Table 1A.3
presents R-squared for the gold–copper blending experiment: as-is (raw),
adjusted and predicted. As you can see by reading across row by row and
noting the decreasing values, this series of calculations becomes more and
more discriminating for model selection.
Table 1A.3 Model Summary Statistics

Source R-Squared Adjusted R-Squared Predicted R-Squared
Linear 0.0955 −0.0176 −0.4734
Quadratic 0.9875 0.9839 0.9729
Cubic 0.9875 0.9813 0.9641
Quartic 0.9883 0.9790 0.9508
WHY R-SQUARED NEEDS TO BE ADJUSTED

We recommend you use either the adjusted or predicted R-squared—the
raw value being flawed due to the fact that it leads to overfitting. The
problem with R-squared is that it always increases. For example, going
down the column of R-squared values in Table 1A.3 leads to the conclu-
sion that the quartic model fits best. However, the adjusted R-squared
increases only when you add good predictors. It decreases when you add
useless terms to your model. For example, notice in Table 1A.3 how it
goes down after adding the cubic term [AB(A−B)], and then it falls further
with next order beyond that—the quartic term [AB(A−B)2]. Therefore,
there is no advantage of putting these complicated terms into the model
for predicting melting point as a function of the relative amounts of cop-
per and gold in the jeweler’s blend.
Another difference of adjusted R-squared versus the raw statistic is
that it can go negative, for example, as it does in this case for the linear
model. That’s not a good sign!
The predicted R-squared provides all the advantages of the adjusted,
plus it serves better for experimenters who will be using their model to
(Continued )
produce better results in the future. This is accomplished by systemati-

cally removing each result, re-fitting the model, and measuring how far
its prediction falls from the removed value.
The adjusted and predicted R-squared statistics usually lead to the
same recommended model. However, when models contain many terms
within a given order that are insignificant, the predicted R-squared can
fall off so much that it leads to under-fitting. In that case, go with what
the adjusted R-squared suggests, but try a model-reduction algorithm
such as a backward stepwise elimination. Later in this book when we
introduce combined mixture-process models we will consider this situa-
tion of discrepant adjusted versus predicted R-squared.
The quadratic model [A, B, and AB] comes out on top overall with the sum-
mary measures of the adjusted and predicted R-squared (never mind the raw
R-squared!).
Chapter 2
Triangulating Your
Region of Formulation
If you don’t know where you are going, you will wind up
somewhere else.
—Yogi Berra
In this chapter, we will build up from the simplicity of dealing only with
two components system and then experiment on three or more. The biggest
step will be recognizing that if you lay this out in rectangular coordinates
then you really do not know where you are going and you will wind up
somewhere else (to paraphrase baseball guru Yogi). You need to get yourself
into the triangular space depicted in Figure 2.1.
The levels of three ingredients can be represented on this two-dimen-
sional graph paper, also known as “trilinear” for the way it’s ruled. It depicts
blends of up to three materials:
1. Vertices are the pure components. For example, pure X1 (or ingredient “A”)
is the point plotted at the top. For the sake of formulators, this paper is
marked off on a zero to one-hundred scale which can be easily trans-
lated to a more mathematically convenient range of zero to one.
2. Sides are binary blends. Being a yardstick on two components, the
sides are also referred to as “q-2 flats” (Myers et al., 2009, p. 570). The
midpoints of these q-2 flats are 50/50 blends of the components at each
end of the side. For example, the point between A and C represents
exactly half of each (and none of material B!).
31
X1 (A)
90
70
50
10
10
30
30
30
50
50
70
70
10
90
90
X2 (B) X3 (C)
Figure 2.1 Trilinear graph paper for mixtures with points to mark pure components,
binary blends, and overall centroid.
3. Mixtures of three components are in the center area. For example, the
point located precisely in the middle of the triangle, called the “centroid,”
represents a blend of one-third each of all three ingredients.
THE INVENTION OF TRILINEAR PAPER

Michael Friendly and Daniel Denis in their “Milestones in the history of
thematic cartography, statistical graphics, and data visualization” (2001)
web document (www.datavis.ca/milestones/) attribute the invention of
trilinear coordinates (graphs of (x, y, z) where x + y + z = constant) to
the American thermodynamicist Josiah Gibbs in 1873, who used them for
phase diagrams. They remain an enigma for many scientists who really
ought to make use of trilinear paper.
Triphase diagrams seem specially designed to confuse those who try

to interpret them.
—Steven Abbott, “Practical Surfactants,”
“Phase Diagram Explorer,”
www.stevenabbott.co.uk/practical-surfactants/pde.php
The neat thing about mapping mixtures to this triangular space is that
once you know two component fractions, the third is determined by the total.
Triangulating Your Region of Formulation ◾ 33
X1 (Cr)
90
70
50
Ni
10
10
8%
30
30
30
50
50
18% Cr
70
70
10
90
90
X2 (Fe) X3 (Ni)
Figure 2.2 Locating the 18-8 composition of stainless steel (for flatware).
To illustrate how this works, consider the stainless-steel flatware—knives,

forks, spoons, and so on—that you keep handy in your kitchen drawer for
everyday eating. A very common metallurgical formulation for this purpose is
18% chromium (Cr) and 8% nickel (Ni) by weight—the remainder being iron
(Fe), of course. Let’s plot this on the trilinear paper laid out in Figure 2.2.
In this case, the metallurgist conveniently chose chromium as the first
component to make the starting point easy—simply draw a line horizontally
18% of the way from the bottom (0% Cr) to the top (100% Cr). Now things
get tricky because you must rotate the graph, so the nickel (Ni) comes out
on top. (Think of these triangular graphs as being “turnery” paper!) Then
it will be easy to draw the 8% line for this metal, designated as the third
component. Now turn the graph, so iron (Fe) is at the top. Notice that the
two lines intersect 74% of the way up from the zero base of iron to its pure
component apex. The three ingredients now add to 100%! This feature of
the ternary graph is very convenient for formulators.
The Simplex Centroid Design

The pattern of points depicted in Figure 2.1 forms a textbook design called
a “simplex-centroid” (Scheffé, 1963; Cornell, 2002). We will introduce a more
sophisticated design variation called a “simplex-lattice” later, but let’s not get
ahead of ourselves. The term “simplex” relates to the geometry—the simplest
figure with one more vertex than the number of dimensions. In this case,
only two dimensions are needed to graph the three components on to an
equilateral triangle. However, a four-component mixture experiment requires
another dimension in simplex geometry—a tetrahedron, which looks like
a pyramid, but with three sides rather than four. To show how easy it is to
create a simplex centroid, here is how you’d lay it out for four components:
1. Four points for the pure components (A, B, C, D) plotted at the corners
of the tetrahedron).
2. Six points at the edges for the 50/50 binary blends (AB, AC, AD, BC,
BD, CD).
3. Four three-component blend points at the centroids of the triangular
faces of the tetrahedron.
4. The one blend with equal parts of all ingredients at the overall centroid
of the tetrahedron.
This totals to 15 unique compositions from the four components. See these
depicted in Figure 2.3.
X1
X2 X4
X3
Figure 2.3 Four-component simplex centroid design.

Although the simplex-centroid is not a very sophisticated design, it does

mix things up very well by always providing a blend of all the components
(the overall centroid). Let’s work through an example of a simplex centroid
for three components, illustrating an entire cycle of mixture experimentation
from design, through actual execution of the runs, statistical analysis, and
finally, optimization of multiple responses with cost taken into account.
The Black and Blue Moon Beer Cocktail

A “cocktail” generally refers to a mixture of hard (high alcohol) liquor. Few
drinkers nowadays would think of blending beers. However, one summer,
Mark was hit by a sudden intuition that it might be very tasty to combine
a black lager with a wheat beer—possibly this might produce a synergistic
sensation. Furthermore, to provide a contrast to this premium pairing and
possibly enable cost savings, he decided to mix in a cheap lager.
Once this idea took hold, Mark knew it must be tested by unbiased tast-
ers with a talent for drinking beer, and that the experiment itself had to be
conducted in a way that would prevent preconceived notions from contami-
nating the results. Let’s see what can be learned via this case study about
the application of multicomponent mixture design aimed at discovering a
sweet spot of taste versus cost.
DO NOT ASSUME A DIRECT CORRELATION

OF COST WITH QUALITY
In the late 1970s Mark took an evening course in marketing en route
to his Masters in Business Administration (MBA). He worked through
a case study showing how, although the national beer brands in the
USA differed very little in their brews, their marketing campaigns divided
drinkers into distinct segments. For example, Miller advertised their high-
priced product as the “champagne” of bottled beer while Old Milwaukee
went for the working man and took the low road on price. Meanwhile,
on Mark’s day job as a chemical process development engineer, an R&D
colleague made a big deal over how one got what one paid for in a beer:
The cheap stuff was simply swilled in his opinion. At this time, Mark
was gaining a great appreciation for experiments based on statistical
principles, such as the use of the null hypothesis for reducing prejudice.
(Continued )
Here was an opportunity to put the beer snob to the test via a blind, ran-
domized, statistically-planned experiment. You can guess the outcome:
He rated the Old “Swill”Waukee (his misnomer) number 1!
It will come to pass that every braggart shall be found an ass.

—William Shakespeare
(from All’s Well that Ends Well)
He was a wise man who invented beer.

—Plato
Here are the beer-cocktail ingredients (prices per 12 ounces, serving shown
in parentheses):
1. Coor’s brand Blue Moon Belgian-style wheat ale ($1.16)

2. Anheuser–Busch brand Budweiser American lager ($0.84)
3. Samuel Adams brand Black Lager ($1.24)
The design of the experiment, based on a simplex centroid, is laid out in

Figure 2.4.
A: Wheat
0
60 1
15
45
2 2 30
30 4 5
2
10
45
15
8 9
2 60
0 2 6 3
0
60 45 30 15
B: Lager C: Black
Figure 2.4 Simplex centroid bolstered with replicates and check blends.
Mark bolstered this design in three ways:
◾ Three added blends midway between the centroid and each of the
vertices (pure beers). In the jargon of mixture design, these are called
“axial check blends.” They otherwise fill empty spaces in the experi-
mental region. The addition of points to a textbook layout like the sim-
plex centroid is called “design augmentation.”
◾ Four point replicates (designated by “2”s)—the three binary blends
(midpoints of sides) plus the centroid. This provides four measures
(“degrees of freedom” in statistical lingo) of pure error. By establishing a
benchmark against which the deviations of actual points from the fitted
line can be assessed, pure error enables the testing of lack of fit—useful
for assessing model adequacy.
◾ Three replications of the entire design were sampled by three tasters.
Although the three subjects were chosen carefully based on their good
taste in beer, they differed in their generosity of rating; that is, tending
to score every brew higher or lower. These individual biases were cor-
rected via a statistical technique called “blocking.”
The 14 blends per person (blocked) were provided in random run order for
these three sensory responses:
Y1. Taste (without looking at the cocktail) on hedonic scale of

1 (worst) to 9 (best)
Y2. Appearance (1–9)
Y3. Overall liking (1–9)
To keep things simple for educational purposes, we will only look at the
overall liking (Y3) and the response of cost, which is determined com-
pletely by the blend’s composition and, of course, the current cost of each
ingredient.
Mark owns a very accurate kitchen scale (pictured in Figure 2.5)
that he uses to weigh out green coffee beans for roasting (another story!)
so it was convenient for him to set the total for each blend by weight
rather than volume—to 60 grams (roughly two fluid ounces). That kept
the total beer consumption per person to a reasonable level—about two
bottles worth. (Mark admits that during the experiments he managed
to drink about the same amount—in the name of science, naturally).
Figure 2.5 Precisely mixing a beer cocktail behind the screen (to keep tasters blind).
Each drinker kept his own beer-shot glass. The plan was for excess material,

beyond what was required for measurement purposes, to be discarded,
but Mark found this very difficult to enforce on the hot summer day of the
experiment—done on his back porch.
EARLY THREE-COMPONENT BEER MIXTURES

According to popular bar lore, eighteenth-century Londoners developed
a liking for a beverage called “three threads”—made by blending a third
of a pint each of ale, beer and “twopenny” (the strongest of these brews,
costing two cents a quart). Another three-component recipe dictated
stale (aged for up to two years), mild and country (pale) ale. It seems
likely that pub keepers experimented on mixtures in the hopes of find-
ing something both tasty and cheap (by diluting costly brews with less
expensive ones).
It is not entirely clear as to why a fashion for mixing beers arose…,

other than a desire to match palate and pocket.
—Beer Before Porter
www.london-porter.com
(Continued )
PS: Many Americans say “cheers” to a binary blend of beers from the

British Isles—thick, dark stout poured on a pale ale. This is com-
monly called a “Black and Tan.” What makes this combination inter-
esting is that with the right order of addition—first the ale and then
the stout, and a steady hand on pouring the drink, exhibits a distinct
layering of black on tan. Evidently, the stout is less dense than the
ale, but the reason remains mysterious—a matter for more research,
no doubt!
Fear not the beer cocktail.
—Stephen Beaumont
World of Beer
See Table 2.1 for the text matrix, laid out by blend type and location,
and the overall liking ratings for the three tasters. The actual order of
presentation was randomized, thus decoupling the cocktail type from
possibly lurking variables such as degrading taste (related to admissions
above), dehydration from exposure to the summertime elements, and
so on.
Be careful about drawing too many conclusions and extrapolating these
very far. However, like all experiments, this one may produce some useful
findings. Let’s see what can be made of it.
Go ahead and look over the results—as Yogi Berra said “you
can see a lot just by looking.” For example, is it possible that some
combinations of beers might be perceived as being unexpectedly tasty?
Or, perhaps, the opposite may be true: Putting individual taste of certain
beers together may not be such a good idea. Keep in mind that this
experiment represents only a sampling of possible reactions by these
particular tasters, who may or may not represent a particular segment of
the beer-drinking market. Does it appear as if any of the three tasters
may have been tougher than the others (hint!)? If so, do not worry; so
long as this individual remains consistent with the others in his or her
relative rankings by blend, then this consistent bias can be easily (and
appropriately) blocked out mathematically, thus eliminating this easily-
anticipated source of variation (person-to-person).
Table 2.1 Results of Beer-Cocktail Experiment (Ratings by Three

Tasters by Blend)
Location A: Wheat B: Lager C: Black Liking Cost $/12
Blend # Type (Coded) (Grams) (Grams) (Grams) 1–9 (ounces)
1 Pure Vertex (1,0,0) 60 0 0 5,5,5 1.16
2 Pure Vertex (0,1,0) 0 60 0 4,4,3 0.84
3 Pure Vertex (0,0,1) 0 0 60 7,6,5 1.24
4a Binary Center Edge 30 30 0 5,5,4 1.00
(0.5,0.5,0)
4b “ “ 30 30 0 5,4,5 “
(0.5,0.5,0)
5b “ “ 30 0 30 7,8,7 “
(0.5,0.5,0)
6b “ “ 0 30 30 4,4,2 “
7 Check Two-third, 40 10 10 6,7,5 1.12
one-sixth,one-
sixth Axial
(0.66 , 0.16,
0.16 )
8 Check One-sixth,two- 10 40 10 5,5,4 0.96
third,one-
sixth Axial
(0.16 , 0.66,
0.16 )
9 Check One-sixth,one- 10 10 40 7,7,6 1.16
sixth,two-
third Axial
(0.16 , 0.16,
0.66)
10a Ternary Centroid 20 20 20 5,7,4 1.08
10b “ “ 20 20 20 6,6,4 “
Diving Under the Response Surface to Detail

the Underlying Predictive Model
The results are very interesting. As you can see by the location of the peak
region in the 3D response plot (Figure 2.6), a blend of Blue Moon wheat
ale (A) and Sam Adams Black Lager (C) really hit the spot for overall liking!
Taste and appearance ratings also favored this binary blend. The tasters all
preferred this combination, which Mark deemed the “Black and Blue Moon”
beer cocktail.
The fitted model in coded units (0–1) that produced this surface is:
Overall Liking = 4.92A + 3.68B + 6.13C + 2.01AB + 7.35AC − 4.65BC

Notice from the coefficients on the pure-component terms (conveniently
labeled alphabetically—A, B, and C, rather than mathematically—X1, X2, X3)
that these beer aficionados liked the Sam Adams Black Lager (C) best and
Budweiser (B) the least: 6.13C > 4.92A > 3.68B. From the second-order
non-linear blending terms one can see the synergism of beer A (wheat)
with beer C (black) by the large positive coefficient: 7.35AC. Do not get
too excited by the very high coefficient: Remember that this value must
be multiplied by one-fourth to calculate the “kicker” for the binary blend.
Nevertheless, this product provides nearly a two-point gain on the hedonic
scale: 7.35 * 1/4 = 1.84.
On the other hand, these tasters’ buds were antagonized by combin-
ing the Budweiser with the Sam Adams Black Lager as evidenced by the
9
8
7
6
Overall liking
5
4
3
A (60)
2
B (0)
1
C (0)
C (60)
A (0)
B (60)
Figure 2.6 Response surface shows peak taste with a synergistic blend of two beers.
negative coefficient (−4.65) on model term BC. You can see this downturn
in the response surface along the BC edge. It is less of a deviation from lin-
ear blending than is observed for the AC binary blend (BC < AC).
That leaves one coefficient to be interpreted—that of term AB. It turns
out that the p-value for the statistical test on this coefficient (2.01) exceeds
0.1, that is, there is more than a ten percent risk that it could truly be zero.
(In contrast, the coefficients for terms AC and BC were both significant at
the 0.01 level). This time around, we did not bother to exclude the insignifi-
cant term (AB) from the model. Removing it would make little difference in
the response surface—just a straight edge between the wheat beer (A) and
lager (B), rather than a slightly upwards curve. We will revisit the issue of
model reduction later. As the number of components increase and modeling
gets more complex, it will become cumbersome to retain insignificant terms.
WOULDN’T IT BE EASIER TO MODEL

MIXTURES IN ACTUAL UNITS?
Statistical software that can fit Scheffé mixture models may offer these
to users in either coded or actual form, or both. In this case, the actual
equation for grams of beer is:
Overall Liking = 0.082031 Wheat + 0.061396 Lager + 0.10214 Black

+ 0.000559187 Wheat * Lager + 0.00204067 Wheat * Black
− 0.001292677 Lager * Black
One advantage you gain from this format is being able to plug in the
actual blend weights and toss out the predicted response for overall lik-
ing. This gets more intense as the order of terms goes up due to the
exponential impact on coefficients—they get really small or very large,
depending on whether your actual inputs are greater than one (as in
this case) or less than one (e.g., if you were serving beer to ants—they
would be happy with very tiny amounts).
Now, look back at the coded equation we provided in the main text and
consider how easy it is to interpret. For example, one can see immediately
what the predicted sensory result will be for each of the pure components
(A, B, and C)—these are the coefficients—no calculating required.
(Continued )
So, here’s the bottom line: For interpretation purposes, always use the
coded equation as your predictive model.
PS: In case you were wondering (?), neither the coded nor the actual
equation features a coefficient for the block effect. These models are
intended for predicting how an elite beer drinker will react to these three
types and their blends. True, some of these individuals will feel compelled
to be snobby and look down on all beers, but this cannot be anticipated
by the formulator, nor controlled once a product goes up for sale. Thus,
the block effect provides no value for predicting future behavior—only
to explain what happened during the experiment.
As discussed in Chapter 1, statisticians like to keep models as parsimonious

as possible, so let’s see if we could get by with only the linear model.
Figure 2.7 provides an enlightening view of the BC edge (two-component)
after the least-squares fit without the non-linear blending terms (AB,
AC, BC).
Notice that all 6 of the actual results (4 of which are at the same level of
liking, as noted by the point on the graph) fall below the predicted value
6
Overall liking
4 4 2
Actual lager 0 15 30 45 60
Actual black 60 45 30 15 0
Figure 2.7 View of BC (Lager-Black) edge after misfit with linear model.
from an over-simplified linear blending model. The cumulative impact of

these deviations far exceeds the value expected based on the pure error
pooled from the four replicate blends tasted by each tester. Thus this linear
model exhibits a significant (p < 0.1) lack of fit. Clearly the surface needs to
dip down at the binary blend of B (lager beer) and C (black), which it does
when fit with the quadratic model (look back at Figure 2.5). Not surprisingly,
this second-order model for nonlinear blending does not exhibit a significant
lack of fit (p > 0.3), that is, it fits!
Taking Cost into Account

Given the expense per 12-ounce serving of each of the three beers as
detailed at the outset of this case, it was a simple matter mathematically
to compute the blended costs shown in the last column of Table 2.1.
(The software that we provide for the practice problems will calculate
this for you: Simply enter the cost equation.) If one wanted to reduce expense,
mixing in a cheap amber-lager like Budweiser (component B) would
help as you can see in the response surface on cost in Figure 2.8.
However, in this case, you get what you pay for—the cheaper the blend,
the less likable it becomes. You can see this in the side-by-side comparison
of contour plots for overall liking versus cost in Figure 2.9a and b. To make
it even more obvious, we flagged the optimums of maximum liking versus
minimum cost.
1.30
1.20
1.10
Cost
1.00
0.90
0.80
A (1) B (0)
C (1)
C (0) A (0)
B (1)
Figure 2.8 Cost as a function of the beer cocktail composition.

A: Wheat A: Wheat
60 60
Y-Pred 7.4 Y-Pred 0.84

X1 (A) 25 X1 (A) 0
X2 (B) 0 X2 (B) 60
X3 (C) 35 X3 (C) 0
0 0 1.15
0 0
1.10
5 7
5.5 6.5
6 1.05
1.00 1.20
4.5 0.95
4 0.90
3.5
60 0 60 60 0 60
B: Lager C: Black B: Lager C: Black
(a) (b)
Figure 2.9 (a) Contour plots for overall liking and (b) cost of blended beers.
SEARCHING OUT THE OPTIMUM NUMERICALLY

Figuring out which blend of beer is cheapest is a “no-brainer”—even a
six-pack drinker could figure out it’s the Budweiser American lager, pure
and simple. That’s because the cost is a simple linear function. Also, it’s
deterministic, that is, not derived empirically. However, it would take
some calculus to find the optimum of the second-order model for overall
liking—perhaps a bit beyond the average beer drinker. As we will discuss
later, things get a lot more complicated when searching out the most
desirable blends on the basis of multiple responses. It turns out that for
computational purposes a hill-climbing algorithm generally works best
for this purpose. That’s how we came to the flagged optimum of 25 grams
of Blue Moon mixed with 35 grams of Black Lager for the (theoretically)
most likable Black & Blue Moon cocktail. As a practical matter, Mark
blends these two premium beers half and half (no cheap lagers allowed!).
Now suppose as a beer-tender you’d be satisfied to serve a cocktail

with an overall liking at or above 5, midway on the hedonic scale of 1 ☹
to 9 ☺. We highlighted this contour in Figure 2.9a. Furthermore, assume that
you need to hold the cost to $1.10 per 12-ounce serving—the highlighted
contour on Figure 2.9b. Years ago, before the onslaught of presentation pro-
grams like Microsoft PowerPoint, statisticians would transfer contour graphs
to individual transparencies, shade out the undesirable regions and overlay
A: Wheat
60
Cost: 1.10
0 0
Y-Pred 5.5
Cost $1.08
Overall liking: 5
60 0 60
B: Lager C: Black
Figure 2.10 Contour plots for overall liking and cost overlaid.
all the transparencies (also known as “view foils”) on an overhead projector.

Ideally, that left a window of opportunity, or “sweet spot,” like that shown in
Figure 2.10—produced directly by modern computer software.
The flag marks the centroid blend with equal measures of all three beers,
which falls inside the window of overall likability at 5 or higher and cost
less than $1.10 per serving. Mark cannot build up much enthusiasm for
this—too much work and not first-class. He is an elitist when it comes to
the finer things in life, such as a cold beer on a hot evening sitting on the
back porch after a long, hard day at the office. Thus, the binary blend of the
“Black and Blue Moon” cocktail gets his nod.
Do Not Put a Square Peg into a Triangular Hole

We hope that by now you see the merit of mapping mixtures to the triangular
space unless the experiment involves only two components, in which case a
simple number line suffices, as illustrated in Chapter 1. Although you are con-
vinced, it may be hard to convert your fellow formulators who remain “square”
by sticking to the factorial space used by process developers. Here’s a post-
script to the beer cocktail case that may help you turn the tide to the triangle.
At this stage, the “Black and Blue Moon” cocktail has achieved a foot-
hold in the drink recipes for those who enjoy a relaxing beverage now and
then. However, some dispute remains on the precise proportions of the
Black Lager versus the Blue wheat ale. An experimenter who has mastered
Blue Moon wheat ale

1
1 2
Black lager
Figure 2.11 Factorial design on formulation of Black and Blue Moon beer cocktail.
two-level factorials, but remains ignorant of mixture design, creates the

experiment laid out in Figure 2.11.
If this graph makes you thirsty, then take a brief break from your read-
ing to pour a little something. OK, now concentrate! What’s wrong with this
picture given that taste is a function of proportions and not amounts of a
beverage? The problem occurs along the diagonal running from lower left
to upper right in this square experimental region. One of each of the beers
versus two of each makes no difference once blended—a sip tastes the
same. All that’s happening in this direction is a scale-up of the recipe. That
would be a waste of good beer!
WHEN YOU KNOW THAT THE BEER

DRINKING HAS GOTTEN OUT OF HAND
Mark and Pat are big fans of baseball and their hometown club—the
Minnesota Twins. Mark shares season tickets with a relative, but he did not
hold seats for a critical year-end game. Thus he was forced to purchase a
spot in the outfield bleachers where the really rabid boosters congregate.
As he worked his way down the row, Mark observed that the portly fellow
in the adjoining seat had lined up 9 beers under foot—one for each inning,
thus avoiding the need to fight the lines at the concession stand. To avoid
such abuse of alcohol, the policy of the Twins is two beers per purchase,
(Continued )
so this guy must have come quite early to stock up so! Such behavior goes
beyond the pale of good taste in our opinion. We urge you to be more
moderate if you try to replicate our beer-blending experiments.
Practice Problems
through the following problems.
Problem 2.1
Normally, we do not recommend the simplex centroid design, especially if
done “by the book,” that is—without check blends. However, it can be use-
ful for three components that cannot conveniently be broken down too far
into fractions. A case in point is the “Teany Beany Experiment” we detailed
in DOE Simplified, 2nd Edition (Chapter 9). Per a randomized plan, a dozen
or so tasters were each required to chew one or more small jelly beans
flavored with apple, cinnamon, or lemon. Biting down on all three at once
presented a challenge, but everyone got the job done. Table 2.2 shows the
experiment design and the taste ratings.
Table 2.2 Teeny-Beany Mixture Design—A Three-Component Simplex Centroid

ID# Blend A: Apple (%) B: Cinnamon (%) C: Lemon (%) Taste Rating
1 Pure 100 0 0 5.1
2 “ “ “ “ 5.2
3 Pure 0 100 0 6.5
4 “ “ “ “ 7.0
5 Pure 0 0 100 4.0
6 “ “ “ “ 4.5
7 Binary 50 50 0 6.9
8 Binary 50 0 50 2.8
9 Binary 0 50 50 3.5
10 Centroid 33.3… 33.3… 33.3… 4.2
11 “ “ “ “ 4.3
This table translates the relative proportions to a percent scale. Beware!

Software that supports mixture design may require you to make each row
add to the specified total, in this case, one hundred. If so, blend number 10
will fail unless you “plug” it by entering 33.34 for one of the three compo-
nents. (A reminder: Statistical software used to do the computations can be
freely accessed via a website developed in support of this book. There you
will also find answers posted. See About the Software for instructions).
Note from the response column that each of the three pure-component
blends was replicated, as well as the overall centroid (the three “beaner”).
The results come from the averaging of several tasters, done for simplic-
ity sake. Analyze this data. What hits the sweet spot for this candy-loving
crowd?
MORE SOPHISTICATED APPROACHES

FOR TASTING TINY JELLY BEANS
To keep things simple, we provide only the average taste rating for this
case study on candy. However, a more precise analysis blocked out the
rating shifts caused by personal preferences—some scoring high across
the board (people with a “sweet tooth”) and others low. We found a
remarkable consistency—all but one of the tasters liked cinnamon best
and lemon least. However, one individual loved lemon far and away. We
excluded her from the average. But do not fear, she remains free to fill
her candy jar with these yellow beans.
I’m not an outlier; I just haven’t found my distribution yet.

—Ronan M. Conroy
Problem 2.2
This is a case where materials can be freely blended, thus the formulators
could augment the simplex centroid with check blends. The experimenters
measured the effects of three solvents known to dissolve a particular family
of complex organic chemicals (Del Vecchio, 1997). They had previously dis-
covered a new compound in this family. It needed to be dissolved for purifi-
cation purposes, so they needed to find the optimal blend of solvents.
Table 2.3 Design Matrix and Data for Solvent Study

A B C
ID# MEK Toluene Hexane Blend Type Solubility (g/L)
1 100 0 0 Pure A 121
2 0 100 0 Pure B 164
3 0 0 100 Pure C 179
4 50 50 0 Binary AB 140
5 0 50 50 Binary BC 180
6 50 0 50 Binary AC 185
7 33.3 33.3 33.4 Centroid 199
8 66.6 16.7 16.7 Check 175
9 16.7 66.6 16.7 Check 186
10 16.7 16.7 66.6 Check 201
Table 2.3 shows the experimental design and results. Remember that the
actual run order for experiments like this should always be randomized to
counteract any time-related effects due to the aging of material, and so on.
Also, we recommend that you always replicate at least four blends to get a
measure of pure error. In this case, it would have been helpful to do two
each of the pure materials and, also, replicate the centroid.
Notice that for the sake of formulating convenience, the interior points
(centroid and check blends) were rounded to the nearest tenth of a percent
so that they always added to one hundred. Which chemical, or a blend of
two or three, will work best as a solvent? (Hint: Read the Appendix before
finalizing your answer.) For extra credit on this problem, determine the
relative costs of each chemical (MEK is methyl ethyl ketone) and work the
material expense into your choice.
Appendix 2A: The Special Cubic (and Advice

on Interpreting Coefficients)
In the Appendix to Chapter 1 we detailed a very sophisticated third-degree
mixture model called the “full cubic.” As we illustrated, this equation
captures wavy response behavior that would be unusual to see in a typical
formulation experiment. As a practical matter (remember the principle of

parsimony) it often suffices to apply a simpler “special cubic” model, such as
this one for three components:
y = β1x1 + β2 x 2 + β3x 3 + β12 x1x 2 + β13x1x 3 + β23x 2 x 3 + β123x1x 2 x 3
As you may’ve already surmised, this modeling option is not applicable to

two components. For three components, the special form saves three model
terms over the full cubic, and it’s easier to interpret. Figure 2A.1 provides an
example of three components of a solid rocket fuel blended for a peak in
the measured response—elasticity (Cornell and Piepel, 2008, p. 55).
Notice the dramatic upward curve along the BC edges—a clear case
of synergism between these two materials (oxidizer and fuel). However, a
further increase occurs with the addition of component A (binder)—a move
from the center of the BC edge to the flagged peak in the interior of the
triangular mixture space. This represents a beneficial non-linear blending
involving three components.
The fitted special cubic model (coded) that generated this response sur-
face is:
Elasticity = 2351A + 2446B + 2653C − 6AB + 1008AC + 1597BC + 6141ABC
The coefficient on ABC seems surprisingly large unless you remember that
the components are on a scale of 0 to 1. For example, recall from the last
chapter that the maximum deflection from linear blending occurs at the
3100
3034
2900
Elasticity
2700
2500
2300
A (1) B (0)
C (1)
C (0)
A (0)
B (1)
Figure 2A.1 Response surface from special cubic model.

1/2–1/2 (“50/50”) point, thus you must multiply the two-component terms,
such as BC, by one-fourth (1/4) to assess the synergism (or antagonism).
However, for the three-component term, the maximum non-linear effect
occurs at the 1/3-1/3-1/3 point (centroid) with a magnitude of 1/27 the
coefficient of that term. Thus, in this case the maximum effect for BC of
approximately 400 (1/4 of 1597) almost doubles the greatest impact of ABC
(1/27 of 6141).
How do higher order terms compare to the linear ones in this case (2351,
2446, and 2653)? Here again, you must be careful not to jump to conclusions
without first considering the meaning of linear coefficients in Scheffé
polynomial mixture models—the difference is what counts, not the absolute
magnitude. The range of linear coefficients is only a bit over 300 (2653 for C
minus 2351 for A), so the tilt in the plane of response (upwards to component C)
is actually exceeded by the synergism of B and C.
Have you had enough of trying to interpret coefficients in these higher
order mixture models? We hope so because it’s not worth belaboring—simply
look at the response surfaces to get a feel for things. Then with the aid of
computer tools, use the model to numerically pinpoint the most desirable
recipe for your formulation.
A FURTHER WRINKLE: THE SPECIAL QUARTIC

Another variation on mixture models is the “special quartic”—shown
here for three components (Smith, 2005):
y = β1x1 + β2 x 2 + β3 x 3 + β12 x1x 2 + β13 x1x 3 + β23 x 2 x 3
+ β1123 x12 x 2 x 3 + β1223 x1x 22 x 3 + β1233 x1x 2 x 32
This model provides additional terms that capture more complex non-
linear blending than the special cubic. However, be forewarned that the
number of unique blends in the mixture design must always exceed the
count of terms in the model you want to fit. As spelled out in Table 2A.1,
the special quartic model requires considerably more blends at four or
more components than the special cubic, which may make the experi-
ment unaffordable.
(Continued )
Table 2A.1 Number of Terms in Full versus Special Scheffé Polynomials

Special Full Special Full
Components (q) Linear Quadratic Cubic Cubic Quartic Quartic
2 2 3 NA 4 NA 5
3 3 6 7 10 9 15
4 4 10 14 20 22 35
5 5 15 25 35 45 70
Chapter 3
Simplex Lattice Designs

to Any Degree You Like
My beloved is like a gazelle or a young stag. Behold, he is standing

behind our wall!
He is looking through the windows. He is peering through the
lattice.
—Song of Solomon 2:9
In this chapter, we will detail the design of lattices within a simplex

geometry. The procedure will work for any number of components, but we
will illustrate simplex lattice design only up to four ingredients, which form
a three-dimensional tetrahedron. That will provide enough of a challenge
for now.
Working with Four Components in Tetrahedral Space

To get a feel for a tetrahedron, make one out of a piece of paper (preferably
ruled for graphs) per the following procedure (Box and Draper, 2007, p. 512).
Start by drawing an equilateral triangle as best as you can (perfectionists see
the sidebar). Then using the midpoints of the edges as vertices, inscribe a
second equilateral triangle as pictured in Figure 3.1a.
55
(a) (b)
Figure 3.1 (a) Laying out the fold lines for building a tetrahedron and (b) the result-
ing 3D paper model.
Label the vertices of the larger triangle “1”—these represent the first
ingredient. Identify the corners of the smaller triangle as 2, 3, and 4 to pinpoint
three more components (thus allowing for four, total). Now cut out the large
triangle and fold the 1’s along lines 4–2, 2–3, and 3–4 to a point, as shown in
Figure 3.1b. There—you’ve made a tetrahedron! Keep this handy to help you
visualize our illustrations of four-component mixture design.
HOW TO DRAW AN EQUILATERAL TRIANGLE

It’s easy to draw a triangle, but surprisingly hard to make it equi-
lateral, that is, with sides of precisely equal length. The trick is to
start with the proper tool—a compass—and knowing how to use
it for drawing arcs and setting fixed distances. We could detail this
here, but it will make a lot more sense for you to see it demon-
strated by The Math Open Reference Project at the internet web page
for “Constructing an Equilateral Triangle”: www.mathopenref.com/
constequilateral.html.
Simplex Lattice Designs to Any Degree You Like ◾ 57
Building a Simplex Lattice Design

Henry Scheffé’s pioneering publication “Experiments with Mixtures” (noted
in a Chapter 1 sidebar) introduced simplex lattice design to industrial statisti-
cians and their client formulators. The name of this mixture design provides
two big clues about its construction:
◾ A simplex is a geometric figure having a number of vertexes or corners

equal to one more than the number of dimensions of the variable space
for n dimensions. For example, when n equals two, the simplex has
three corners—an equilateral triangle.
◾ A lattice is an arrangement in space of isolated points in a regular
pattern, such as atoms in a crystalline solid.
The simplex lattice design comprises m+1 equally spaced values from 0 to 1,
thus
x i = 0, 1/m,2/m, ,1
where x is the individual (i) component in coded levels and m represents

the degree of the polynomial that the formulator feels will be needed to fit
the experimental response. The number of blends in a simplex lattice design
depends on both the number of components (q) and the degree of the poly-
nomial. The simplest (overly so!) simplex lattice is this one designed for two
components at a degree of one (m = 1):
x i = 0, 1/m, 2/m , ...
x i = 0, 1/1
x i = 0, 1
There are only two points, 0 and 1! Going beyond 1 is not allowed, so the
design must stop there. It’s designated as “(2, 1)” based on the number of
components and degree; respectively. We do not recommend this (2, 1)
simplex lattice as-is, there being too few points for any appreciable power.
However, it serves as a launching pad to designs for three components that
are not that bad – for example, the two pictured in Figure 3.2a and b for
second and third-degree modeling.
X1 = 1 X1 = 1
X3 = 1 X2 = 1 X3 = 1 X2 = 1
(a) (b)
Figure 3.2 Three-component simplex lattices of 2nd (a) and 3rd (b) degree.
If you are a formulator, it will seem odd to experiment with several com-
ponents but never a complete blend, which is exactly what happens with
the (3, 2) design depicted in Figure 3.2a—its interior remains devoid of points.
However, keep in mind that you need not adhere strictly to this template—
strongly consider adding the centroid and, if not impractical, additional
check blends inside the simplex formulation space. This will be demon-
strated by example later in this chapter.
Two more simplex lattice designs are shown in Figure 3.3a and b. Notice
by their geometry—tetrahedral—that these encompass four components.
X1 = 1 X1 = 1
X4 = 1 X2 = 1 X4 = 1 X2 = 1
(a) X3 = 1 (b) X3 = 1
Figure 3.3 Four-component simplex lattices of 2nd (a) and (b) 3rd degree.
Also, evident at-a-glance is the increase in the degree of the lattice from left
to right of two to three; respectively.
An easy way to infer the degree is by the number of design points along
the edges; when broken in half, the degree is two—whereas a fragmentation
by thirds indicates a third-degree lattice.
AN EXCITING (!) FORMULA FOR THE NUMBER

OF DESIGN POINTS IN A SIMPLEX LATTICE
The number of design points N in a simplex lattice depends only on the
number of components q and the degree of polynomial m. The equa-
tion is:
N = (q + m − 1)! /(m! (q − 1)!)
You may recall from math and/or stats class that the exclamation marks
denote a factorial notation. This will come back to you more quickly by
following this example calculation on the number of design points for a
four-component simplex lattice designed to the third degree—a (q, m) of
(4,3), which computes as:
= (4 + 3 − 1)!/(3!(4 − 1)!) = (6!)/(3!3!) = (6 × 5 × 4 × 3 × 2 × 1 )/[(3 × 2 × 1)( 3 × 2 × 1 )]
Notice that the multiplication of 3 × 2 × 1 appears both in the numerator

(top) and denominator (bottom) of this equation. Thus, these numbers
cancel, and the equation simplifies to:
N = (6 × 5 × 4)/(3 × 2 × 1) = ( 6 × 5 × 4)/ 6 = 5 × 4 = 20points
Check this calculation by counting the number of points in Figure 3.3b,

comprised of 4 pure blends at the four vertices, 12 two-part blends along
the six edges and 4 three-part blends at the centers of the four faces. That
comes to 20. Yes?
PS: For purposes of building simplex lattice designs, the special model
forms are considered to be full degree. So, for example, the degree (m)
for the special cubic is 3—the same as it would be for the full cubic
Scheffé polynomial.
Augmented Simplex Lattice: When in Doubt,

Build Them Stout
As we’ve shown, the simplex lattice design may not provide a complete
mixture—in fact, this glaring deficiency occurs whenever the number of compo-
nents exceeds the degree (q > m). Thus the first step for a practical augmenta-
tion will be to add the overall centroid—a complete blend with equal amounts
(1/q) of each component. However, there’s a less obvious drawback to run-
ning the raw simplex lattice as spelled out by textbook—they often use up all
the degrees of freedom for model-fitting, with none left over to provide check
blends for lack of fit test. For example, notice how in Table 3.1 the number of
points in the second-degree (m = 2) simplex-lattice designs equals the number
of terms in the quadratic mixture model (second-order Scheffé polynomial).
These are said to be “saturated” designs. If the formulator can be certain of the
degree, a saturated experiment would be most efficient. But that level of knowl-
edge usually comes after all the research, not before. Otherwise, why bother
experimenting?
In case you may be tempted to fall back on the less sophisticated simplex
centroid, which does guarantee inclusion of a complete blend (by defini-
tion), notice how the number of points increases exponentially—going far
beyond the minimum required for the quadratic mixture model. Thus, we
advise the simplex centroid only for exceptional cases for three components
only, such as the taste-testing experiment presented in Problem 2.1.
Instead, we recommend an augmented simplex lattice (ASL) that incorpo-
rates these added points:
◾ 1 centroid (if needed)
◾ q axial check blends
◾ At least 3 replicates
Table 3.1 Number of Points in Textbook (Raw) Simplex Designs versus What’s
Required by Model
Simplex Simplex Quadratic
Components (q) Centroid Lattice (m = 2) Model Terms
3 7 6 6
4 15 10 10
5 31 15 15
6 63 21 21
We detailed axial check blends via the beer blending case study presented
in Chapter 2, so you’re aware of how these points fill the gaps between the
centroid and each of the q vertices. This design also featured replicates and
how they provide measures of pure error, which, in conjunction with check
blends, facilitate testing for lack of fit. Now we are establishing this augmen-
tation as the standard procedure for simplex designs—centroid or lattice.
HOW MANY POINTS TO REPLICATE AND WHICH ONES

We’re trying to be sensible in advising how many points need to be
replicated. Having none provides zero degrees of freedom (df) for
estimating pure error, so adding one replicate is infinitely better from
a mathematical perspective (one divided by zero). However, when
computing the lack-of-fit test, a very large ratio of variance from the
benchmark pure error would be required. For example, going back to
the gold-copper mixture experiment in Chapter 1, here are the critical
values of this ratio at a p-value of 0.05 for df’s of 1, 2, 3, 4, and 5,
respectively: 199.5, 19.0, 9.5, 6.9, and 5.8.* Notice how quickly this value
falls but then it starts leveling off after 3 df. By 5 df the critical F
becomes stable. Thus, we recommend that at least five points be rep-
licated in any mixture design.
Now which points will be best to replicate? We suggest picking points
that exhibit the highest leverage. You may be familiar with this statisti-
cal term: It’s simply a measure of potential influence for any given point.
Consider again, for example, the gold-copper blending experiment. If the
goldsmith only checked the melt point on one sample each of pure gold
and pure copper, these two results would be entirely influential on a linear
model, that is, both design points exhibit a leverage of one. However,
replicating these points—increasing the experimental runs from two to
four—reduces the leverage to 0.5 for each.
Computing leverage becomes more complicated for higher-order
models,** but let me give you some feel for how it works—Here are the
measures for the gold–copper experiment without any replicates: 0.89 at
the vertices (pure metal), 0.49 at the centroid (“50-50” gold-copper) and
0.37 for the axial check blends (1/4th-3/4th and 3/4th-1/4th). The three
highest leverage points are the two vertices and the centroid, so these
were the ones chosen for replication.
(Continued )
*(From 5% table provided in Appendix 1–3 of DOE Simplified with

2 df in the numerator of the F-ratio—starting from the top of the second
column and looking down from 1 df in the denominator—the first row.)
**(For more details on leverage, see the sidebar titled “Assessing the
Potential Influence of Input Data via Statistical Leverage” provided in
Chapter 1 of RSM Simplified.)
Second-order designs augmented per the guidelines we’ve provided are suit-
able for producing a “simplex response-surface” (Smith, p. 49). This leads to
an important insight: Mixture design for an optimal formulation is a close
cousin to response surface methods (RSM) for process optimization.
Using Augmented Simplex Lattice Mixture Design

to Optimize an Extra-Virgin Olive Oil
Olive oil, an essential commodity of the Mediterranean region and the main
ingredient of their world-renowned diet (see sidebar), must meet stringent
European regulatory guidelines to achieve the coveted status of “extra
virgin.” Oils made from single cultivars (a particular cultivated variety of the
olive tree) will at times fall into the lower “virgin” category due to seasonal
variation. Then it becomes advantageous to blend in one or more superior
oils based on a mixture design for optimal formulation. For example, a team
of formulators (Vojnovic et al., 1995) experimented on four Croatian olive
oils—Buza, Bianchera (pronounced “be an kay ra”), Leccino (pronounced
“la chee no”) and Karbonaca—to achieve an overall sensory rating of at least
6.5 on a 9-point hedonic scale, thus easily exceeding the cut off for “extra
virgin” (5.5 considered to be “virgin”). The ratings were done by ten asses-
sors trained on fundamental tastes (sweet, salt, sour, and bitter) and defects
of virgin olive oils, such as rancidity (Figure 3.4).
We’ve adapted a bit of the original study, simplified it, and made it more
educational while capturing the essence of how these formulators made use
of mixture design and what they discovered.
Thus the assessors can discern very tiny differences in sensory attributes
that may depend on the subtle nonlinear blending of two or more oils.
Therefore, the four-component (q = 4) simplex lattice is set up to the third
degree (m = 3). That produces 20 unique blends as shown in Figure 3.3b.
Figure 3.4 Olive oil. (Royalty-free Internet post 670878 by Stockxpert.)
To augment this lattice, the formulators add 4 axial check blends to the overall
centroid. They then specify that the four vertices (chosen for their high lever-
age) and centroid be replicated (for added pure error measure) at random
intervals. (Always randomize!) Assume that the formulators use a 1-liter blender
to mix the oils—30 blends in total after the augmentation. This ASL design and
the results for overall sensory ratings are shown in Table 3.2. (Note that, for the
Table 3.2 ASL Design for Blending Four Olive Oils and Their Sensory Results
A: B: C: D: Sensory
# Point Type Buza Bianchera Leccino Karbonaca Rating
1 Vertex 1 0 0 0 6.98
2 “ 1 0 0 0 6.84
3 Vertex 0 1 0 0 6.49
4 “ 0 1 0 0 6.45
5 Vertex 0 0 1 0 7.25
6 “ 0 0 1 0 7.30
7 Vertex 0 0 0 1 5.88
8 “ 0 0 0 1 5.95
9 Third edge 0.667 0.333 0 0 7.38
10 Third edge 0.333 0.667 0 0 7.12
(Continued)
Table 3.2 (Continued) ASL Design for Blending Four Olive Oils and Their Sensory
Results
A: B: C: D: Sensory
# Point Type Buza Bianchera Leccino Karbonaca Rating
11 Third edge 0.667 0 0.333 0 6.87
12 Third edge 0 0.667 0.333 0 6.84
13 Third edge 0.333 0 0.667 0 6.95
14 Third edge 0 0.333 0.667 0 7.17
15 Third edge 0.667 0 0 0.333 7.36
16 Third edge 0 0.667 0 0.333 7.14
17 Third edge 0 0 0.667 0.333 7.50
18 Third edge 0.333 0 0 0.667 7.16
19 Third edge 0 0.333 0 0.667 6.95
20 Third edge 0 0 0.333 0.667 7.00
21 Triple blend 0.333 0.333 0.333 0 7.56
22 Triple blend 0.333 0.333 0 0.333 7.53
23 Triple blend 0.333 0 0.333 0.333 7.29
24 Triple blend 0 0.333 0.333 0.333 7.28
25 Axial CB 0.625 0.125 0.125 0.125 7.41
26 Axial CB 0.125 0.625 0.125 0.125 7.37
27 Axial CB 0.125 0.125 0.625 0.125 7.50
28 Axial CB 0.125 0.125 0.125 0.625 7.19
29 Centroid 0.25 0.25 0.25 0.25 7.58
30 “ 0.25 0.25 0.25 0.25 7.55
sake of simplicity, the one-third and two-thirds levels are rounded to 0.333 and
0.667, respectively—thus adding to the total of 1.)
The statistical analysis of this data is detailed via Problem 3.2. The chosen
model is a reduced special cubic:
Sensory Rating = 6.91A + 6.47B + 7.29C + 5.93D + 2.51AB − 0.91AC
+ 3.70AD + 0.54BC + 3.75BD + 2.78CD + 11.65ABC

The presence of the ABC nonlinear blending term supports the choice of a
third-degree lattice design. The other three special-cubic terms (ABD, ACD, and
BCD) were insignificant (p > 0.1) so we chose to remove them from the final
model. Rather than laboriously dissecting the model by its remaining terms,
let’s focus on the response surface graphics: The pictures will tell the story.
Unfortunately, now that we’ve gone to the third dimension the imaging
gets trickier—for example, only three out of the four components can be
depicted on a contour plot. This complication provides the perfect oppor-
tunity to present the “trace” plot—a way to view the relative effects of any
number of components. A trace plot for the olive-oil mixture experiment is
shown in Figure 3.5.
The traces are drawn from the overall centroid—all components at equal
volume within the 1-liter vessel. This is called the “reference blend.” Each
component alone is then mathematically varied while holding all others in
constant proportion. This reveals, for example, that the predicted sensory
evaluation falls off dramatically as the Karbonaca oil (D) is increased relative
to the three alternatives.
To give you a better feel of how the trace plot is produced, let’s consider
the simpler case of only three components. Figure 3.6 shows the paths of
the three traces.
The trace for x1 starts at the overall centroid where it amounts to
one-third of the three-component blend. The other two components are
7.70
D
C
A
C
B
7.23
A
Sensory rating
6.75
B
Reference blend
6.28 A: Buza = 0.250
B: Bianchera = 0.250
C: Leccino = 0.250
D: Karbonaca = 0.250
D
5.80
−0.250 0.000 0.250 0.500 0.750

Deviation from reference blend
Figure 3.5 Trace plot for olive-oil mixture experiment.

X1 = 1
90
70
X3 = 0 50 X2 = 0
10
10
30
30
30
50
50
70
70
10
90
90
X2 = 1 X1 = 0 X3 = 1
Figure 3.6 Traces for three components only.
also at one-third. Thus their ratio is one-to-one. Tracing x1 from the centroid
down to the base of the ternary diagram reduces the amount of this individ-
ual component to zero. At this point the amounts of the other two compo-
nents become one half each—thus their ratio remains one-to-one. In fact, if
you pick any point along any of the three traces, the other two components
remain at constant proportions! Try working this out for yourself—it will be
good practice for reading off coordinates on the ternary diagram.
TRACE VERSUS PERTURBATION

Those of you who are knowledgeable about response surface methods
(RSM) for process optimization may be familiar with the “perturbation”
plot—the RSM equivalent of the mixture trace. This plot predicts what
will happen if you perturb your process by changing only one factor at a
time, for example, by first varying time and then varying the temperature
of a chemical reaction. The perturbation plot generally emanates from the
center point—all process factors at their middle level. It looks the same
as a trace, but the difference is that, given a fixed total of the amount,
no single component can change without one or more of the others
taking up the slack. Creating a trace as we’ve detailed is a workaround
that provides the same benefits as a perturbation plot for RSM, that is,
graphically depicting the relative effects of individual components as they
(Continued )
become more, or less, concentrated in your formulation. However, keep

in mind the one-dimensional nature of the trace, which cannot substitute
for contour plots or 3D views of the surface as a function of any two
components. Only then will you see an accurate picture of non-linear
blending effects.
PS Warning: The trace (and perturbation) plot can change dramatically
with a change in the reference point. Suggestion: Once you settle on the
maximum draw of the trace (or perturbation) plot from that location, it
provides perspective on how robust the solution may be for undesirable
changes caused by variations in the inputs.
Now that we’ve been provided with clues on the non-linear blending behavior
of the four olive oils, it seems sensible to study the response surfaces of
the three good components “sliced” at varying levels of the inferior fourth
component. For example, Figure 3.7a and b show the sensory results at the
overall centroid (all components, including D, at 0.25 concentration) versus no
Karbonaca oil (D = 0). If anything, it’s the Bianchera (B) oil that creates the
greatest effect on taste—very noticeably on these slices with D held fixed at
two specific levels (0 and 0.25).
These response surface graphs are very illuminating! It appears as if the
complete four-part blend at the centroid, shown on the left (Figure 3.7a),
will be most robust to variations in olive oil concentrations and deliver a
8 8
7.5 7.5
Sensory rating
Sensory rating
7 7
6.5 6.5
6 6
B (0) B (0)
A (1) A (1)
C (1) C (1)
C (0) A (0) C (0) A (0)
(a) B (1) (b) B (1)
Figure 3.7 (a, b) Sensory results at the overall centroid versus no Karbonaca oil (D = 0).
superior sensory rating for the most part. A more comprehensive computer-
aided search of the entire tetrahedral formulation space produced the blend
detailed below and depicted in Figure 3.8:
1. 0.333 Buza
2. 0.299 Bianchera
3. 0.189 Leccino
4. 0.179 Karbonaca
This is predicted to produce a sensory rating of 7.63—higher than any of

the actual test results. However, any such prediction must be subject to
verification.
The upward ramp for the sensory response shows how a rating below
6.5 will be completely undesirable, whereas a rating of 9 represents the
peak of desirability (prima!). The predicted optimum falls 0.454 up the scale
from zero to one on desirability. This may be the best the blenders can do
0.000 1.000 0.000 1.000
A: Buza = 0.333 B: Bianchera = 0.299
0.000 1.000 0.000 1.000
C: Leccino = 0.189 D: Karbonaca = 0.179
9.00
6.50
Desirability = 0.454
7.58
5.88
Sensory rating = 7.63
Figure 3.8 Most desirable blend via computer-aided numerical search.

A: Buza
0.821
Prediction 7.63
95% PI Low 7.51
95% PI High 7.76 7.3
X1 0.333
X2 0.299
X3 0.189 7.4 0.000
7.6
7.5
7.4
7.3
7.2
7.0 7.1
0.821 0.000 0.821

B: Bianchera C: Leccino
Figure 3.9 Most desirable blend flagged on contour plot (D sliced at 0.179).
with these particular varietals. A sensory outcome of 9 on the hedonic scale

remains in the province of the gods who enjoy only the best of the best.
Figure 3.9 flags the attainable optimal blend of these four olive oils. It
displays the prediction interval between 7.51 and 7.76, based on 95%
confidence. Ideally, the verification blend will be rated by the sensory
panel within this range. A result outside of the prediction interval would
cast doubt on the validity of the model.
Remember that these plots are derived from the Scheffé-polynomial
predictive-model fitted to the actual experimental data and validated sta-
tistically. However, only by producing an independent confirmatory blend
and subjecting it to sensory evaluation will this be verified for all practical
purposes.
The optimum sensory rating comes from a blend that goes light on the
Karbonaca oil. But what if this inferior oil can be bought very cheaply?
Then a blend of it being maximized could be made at a sensory rating
just good enough to pass at the extra-virgin level. This might be worth
a try!
MEDITERRANEAN DIET
The key components of the Mediterranean diet include:
◾ Olive oil as an important source of monounsaturated fat

◾ A generous amount of fruits and vegetables
◾ Red wine in moderation
◾ Fish on a regular basis
◾ Very little red meat
Many benefits have been attributed to this diet, including a reduced rate
of coronary events and weight loss. See the American Heart Association’s
internet post on the “Lyon Diet Heart Study” at http://circ.ahajournals.
org/content/103/13/1823 for details on a randomized, controlled trial with
free-living subjects (Kris-Etherton et al., 2001).
Two themes characterize people who have lost a significant amount

of weight and kept it off long-term: (1) they don’t eat as much as in
the past, and (2) they exercise more. Look for these clues when you
search for effective weight-loss programs.
—Dr. Steve Parker, author of The Advanced
Mediterranean Diet (Vanguard Press, 2008)
By employing mixture design, these formulators learned how to blend vari-

ous olive oils to achieve similar qualities. This allowed them to work around
limited supplies of specific varieties and adapt their blends to prevailing
preferences in the market.
Practice Problems
Problem 3.1
Experience how easy it will be to design a mixture experiment and analyze
the results by using a computer tool specialized for this purpose. It can
be freely accessed via the web site developed in support of this book:
See About the Software for the path. When you arrive at the internet page,
follow the link to the accompanying tutorials. Then download and print
the Mixture Design Tutorial (Part 1—The Basics). It details a case study on
a detergent product that introduces some very practical aspects on how to
experiment on only a portion of an entire formulation. If you come across
a few new concepts, follow the advice in the Introduction of the book,
just keep moving, and worry later about the explanations—these will be
forthcoming. The primary purpose of this exercise is to get a feel for how
dedicated DOE software can facilitate your design and analysis of mixture
experiments.
Problem 3.2
Via the web site developed in support of this book, go to the answer posted
for this problem—a follow up on the olive oil mixture experiment. It details
all the important statistical models and validation of the chosen one by
analysis of variance (ANOVA) and residual diagnostics. Your assignment will
be to reproduce these results using the software made available to you for
this purpose.
Chapter 4
Mixture Constraints That

Keep Recipes Reasonable
The more constraints one imposes, the more one frees one’s
self. And the arbitrariness of the constraint serves only to obtain
precision of execution.
—Igor Stravinsky
In Chapter 1, we said that fixing the total amount of a mixture facilitates

modeling of the response as a function of component proportions.
Mathematically this translates to an equality constraint, that is, the sum of
all ingredients comes to a fixed total. Now we detail how to contend with
further constraints on individual ingredients.
Setting Minimum Constraints

The fewest complications come from setting minimum constraints only. For
example, let’s say that after seeing our sidebar on Cornell’s famous Harvey
Wallbanger’s experiment you ask a bartender to mix one up for you. The
three ingredients are orange juice, vodka and vanilla Galliano (a sweet
Italian liqueur). This mixed drink is traditionally served in a 10-fluid ounce
(fl. oz.) highball glass in proportions of 6, 3 and 1; respectively, by the speci-
fication of the International Bartenders Association (IBA). However, suppose
73
that the bartender runs low on the alcoholic ingredients and, furthermore,
you appear to be someone who doesn’t drink heavily. That leads to him
establishing these minimums:
1. Orange juice (OJ), ≥6 fl. oz. (the more the better)

2. Vodka, ≥1.5 fl. oz. (a “jigger”)
3. Galliano, ≥0.5 fl. oz. (a “splash”)
The bartender starts by pouring at least 6 ounces of orange juice into the
highball glass. This constraint is labeled “1” (segment 1–1) in the ternary
diagram (Figure 4.1). The “mixologist” follows with a jigger of vodka—
constraint 2 (segment 2–2)—and garnishes the drink with a splash of
Galliano—constraint 3 (segment 3–3). This shrinks the original simplex
region to the area labeled “A-B-C.” The point pins down the IBA’s ideal
recipe from which the bartender allows himself some variation—providing
a triangular zone of acceptable Harvey Wallbanger’s drink mixes.
Notice that the addition of lower constraints does not affect the shape
of the mixture space: It remains a simplex region. However, the shrinkage
makes it very inconvenient to work graphically within this space. We need
a way to blow it back up—like a photo enlarger. Fortunately, experts in
mixture design developed mathematical tools to accomplish this.
OJ (A)
9
2
3
A
7
1 B C 1
3 2
5
1
1
3
3
3
5
5
7
1
9
Vodka (B) Galliano (C)
Figure 4.1 Constrained region for Harvey Wallbanger mixed drink.

Mixture Constraints That Keep Recipes Reasonable ◾ 75
GOING REALLY RETRO: HARVEY WALLBANGER BUNDT CAKE

A 1966 baking contest sponsored by Pillsbury, a food company based in
Minneapolis, Minnesota, featured a “Tunnel of Fudge” cake baked in a
circular “Bundt” pan invented locally by H. David Dalquist, founder of
Nordic Ware. This creative combination of fudge-cake recipe (provided at
the Pillsbury website) and nonorthogonal geometry created a sensation—
surpassed only when someone came up with the idea of a Harvey
Wallbanger bundt cake. This makes a great dessert for a retro 1970s
party—it goes well with leisure suits, hot pants, and disco music.
If you look over the years, the styles have changed—the clothes,
the hair, the production, the approach to the songs. The icing to the
cake has changed flavors. But if you really looked at the cake itself,
it’s really the same.
—John Oates
Expanding the Constrained Space via Mathematical Coding

This might be a good time for you to take a break and prepare yourself for a
mathematical sprint by drinking a suitable beverage—caffeinated rather than
alcoholic (tempting as that may be!). The objective of this number-smithing
will be to rescale the mixture space for the sake of graphical convenience.
This will be far easier to explain by an example. Let’s use the one detailed
in Problem 3.1, which showed how to optimize a detergent formulation with
the aid of computer software for mixture design and analysis.
We will presume you’ve worked through the tutorial on this case. What’s
most relevant at the moment is how the experimental focus remains on only
the three key components which comprise 9% by weight of the total mixture.
Thus, it becomes convenient to rescale or “code” the actual component levels
to the actively varied ingredients—designated as “real” values (x′):
Xi X
x′i = = i
∑ Xi 9
where “i” represents an individual component. The coding from actual to

real levels is very simple—it’s just a conversion to fractions, in this case
Table 4.1 Coding of Actual to Real Values

Actual Real
Li Ui l’i u’i
1. Water 3% 5% 0.333 0.556
2. Alcohol 2% 4% 0.222 0.444
3. Urea 2% 4% 0.222 0.444
l′2
X′1
:A
2
lc
22
o
0.
ho
≥
l(
C)
B)
.9
a(
u′
re
4
2:
0.
44
:U
22
A
0.
lc
l′3
oh
≤
C)
ol
.7
a(
(B
re
)≤
U
0.
3:
44
u′
u′1: Water (A) ≤ 0.556

4
.5
.1
.1
l′1: Water (A) ≥ 0.333

.3
.3
.3
.5
.5
.7
.7
.1
.9
.9
X′2 X′3
Figure 4.2 Real region for detergent experiment.
based on 9% as the formulator’s total. Table 4.1 shows the ranges, lower
(L and l′) to upper (U and u′), for the three detergent ingredients in actual
percents (by weight) and their real values scaled 0 to 1 (dimensionless).
Figure 4.2 maps out the real boundaries of a ternary mixture diagram.
Notice how the space, a simplex triangle, is defined by the lower (l′)
limits of the three components. However, it will be hard to do design-
work within such a small region and, ultimately, visualize the modeled
response. We must perform one more mathematical transformation, called
“pseudo” coding, to expand the restricted reals into a maximum range
from 0 to 1:
Real − l ′i
L_Pseudo =
1 − ∑ l ′i
x i − l ′i
xi =
1 − ∑ l′
x′1 − 0.333
x1 =
1 − 0.778
x′2 − 0.222
x2 =
1 − 0.778
x′3 − 0.222
x3 =
1 − 0.778
The prefix “L” indicates that this coding has been done on the basis of
the lower boundaries (for more explanation, see the appendix on upper-
bounded pseudos). We’ve removed the prime (′) mark to differentiate the
pseudos from the reals. Table 4.2 provides the amazing results of this trans-
formation: The components now range from 0 to 1!
The payoff to pseudo coding can be seen in Figure 4.3—the entire
ternary graph is now utilized to map out the design points for the detergent
experiment.
You can see in Figure 4.3 that we’ve enlarged the space via this coding.
It now becomes far easier to work with for design purposes. Notice that the
points are now shown—two each (designated by the number “2”) at the
three corners of the triangular experimental space, for example. (The overall
centroid is also replicated.) Refer to Problem 3.1 and the associated online
tutorial for more details on this design.
Table 4.2 Coding of Real to Pseudo Values

Real Pseudo
l’i u’i li ui
1. Water 0.333 0.556 0 1
2. Alcohol 0.222 0.444 0 1
3. Urea 0.222 0.444 0 1
X′1
.9
X1
.7 2
90
.5
.1
.1
.3
.3
.3
70
.5
.5
.7
.7
.1
.9
.9
X′2 X′3
50
10
10
30
30
2
30
50
50
70
70
10
90
90
2 2
X2 X3
Figure 4.3 L-pseudo-coded region for detergent experiment.
Note this graph is ruled by percent—the standard for an “off-the-shelf” ter-

nary graph. Don’t let this throw you off. When looking at the lines labeled
10, 20, 30… (e.g., from bottom to top), read these off as 0.1, 0.2, 0.3… in
coded level; respectively.
HOW TO EXPLAIN THE RELATIVE COMPONENT

LEVELS DISPLAYED BY CODED VALUES
If a colleague or superior asks you to explain a graph such as that of
Figure 4.3, be sure to remember that the 0 coded level may not mean
that none of the components is present in the blend. For example, our
detergent experiment always has some of each ingredient. One way to
present this is to say that as one moves from any side of the ternary
graph in L-pseudo-component coding, an ingredient goes from its “lean-
est” (lowest) level to its “richest” (highest) level. Consider, for example,
the urea that’s varied experimentally for the detergent. It goes from an
actual low of 2% by weight to a maximum of 4%. However, it may suffice
to say that the left side of the triangle in Figure 4.3 represents mixtures
with urea being leanest and the lower right corner (labeled X3) is the
point where this chemical is richest.
Why It Was Worth Reading This Chapter

and What’s in It for You as a Formulator
Although we tried to keep this chapter short and sweet, you may be won-
dering what’s really in it for you. First, consider that as a practical matter, it
may be relatively uncommon to experiment on components that can range
from none to all, that is, zero to one-hundred percent (0%–100%) of the
mixture. For example, a coatings chemist would not dream of making up a
can of pure pigment. Or, how about this—a pound cake comprised only of
flour—no butter, sugar or eggs. Thus it becomes mandatory to impose con-
straints on individual ingredients.
This provides a practical perspective on the first part of this chapter, which
details an example of a constrained mixture design. The remainder seemingly
devolves into mathematical detail on component coding. Other than statisti-
cians, who needs that? In fact, computer software that handles the design and
analysis of mixtures may shield you from anything other than the actual levels
of components, at least in the recipe sheet (ideally in random order, of course)
and the outputs that you will present to a nonstatistical audience. However,
if you make the effort to see how coding greatly simplifies laying out designs
via expansion of the experimental region, this will at least provide an appre-
ciation for how statisticians do the calculations (or program software to do
so). Furthermore, when you see some of the model details reported, such as
“Final Equation in Terms of L-Pseudo Components,” the prior exposure to this
standard mathematical methodology for mixtures may prevent the perfectly
normal adverse reaction to statistical mumbo-jumbo. We hope so!
A TRICK TO SPELLING “PSEUDO”

Mark unabashedly declares his coauthors Pat and Martin to be superior
“mathletes.” However, Mark takes great pride in his word skills, espe-
cially spelling. Thus it was embarrassing for him to continually misspell
“pseudo” as “psuedo” (sic). Although he maintains a standing protest
that it is unfair to put a “p” before an “s” and combine the two vowels
of “e” and “u,” Mark had to come up with a way to remember the cor-
rect spelling of the essential statistical term. Finally, he realized that the
only way was to first spell it in his head with the “e” and “u” transposed
as “sued”—that being related to onerous legal matters is bad. Mark then
(Continued )
silently hums the famous Elvis Presley song about his blue suede shoes
and knows that this is all wrong as well. That puts him back on track for
“pseudo”—the correct spelling. Such are the lengths one must go through
to contend with peculiar scientific jargon.
Drink my liquor from an old fruit jar. … You can do anything but lay
off of my blue suede shoes.
—Excerpt from lyrics by Carl Perkins (music also
originated by him—recorded in 1956)
Practice Problem
through the following problem.
Problem 4.1
Two Portuguese food scientists, Margarida Vieira and Cristina Silva (2004),
applied mixture design to optimize the taste of an exotic nectar based on
the Cupuacu (pronounced “koo-poo-a-soo”) fruit from the Amazon jungle.
They established the following ranges (coded on a 0 to 1 scale) for their
mixture components (in percentages by weight):
1. (X1) Sugar, 10%–25%

2. (X2) Cupuacu, 15%–30%
3. (X3) Water, 60%–75%
The two fruit-juice formulators then set up a completely replicated second-

degree simplex-lattice design. They augmented their design by adding three
axial check blends (unreplicated), and the overall centroid replicated three
times.
Pencil in on Figure 4.4, a blank ternary diagram, with the upper and
lower constraints for each of the three components in their actual levels,
expressed as weight percent. You should find that this space remains a
simplex, albeit much smaller than the original. In the answer posted for
this problem, we provide a link to an article (Anderson, 2005) that details
the computer-aided design and analysis of this experiment, including the
final formulation of the Cupuacu nectar that matched what the original
experimenters recommended as the most desirable sensorially.
X1
90
70
50
10
10
30
30
30
50
50
70
70
10
90
90
X2 X3
Figure 4.4 Blank ternary diagram to draw in Cupuacu component constraints.
A SEVEN-COMPONENT (?) BEVERAGE: “7-UP®”

An early sales bulletin (~1929) for 7-Up told consumers that the drink
contained seven natural flavors “so blended and in such proportions that,
when bottled, it produces a big natural flavor with a real taste that makes
people remember it.” The original name was Bib-Label Lithiated Lemon-
Lime Soda. The new moniker coined by advertising expert C. L. Grigg,
“7-Up,” made the drink seem more palatable.
From the story, above, one would conclude that the “7” in “7-Up”
relates to the number of ingredients. However, per the website Wikipedia,
Donald R. Sadoway, a professor of materials chemistry at the Massachusetts
Institute of Technology, contends that the name derives from the atomic
mass of lithium, 7, which was originally one of the key ingredients of
the drink (as lithium citrate—purported to be a mood-stabilizing drug).
In any case, evidently during one of the many reformulations of 7-Up in
1950, lithium citrate was removed.
Here’s a thought for further experimentation by the Portuguese food
scientists—punch up the Cupuacu blend with some bubbly such as 7-Up.
(The rights to the 7-Up brand are held by the Dr Pepper Snapple
Group in the United States, and PepsiCo (or its licensees) in the rest of
the world.)
Appendix 4A: Upper (“U”) Pseudo

Coding to Invert Mixture Space
At the outset of this chapter, we suggested that setting lower
constraints keeps things simplest. When upper constraints define the
design space, it may become an inverted simplex. For example, let’s say
that, while keeping the overall total at 1, we constrain three-components
as follows—upper only:
1. X1 ≤ 0.4
2. X2 ≤ 0.6
3. X3 ≤ 0.3
As Figure 4A.1 illustrates, this upper-only approach to setting component

constraints creates an upside-down triangle within the real-scaled ternary
diagram! Unfortunately, as you can see in Figure 4A.2, the lower-bounded
pseudo values for an inverted simplex do not vary from 0 to 1, which would
be most advantageous graphically and mathematically. Rather, they range
from 0–0.5 in this coded scale.
This geometry can be likened to trying to put a square peg in a round
hole. Luckily for all of us formulators, an expert in the field (Crosier, 1984)
X1
90
70
50
10
10
30
30
30
50
50
70
70
10
90
90
X2 X3
Figure 4A.1 Inverted simplex (real) created by entering only upper component
constraints.
A: X1
0.000
1.000
0.200
0.800
0.400
0.600
0.600
0.400
0.800
0.200
1.000
0.000
1.000 0.800 0.600 0.400 0.200 0.000
B: X2 C: X3
Figure 4A.2 Inverted simplex in L-pseudo-coded mixture space.
came up with a simple solution: Code pseudo-components from the upper,

not the lower, constraint. The key equation is:
u′i − Re al
U_Pseudo =
∑ u′i − 1
If you are into math, compare this to the equation for L-pseudo coding we
provided in the core of the chapter. Software designed for mixture experi-
mentation, such as the program provided with this book, can detect when
upper (“U”) coding will be advantageous. Then, if the user elects to go this
route, it can do the necessary calculations.
Let’s see how U-pseudo coding works on a real-life example. A coatings
chemist designed a mixture experiment on a chemical paint remover for
an aerospace application (Hensley, 2008). The chemical supplier’s material
safety data sheet (MSDS) specified the following constraints for the three
key active ingredients for their recommended formulation, which totaled
12% by weight:
1. 0%–5%
2. 0%–5%
3. 2%–7%
WHAT GOES ON MUST COME OFF

The coatings removal process is often referred to in the field as
“depainting.” Evidently, this sounds more proper than paint “stripping.”
Similarly, when an old structure must be removed before putting up a
new facility, would this be called “debuilding?” Perhaps this terminol-
ogy could become the razer’s edge (pun intended).
Figure 4A.3 shows the resulting region of experimentation for a mixture

design developed with the usual L-pseudo coating.
Aided by computer software the coatings chemist redesigned his experi-
ment using U-pseudo coding, which enabled the application of a second-
degree simplex-lattice, augmented with the overall centroid plus three check
blends—illustrated in Figure 4A.4. Notice that the vertices and centroid are
replicated (designated by the number “2”).
This “flip” from lower to upper coding doubled the ranges—they had
been 0 to 0.5 in L-pseudo, but now (after converting to U-pseudo) all the
components go from 0 to 1. Statisticians delving into an evaluation of
this design on a “before and after” basis will be most impressed by the
decreased collinearity in the matrix and among the model coefficients
A: A
0.000
1.000
0.200
0.800
0.400
0.600
0.600
0.400
0.800
0.200
1.000
0.000
1.000 0.800 0.600 0.400 0.200 0.000
B: B C: C
Figure 4A.3 Paint remover experiment constrained within L-pseudo-coded mixture

space.
A: A
0.000
1.000 2
0.250
0.750
0.500
0.500
2
0.750
0.250
2 2 1.000
0.000
1.000 0.750 0.500 0.250 0.000
B: B C: C
Figure 4A.4 Paint-remover experiment converted to U-pseudo-coded mixture space.
(see sidebar for details). However, the payoff will be far more visible via the
larger region of experimentation displayed in the response graphics.
Beware when working in this upside-down world of U-pseudo coding:
Low becomes high, and high becomes low on the ternary diagrams. For
example, look carefully at the triangle in Figure 4A.5, which displays the
A: A
0.00
2
2
7.00 115 5.00
120
110 2
125
105 130
X1 5.00
X2 1.93
X3 5.07
100
2 2
2
0.00 5.00 2.00
B: B C: C
Removal
Figure 4A.5 Paint-remover results.

response surface contours that resulted from the paint-remover experiment.

(They predict a time in minutes to take off the coating entirely—the lower
the value, the better.) Notice how ingredient A goes from 0% at the top to
5% at the bottom. The other component axes are also reversed.
The happy ending to this story is that the formulator came up with an
optimal recipe (flagged in Figure 4A.5) that dramatically reduced the pro-
cessing time for stripping the paint, thus minimizing hazards due to chemi-
cal exposures.
STATISTICAL PROPERTIES GAINED BY U-PSEUDO CODING

The variance inflation factor (VIF) quantifies the inflation in error for
estimating model coefficients that are caused by the correlation of any
individual model term with all the others (see RSM Simplified, 2nd ed.,
Chapter 2, p. 33–34). Ideally, all VIFs equal 1, that is, there is no infla-
tion of variance whatsoever. When this occurs, the experimental design
matrix is said to be orthogonal, which provides maximum power for
estimating each coefficient. Unfortunately, due to the imposition of a
fixed total of the ingredients, mixture designs can never be orthogonal.
However, VIFs can still be useful for comparative purposes.
For example, let’s look at the paint remover case. In actual units,
the VIFs range from 55 to 162 per a computer-aided design evaluation.
Coding the matrix to L-pseudo-components helps–this reduces the VIF
range to 54–91. However, the significant improvement comes from recod-
ing to the upper (U) pseudo-components—a matrix with values of 1.55–
1.72 for the VIFs.
A forewarning: Tight constraints on certain ingredients, such as a
potent catalyst, will create severe multicollinearity, causing VIFs to soar.
Due to physical necessities, nothing can be done to alleviate this. Even
so, do not worry; we have an ace up our sleeve to work around the dif-
ficulty this causes for assessing design power. This card is a graphical
tool called “fraction of design space” (FDS). However, it will be best to
hold off on playing out this hand until we provide a full deck. This will
come under the topic of “sizing” in the final chapter of this book, which
provides practical aspects of mixture experiments.
Chapter 5
Optimal Design to Customize

Your Experiment
The computer can’t tell you the emotional story. It can give you the
exact mathematical design, but what’s missing is the eyebrows.
—Frank Zappa
In Chapter 4, we detailed a set of constraints to keep individual ingredients

within reasonable limits. For example, the recipe for the Harvey Wallbanger’s
cocktail requires at least 6 ounces of orange juice in the 10-ounce highball,
thus limiting the alcoholic content coming from the vodka and Galliano.
Fortuitously, after accounting for constraints on these other two ingredients,
the resulting experimental region remains a simplex, thus accommodating
standard mixture designs. In most cases, though, the limits imposed by
formulators create a nonsimplex shape such as the one in Figure 5.1, where
ingredient A is cut off at 60%.
Dealing with nonsimplex constraints is our focus for applying optimal
designs. However, these designs, typically computer-generated, can be
customized to any shape for any order of polynomial desired by the
experiment. Thus, for example, with the proper software, you could build a
special cubic “custom design” within a simplex geometry. Nevertheless, to
keep things more interesting, we will restrict our discussion in this chapter
to optimal design within nonsimplex regions.
87
A: A
0
100
25
75
2 2
50
50
2
75
25
2 2 100
0
100 75 50 25 0
B: B C: C
Figure 5.1 Nonsimplex created by truncating component A.
THE TRAPPINGS OF OUR TRAPEZOID

The truncated triangle in Figure 5.1 forms a trapezoid or trapezium,
depending on whether you go by American or British terminology,
respectively. We decorated it with 18 points to provoke thought on how
a mixture design might be fitted to this nonsimplex shape. It pays first to
cover the corners, or vertices, in this case with 2 points apiece to provide
estimates of pure error. We also replicated a blend at the middle of the
experimental region—the centroid. The next step is considering the mid-
points of the edges. As a final touch, we put in points midway between
the centroid and the vertices—these serve as check blends to fill out the
space. This handmade layout of experimental formulations is simple, but
it literally (by the definition below) provides a table upon which to build
this chapter on optimal design. Things will get more complicated from
here on, particularly beyond three-component designs.
From Late Greek trapezoeides, noun use by Euclid of Greek trape-

zoeides ‘trapezium-shaped,’ from trapeza, literally “table.”
—trapezoid (n.)—Online Etymology Dictionary
www.etymonline.com
Optimal Design to Customize Your Experiment ◾ 89
Extreme Vertices Design: Shampoo Experiment

Now we will explore data from a case with complex constraints, that is, nonsim-
plex shaped. This design was created in the early 1980s by one of the coauthors
(Pat) without the aid of computer software. You will get to share the pain of
working through, like him, by hand—all in the interest of education, of course.
Cosmetic chemists needed help for an experiment to optimize foam
height in a shampoo. They identified three key surfactants and specified
constraints (in percentages by weight) on each of them, as follows:
1. Triethanolamine (TEA)–Lauryl Sulfate: 20%–30%

2. Cocamide DEA (diethanolamine): 1%–7%.
3. Lauramide DEA: 1%–3%.
They required that the sum of these three components come to a constant
of 30% of the shampoo. In other words, 70% of the formulation (water, thick-
eners, preservatives, etc.) remained fixed. The chemists focused only on one
response, which is the foam height—measured in millimeters (mm).
WASH THAT FOAM RIGHT OUT OF YOUR HAIR

The ultimate test on foam is the “half-head test” because it provides a
real-live one-to-one comparison of a newly developed shampoo versus a
reference product. A representative subject applies the two products on
either side of their head. They then evaluate on a sensory scale (typically
1–9) the feel of the foam, how it looks and what it takes to rinse off the
shampoo, all under the watchful eye of the cosmetic chemists.
Source: Ansmann, A. et al., Personal care formulations. Chapter 7 of the

Handbook of Detergents: Part D: Formulation, M. S. Showell (Ed.). CRC
Press, 2005, p. 248. This book also provides an excellent overview of
“Statistical mixture design for optimization of detergent formulations,” in
Chapter 2 by S. S. Ashrawi and G. A. Smith.
When the ranges of the factors are not equal such as in this case with low to
high differences of 10%, 6%, and 2% for A, B, and C, respectively, the design
space will not be a simplex. It turns out that the constraints for the shampoo
experiment create a four-cornered shape defined by the vertices laid out in
Table 5.1, which we identified in actual component levels via the algorithm
Table 5.1 Vertices for Shampoo Experiment

Vertex Actuals (%) Reals (Fraction of Total)
# A(X1) B(X2) C(X3) x1 x2 x3
1 28 1 1 0.933 0.033 0.033
2 22 7 1 0.733 0.233 0.033
3 20 7 3 0.667 0.233 0.100
4 26 1 3 0.867 0.033 0.100
detailed in Appendix 5A. This table also provides the compositions as a frac-
tion of the total (30%)—then referred to a “reals.”
When scaled to reals, the vertices can be plotted on ternary paper so the
experimental region can be laid out and additional points around and inside
it plotted. Figure 5.2 provides the picture of the plot.
This experiment design looks a bit cramped in the rectangular
“penthouse” of the triangular structure. It needs to be converted to pseudo-
component coding, which we introduced in Chapter 4. See Figure 5.3 for
the resulting expansion of space on the ternary plot and blends filled in
along the edges (the four midpoints) and the interior (overall centroid and
four axial checkpoints). For added precision in fitting, all the blends were
replicated in the final randomized plan.
The result of this experiment, for what it’s worth (a lot to the formula-
tors, but not for our discussion on optimal design), is displayed in Figure 5.4,
x1
1
90
4
2
70
3
50
10
10
30
30
30
50
50
70
70
10
90
90
x2 x3
Figure 5.2 Extreme vertices for shampoo formulation plotted in reals.

A: TEA-LS (wt%)
0.000
1.000 2
2 0.250
0.750 2
2 2
2 0.500
0.500
2 2
2 2 0.750
0.250
2 1.000
0.000
1.000 0.750 0.500 0.250 0.000
B: Cocamide (wt%) C: Lauramide (wt%)
Figure 5.3 Shampoo design in pseudo-component space.
A: TEA-LS (wt%)
1
28 2
150
160 2
2 170
3
26 2
Prediction 177.1
X1 24
2 X2 3
X3 2
5
24 2
2 2
160 7
22 2 2
150
2
140
9
20 2
9 7 5 3 1
Figure 5.4 Contour plot of foam height (mm) from shampoo experiment.
with the axis labeling and gridlines reverted conveniently to actual scale.
The optimum foam height in mm is flagged and identified with the compo-
sitional coordinates.
This case study on the shampoo experiment provides a good start on
how to contend with constraints that do not form a simplex. Keep in mind
that, this being done before the availability of software for mixture DOE,
the cosmetic chemists had to go through all the work to lay out the design
graphically, decode the point coordinates back to actual component levels,
and randomize the order before embarking on the experiment. Nowadays
specialized programs, such as the one that accompanies this book, provide
optimal custom designs that do all of this for formulators. Figure 5.5 presents
the shampoo design the way it would be done with these state-of-the-art
computerized tools. In this case, we kept it simple by doing the building
via an algorithm called “point exchange.” We will provide details on its
construction, and a more-sophisticated option called “coordinate exchange,”
later—after covering optimal design.
This may look a bit odd, compared to the original handmade
experiment design. Keep in mind, though, that computers care nothing for
what pleases the eye—only what the algorithmic code drives them to do.
Without the utilization of software tools like this, it is daunting to design
A: TEA-LS (wt%)
1
28 2
3
26 2
6
5
24
2
7
22
2
9
20
9 7 5 3 1
Figure 5.5 Shampoo experiment rebuilt using optimal custom-design tools in

computer software.
mixture experiments with three components, unreasonable to embark

upon with four, and silly to even contemplate beyond that. Let’s now turn
our attention to what makes a good custom design and the generally
accepted statistical criteria for optimality.
Optimal Designs Customize Your Experiment

as You See Fit for Any Feasible Region
We advise using ASL designs, spelled out in Chapter 3, for all mixture exper-
iments involving 5 or fewer components that form a simplex as-is or inverted
(detailed by Appendix 4A). If you want to study 6 or more components, we
suggest that, before embarking on optimization, you consider first doing a
screening experiment using designs that will be provided later in the book.
For all other experiments (i.e., nonsimplex and/or not models of degree 1, 2,
or 3), use a statistically optimal design—in particular, a specific type called
“I-optimal” that is best suited for formulation improvement.
IF YOU WANT TO LEARN THE ALPHABET

OF OPTIMALITY CRITERIA
Since the beginnings of optimal design half a century ago by Box and
Lucas, (1959) industrial statisticians have developed many matrix-based
criteria which they label alphabetically. For DOE and RSM the favored
few are D and I, respectively. Since mixture design is mainly used to
optimize formulations, we will use I-optimality exclusively. This crite-
rion minimizes the average prediction error of the design across the
region of experimentation—a very suitable goal for mapping out a for-
mulation space.
For details on D versus I refer to “Practical Aspects for Designing
Statistically Optimal Experiments” (Anderson and Whitcomb, 2014). Also,
see Chapter 4 of RSM Simplified, 2nd ed., for in-depth coverage on these
computer-generated run-layouts. There you are provided with details
on “Optimality from A to Z,” a checklist on what these design-building
(Continued )
criteria aim to do, and helpful hints on Getting the Computer to be “More
Open-Minded About What’s Optimal.”
Our mission now is to provide formulators like you just enough infor-
mation (i.e., a “need-to-know” basis) to feel comfortable using optimal
design. You are likely to discover that your constraints will not create
a simplex. Thus these custom design tools will be a godsend. Just do
not get hung up on the details of their development and mathematical
construction.
No postulated model can ever be assumed to be fully correct [there-

fore], regarding the basic assumptions underlying the alphabetic-
optimality approach, which are often unrealistic from the practical
viewpoint of actually designing real experiments.
—Draper and Guttman (1988)
Take This Tableting Case-Study Now and Call

the Statistician in the Morning
Ever since Bayer invented aspirin in 1897, medical doctors have been telling
their patients to “take two and call me in the morning.” We hope to keep
you from getting a headache by detailing optimal design for mixture experi-
ments via a case study, which, as it happens, involves the formulation of a
tablet.
This case study on optimal design is loosely based on an example by
Lewis et al. (Lewis, G.A., D. Mathieu, and R. Pahn-Tan-Luu, Pharmaceutical
Experiment Design, Chapter 10, “Mixtures in a constrained region of interest,”
Section II, “Second example: A hydrophilic matrix tablet of 4 variable
components,” 2008, Informa Healthcare, USA, Inc.). The formulators hoped
to achieve a sustained release of an active pharmaceutical ingredient
(API) tablet. The key ingredient is a hydrophilic polymer that swells in
the presence of water and, thus, impedes drug release. By experimenting

on the proportion of this polymer to three excipients—inactive substances
that served as the drug vehicle—they developed a handle on the rate of API
release, as well as vital physical properties of the tablet.
Ranges regarding weight percent are laid out here for the four compo-
nents to be varied:
1. Lactose (“Lact”), 5%–42%

2. Calcium phosphate (“CaPh”), 5%–47%
3. Microcrystalline cellulose (“Cell”), 5%–52%
4. Hydrophilic polymer (“Poly”), 17%–25%
These ingredients were kept to a total of 98% to accommodate a fixed

concentration of the API (“Drug”) at 2%.
The first step for designing a mixture experiment is to define the
experimental region by running the ranges through an extreme-vertices
algorithm, such as the one embedded in the software accompanying this
book. We advise entering them in the simplex-lattice design-builder, which,
in this case, recognizes that the mixture space does not form a simplex.
The program then suggests an optimal design for customizing the experiment
to the nonsimplex space.
Optimal Design Simplified

With the aid of software (plus being able to slide by the mathemati-
cal and computational details), it is quite easy to generate an optimal
experiment design. All you need to do is specify the model that
the design must be capable of fitting. For the most part you will do
well by choosing a quadratic—a model that characterizes nonlinear
blending, that is, antagonism or synergism between components. Your
chosen model then dictates how many unique blends will be required
in the mixture design—for the most part, one for each term that must
be estimated.
In this case, the tablet formulators went with this standard selection
of a quadratic model, which featured 10 terms—4 for the main compo-
nent effects (A, B, C, and D) plus 6 for nonlinear blending (AB, AC, AD,
BC, BD, and CD). To fit these 10 terms, 10 unique blends are required.
This is where the I-optimal criterion comes into play. It will be the judge
by which blends to run from within the mixture is spaced. Although
this can be done with few geometric restrictions via a method called
“coordinate exchange” (detailed in sidebar “Candidate-free Approaches for
Exact Optimal Designs,” Appendix 7C, RSM Simplified, 2nd ed.), let’s keep
things simple by identifying a discrete number of points as candidates for
the optimal selection. This is called the “point exchange” method. The
only requirement is that the candidate set exceeds the number of blends
required to fit the model.
The starting point for building up a good candidate set is the extreme
vertices. These being far apart will provide the highest leverage for
fitting the main component effects. There are 12 extreme vertices in the
mixture space for the tablet experiment. To fit the nonlinear terms, it is
helpful to also include centers of edges (binary blends)—18 in this case.
One more point to include is the overall centroid—a complete blend.
This brings the total to 31. Furthermore, to fill in the gaps between the
centroid and vertices, let’s add the 12 axial check blends. This brings
the total to 43 candidate points—well beyond the 10 needed for the
quadratic mixture model.
From this set of candidate points, we laid out (with the aid of the soft-
ware) the 20-run custom design shown in Table 5.2. (Note that, to make it
simpler to see that the components add up to 100%, we included the API
(shaded to highlight it being fixed) and expressed the fractions). It is com-
prised of the following blends:
◾ 10 for the model chosen I-optimally.

◾ 5 to test lack of fit (“LOF”) selected by a distance-based algorithm (refer
to the sidebar “Getting the Computer to be More Open-minded About
What’s Optimal” in Chapter 7 of RSM Simplified, 2nd ed.) to fill the
component space.
◾ 5 replicates (“Rep”) for estimation of pure error (needed for the LOF
test) chosen by the same optimal criterion as that used for selecting the
model points—these being the points creating the maximum impact for
improving the fit.
Table 5.2 Custom Design of Experiment on Tablet Formulation

A: Lact B: CaPh C: Cell D: Poly E: Drug
# Build Space (wt%) (wt%) (wt%) (wt%) (wt%)
1 Model Vertex 42 34 5 17 2
7 Model CentEdge 42 5 30 21 2
10 Model Center 231/6 252/3 281/6 21 2
11 LOF CentEdge 42 191/2 191/2 17 2
12 LOF CentEdge 5 311/2 361/2 25 2
13 LOF AxialCB 261/12 364/12 167/12 19 2
14 LOF AxialCB 141/12 364/12 247/12 23 2
15 LOF AxialCB 197/12 154/12 401/12 23 2
16 Rep Vertex 42 26 5 25 2
17 Rep CentEdge 42 5 30 21 2
18 Rep CentEdge 25 47 5 21 2
19 Rep CentEdge 5 20 52 21 2
20 Rep Center 231/6 252/3 281/6 21 2
Take-Home Advice on Deploying Optimal Design

This concludes our briefing on optimal design. To keep things simple,
when you set up a mixture experiment with the aid of software that
provides the proper tools to get this job done, start by setting it up as
a simplex. When warned that your constraints make this impossible, shift
to the program’s custom builder using optimality (I-best, but D-acceptable)

for selection of your model points. The point exchange option provides
compositions from specific geometric locations within your experimen-
tal space, whereas coordinate exchange roams relatively freely—either
selection serves well, so we suggest trying them both and picking
whichever layout works best as a practical matter. Provided the minimal
design (model points only) does not exhaust your budget of time, cost
or material consumption, always add 3–5 extra “check” blends for LOF
testing and replicate 3–5 of these and the others (picked for the model)
to provide an estimate of pure error (needed for LOF).
SPECIFICATIONS FOR CUSTOM DESIGN ON SHAMPOO

Now that we’ve covered optimal design, here are details on how we
produced the layout of points in Figure 5.5. It was set up with I-optimal
criterion for a quadratic model via point exchange. To match the original
design for the precision of fit, we added 8 model points beyond the 6
required to fit the model. The design was then augmented with 5 lack-
of-fit points and 5 replicates. Due to random elements in optimal algo-
rithms, you may not get an exact match to our design, but it should be
more like it than the original handmade experiment.
Practice Problem
Problem 5.1
For a fun do-it-yourself experiment that involves optimal design for
mixtures, we recommend you make play putty per the recipe posted at
www.statease.com/publications/marks-play-putty-experiment (also refer to
Anderson and Whitcomb, 2002). We chose the following components and
levels to make up 100 milliliters (mL) of this silly stuff for our experiment:
1. Glue, 40–59 mL
2. Water, 40–59 mL
3. Borax, 1–3 mL
Enter these constraints in your mixture-design software or plot them

out by hand on a ternary diagram. Do they form a simplex? Set up a
design and compare it to the one in the referenced publication. Follow
up by developing your own recipe for play putty. That will be fun and
educational.
SILICONE SUPER BALL

A primary attribute of good play putty is a high bounce. The best we
achieved was a bit over one foot in height. If you want to make a far
bouncier ball, mix 5–10 mL (experiment!) with 20 mL of sodium silicate
solution (water glass) and follow the instructions posted at www.flinnsci.
com/silicone-super-ball/dc07585/ by a supplier of these chemicals—Flinn
Scientific. A variation on the recipe (http://chemicalrecipes.blogspot.
com/2010/05/making-superball.html) calls for replacing the alcohol with
phenolphthalein to color the ball bright pink.
Appendix 5A: An Algorithm for Finding Vertices

Here is a relatively simple way to identify vertices in a nonsimplex mixture
experiment—McLean and Anderson’s extreme vertices algorithm (McLean
and Anderson, 1966):
1. Make a table of all possible combinations of the lower individual (Li)

and upper individual (Ui) limits (as you would do for a two-level full
factorial) for one less than the number (q) of the components (i.e., q−1).
Leave the value for the remaining component (the qth one) blank. This
table will consist of q*2q−1 possible combinations.
2. Fill in the acceptable combinations, that is, those that don’t violate indi-
vidual or total constraints (TC).
3. Eliminate duplicate combinations.
For the three-components shampoo case, the first step of this algorithm
dictates the construction of three two-by-two tables that spell out the four
extreme combinations for A versus B, A versus C, and B versus C. These are
shown from left to right, respectively, in Table 5.3.
Table 5.3 Candidate Vertices for Shampoo Experiment

A B C A B C A B C
20 1 >U3 20 >U2 1 28 1 1
30 1 >TC 30 >TC 1 22 7 1
20 7 3 20 7 3 26 1 3
30 7 >TC 30 >TC 3 20 7 3
Six combinations out of the twelve meet all constraints, of which four are
unique (blend (20, 7, 3), which are repeated several times)—the ones boxed
by thicker lines. The combinations fail either due to being outside of the
individual component limits (Li−Ui) or the total constraint (TC). For example,
if you blended up 20% of A with 1% of B, then, to meet the total of 30%, 9%
of C would be needed to make up the difference. However, the cosmetic
chemists restricted C to a maximum of only 3%. Therefore, this combination
does not satisfy the rules of the algorithm. Other combinations get thrown
out straightaway due to one component taking up the entire 30% allowable
total—leaving no room for anything else.
More efficient algorithms have been developed since McLean and
Anderson invented theirs, most notably XVERT by Snee and Marquardt
(1974) and Piepel’s CONVRT algorithm (Piepel, 1988). The code is available
in R to compute extreme vertices and other points in nonsimplex spaces—
see Lawson and Cameron’s “Mixture Experiments in R Using mixexp”
(Lawson and Willden, 2016. Code Snippet 2).
Chapter 6
Getting Crafty with

Multicomponent Constraints
No great discovery was ever made without a bold guess.

—Isaac Newton (attribution: p. 298, Presidential
Address: The Nature of Discovery, A. Blalock—
Annals of Surgery, 1956. www.ncbi.nlm.nih.gov/pmc/
articles/PMC1465405/pdf/annsurg01270-0005.pdf)
Being bold in setting-up component levels pays off by generating more

significant effects from mixture experiments. However, extreme combinations
may create an infeasible product, for example, bread that fails to rise due to too
little yeast at high salt levels. By imposing multicomponent constraints (MCCs),
formulators can lay down boundaries that keep their region of experimentation
within reason, but as broad as possible.
How Multicomponent Constraints Differ from Simple Ones

Ron Snee in a pioneering publication (Snee, 1979) presented this set of simple
limits on three ingredients:
1. 0.1 ≤ X1 ≤ 0.5
2. 0.1 ≤ X2 ≤ 0.7
3. 0.0 ≤ X3 ≤ 0.7
101
These constraints translate to the following five boundaries on the ternary

diagram shown in Figure 6.1:
1. 0.1 ≤ A
2. A ≤ 0.5
3. 0.1 ≤ B
4. B ≤ 0.7
5. C ≤ 0.7
Note that the boundaries of this experimental region (in grey) are parallel to
the triangular perimeter of the ternary plot or on its edge. Now comes the
tricky part—two multicomponent constraints that depend on a combination
of ingredients:
◾ 0.90 ≤ 0.85A + 0.90B + 1.00C ≤ 0.95
◾ 0.4 ≤ 0.7A + 1.0C ≤ 1.0
These additional (MCC) constraints translate to the following three additional
boundaries (numbers 6, 7, and 8) displayed in Figure 6.2.
6. 0.90 ≤ 0.85A + 0.90B + 1.00C
7. 0.85A + 0.90B + 1.00C ≤ 0.95
8. 0.4 ≤ 0.7A + 1.0C
These new boundaries are not parallel to the sides of the ternary diagram.
Thus, they allow for a more conservative experimental region that is crafted
by the subject matter expert, in this case, a chemist.
X1
B
.1 ≤
90
3. 0
70
50 2. A ≤ 0.5
10
10
.7
≤0
30
30
4. B
30
50
50
70
70
10 1. 0.1 ≤ A
90
90
X2 5. C ≤ 0.7 X3
Figure 6.1 Simple constraints for Snee example.

Getting Crafty with Multicomponent Constraints ◾ 103
X1
3 90
70
6
8
50 2
10
10
30
30
7
4 30
50
50
70
70
10 1
90
90
X2 5 X3
Figure 6.2 Multicomponent constraints added for Snee example.
HOW MULTICOMPONENT CONSTRAINTS

ARISE IN REACTIVE FORMULATIONS
Cornell (2002) in his detailing of MCCs (Section 4.10, p. 180), lays out
this one for a metallurgy experiment that required maintaining at least
46 wt% iron in the furnace. However, five different feedstocks contained
at least some iron, ranging from 20% to 60%. This limitation imposed on
the total iron was maintained via the following constraint:
0.46 ≤ 0.6x1 + 0.5x 2 + 0.35x 3 + 0.2x 4 + 0.7x 5
Along similar lines, a petroleum engineer must maintain carbon within a

certain range for distillation of fuel oils from many feedstocks that vary
in carbon content.
The three MCCs (6, 7, and 8) presented by Snee probably stemmed
from comparable requirements for maintaining a proper chemical reaction
by controlling particular elements within specific bounds.
There exists everywhere a medium in things, determined by

equilibrium.
—Dmitri Mendeleev (The Principles of Chemistry,
1891, 1, 257)
MCCs Made as Easy as Making a Pound Cake

The Pound Cake arose (pun intended) in Britain in the early 1700s (“Pound
Cake History,” What’s Cooking America, https://whatscookingamerica.net/
History/Cakes/PoundCake.htm). Being baked from one pound each of butter,
sugar, eggs, and flour, it created a very large confection—serving multiple
families.
After teaching statistical-process-control to master-bakers of Sara Lee®
brand Pound Cakes (all butter, lemon, and so on), Mark resolved to experi-
ment on the composition, not believing it is needed to be so rigidly bal-
anced between the four basic ingredients. Perhaps a slight variation might
lead to better taste.
Mark’s quest for perfect Pound Cake began with research into The Cake
Bible, first published by Rose Levy Beranbaum in 1988 (William Morrow
Cookbooks). Straightaway, he learned that low-protein “cake” flour creates a
desirably lighter consistency, that is, an airier texture, than all-purpose (AP)
flour. However, this special flour costs about twice as much per pound as
its more common AP cousin. If some, or, ideally, all, of the cake flour, could
be replaced by AP flour without anyone noticing the difference, then Pound
Cakes could be made far cheaper. This made a great case for application of
MCCs, as you will soon see.
A DIFFERENT SPIN ON THE DANGERS OF GLUTEN IN FLOUR

Mark started off his career working for General Mills in Minneapolis,
in the center of the United States’ “wheat belt.” During this time, he
worked for some time on the startup of a starch–gluten plant, which fed
off coarsely ground spring wheat from the western part of Minnesota
and the vast fields of South and North Dakota. His primary objective
was to increase the yield of the protein-rich gluten. If you take a hand-
ful of flour under running water, the starch runs off, leaving sticky
gluten behind. Similarly, by mixing flour with water and running it
over a stainless-steel bar screen, the starch goes through, while the
gluten agglomerates on the surface and, in theory, rolls nicely down a
chute to be dried into a high-protein powder. Unfortunately, the gluten
did not always cooperate. It tended to stick together into evergrowing
blobs that at times got away from Mark and his crew by rolling off the
(Continued )
bar-screen balcony. This created a crisis when the gluten blob got going
toward the waste-water treatment plant and threatened to overwhelm
it with protein. Only by the last stand of push-broom wielding workers
was the disaster averted.
I think you should send us the biggest transport plane you have, and
take this thing to the Arctic or somewhere and drop it where it will
never thaw.
—Lieutenant Dave, The Blob, 1958, Source: IMDB
http://www.imdb.com/title/tt0051418/quotes
Mark laid out these ingredients and ranges for the Pound-Cake mixture
design, all in ounces (oz.) by weight:
1. Cake flour, 3–5 oz

2. AP flour, 3–5 oz
3. Sugar, 3–5 oz
4. Butter, 3–5 oz
5. Eggs, 3–5 oz
He kept the total 16 ounces, that is, one pound (1 lb), which, being
one-fourth the size of the traditional Pound Cake, fit nicely into four-cavity,
nonstick, mini-loaf pans for baking in his kitchen oven. But most importantly,
to avoid an overdose of flour, Mark specified the following MCC:
3≤ A+B≤5
This created a little experiment with flour types (cake versus AP) within the
greater experiment on the recipe for Pound Cake (relative amounts of flour,
sugar, butter, and eggs). Table 6.1 lays out 12 out of 24 mixtures from an
augmented I-optimal design to fit a quadratic model. It provides 15 blends
to fit the model, plus 4 lack-of-fit points (check blends) and 5 replicates (for
pure error estimation). The total of 24 runs breaks down conveniently into 6
four-cavity pans that can be made in one oven batch.
Notice from the selected runs under columns A and B in Table 6.1 how
the MLC keeps the total of the cake and AP flour within the bounds of
3–5 oz. Success!
Table 6.1 Selected Runs (Weights in Ounces) from Pound-Cake

Experiment-Design
ID Build Type A: Cake B: AP Flour C: Sugar D: Butter E: Eggs
1 Model 2.5 2.5 5.0 3.0 3.0
2 Model 0.0 5.0 4.0 4.0 3.0
3 Replicate 4.0 0.0 5.0 4.0 3.0
3 Model 4.0 0.0 5.0 4.0 3.0
4 Model 5.0 0.0 3.0 5.0 3.0
5 Lack of fit 1.5 1.5 5.0 5.0 3.0
~
14 Model 0.0 5.0 3.0 3.0 5.0
15 Model 4.0 0.0 4.0 3.0 5.0
16 Lack of fit 1.5 1.5 5.0 3.0 5.0
17 Model 4.0 0.0 3.0 4.0 5.0
18 Model 0.0 3.0 4.0 4.0 5.0
19 Lack of fit 1.5 1.5 3.0 5.0 5.0
RESULTS FROM THE POUND-CAKE EXPERIMENT

Although our job is done for laying out a case for MLCs for substituting
materials, you may be curious to know what Mark came up with for an
ideal recipe. The published* one calls for only AP flour (4.3 oz), with lots
of sugar (4.9 oz), a generous amount of butter (4.3 oz) and just enough
eggs (3.8 oz) to make up the remainder of the 16-ounce Pound Cake.
This hits the spot of taste, density, and color, albeit falling well short of
Sara Lee for overall scrumptiousness, heft, and appearance. A recipe that
maximizes the taste alone is provided in Table 6.2. It minimizes the eggs
and flour (AP only) while maximizing butter and sugar. Yum!
By the way, neither recipe requires any of the expensive cake flour.
That’s a bonus!
*Anderson, M., and P. Whitcomb. Mixing it up with computer-aided

design. Today’s Chemist at Work, November 1997b, p. 34.
(Continued )
Table 6.2 Recipe for Pound Cake That Maximizes Taste

Flour (AP Only) 3 oz
Sugar 5 oz
Butter 5 oz
Eggs 3 oz
Ratio Constraints
In some mixture problems, the ratios of components must be carefully
controlled. For example, bakers of bread do well by blending a ratio of
5 parts of flour to 3 parts of water (Ruhlman, 2009). In some cases, these
ratios relate to the ideal stoichiometry for chemical reactions, such as the
air-to-fuel ratio in a combustion engine, which, depending on the grade
of gasoline, runs at about 15-to-1 [“Air-fuel Requirement in SI Engines
(Automobile)],” what-when-how, (http://what-when-how.com/automobile/
air-fuel-requirement-in-si-engines-automobile/). These ratios generally can
be mathematically converted to MLCs as you will see now via a real-life
example.
An adhesive’s chemist ran an optimal mixture design to model and
control gel time for a liquid epoxy while maintaining other functional prop-
erties (Roesler, 2004). He varied three key components in a mixture, totaling
100% by weight:
1. Amine cross-linker: 20%–100% by weight

2. Plasticizer: 0%–70%
3. Catalyst: 0%–35%
The chemist imposed an additional constraint: The catalyst to plasticizer

does not exceed a 4:1 ratio, that is, C/B ≤ 4. This can be rearranged into an
additive (non-ratio) MLC by multiplying both sides by A and subtracting one
side from the other:
C ≤ 4B
0 ≤ 4B − C or, alternatively, − 4B + C ≤ 0
A: Amine cross-linker (wt%)

100
2
2
2 X2 8.75
X3 35
0 0
2 2
80 20 80
B: Plasticizer (wt%) C: Catalyst (wt%)
Figure 6.3 Adhesive experiment with ratio constraint.
An experiment fitting within these individual component constraints and

the MLC is pictured in Figure 6.3. It was customized via I-optimal point
exchange to fit a quadratic mixture model.
The edge at the upper right is flagged to show that this run (35 and 8.75
for X2 and X3, respectively) marks the edge of the feasible region where the
B/C ratio hits its limit of 4.
Appendix 6A: Combining Components

Components having similar effects are not unusual, particularly when an
experiment, facilitated by multicomponent constraints, includes substi-
tutes. In that case, combining components will simplify the modeling and
make the presentation of results more compelling. This turned out to be
the case for the Pound-Cake experiment where two flours came into play.
To illustrate why combining the flours makes sense and how this clari-
fies the picture, it will be helpful to introduce the “trace” plot. This plot,
which we will feature in Chapter 8 on screening designs, displays the
6
C
Taste (rating)
5 A
B
D
4 E
E
D B
3 A
C
2
−0.300 −0.200 −0.100 0.000 0.100 0.200 0.300

Figure 6A.1 Trace plot of taste from a Pound-Cake experiment with both flours
shown.
predicted response as any given component deviates from any chosen

reference point (typically the centroid) while holding all other compo-
nents in constant proportion. Figure 6A.1 displays the trace plot for the
Pound-Cake experiment.
Notice how the tracks for components A and B, the two flours, run along-
side each other, which makes sense from a culinary perspective and it was
what Mark expected to find. The same overlay of flours occurred for other
attributes he measured—density and color. Seeing his hypothesis that flours
make little difference confirmed, he then combined the two components
into one, as shown in Table 6A.1.
This simplifies the story as illustrated by the revised trace plot in
Figure 6A.2 (we expanded the axis to make the differing impacts of each
main ingredient more obvious).
Now it becomes crystal clear that more sugar drives up the taste.
Furthermore, flour, regardless of its type, is a “downer,” as well as eggs, but
to a lesser extent. Butter falls in a neutral zone, but, perhaps it may be good
to be somewhat liberal in its addition to the recipe.
Table 6A.1 Selected Runs (Weights in Ounces) from Pound-Cake Experiment-Design

ID Flour Sugar Butter Eggs
1 5.0 5.0 3.0 3.0
2 5.0 4.0 4.0 3.0
3 4.0 5.0 4.0 3.0
3 4.0 5.0 4.0 3.0
4 5.0 3.0 5.0 3.0
5 3.0 5.0 5.0 3.0
~
14 5.0 3.0 3.0 5.0
15 4.0 4.0 3.0 5.0
16 3.0 5.0 3.0 5.0
17 4.0 3.0 4.0 5.0
18 3.0 4.0 4.0 5.0
19 3.0 3.0 5.0 5.0
Sugar
5
Taste (rating)
Butter
4
Eggs
3 Flour
−0.300 −0.200 −0.100 0.000 0.100 0.200 0.300

Figure 6A.2 Trace plot redone after combining the two flours.
Keep in mind that this trick of combining components only comes into
play in the final analysis as a simplification. It should be done judiciously
on the basis of subject-matter-knowledge, that is, not on disparate materials
that coincidentally create similar effects, but only on ones that are chemically
similar.
Chapter 7
Multiple Response
Optimization Hits the Spot
If your sweet spot is in the right place, there will be no vibration

through the handle, and you will hear the sweetest sound in cricket,
the Crack! of leather on willow.
—Blackledge (2012)
Formulating a functional product can be extremely challenging, due to

competing specifications. For example, consider a coating that must be soft
to the touch, yet resistant to abrasion. Paint chemists find that reducing
the amount of cross-linker achieves a velvety feel. However, it significantly
degrades the durability of the coating (Spagnola et al., 2016). By deploying a
mixture design that provides adequate models for critical attributes (such as
the aforementioned), formulators often discover a region where components
combine to meet all specifications. This is known as the “sweet spot.” A sim-
ple, yet effective, way to gauge when you hit this spot for multiple responses
is by scaling them all to one objective function called “desirability.”
113
Desirability Simplified
Being on a roll with paint (pun intended), let’s see how desirability works
by way of a case study from the coatings industry. In this case (Anderson
and Whitcomb, 1997a), experimenters explored the impact of three rhe-
ology modifiers (“RM”s) on the viscosity and flow of an architectural
coating.
Each of the three components was varied from zero to one-hundred
percent (0%–100%) via a second-degree simplex-lattice design augmented
with the overall centroid and axial check blends. The three vertices were
replicated to provide a measure of pure error. Table 7.1 lays out all of these
blends and the test results. Cost is also included based on a hypothetical
spread of pricing per kilogram of $7.50, $10.00, and $15.00 for RMs A, B,
and C, respectively.
Fitting the rheological data to Scheffé polynomials produces these highly
significant models, quadratic and linear, respectively:
Table 7.1 Experiment on Mixtures of Rheology Modifiers

A:RM-A B:RM-B C:RM-C Visc Flow Cost
ID Space Type (wt%) (wt%) (wt%) Poise Units ($/kg)
1a Vertex 100.00 0.00 0.00 0.76 8 5.00
1b Vertex (rep) 100.00 0.00 0.00 0.75 8 5.00
2 Center edge 50.00 50.00 0.00 1.40 7 7.50
3 Center edge 50.00 0.00 50.00 0.55 8 10.00
4a Vertex 0.00 100.00 0.00 4.10 4 10.00
4b Vertex (rep) 0.00 100.00 0.00 4.40 4 10.00
5 Center edge 0.00 50.00 50.00 0.90 7 12.50
6a Vertex 0.00 0.00 100.00 0.42 9 15.00
6b Vertex (rep) 0.00 0.00 100.00 0.40 10 15.00
7 Axial check 66.67 16.67 16.67 0.80 7 7.50
8 Axial check 16.67 66.67 16.67 1.70 7 10.00
9 Axial check 16.67 16.67 66.67 0.55 8 12.50
10 Centroid 33.33 33.33 33.33 0.80 8 10.00
Multiple Response Optimization Hits the Spot ◾ 115
Viscosity = 0.77A + 4.21B + 0.423C − 4.33AB + 0.31AC − 5.6BC

Flow = 7.97A + 4.57B + 9.37C
The cost is calculated deterministically via a linear equation based on the
composition of each blend.
A BRIEFING ON RHEOLOGICAL PROPERTIES OF PAINT

Rheology studies the flow of liquids, primarily via measurement of vis-
cosity. For fluids that become thinner or thicker under shear (a behavior
known as non-Newtonian), other attributes must be characterized, in this
case, the “process by which a wet paint flows out to a smooth film free of
irregularities, that is, brush marks and ridges.”* The units for flow, shown
in Table 7.1 provide an indication of brush-out leveling on a scale of 1 to
10—the higher the better. For details on the actual measurement of flow,
done via ASTM D4062 “Standard Test Method for Leveling of Paints by
Draw-Down Method,” see www.leneta.com/leveling-test.html.
Rheology modifiers reduce dripping and spattering of paint during

roller or brush application. Sag resistance of paint is improved by
a rapid but controlled viscosity increase after application. During
transport and storage of the paint, the rheology modifiers prevent
sedimentation of the pigments within a formulation.
—Practical guide to rheology modifiers. BASF. www.
dispersions- pigments.basf.com/portal/load/fid793184/
BASF%20Rheology%20Modifiers%20Practical%20Guide.pdf
*Source: Private correspondence to authors on 5/14/01 from Coatings
Chemist Sol Vincent, Rohm and Haas, Spring House, PA.
These models for viscosity and flow provide complete control over the
rheology required for any application needs. For example, assume that the
following specifications must be met:
1. Viscosity: 0.5 to 0.7 poise (target 0.6)
2. Flow: 8.0 units or better on a 10-point scale
3. Cost: $10.00/kg or less ($5.00/kg the cheapest possible)
Figure 7.1a–c shows contour plots for viscosity, flow and cost, respectively.
Each of the three plots displays two or more blend-points that fall within
specification. However, none of them meet all the requirements, the center
of edge A–C blend—ID #5 at composition (50, 0, 50) being a near miss.
A: RM-A (wt%) A: RM-A (wt%)

100.00 100.00
2 2
Prediction 0.68
Prediction 0.67 Prediction 8.67

X1 50.00 X1 50.00
0.7 X2 0.00 X2 0.00
X3 50.00 X3 50.00
8
0.00 0.00 0.00 7.5 0.00
0.9 7
1.1 0.5 8.5
1.3 6.5
1.9 1.5 6 Prediction 8.34
1.7
2.5 2.1
5.5 9
2.7 2.3
3.3 2.9
3.5 3.1 5 Prediction 9.37
3.7 0.3
2 2 2 2
100.00 0.00 100.00 100.00 0.00 100.00
B: RM-B (wt%) C: RM-C (wt%) B: RM-B (wt%) C: RM-C (wt%)
(a) (b)
A: RM-A (wt%)
100.00
2
Observed 5.00
6
Observed 7.50
Observed 10.00
8 X1 50.00
X2 0.00
Observed 7.50 X3 50.00
0.00 0.00
Observed 10.00
10
Observed 10.00
12
Observed 10.00 14
2 2
100.00 0.00 100.00
B: RM-B (wt%) C: RM-C (wt%)
(c)
Figure 7.1 Contour plots for viscosity, flow, and cost (a, b, and c).
A: RM-A (wt%)
100.00
2
Viscosity: 0.64
Flow: 8.30
Cost: 9.05
X1 56.56
X2 5.79
X3 37.64
Viscosity: 0.7
Flow: 8
0.00 0.00
Viscosity: 0.5
Cost: 10
2 2
100.00 0.00 100.00
Figure 7.2 Overlay plot, showing the sweet spot for a blend of rheology modifiers.
The area meeting both specifications becomes far clearer by overlaying the
three plots and then shading out the regions going out of bounds. As shown
in Figure 7.2, this overlay plot provides a window that frames the sweet spot.
The flag marks a better blend than any made for the experiment: 56.56%
RM-A, 5.79% RM-B, and 37.64% RM-C. (As a practical matter, the coatings
chemist would do well by rounding these levels to 56.5%, 5.8%, and 37.7%,
respectively.) This optimal blend beats mixture #5 on two out of the three
specifications—viscosity (a bit closer to target) and cost (much cheaper).
Only on flow does it fall off somewhat—8.30 versus 8.67 for blend #5.
To achieve this optimal blend, all three response measures were rescaled to
one objective function called “desirability” going from 0 (none) to 1 (complete).
Figure 7.3 shows how desirability ramps up and down for viscosity (a target),
up for flow (maximized) and down for cost (minimized). (The number lines
are bounded by the observed extremes printed in smaller font size.)
0.6 (d = 1)
8 (d = 0) 10 (d = 0)
0.5 0.7 (d = 0)
10 (d = 1) 5 (d = 1)
0.40 4.40 4 15.00
Viscosity = 0.64 Flow = 8.30 Cost = 9.05
Figure 7.3 Ramps view of the most desirable blend of rheology modifiers.
An individual’s desirability (small case “d”) and the overall desirability

(upper case “D”) at this optimal blend point are:
d1 = (0.70 − 0.64)/(0.70 − 0.60) = 0.6
d 2 = (8.3 − 8.0)/(10.0 − 8.0) = 0.3/2.0 = 0.15
d 3 = (10.00 − 9.05)/(10.00 − 5.00) = 0.95/5.00 = 0.19
D = (0.6 × 0.15 × 0.19)1/3 = (0.0171)1/3 = 0.26
DETAILS ON DESIRABILITY
For mathematical details on calculating desirability and the attendant
search algorithms for numerical optimization, refer to Chapter 6 of RSM
Simplified, 2nd ed. There, you will also find a background on how to pri-
oritize responses by scaling them in terms of “importance.” For example,
in this case, a coatings chemist could push the selection of the most
desirable blend of rheology modifiers closer to the target on viscosity
by making the response most important—5 on the 5-point scale, while
setting both flow and cost at the least important level of 1.
Whereas economic man maximizes, selects the best alternative from

among all those available to him, his cousin, administrative man,
satisfices, look for a course of action that is satisfactory or “good
enough.
—Simon (1997)
As evidenced by Figure 7.4, the D of 0.26 is the best that can be achieved
for the goals set on the responses. Its absolute value, that is, the overall
desirability not achieving a perfect value of 1, is of no concern.
When every goal is at least minimally met, that is, just inside the lower
and or threshold level, an overall desirability above zero is achieved. This
“satisfices” (a la the quote by Simon in the sidebar), rather than optimizes.
“Good enough.”
A: RM-A (wt%)
0.1
Desirability 0.26
0.25
0.2 0.15
Figure 7.4 Contour plot of desirability (magnified view).
Framing the Sweet Spot and Narrowing It Down

to Achieve Quality by Design (QbD)
The overlay plot provides a window into regions where experimenters can
meet specifications for multiple responses. Thus, it facilitates the develop-
ment of a QbD “design space,” which is defined by the Food and Drug
Administration (FDA) as the “multidimensional combination and interac-
tion of input variables (e.g., material attributes) and process parameters that
have demonstrated to provide assurance of quality” (Guidance for Industry,
2009). However, operating at the outer edges of the overlay window, that
is, its frame, cannot assure quality due to it being only a 50/50 chance of
that response going out of bounds. To improve these odds, we advise you
impose a confidence interval (CI) or, better yet, a tolerance interval (TI). The
following case study illustrates the application of the CI and, ultimately the
TI, to achieve a QbD design space.
Pharmaceutical formulators sought the sweet spot for the following com-
ponents in a hydrophilic tablet with the ranges listed by weight percent:
1. Lactose, 5%–42%
2. Phosphate, 5%–47%
3. Cellulose, 5%–52%
4. Polymer, 17%–25%
5. Drug, 1%–2%
They wanted their final formula to achieve the following two specifications
reliably:
1. Dissolution time (T) in hrs to dissolve 50%, 5–11 (target 8)

2. Hardness in kiloponds (kP), maximize from 4, at least, with 8 being
good enough
Assuming quadratic mixture behavior, that is, nonlinear blending effects, the
chemists set up an I-optimal design with 15 model points. They augmented
this base experiment with 5 check-blends, and 5 replicates chosen optimally.
Table 7.2 lays out the resulting 25 runs.
Table 7.2 QbD Experiment on Tablet Formulation

Dis @
A: Lact B: Phos C: Cell D: Poly E: Drug t (50%) Hard
ID (wt%) (wt%) (wt%) (wt%) (wt%) hrs kP
0 23.33 25.83 28.33 21.00 1.50 8.23 7.56
1 42.00 20.00 20.00 17.00 1.00 6.27 8.24
2 22.33 5.00 52.00 19.67 1.00 6.14 6.48
3 5.00 47.00 26.00 21.00 1.00 11.61 9.44
4a 42.00 29.67 5.00 22.33 1.00 7.89 8.30
4b 42.00 29.67 5.00 22.33 1.00 8.42 6.43
5 5.00 32.00 37.00 25.00 1.00 14.12 3.97
6 29.67 33.42 16.67 19.00 1.25 8.18 7.35
7 20.17 15.42 40.17 23.00 1.25 8.99 4.13
8 29.67 47.00 5.00 17.00 1.33 6.33 6.09
9 42.00 5.00 29.33 22.33 1.33 5.69 8.69
10a 5.00 24.33 52.00 17.00 1.67 7.56 6.61
10b 5.00 24.33 52.00 17.00 1.67 8.27 6.48
11 14.17 36.42 24.67 23.00 1.75 9.16 8.39
12a 17.00 47.00 17.00 17.00 2.00 6.02 7.22
12b 17.00 47.00 17.00 17.00 2.00 4.66 6.10
13a 30.00 5.00 46.00 17.00 2.00 4.53 8.04
(Continued)
Table 7.2 (Continued) QbD Experiment on Tablet Formulation

Dis @
A: Lact B: Phos C: Cell D: Poly E: Drug t (50%) Hard
ID (wt%) (wt%) (wt%) (wt%) (wt%) hrs kP
13b 30.00 5.00 46.00 17.00 2.00 5.61 9.17
14a 42.00 30.00 5.00 21.00 2.00 3.94 6.20
14b 42.00 30.00 5.00 21.00 2.00 4.56 7.79
15 22.67 40.00 10.33 25.00 2.00 7.98 8.86
16 5.00 47.00 21.00 25.00 2.00 10.45 10.54
17 33.33 12.00 27.67 25.00 2.00 9.21 9.47
18 29.00 5.00 39.00 25.00 2.00 7.64 9.83
19 8.67 12.33 52.00 25.00 2.00 9.84 5.09
After fitting the full quadratic model (A, B, C, D, E, AB, AC, AD, AE, BC,
BD, BE, CD, CE, DE) to each response, the following terms were removed via
backward elimination at p of 0.1 (for background on this tool, see “A Brief
Word on Algorithmic Model Reduction”, RSM Simplified, 2nd ed., pp. 30–32.):
1. Dissolution: AB, AC, AE, CE, DE

2. Hardness: AC, BC, CD, DE
Based on these models, a numerical search found the optimum depicted

by the desirability ramps in Figure 7.5. Factors A through D ranged freely
within their initial design constraint. However, factor D—the drug—was set
at 1% for this optimization, that is, the low dosage of active pharmaceutical
ingredient (API). This is an equality constraint. It can be applied only to fac-
tors when searching for most desirable formulations (Figure 7.5). (A separate
optimization was done for the 2% API tablet.)
This most desirable formulation for the 1% API tablet is flagged on the
overlay plot of Figure 7.6a (first plot or plot on the left), which depicts A, B
and C with component D (polymer) sliced at its optimal level (17%).
Figure 7.6b (second plot or middle plot) shows a more conservative
sweet spot framed by the CI’s. The flagged formulation still falls well within
boundaries, despite them being narrowed to account for the uncertainty that
naturally stems from limited experimental data. However, FDA regulators
want assurance that, all or nearly all, individual units (not just on average!)
5.00 42.00 5.00 47.00
A: Lactose = 28.47 B: Phosphate = 24.58
5.00 52.00 17.00 25.00
C: Cellulose = 28.94 D: Polymer = 17.00
8.00
1.00 8.00
5.00 11.00 4.00
1.00 2.00 10.54

3.94 14.12 3.97
E: Drug = 1.00 Dissolution t(50%) = 8.00 Hardness = 7.97
Figure 7.5 Desirability ramps for tablet formulation.
A: Lactose (wt%) A: Lactose (wt%)

5.00 5.00
72.00 72.00
21.75 21.75
55.25 55.25
Dissolution t(50%):
38.50 8.00 Dissolution t(50%):
38.50 8.00
38.50 Hardness: 7.97 38.50 Hardness: 7.97
X1 28.47 X1 28.47
X2 24.58 X2 24.58
X3 28.95 X3 28.95
Dissolution t(50%) CI: 5
55.25 55.25
21.75 21.75
Dissolution t(50%) CI: 11
Dissolution t(50%): 11 72.00 72.00

5.00 5.00
72.00 55.25 38.50 21.75 5.00 72.00 55.25 38.50 21.75 5.00
B: Phosphate (wt%) C: Cellulose (wt%) B: Phosphate (wt%) C: Cellulose (wt%)
(a) (b)
Figure 7.6 Overlay plots for tablet formulation: (a) as-is and (b) with CI. (Continued )
A: Lactose (wt%)
5.00
72.00
21.75
55.25
Dissolution t(50%): 8.00

38.50 Hardness: 7.97
X1 28.47
Hardness TI: 4 X2 24.58
X3 28.95
Dissolution t(50%) TI: 5
Dissolution t(50%) TI: 11
55.25
21.75 Hardness TI: 4
72.00
5.00
72.00 55.25 38.50 21.75 5.00
(c) B: Phosphate (wt%) C: Cellulose (wt%)
Figure 7.6 (Continued) Overlay plots for tablet formulation: (c) with TI.
will meet specification. This requires imposing a TI that provides a high

level of confidence, say 95%, out of 99% (a generally acceptable threshold) of
the population meet (or exceed) specifications.
Figure 7.6c (third plot or plot on the right) shows how the TI pushes the
dissolution boundary well beyond that created by the CI. Also, the hardness
specification now imposes a restriction. However, the flagged formulation
remains acceptable. Thus, this qualifies as a QbD design space, within which
the formulation can float with little risk of any tablets going out of specification.
DETAILS ON COMPUTING CI AND TI AND

THEIR IMPACT ON DESIGN SIZING
To develop assurance for the quality of their products, manufacturers
must back off from their specification boundary by the CI, or, more con-
servatively for QbD, the tolerance interval (TI). This requires computation
of the intervals’ half-width. For references to the required calculations,
included specialized ones for designed experiments, see “Using DOE with
Tolerance Intervals to Verify Specifications” (Whitcomb and Anderson,
2011). In general, the TIs widen directly with confidence level and propor-
tion of in-specification product. Thus, the more demanding one gets for
(Continued )
confidence and/or the nearer to 100% the requirements reach, the smaller
the QbD design space becomes—all else equal. In fact, we often see the
space covered up entirely by the TI, due to the experiment being sized
too small to support the imposition of the interval. For example, adding
the TI (or even the CI) to the viscosity specification closes the window
in the overlay plot depicted by Figure 7.2 for the rheology-modifier
formulation—this 10-blend design being far too meager for QbD.
A QbD design-space determination maps out a region where the

in-specification product can be reliably manufactured. We advise that
you size your experiment appropriately for the purpose:
◾ For “functional” design, impose CIs, which reduce the risk of mean
results falling outside the allowable operating boundaries.
◾ To verify that your product falls within final manufacturing speci-
fications, for example, to achieve a QbD design space, enforce
tolerance intervals (TI).
Keep in mind that TIs range far wider than CIs. Therefore, a relatively
large experiment will be required to create any design space meeting such
a rigorous statistical requirement.
Practice Problem
Problem 7.1
Follow up on the detergent case from Problem 3.1 by completing the
multiple-response optimization. Is there a sweet spot for the formula-
tion that achieves all specifications? Find out by using the computer tool
specialized for this purpose—see About the Software for the path to the
program and the link to accompanying tutorials. Download and print
the Mixture Design Tutorial (Part 2/2—Optimization). Follow it through
to the conclusion of this case study.
PROPAGATION OF ERROR (POE)

As detailed in the tutorial, if you enter the variation in your component
amounts (e.g., due to how they are weighed out), the software can gener-
ate POE plots showing how that error transmits to the response. Look for
compositions that minimize POE, thus creating a formula that’s robust to
variations in the measured amounts. For details on POE refer to Chapter 9
in RSM Simplified, 2nd ed.
Chapter 8
Screening for Vital

Components
Nearly all the grandest discoveries of science have been but the
rewards of accurate measurement and patient long-continued labor
in the minute sifting of numerical results.
—Lord Kelvin (Presidential Address to Royal
Society, 1871, quoted p. 940 in Life of Lord
Kelvin, Silvanus Phillips Thompson, 1910)
For process troubleshooting, a tried-and-true strategy of experimentation

begins with screening previously unknown factors, from which engineers
often find a vital few to carry forward to the next stages—the characterization
of interactions and, finally, optimization. Chemists usually jump directly into
optimization when experimenting on formulations, relying on their subject-
matter knowledge to narrow down the field of components. However, as this
chapter will demonstrate, screening for main-component effects can uncover
vital ingredients that might otherwise be overlooked.
Screening designs only fit the linear model. We recommend them
for 6 or more components, for which higher order models become too
big to accommodate in an affordable experiment. This becomes clear
127
Table 8.1 Number of Terms in Scheffé Polynomials for 5–10 Components

Components (q) Linear Quadratic Special Cubic
5 5 15 25
6 6 21 41
7 7 28 63
8 8 36 92
9 9 45 129
10 10 55 175
from Table 8.1, which, starting with 5 components (where a similar

accounting in Table 2A.1 ends), enumerates terms for mixture models of
degrees 1 through 3.
Keeping in mind that when building a stout experiment design, it
pays to add another 10 blends (5 to check lack of fit, plus 5 replicates),
a robust 6-component design for the special cubic model comes to
51 blends (41 + 10). This goes well beyond the scope of most formula-
tors in our experience. On the other hand, a screening experiment, aimed
at estimating only the linear terms, requires at a minimum, 6 blends, and,
at most, 23 blends per generally acceptable practice, which we will now
detail.
We recommend that, if you wish to experiment on 6 or more components,
you first screen them down and then follow up with an optimization design
that models quadratic (second degree) or special cubic (third degree)
behavior.
High-Octane Simplex Screening Designs

We begin the detailing of screening designs with ones that cover a simplex
region, for example, when the amounts of each component vary from
zero to one-hundred percent (0 ≤ xi ≤ 1 in coded scale). This most-basic
scenario will be illustrated by a case study on octane-improvement
additives. The data—modified somewhat for explanatory purposes—come
from Snee and Marquardt (Snee and Marquardt, 1976).
Screening for Vital Components ◾ 129
DON’T KNOCK IT
The octane number characterizes a gasoline’s antiknock quality, that is,
its capacity to withstand damaging premature detonation in an engine’s
combustion chamber. Isooctane was the original antiknock additive—
hence the name “octane” for the rating. The higher the octane num-
ber, the more compression the fuel can handle. When gasoline is first
distilled from oil, it has an octane number of about 70, which would
severely limit engine efficiency. An octane rating of 87 meets the needs
for regular gasoline, a level that correlates directly to a mixture of 87%
isooctane and 13% heptane. (A 100% heptane fuel is the zero point of
the octane rating scale.)
In 1921, a team of General Motors chemists discovered that tetraethyl
lead worked wonders for octane ratings. However, due to health risks,
the Clean Air Act of 1996 banned the sale of leaded gasoline in the U.S.A.
One of the cheapest and most common replacements for the lead as an
octane booster is ethanol. However, its energy content is only 70% of
gasoline. Also, ethanol creates a host of engine issues. Currently, the best
option to ethanol is a toxic mixture of benzene, toluene, ethyl–benzene,
and xylene (BTEX).
The search for safe and effective octane-additives continues.
A difference of only 13 points in octane number made possible the

defeat of the Luftwaffe by the R. A. F. in the fall of 1940.
—Kalichevsky (1943)
Petroleum chemists want to screen 10 octane additives that will be

blended in a fixed amount to a specific quantity of gasoline. Blends of
these additives will not be ruled out, but that can be studied later in
a follow-up experiment. For now, the objective is to identify which, if
any, components generate promising linear effects. A design that boldly
spans the ranges will be most effective for this screening experiment.
Figure 8.1 shows a good choice—a “simplex screening” design (Snee
and Marquardt, 1976).
X1 = 1
X3 = 0 X2 = 0
X2 = 1 X3 = 1
X1 = 0
Figure 8.1 Simplex screening design.
Keep in mind that along the axes of a simplex (e.g., X1 from 0 to 1),
the proportion of each component spans its entire range while the
relative proportions of the other components to one another remain
unchanged. At a minimum, a simplex screening design should be
comprised of the q vertices (10 in the case of the octane additives).
Next, if the experimental budget allows, we advise you add the centroid
and replicate it 5 or so times to provide an estimate of pure error.
Beyond that, consider adding q axial points (located midway between
the centroid and the vertices) to fill the gaps and provide more power.
This brings the total blends to 2q + 1. Finally, another q “end effect”
blend (Snee and Marquardt, 1976, p. 22) can be added—these being
the points along the sides in Figure 8.1 where one component goes to
zero (e.g., X1 = 0). These latter blends, expanding the design to 3q + 1,
measure the effect from the complete elimination of any single
ingredient.
The petroleum chemists chose the middle-sized 2q + 1 simplex screen-
ing design shown in Table 8.2, which included 5 replicates of the centroid
Table 8.2 Octane-Additive Design and Results

# A:X1 B:X2 C:X3 D:X4 E:X5 F:X6 G:X7 H:X8 J:X9 K:X10 Octane
1 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 62.30
2 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 82.90
3 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 86.00
4 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 77.50
5 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 79.10
6 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 74.90
7 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 80.20
8 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 78.30
9 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 77.30
10 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 84.40
11 0.55 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 74.20
12 0.05 0.55 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 83.60
13 0.05 0.05 0.55 0.05 0.05 0.05 0.05 0.05 0.05 0.05 83.60
14 0.05 0.05 0.05 0.55 0.05 0.05 0.05 0.05 0.05 0.05 79.00
15 0.05 0.05 0.05 0.05 0.55 0.05 0.05 0.05 0.05 0.05 79.90
16 0.05 0.05 0.05 0.05 0.05 0.55 0.05 0.05 0.05 0.05 77.90
17 0.05 0.05 0.05 0.05 0.05 0.05 0.55 0.05 0.05 0.05 78.00
18 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.55 0.05 0.05 82.50
19 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.55 0.05 79.90
20 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.55 81.90
21 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 75.50
22 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 79.50
23 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 78.70
24 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 80.30
25 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 81.70
90
C
85 KB
A K HG E
80 H
B C DE F G J
Octane (number)
JD
F
75
70
65 A
60
−0.2 0.0 0.2 0.4 0.6 0.8 1.0

Figure 8.2 Trace plot of component effects on octane number.
(blends 21–25) and the 10 axial check blends (#s 11–20), as well as the
10 vertices (#s 1–10) that must be tested at a minimum.
The resulting octane numbers, which range from 62.3 to 86.0,
produce a significant fit to the linear mixture model. The most
important view of the relative effects is provided by the trace plot
depicted in Figure 8.2.
As we explained in Appendix 6A, this plot displays the predicted
response as any given component deviates from any chosen reference
point (typically the centroid), while holding all other components in
constant proportion. For example, follow component A (X1) from the nexus
of all 10 ingredients at the zero point on the abscissa (0.0 deviation from
reference blend). As indicated by the line going up to the left, reducing
A to its minimum of zero (−0.1 deviation), that is, taking it out of the
mixture, increases the octane. As more and more of component A goes
into the blend, it steadily degrades the octane. The lowest response
prediction comes when the additive is comprised only of A (0.9 units
above the centroid reference blend). Clearly, ingredient X1 (component A)
should be discarded.
VARIATIONS IN HOW TO DRAW A TRACE PLOT

As detailed in the Appendix, two methods prevail for the direction of
the trace—Cox versus Piepel. In this case, because all components go
from zero to one-hundred percent (0%–100%), the plots are identical.
Otherwise, though, you will see differences between the Cox and Piepel
traces that can lead to ambiguous interpretation for all but very active
components.
PS: Perhaps you noticed that letter “I” is missing from the design layout
in Table 8.2 and the trace plot—the components skip from “H” to “J.” This
is done because “I” is reserved as a symbol for the intercept in conven-
tional polynomial models.
On a positive note, components C (X3), K (X10) and B (X2) increase

octane numbers dramatically—they are keepers. The other 6 components
create little effect, one way or the other, so depending on the whims
of the petroleum chemists, they can be kept or tossed. These subject-
matter experts may also make something out of those components
whose responses follow each other very closely, for example, J (X9)
and D (X4)—perhaps they are similar chemically or the same, but from
differing suppliers. That might be good to know. Furthermore, such
ingredients could be combined to simplify follow up experiments.
Measuring the Effect of a Component

Although many inferences can be made from a trace plot, it will not suffice
for those who like numerical results. To satisfy these “quants,” effects (E) can
be measured by the following equation, which takes the difference from the
model coefficient of any individual (i) component to the average of all the
others:
q
∑β
j≠ i
j
E i = βi −
( q − 1)
Table 8.3 Octane-Additive Coefficients and Effects

Comp Beta Effect p-Value
A-X1 64.03 −16.81 <0.0001
B-X2 84.27 5.68 0.0071
C-X3 86.75 8.43 0.0003
D-X4 78.11 −1.17 0.5307
E-X5 79.75 0.65 0.7248
F-X6 75.59 −3.97 0.0457
G-X7 79.87 0.79 0.6719
H-X8 80.15 1.10 0.5557
J-X9 78.31 −0.95 0.6108
K-X10 84.79 6.25 0.0037
where ß (beta) represents the model coefficients, q is the number of compo-

nents and j is an index for all the other ingredients aside from the one being
measured for effect. Table 8.3 provides the linear model coefficients for the
10 octane additives as well as their linear effect. For good measure, p-values
are also displayed.
In this case, the coefficients are simply the predicted octane number for
the gasoline with only that additive. For example, component A (X1) can be
expected to produce a meager octane of 64.03. Its effect computes extremely
negative:
E A = 64.03 − 80.84 = −16.81
where 80.84 is the octane number on average for the nine other additives.
Extreme Vertices Design for Non-Simplex Screening

When component constraints violate the simplex geometry, we recommend
extreme-vertices screening design (EVD), which comprises as follows:
◾ 2q vertices chosen optimally

◾ n replicates (5 generally suffices) of the centroid.
Because some components in a nonsimplex design change less than others,

their effects must be adjusted accordingly and it becomes helpful to look
at their gradient. We will detail this by the following case study drawn in
most part (some changes made for the sake of simplicity) from Snee and
Marquardt (1976, p. 26).
Chemists crafting a formulation came up with eight candidate compo-
nents that can provide a key attribute—the higher, the better—for a new
product. They are determined to screen them down to a vital few. Due to
varying potencies, the ranges must differ as follows:
1. X1, 0.10–0.45
2. X2, 0.05–0.50
3. X3, 0.00–0.10
4. X4, 0.00–0.10
5. X5, 0.10–0.60
6. X6, 0.05–0.20
7. X 7, 0.00–0.05
8. X8, 0.00–0.05
These ranges do not form a simplex. Therefore, an EVD is required. The

effects of the components are unknown. With the aid of a computerized
algorithm, 182 vertices are identified as extremes in the seven-dimensional
(q − 1) mixture space. The program applies the D-optimal criterion to select
the best 16 vertices (2q). It then locates the centroid and replicates it 5 times
at random intervals in the resulting 21-blend layout for the experiment.
The resulting EVD is laid out in Table 8.4 with the vertices first and then the
5 replicates of the centroid (rounded to nearest one-hundredth for the sake
of convenience).
Figure 8.3 displays the trace plot for the eight-component ESD. It’s
magnified to give a better view of the four highly-restricted components—
C (X3), D (X4), G (X 7), and H (X8). Notice that the other, relatively wide-
ranging components go off the chart at left and/or right.
Table 8.4 Eight-Component EVD

Std A:X1 B:X2 C:X3 D:X4 E:X5 F:X6 G:X7 H:X8 Y
1 0.10 0.05 0.10 0.10 0.55 0.05 0.00 0.05 118
2 0.15 0.50 0.00 0.00 0.10 0.20 0.00 0.05 17
3 0.10 0.50 0.10 0.00 0.20 0.05 0.00 0.05 34
4 0.10 0.50 0.00 0.00 0.10 0.20 0.05 0.05 29
5 0.10 0.05 0.00 0.00 0.55 0.20 0.05 0.05 114
6 0.45 0.15 0.10 0.10 0.10 0.05 0.05 0.00 11
7 0.10 0.50 0.00 0.00 0.20 0.20 0.00 0.00 22
8 0.45 0.05 0.10 0.05 0.10 0.20 0.00 0.05 21
9 0.45 0.05 0.00 0.00 0.40 0.05 0.00 0.05 42
10 0.20 0.50 0.00 0.10 0.10 0.05 0.00 0.05 13
11 0.10 0.05 0.00 0.10 0.55 0.20 0.00 0.00 79
12 0.10 0.10 0.10 0.00 0.60 0.05 0.05 0.00 115
13 0.45 0.05 0.00 0.00 0.45 0.05 0.00 0.00 48
14 0.45 0.20 0.00 0.10 0.10 0.05 0.05 0.05 20
15 0.45 0.10 0.10 0.00 0.10 0.20 0.05 0.00 9
16 0.10 0.50 0.00 0.10 0.20 0.05 0.05 0.00 29
17 0.24 0.22 0.05 0.05 0.28 0.12 0.02 0.02 35
18 0.24 0.22 0.05 0.05 0.28 0.12 0.02 0.02 54
19 0.24 0.22 0.05 0.05 0.28 0.12 0.02 0.02 35
20 0.24 0.22 0.05 0.05 0.28 0.12 0.02 0.02 42
21 0.24 0.22 0.05 0.05 0.28 0.12 0.02 0.02 37
Components E (X5), G (X 7), and H (X8) stand out as “keepers” by their

steeply upward slopes. Table 8.5 lists the gradients, calculated by dividing
the component effect by the difference between the highest and lowest com-
ponent value on the trace. They reinforce the value of E, G and H, while, on
the other hand, accentuating the deleterious impacts of components A (X1)
and B (X2).
60
A
E
55
H
B G
50
F C
D
Y 45 D
C F
40 G
B
H
35
E A
30
−0.100 −0.050 0.000 0.050 0.100

Figure 8.3 Trace plot of component effects from eight-component EVD case.
Table 8.5 Gradients for Eight-Component EVD

Comp Gradient p-Value
A-X1 −101.21 <0.0001
B-X2 −70.02 <0.0001
C-X3 40.27 0.1664
D-X4 −12.39 0.6655
E-X5 102.05 <0.0001
F-X6 −28.66 0.1321
G-X7 160.75 0.0096
H-X8 210.55 0.0015
Components C (X3), D (X4), and F (X6) do not affect the response much
one way or the other, as evidenced by their insignificant (p > 0.05) gradi-
ents. If any or all of these ingredients come cheap, these might be retained
in the formula. Otherwise, they should be taken out.
ADJUSTMENTS TO LINEAR EFFECTS

DUE TO VARYING RANGES
The interpretation of linear effects must be modified to account for com-
plex constraints. This can be done by simply adjusting each effect by the
relative range of the given component. Doing so in the eight-component
EVD case reveals component E as, by far, the most effective for increas-
ing the response (as desired) due to it having such a wide range, rela-
tive to higher-gradient components G and H. However, the gradients tell
the story best in these nonsimplex mixture experiments, and they keep
things simple, so we will focus on this statistic for interpretation of results
from nonsimplex screening designs.
Moving on from here, the formulators will do well by a follow-up optimiza-

tion experiment on components E, G, and H, aimed at detecting nonlinear
blending effects, for example, via an I-optimal custom design for a quadratic
or special-cubic model. If the ranges on G and H can be expanded without
dire consequence, that will be all the better.
A word of caution—the cases we made for screening only included one
response, whereas in real life a formulation generally must provide many
functions. When you have multiple responses like this, things get far more
complicated when trying to eliminate ingredients, because any one may be
good for one attribute, while bad for another. It may be difficult to find a
single component that is entirely superfluous. Nevertheless, when recipes
become bloated or many ingredients can be considered, a screening design
might be just the ticket.
Appendix 8A: Trace Plots—Cox versus Piepel Direction

The trace plot shows the effects of changing each component along an
imaginary line from the reference blend (defaulted to the overall centroid)
to the vertex. In the Cox direction (Cox, 1971), done in the real-coded
space, as the amount of this component increases, the amounts of all other
components decrease, but their ratio to one another remains constant.
Chemists may like this because it preserves the reaction stoichiometry.

However, when plotted in this direction, traces of highly constrained mix-
ture components, such as a catalyst for a chemical reaction, become trun-
cated. Thus although it no longer holds actual ratios constant, we feel that
Piepel’s direction (Piepel, 1982), done in pseudo-coded space, provides a
more helpful plot by providing the broadest coverage of the experimental
space. For example, Figure 8.3 is done by the Piepel method, as well as the
associated gradients in Table 8.5.
How these two options for trace plots differ becomes clear with a few
pictures from an experiment to adulterate an acrylonitrile–butadiene–
styrene (ABS) plastic pipe with otherwise unwanted coal-tar pitch (Koons
and Wilt, 1982). The other two ingredients (besides the pitch) were grafted
polybutadiene (graft) and styrene–acrylonitrile (SAN). Table 8A.1 lays out
the experiment—a second-degree simplex lattice augmented with the
centroid and axial check blends (rounded a bit for the convenience of the
formulation).
Figure 8A.1a and b shows how the directions of Cox and Piepel diverge.
Both begin at the same point—the centroid. Cox goes from there to the
Table 8A.1 Pipe Experiment

Std A: Graft (Wt%) B: SAN (Wt%) C: Pitch (Wt%) Izod (ft-lb/in)
1 70.0 30.0 0.0 7.4
2 45.0 55.0 0.0 6.1
3 45.0 30.0 25.0 0.8
4 57.5 42.5 0.0 7.3
5 57.5 30.0 12.5 3.9
6 45.0 42.5 12.5 2.1
7 60.0 35.0 5.0 6.1
8 50.0 45.0 5.0 4.9
9 50.0 35.0 15.0 2.5
10 53.5 38.0 8.5 4.4
A: graft (%) A: graft (%)

1 1
0.7 0.7
0 0.3 0 0.3
0 0 0 0
0.55 0.45 0.25 0.55 0.45 0.25
1 0 1 1 0 1
B: SAN (%) C: pitch (%) B: SAN (%) C: pitch (%)
(a) (b)
Figure 8A.1 Directions of trace for Cox (a) versus Piepel (b) in pipe case.
corners of the big triangle (coded in reals), whereas Piepel steers for
the vertices of the small triangle (coded in pseudos’).
The resulting trace plots are displayed in Figure 8A.2a and b.
The advantage provided by Piepel now becomes apparent by its tracks
being of a consistent length and, for the most part, longer than Cox. Do
not let the change in slope for component B from negative (Cox) to positive
8 8
C C A
B
6 A 6
B
Izod (ft-lb/in)
Izod (ft-lb/in)
B
4 B 4
A
A
2 2
C C
0 0
0
0
0
0
40
20
40
20
00
20
40
60
80
00
20
40
60
80
.
.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
−0
−0
−0
−0
(a) Deviation from reference blend (b) Deviation from reference blend
Figure 8A.2 Trace plots by Cox (a) versus Piepel (b) in pipe case.
(Piepel) put you off—this wavering indicates a linear effect that may not
be significant and, indeed, pales in comparison to the impacts of A and C,
which, throughout, remain consistent.
Consider that the traces are one-dimensional only, and thus, cannot
provide a beneficial view of a response surface, especially with a nonlinear
blending model. Furthermore, they depend not only on direction but also
on the point of origin.
Chapter 9
Working Amounts, Categorical
and Process Factors into
the Mix
The universe is like a safe to which there is a combination. But the

combination is locked up in the safe.
—DeVries (1965)
This chapter lays out designs that combine varying compositions of mixtures
with changes in levels of process factors and other variables. The frosting on
the cake, perhaps literally, is to experiment on two mixtures simultaneously,
that is, a “mix–mix.” These combined designs unlock a universe of potential
synergisms. However, they can quickly expand far more runs than can be
afforded. Therefore, they must be used judiciously.
Mixture-Amount Experiments—Not Just

the Composition but How Much of It
We begin with the simplest combined experiment—one that incorporates
a single process factor. A special case of this, first described by Piepel and
Cornell (Piepel and Cornell, 1987), is a “mixture-amount” design (MAD).
143
X1 = 1
X1 = 1
Z1 = −1
Z1 = +1
X2 = 1 X3 = 1
X2 = 1 X3 = 1
Figure 9.1 Three-component mixture-amount design.
A standard mixture experiment assumes that the response depends only

on the proportions of the components present, for example, when tasting the
food, a nibble will do. However, for such things as the effectiveness of fertilizer,
the amount matters, as anyone with a lawn realizes after their grass comes out
striped from variations in application. How much to use usually depends on
the composition and vice versa—these two variables often interact.
Figure 9.1 depicts a combined design with three components and its amount
(or one process factor) at two levels. Geometrically this three-component MAD
forms a triangular prism. The mixture components define a (q−1) dimensional
simplex: X1, X2,..., Xq. The amount contributes to additional unconstrained
dimension: Z1. The combined region of interest has (q−1 + 1) dimension—three
in this case (3−1 + 1).
We will illustrate the structure of MAD and modeling of results via a case
study on a controlled-release ibuprofen tablet (adapted from a study by
Singh et al., 1995). The goal was to deliver just the right amount of this pain
reliever over a period of 12 hours. The release properties depend both on
the composition of the coating and its amount. Too little of a “weak” coat-
ing releases ibuprofen more quickly than desired. Conversely, an excessive
amount of a “strong” coating falls short on immediate pain relief—measured
at the first hour. The coating comprises three ingredients that total to 31%
by weight/volume (w/v) in grams per milliliter (g/mL) of the solution:
1. Ethyl acrylate (EA) copolymer: 7–20 w/v%

2. Methyl methacrylate (MMA) copolymer: 7–20 w/v%
3. Triethyl Citrate (TEC) plasticizer: 4–14 w/v%
Working Amounts, Categorical and Process Factors into the Mix ◾ 145
The thickness of the film is controlled by volume—the final factor in the

combined design:
4. Coating amount, 15–25 mL
RELIEF OF THE PAIN OF PHARMACHEMICAL JARGON

Our mixture-amount case study involves poly(methyl) methacrylates
(PMMAs), which are glassy copolymers that serve well as films for con-
trolling the release of drugs. The polymerization occurs in an aqueous
medium laden with surfactants to make the ingredients miscible. It all
must be mixed very intensely to achieve emulsification (Roudsari et al.,
2014). The result is a stable dispersion called “latex”—a term you may be
familiar with in the context of water-based paints.
Incomprehensible jargon is the hallmark of a profession.

—Kingman Brewster, Jr. (Address to the British
Institute of Management, December 13, 1977)
Two levels of amounts, as depicted in Figure 9.1, will be simplistic in this

case—the rate of release is likely nonlinear as a function of coating thick-
ness. Thus, the chemists decided to allow the amount to vary over three
or more levels on a continuous scale, as dictated by an I-optimal selection
based on the quadratic model:
Y( z ) = α 0 + α1z1 + α11z12
This three-term model, when crossed with a quadratic mixture polynomial

with six coefficients, creates an equation with 18 terms, which requires a
like number of model points. Adding five checkpoints for testing lack of
fit (chosen to fill space) and five replicates (selected optimally) brings the
total of the MAD to 28 runs. The release results showing cumulative drug
dissolved (“Dis”) at two-time intervals are shown in Table 9.1. Underlined
runs are replicates—these, along with all the others, being done in a
randomized order.
Table 9.1 Results of Experiment on the Controlled Release of a Drug

A: EA B: MMA C: TEC Dis 1 hr Dis 12 hr
ID (w/v%) (w/v%) (w/v%) D: Coat (ml) (cum%) (cum%)
1 20.0 7.0 4.0 15.0 59.2 93.5
2 13.5 13.5 4.0 15.0 37.9 99.3
3 7.0 20.0 4.0 15.0 64.8 88.3
4 15.0 7.0 9.0 15.0 77.9 100.0
5 7.0 15.0 9.0 15.0 76.9 80.9
6 10.0 7.0 14.0 15.0 82.6 100.0
7 10.0 7.0 14.0 15.0 75.4 91.7
8 20.0 7.0 4.0 17.5 42.0 77.3
9 7.0 20.0 4.0 17.5 57.0 74.0
10 11.0 11.0 9.0 17.5 51.6 69.6
11 20.0 7.0 4.0 20.0 40.2 74.3
12 13.5 13.5 4.0 20.0 9.8 87.3
13 7.0 20.0 4.0 20.0 39.1 57.0
14 15.0 7.0 9.0 20.0 41.1 79.9
15 7.0 15.0 9.0 20.0 47.3 64.0
16 10.0 7.0 14.0 20.0 49.5 76.7
17 10.0 7.0 14.0 20.0 53.7 69.9
18 20.0 7.0 4.0 22.5 31.2 54.6
19 11.0 11.0 9.0 22.5 32.6 54.1
20 20.0 7.0 4.0 25.0 25.6 54.0
21 13.5 13.5 4.0 25.0 4.0 55.1
22 13.5 13.5 4.0 25.0 3.1 54.1
23 7.0 20.0 4.0 25.0 1.1 48.0
24 7.0 20.0 4.0 25.0 1.9 40.7
25 15.0 7.0 9.0 25.0 14.0 42.9
26 7.0 15.0 9.0 25.0 15.5 20.2
27 10.0 7.0 14.0 25.0 13.3 42.1
28 10.0 7.0 14.0 25.0 27.2 43.8
Due to the size of the crossed models for combined designs such as this,
we recommend that they be reduced. One of the better ways to eliminate
terms is to take out any that fall below a specified p-value (typically 0.1)
but do so in step-wise fashion, starting with the least significant one and
going on from there until only the vital ones remain, that is, when no
further reduction could be made. As noted in Chapter 7 for the tableting
case, this is called the “backward” selection method (as opposed to
“forward”—a method starting from a core model and adding the most
significant term, next most, and so on). The two reduced models are shown
below—all terms being significant but for AC in Y1, which comes in due
to it being needed to support the hierarchy of ACD. (We detailed hierarchy
in Appendix 1B. However, you may not have realized its repercussions
on third-order terms such as ACD, which require not only the three main
components (A, C, and D) be modeled, but also the three nonlinear
blending terms: AC, AD, and CD.)
Y1. Dis 1 hr = 39.62 A + 37.01 B + 57.66 °C

Y2. − 78.63 AB − 10.66 AC − 15.98 AD − 33.93 BD − 21.48 CD
Y3. + 35.76 ABD − 53.01 ACD
Y4. Dis 12 hr = 71.52 A + 64.23 B + 70.92 °C
Y5. + 42.97 AB − 21.49 AD − 56.23 BC − 22.84 BD − 29.94 CD
The lack of fit on both models is insignificant (p > 0.1).
HOW LESS PROVIDES MORE FOR PREDICTED R-SQUARED

We promised, in the Chapter 1 sidebar “Why R-Squared Needs to Be
Adjusted” that we would discuss using model reduction if the adjusted
R-squared (R2adj) differs appreciably from the predicted R-squared (R2pred).
Table 9.2 spells out the benefits of doing so in the case of the experiment
on controlled release of ibuprofen. Notice how much the R2pred improves,
especially on the second response (Y2).
Other model statistics benefit from the model reduction but less dra-
matically. One exception is the raw R-squared, for which the larger mod-
els win, as always. That is simply wrong, as the case makes clear; thus,
(Continued )
Table 9.2 Statistics on Full versus Reduced Models

Statistic Y1 Full Y1 Reduced Y2 Full Y2 Reduced
Terms 18 10 18 8
Overall F 35.45 70.75 13.83 44.10
LOF p-value 0.4800 0.5754 0.1608 0.0582
R2 0.9837 0.9725 0.9592 0.9391
R2adj 0.9559 0.9588 0.8898 0.9179
R2pred 0.4736 0.9236 −1.7614 0.8861
reinforcing that this unadjusted form of coefficient of determination can-

not be relied upon for model selection.
Generally, we advise that when the R2pred falls to 0.2 or more units
below R2adj, you try a model reduction. If that does not help, then con-
sider a response transformation such as taking and converting them
into logarithm scale (see, for example, Problem 5.1 in the 2nd edition
of RSM Simplified). Another possible cause for such a discrepancy on
the R-squared is a substantial block effect, in which case it cannot be
rectified.
PS: A negative R-squared would normally be impossible, but in the
predicted form, it can happen, for example, the R2pred for Y2. This hap-
pens when the predicted residual sum of squares (PRESS) exceeds the
residuals for the mean model, that is, less variation occurs around the
average of the response than when it is fitted to the input variables. That
is not acceptable. However, in this case, taking out the insignificant terms
brings the R2pred well into positive territory.
Being significant terms for both responses that include factor D indicates
that amount does affect the release of the drug as one would assume. The
pharma chemists hoped to achieve 10%–30% dissolution at one hour, for
the same composition and amount of coating, a release of 80%–100% at
12 hours. They succeeded as seen by the window that appears in the graph-
ical overlay depicted in Figure 9.2.
A: EA (w/v %)
20
Dis 12 hr: 80
4 7
Dis 1 hr: 30
20 7 17
B: MMA (w/v %) C: TEC (w/v %)
Figure 9.2 Sweet spot for the release of ibuprofen at coating amount (D) of 19 mL.
This elusive sweet spot occurs only between 16.5 and 19.7 mL of the
coating and only for a particular combination of ingredients. Only by a well-
conceived combined DOE can such desirable outcomes be found in a rea-
sonable number of runs.
Contending with Categorical Variables

Categorical variables come up frequently in industry, for example, when
a purchasing agent seeks competitive vendors, or a chemist wants to try
substitute materials. An adjustment to the mixture recipe may be required
to facilitate these alternatives. To illustrate how to set up such a combined
design, let’s look at a hypothetical case on the material science of a compos-
ite material used in aerospace. The mission is to:
1. Select one of two elastomers.

2. Choose a fiber from three alternatives.
3. Optimize the recipe of elastomer, fiber, hardener and epoxy resin.
Elastomer and fiber types are categoric factors; that is, only one of each can
be present in any given mixture.
THE ESSENTIALS FOR A HIGH-STRENGTH COMPOSITE

In composites, one material, such as epoxy resin, serves as a “matrix” to
hold everything together. Fibers, such as glass or carbon, when embed-
ded in the matrix provide reinforcement. Elastomers are added to make
the whole composite less brittle.
PS: Epoxies require a hardener to initiate curing.
The ranges in weight percent for the elastomer and fiber are 5%–10%
and 54%–62%, respectively. The ratio of epoxy to hardener must be
maintained from 1.8 to 2.1 (refer to Chapter 6 for detail on how to convert
this into a multicomponent constraint). These specifications set the ranges
for hardener and epoxy levels somewhere between zero and one-hundred
percent (0%–100%)–an algorithm such as that detailed in Appendix 5A will
get this job done. Nothing about the mixture part of the combined design
is new at this stage in the book—the only aspect of it is being combined
with categorical variables. Figure 9.3 shows how this complicates the
structure.
The minimal number of points for an optimal design expands
exponentially as more categorical variables are added, especially,
when they expand beyond two types each, as in the case of the fiber
alternatives. To keep the design to a manageable level, this experiment
on composites is set up for a quadratic mixture model with 10 terms
(the 4 main components plus 6 second-order nonlinear blending
combinations) crossed by a two-factor interaction equation for the
Z1
Z2
Figure 9.3 Four-component mixture combined with two categorical variables.

5 10 54 62 9.03226 14.6429
A: Elastomer = 7.78121 B: Fiber = 60.5757 C: Hardener = 10.2075
18 27.7742 1 2 1 2 3
Treatments Treatments
D: Epoxy resin = 21.4357 E: Elastomer (type) = Elast-2 F: Fiber (type) = Fiber-3
100
54.6 149 1009 1639
Impact = 103.258 Tensile = 1353.53
Figure 9.4 An optimal composite recipe made with selected types of elastomer
and fiber.
categorical variables (Z1 and Z2) with 6 coefficients (c 0, c1Z1, c2[1]Z2 , c2[2]Z2 ,

c12[1]Z1Z2 , c12[2]Z1Z2), where the square-bracketed numbers deal with the
two contrasts between the three types of fiber. This is complicated! The
design then must provide a base of 60 model points (10 mixture times
6 categorical). Adding 5 checkpoints for lack of fit and replicating 5
combinations for pure error brings the design to 70 points in total.
This example is made up, but to provide an idea of where it might lead,
we offer Figure 9.4 to spell out the feasible ranges for the hardener and
epoxy (these were allowed to float initially), and a hypothetical solution
where impact strength of the composite must be at least 100 kilojoules per
meter squared and tensile properties maximized.
This material-science example only scratches the surface (pun intended)
of what can be done and how much work will be required when combining
mixtures with categorical variables. Before going down this road, consider
simplifying things by first sorting out which materials to use, whom to get
them from and so forth—using a standard recipe, and then reoptimize the
formulation once you settle on the categorical variables.
ICING ON THE CAKE MIX–MIX AND

OTHER COMBINED DESIGNS
We could continue with many other combinations—mixtures on top of
mixtures (e.g., multilayer films) crossed with process factors and/or cat-
egorical variable. However, except those who work for their institution’s
Department of Redundancy Department, we expect you to get the gist of
how this can be done from the examples we provided.
Making the simple complicated is commonplace; making the compli-
cated simple, awesomely simple, that’s creativity.
—Charles Mingus
(quoted in Creativity and the writing process
by Olivia Bertagnolli, p. 182, 1982)
Practice Problem
Problem 9.1
Experience how easy it will be to design a mixture experiment, combined
with multiple factors (one of the combinations we skipped over) and
analyze the results by using a computer tool specialized for this purpose.
It can be freely accessed via the website developed in support of this book:
See About the Software for the path. When you arrive at the internet page,
follow the link to the accompanying tutorials. Then download and print the
one on Combined Mixture-Process. It explains how food scientists came up
with just the right texture (never mind the taste!) for fish sticks made from
an optimal blend of mullet, sheepshead, and croaker (yuk!). They were then
deep-fried for an ideal period before being baked to perfection at the ideal
temperature and time settings. You will be amazed.
Appendix 9A: Alternatives for Modeling Results

from Combined Designs
The crossed models we detailed in this chapter are straightforward to
create but, as the number of components and factors increase, they
become very large with many terms going beyond second order that
end up being taken out due to their insignificance. Chapter 11 in RSM

Simplified provides one good alternative—model the mixture compo-
nents in terms of ratios and simply incorporate them into factorial or
RSM designs alongside the process factors. However, the mathematics
required by the ratio approach make this impractical. An equally efficient
(for a number of model terms), but less intense approach, is to selectively
remove terms from the crossed model, as advised by Kowalski, Cornell,
and Vining (Kowalski et al., 2000). These “KCV” models are geared to a
second-order approximation via an appealing “intuitive balance” (p. 2279).
They preserved the view of mixture-process interactions that experiment-
ers are seeking when they deploy combined designs. For these reasons,
we feel that KCV models deserve detailing.
The advantage of KCV models comes from adding, rather than cross-
ing, higher order terms from the equations for the mixture and process
variables. To illustrate the difference, let’s apply the KCV model to the fish-
patty experiment of Problem 9.1, which involves three mixture components
(q = 3) and three process factors (k = 3). The crossed quadratic-by-quadratic
(QxQ) model requires 60 terms from the multiplication of the six in the
mixture (X1, X2, X3, X1 X2, X1 X3, X2 X3) and ten for the process (Intercept,
Z1, Z2, Z3, Z1Z2, Z1Z3, Z2Z3, Z12, Z22, Z32). It includes 18 fourth order terms
(e.g., X1 X2 Z1Z2), which can be safely removed by capping the crossed
model at cubic. This reduces the model to 42 terms. The quadratic KCV
model (and ratio RSM) requires only half of the reduced crossed QxQ—21
terms (X1, X2, X3, X1 X2, X1 X3, X2 X3, X1Z1, X1Z2, X1Z3, X2Z1, X2Z2, X2Z3, X3Z1,
X3Z2, X3Z3, Z1Z2, Z1Z3, Z2Z3, Z12, Z22, Z32). As shown in the following general
equation, it is generated by crossing the linear models and then adding the
second-order terms:
q q q k k k
η( x,z ) = ∑β x + ∑ ∑β x x + ∑ ∑ γ x z + ∑ ∑ α
i =1
i i
i< j
ij i j
i =1 n =1
ik i n
n< m
zz +
nm n m ∑α
n =1
2
nn nz
where q and k are the indices for the process factors and the mixture
components, respectively. The interactions between the linear blending
terms in the mixture and main-effect process variables (e.g., X1Z1) pro-
vide the core value for the crossing of models, that is, an opportunity to
find just the right combination of component levels at a particular set of
factors.
Table 9A.1 Number of Terms in Various Models for Combined

Mixture-Process Experiments
Mixture Process Crossed QxQ Max Ratios (#)
Components Factors QxQ All Cubic RSM Model KCV
2 2 18 15 10 (1) 10
3 2 36 27 15 (2) 15
4 2 60 42 21 (3) 21
3 3 60 42 21 (2) 21
4 3 100 64 28 (3) 28
5 3 150 90 36 (4) 36
4 4 150 90 36 (3) 36
5 4 225 125 45 (4) 45
6 4 315 165 55 (5) 55
Table 9A.1 shows, for increasing numbers of mixture components (2–6)

and process factors (2–4), how KCV (hand-in-hand with the ratio approach)
fares far better than the crossed QxQ model for combined designs, even
when the terms are kept to a maximum of cubic (3rd order). As the designs
build beyond the underlined combination for the fish-patty case (q = 3,
k = 3), the practical advantage of KCV becomes pronounced.
NONLINEAR COMBINED MIXTURE MODELS

Ronald D. Snee, a pioneer in the field of mixture design, suggests using
nonlinear estimation techniques to fit models that are “nonlinear in
the parameters” (private correspondence to MJA 6/10/15), for example,
exponential [Y = f(X k)]. However, application of nonlinear modeling to
combined mixture-process experiments (Snee et al., 2016) requires users
to specify the equation that relates the coefficients to the response to
understand and use, without providing any advantage for reducing runs,
but this strategy is more complicated than KCV.
Chapter 10
Blocking and Splitting Designs

for Ease of Experimentation
The generation of random numbers is too important to be left to

chance.
—Robert R. Coveyou, title of article in Studies
in Applied Mathematics, III (1970)
Randomization protects experimenters against confounding caused by lurk-

ing variables such as rising ambient temperatures run-by-run. However, it pays
to restrict randomization under the following circumstances:
◾ Limited availability of resources such as material, time, equipment or

operators, which then must be carefully managed so the variability
caused by shifting them over in mid-experiment does not create bias in
effect estimates.
◾ Input variables being very hard to change (HTC).
These limitations lead to groupings of runs via the block and split plots,
respectively.
Blocking to Remove Known Sources of Variation

We introduced blocking in Chapter 2 for the beer-blending case where it fil-
tered out differences between the three tasters. This experiment being fully
replicated by taster presented no difficulty in design. Blocking becomes far
155
more problematic when it involves a breakdown of a given mixture design

into subgroups. If it is done without careful thought, model parameters
can become highly correlated with the block effect. To avoid such biasing,
Draper et al. and Lewis et al. developed templates for orthogonal blocking
of simplex designs for a limited number of components (Draper et al., 1993)
that they then generalized (Lewis et al., 1994). We will lay out one of these
Draper–Lewis designs to provide a picture of well-blocked simplex. This
sets the stage for a more-flexible method of blocking mixture experiments—
applicable to simplex or nonsimplex regions—which makes use of modern
optimal-design tools.
LATIN SQUARES FRAME OUT BLOCKS ORTHOGONALLY

The Draper–Lewis blocking schemes use Latin squares as their basis.
These are square layouts with Latin letters with a number of rows and
columns equal to the number of treatment levels. They must be con-
structed such that each treatment occurs only once in each row and once
in each column. For example, see Table 10.3 (p. 218) in RSM Simplified,
2nd edition, which lays out a balanced plan to test 4 tires being rotated
around the 4 positions in a car.
For more details on constructing orthogonal blocks in mixture designs
using Latin squares see Chapter 8.4 (p. 447) of Experiments with Mixtures
(3rd ed., Wiley, 2002) by John Cornell.
The blocked mixture experiment cited by Draper et al. and Lewis et al. in their
1993 and 1994 publications, respectively, was conducted by an English miller
on four flours, each derived from a different variety of wheat. The food chem-
ists blended these four components into doughs in various proportions, which,
when baked into bread, created the varying results (simplified somewhat) in
specific volumes shown in Table 10.1. The higher the volume the better for it
to produce loaves of least density, as desired by consumers in Great Britain. By
breaking this experiment into two blocks, it lessened the impact of unknown
time-related variables that might otherwise be confounding. The block size of
9 runs, each fell well within the capacity of the pilot-scale bakery.
After removing the variation due to the blocks, the regression produces
a model that exhibits strong nonlinear blending of flours 1–3 with flour 1,
Blocking and Splitting Designs for Ease of Experimentation ◾ 157
Table 10.1 Blocked Mixture Experiment on Bread-Flours

Run Block A: Flour 1 B: Flour 2 C: Flour 3 D: Flour 4 Volume (ml/100 g)
1 1 0 0.25 0 0.75 392.0
2 1 0.25 0 0.75 0 423.5
3 1 0 0.75 0 0.25 427.0
4 1 0.75 0 0.25 0 423.0
5 1 0 0.75 0.25 0 421.5
6 1 0.25 0 0 0.75 423.5
7 1 0 0 0.75 0.25 378.0
8 1 0.75 0.25 0 0 426.5
9 1 0.25 0.25 0.25 0.25 413.0
10 2 0 0.25 0 0.75 389.5
11 2 0.25 0 0.75 0 421.0
12 2 0 0.75 0 0.25 413.5
13 2 0.75 0 0.25 0 416.5
14 2 0 0 0.25 0.75 350.0
15 2 0.25 0.75 0 0 431.0
16 2 0 0.25 0.75 0 384.5
17 2 0.75 0 0 0.25 408.5
18 2 0.25 0.25 0.25 0.25 429.0
that is, terms AB, AC, and AD. The strong influence of this first flour (A) can
be seen in the trace plot shown in Figure 10.1.
The mean square for the blocks exceeds the residual mean square
by nearly fourfold, which is appreciable, that is, only by filtering out this
variation over time was it possible for this design to reveal the formulation
behavior exhibited in the trace plot.
This experiment on bread flour pioneered the application of blocking
to mixture designs. However, modern-day computational tools (such those
provided by the software accompanying this book) provide more flexible
(not restricted to simplex regions) and better (statistically optimal) designs
for any number of blocks within reason. Applying optimal design, in
440 D B
C
420 B
Volume (ml/100 g)
400
A
380 A
D C
360
340
−0.400 −0.200 0.000 0.200 0.400 0.600 0.800

Figure 10.1 Trace plot of effects of flours on baked volume of bread.
retrospect to the flour-blending case, we produced the experiment shown

in Table 10.2. It incorporates four pure replicates, which enables a test lack
of fit—a shortcoming of the original design by Draper and Lewis. This
design improvement comes at a small price—one extra run per block.
THE POSSIBILITY OF THIRD-ORDER NONLINEAR BLENDING

As you may have noticed, beyond the overall centroid, the original flour-
blending design restricts the blending to only two components at a time,
thus it cannot estimate the third-order special cubic term in the mixture
model. Also, as we’ve already mentioned, this experiment lacked the nec-
essary elements (check blends plus replicates) to provide a test for lack
of fit. Given the strong nonlinear blending impact of flour #1 (A) with all
other flours, perhaps this might be better modeled by 3rd order special
cubic terms (ABC, ACD, ABD). We will never know.
The rebuilt design includes numerous blends of three or more compo-
nents and it does support the fitting of all special cubic terms. It also tests
lack of fit beyond that. Thank goodness for advancements in program-
ming along with the leaps in computing power since 1993 when blocked
(Continued )
mixture designs were still in their infancy. To be fair, all of these, coming
at a very slight cost-optimal designs, such as those we laid out, are not
orthogonally blocked, only nearly so (e.g., VIF of 1.4 in our design—a
value of 1.0 being completely orthogonal).
Orthogonally blocked mixture experiments [such as the bread-baking

one] are highly inefficient compared to optimal designs.
—Goos and Donev (2003)
Table 10.2 Flour Experiment Rebuilt with Modern Tools for Optimal Design
A: B: C: D:
ID Block Build Type Space Type Flour 1 Flour 2 Flour 3 Flour 4
1a 1 Model CentEdge 0.500 0.500 0.000 0.000
1b 1 Replicate CentEdge 0.500 0.500 0.000 0.000
2 1 Model Vertex 0.000 1.000 0.000 0.000
3a 1 Model CentEdge 0.500 0.000 0.500 0.000
4 1 Model CentEdge 0.000 0.500 0.500 0.000
5 1 Model Vertex 0.000 0.000 1.000 0.000
6 1 Model PlaneCent 0.000 0.33̅3̅ 0.33̅3̅ 0.33̅3̅
7 1 Model Interior 0.375 0.125 0.125 0.375
8 1 Lack of fit AxialCB 0.125 0.125 0.125 0.625
9 2 Model Vertex 1.000 0.000 0.000 0.000
10 2 Lack of fit PlaneCent 0.33̅3̅ 0.33̅3̅ 0.33̅3̅ 0.000
13a 2 Model CentEdge 0.500 0.000 0.000 0.500
14a 2 Model CentEdge 0.000 0.500 0.000 0.500
15 2 Model CentEdge 0.000 0.000 0.500 0.500
16 2 Model Vertex 0.000 0.000 0.000 1.000
D D
8
13 15
7
10 14
6
12
A 3 5 C A 9 C
1 11
4
2
(a) B (b) B
Figure 10.2 Point locations in blocks 1 (a) and 2 (b) for optimal design on flours.
Figure 10.2 shows the optimal design block by block within the tetrahedral
space. The double-circled points are replicated.
Note how the point allocation in our optimal design is mutually exclusive,
that is, they complement each other by block, whereas in the original
experimenters reran blends 1–4 and 9 in their second block. If you do want a
“control” blend across both blocks, we recommend it be the overall centroid.
Also, since three out of the four vertices made it into this algorithmic design,
adding the missing one (D—the pure blend of flour 4) would be sensible.
To summarize, using standard templates, or more-versatile computer
tools for optimal design, blocks should be set up to be orthogonal, or nearly
orthogonal, to the component effects you want to estimate. Having done so,
the breakdown of runs into subgroups will not make much of a change to
the model-coefficient estimates, if at all. Then if blocks differ in an apprecia-
ble degree, you need not worry, because this variation will be filtered out of
the overall system noise, thus making it easier to detect significant changes
to the response caused by the components themselves.
Split Plots to Handle Hard-to-Change Factors

or Components in Combined Designs
To satisfy statistical assumptions, properly executed experiments require
runs to be completely independent. This can be very daunting for researchers
running combined designs such as those detailed in Chapter 9, because:
◾ A new blend must be prepared even if the recipe listed in the experi-
mental plan does not change.
◾ The process must be reset, even though all of the factors remain at the
same level in the design from one run to the next.
If all the variables are easy to change (ETC), this complete randomization
with resets should be done. However, it is not often practical, or even pos-
sible, to perform an experiment in this way. Consider, for example, a food
scientist mixing up a single cookie by the recipe specified in the design
layout and then baking it all by itself per the plan for that run. That would
be ridiculous. A more sensible approach in this case would be to mix a
batch of cookies with varying recipes and then bake a tray of them. This
introduces a restriction in randomization that creates a designed-experiment
called a “split plot.”
APPLYING A SPLIT PLOT TO MIXTURE ONLY

The big advantage of split plots comes with combined experiments
as we’ve discussed. However, it is conceivable that some components
might be harder to change than others, for example, in a heterogeneous
reaction involving a liquid reagent catalyzed by a mixture of solid ingre-
dients. In this case, it would be sensible to create master batches of the
liquid components and then add various blends of the solid catalysts.
This would be a mix-by-mix split plot.
A silicone base (Mixture 1-HTC) is prepared and split up, after which
a different catalyst package (Mixture 2-ETC) is added to each piece.
For background on split plots see Chapter 11 of DOE Simplified, 3rd edition.
Here we will detail its application to an experiment done at Stat-Ease head-
quarters to improve the office coffee by a better blend of beans, combined
with improved methods for grinding and brewing (Bezener and Anderson,
2016).
The coffee provided by the vendor of the Stat-Ease brewing machine
created a backlash from our programmers who declared it to be “disgusting
and unacceptable.” They questioned not only the type of coffee being used
but also the fineness of its grind and how much of it to be brewed per pot.
The team agreed that changes would be made only if the new coffee be
judged better on average by a core group of 5 drinkers, and no one should

hate it. This second requirement essentially gave each of the individual
testers the power to veto the majority’s rule.
The experiment combined a mixture-amount design (discussed
in Chapter 9) with a categorical factor—the grind size. The input
variables were:
◾ Components a, b, and c*: Light, Medium, and Dark roast

*(lower-case lettered to denote them being HTC and thus grouped by
blend)
◾ Amount (factor D): 2.5–4.0 ounces of beans (continuous)
◾ Grind (E): Fine, Medium, or Coarse (categorical)
The fully randomized experiment required freshly blended beans for each
pot. However, it being far more convenient to mix these up in quantity,
the blends were restricted to 16 groups using a 74-run split-plot design.
Table 10.3 shows the first 9 runs of the experiment to illustrate the grouping
(3 shown) of the mixtures.
Table 10.3 First 9 Runs of Coffee Experiment

Avg Min
A: B: C: D: E: Liking: Liking:
Group Run Light Medium Dark Amount (oz.) Grind (1–9:) (1–9:)
1 1 0 0 1 4.00 Medium 6.4 5
1 2 0 0 1 2.50 Coarse 5.2 4
1 3 0 0 1 3.25 Fine 5.2 3
1 4 0 0 1 2.50 Medium 5.0 4
2 5 0.5 0.5 0 4.00 Fine 4.2 3
2 6 0.5 0.5 0 3.25 Medium 5.6 5
2 7 0.5 0.5 0 4.00 Coarse 5.4 4
2 8 0.5 0.5 0 2.50 Fine 5.2 3
3 9 0.5 0 0.5 3.18 Fine 5.4 4
~ ~ ~ ~ ~ ~ ~ ~ ~
Each blend of coffee beans was, on average, tested at 4 amount-grind

size combinations. The barista (Martin) randomly interspersed six runs
of the current office coffee throughout the experiment to serve as a
control.
LIGHT COFFEE HEAVIEST FOR CAFFEINE?

It is natural to assume that dark roast has more caffeine than a lighter
coffee. The true “caffiends” argue that roasting breaks down caffeine
so they go for the light. For that reason breakfast blends are generally a
lighter roast. However, both of those schools of thought may be wrong
because all roasts of the same bean have basically the same amount
of caffeine, at least according to one expert (Thomson and Thomson,
2016).
Another variable involves how you measure your coffee. If by scoops,
light coffee will provide a bit more caffeine, because roasting drives
out water and thus reduces the density of darker beans (Kicking Horse
Coffee. Caffeine myths: Dark vs. light. www.kickinghorsecoffee.com/en/
blog/caffeine-myths-dark-vs-light). We advise you weigh out your coffee
and thus avoid this confounding variable.
Perhaps the best advice is to drink what you find tastiest. Then you
will imbibe more and dose up on caffeine more heavily, as needed for
whatever task you have at hand.
A mathematician is a device for turning coffee into theorems.
—Alfréd Rényi, Probability Theorist (Suzuki, 2002)
A supplement to the book details the modeling and statistical analysis of this
combined split-plot design. A subsequent numerical search came up with
a most desirable 50–50 blend of medium and dark beans (no light ones)
when ground to the fine level and brewed at a loading of 2.5 oz. This hit the
“sweet spot” for the taste testers as evidenced by 10 follow-up runs—4 of the
chosen blend, 2 with the standard office coffee, and 4 at various other com-
binations of beans ground in different ways and produced with changing
amounts, none of which deviated significantly from the model predictions.
A Case Where the Process, Rather Than

the Mixture, Is Hard to Change
Kowalski et al. the inventors of the KCV model we touted in Appendix 9A,
provide an example (Section 5 of Kowalski et al., 2002) of process factors
being HTC in a design combined with ETC variations in mixture components.
This contrasts with the coffee case where the blending was HTC. Figure 10.3
illustrates the primary options for a split-plot, with combined design either
being mixture HTC—a triangle with squares at the vertices or process
HTC—a square with triangles at the corners.
The KCV case, pictured by Figure 10.3b, was aimed at modeling the thick-
ness of a vinyl car-seat covering as a function of the plasticizer composition
and processing conditions. The mixture side of the experiment studied pure
blends, binary blends, and the centroid of three plasticizers (P1, P2, and P3)
randomized (not a problem being ETC) within groups. The process factors
were extrusion rate and drying temperature set at two levels, each in 4 groups
(these being HTC). Three more groups were run at the center point of the
process factors with the centroid blend replicated 4 times. This is a good
example of a split plot being put to good use for the practicality of processing.
In conclusion, split-plot designs provide accommodating layouts, espe-
cially when experimenters want to combine mixture components with pro-
cess factors and one or the other is hard to change. Computer-aided optimal
selection of runs facilitates customization of group sizes to make things even
more convenient.
P1 P1
P2 P3 P2 P3
Extrusion rate
P1 P1
Drying temp
P2 P3 P2 P3
(a) (b)
Figure 10.3 Split plots for mixture HTC (a), versus process HTC (b).
Chapter 11
Practical Magic for Making

the Most of a Mixture
Eye of newt and toe of frog, wool of bat and tongue of dog.
Adder’s fork and blindworm’s sting...Barbados lime is just the
thing. Cragged salt like a sailor’s stubble! Flip the switch and let
the cauldron bubble!
—Owen’s sisters (witches) reciting recipe for
margaritas (Practical Magic, Warner Brothers, 1998)
In this last chapter, we come back to what defines a mixture and provide
practical advice on how to put the statistical design of experiments to good
use for optimizing your formulation. To begin, let’s dispel various excuses
to avoid the peculiarities of mixture design and analysis by denying the true
nature of the experimental variables.
The Failure of Fillers and Perils of Parts

As we advised in Chapter 1, you’d best deploy the statistical tools for
mixture design and analysis when you experiment on a formulation
for which the response depends on proportions, not amounts, of
165
ingredients. For this to work, you must fix the total of your components
that will be varied, for example, a 20-gallon cauldron for a magical
concoction. Unfortunately, many formulators, particularly those who are
only taught the standard DOE tools for factorials and response surface
methods, do not take to mixture design. They excuse themselves along
the following two lines epitomized by these actual quotes from clients
who must remain nameless:
1. Filler (ignoring the main ingredient): “The process engineers do not

see a clear reason why their 23 factorial was inappropriate. They
told me that the filler was “like the ocean” compared to the amounts
of the other three components, so they didn’t see it as a variable to
have in the model.”
2. Parts (not including the base as a component): “I am doing a
chemical formulation DOE study. I have two ingredients to study in
relation by parts to the base polymer (100 parts). So, I don’t think it is
a mixture design.”
The first objection is easiest to address because “filler” is an oxymoron.

These materials would not be included in the recipe if they did not
perform some function if only to keep extremely potent “active”
ingredients diluted. Beyond that, allowing the studied components
to vary in total creates redundancies in how the ingredients relate to
each other by proportion. To illustrate this latter issue, consider the
following failed experiment brought to one of the authors (Pat) for a
postmortem. It involved the study of a combination of chemicals called
“dopants” that created various colors on a cathode ray tube (CRT),
such as those used in televisions before flat panels took over. These
were mixed in quantities of parts per billion into inert filler. Naturally,
the experimenters figured they’d simply run a factorial design on the
dopants. However, due to the dependency of color on proportions of
these ingredients, this approach failed miserably. Consider mixing one
blue and one yellow, versus two blues and two yellows—either way,
the picture goes green. The ratios of 1:1 versus 2:2 create similar effects.
Unfortunately, without employing the right tool for this problem—a
mixture design—the experimenters unwittingly, are limited, relative
to their exploration of possible proportions of incredibly potent
components.
Practical Magic for Making the Most of a Mixture ◾ 167
APPLYING FACTORIALS AND RESPONSE

SURFACE METHODS (RSM) TO MIXTURES
We layout at the outset of Chapter 11 of RSM Simplified, 2nd ed. on “Applying
RSM to Mixtures” how a factorial design fails for a smoothie comprised of
apples and bananas. This experiment features both a 2-to-1 and a 4-to-2
ratio of these two fruits, which simply scales up the amount of beverage
without changing its taste. Factorials and RSM can work well for cases like
this, but only if the components are varied by ratios. For example, a one-
factor experiment on the ratio of apple to banana ranging from 1 to 2 would
be good. Beyond two components this approach becomes more complex,
as we spell out later in Chapter 11. Using mixture design, when supported
by the proper computing tools, is a lot easier than going the ratio route.
The second objection to mixture design, the one based on formulating

by parts, merits an in-depth rebuttal because of this seductively simple
methodology being so widespread.
Many years ago, we worked with a client who learned RSM. Naturally,
they were very excited by the possibilities of this powerful statistical
tool and applied it immediately to the following recipe for a silicone
sealant:
1. Plasticizer varied from 50 to 100 parts

2. Filler (inorganic) varied from 100 to 250 parts
3. All other ingredients held at 57 parts
4. Polymer set at constant level of 100 parts
The adhesive chemists set up a 13-run central-composite design on

the filler and plasticizer. Figure 11.1 shows the experiment in a standard
order (left to right) with the four two-level (22) factorial core runs
first (1–4), then the four axials (5–8), and finally, the five center
points (9–13).
Unfortunately, as you now can plainly see by the differing bar-
heights, the total amount going into the reactor changed with every new
combination of ingredients. The proportions of the other ingredients (C)
and polymer (D) also vary. Their relative concentrations range up and
500
450
400
350 Plasticizer
300 Filler
Other
250 Polymer
200
150
100
50
0
1 2 3 4 5 6 7 8 9 10 11 12 13
Figure 11.1 Sealant experiment laid out by parts.
100%
90%
80%
70%
Plasticizer
60%
Filler
50% Other
40% Polymer
30%
20%
10%
0%
1 2 3 4 5 6 7 8 9 10 11 12 13
Figure 11.2 Sealant experiment laid out by percentages.
down in a very disorganized fashion as illustrated in Figure 11.2, which

redraws the bar chart by percentages.
The ingredient ranges by weight percent were:
1. Filler, 23%–55%
2. Plasticizer, 11%–28%
3. Other, 11%–19% (not a constant!)
4. Polymer, 19%–33%
The factors studied were ingredients and performance as a function of pro-

portions: This is a mixture! Fortunately, this persuasive presentation helped
the formulators see the light before they wasted a lot of time. With our
guidance, the chemists redesigned their experiment as a three-component
mixture:
1. Filler 25%–55%
2. Plasticizer 11%–28%
3. Polymer 20%–33%
The remainder of 14% was held constant in a thirteen-blend design.

To wrap up this case of the perils of parts, check out the bar chart in
Figure 11.3 displaying the final mixture design.
Notice how the bottom bar remains constant as originally intended,
the chemist only wanting to vary the three components that make up
86% of the fixed total. This is a good case against the use of parts in a
factorial or RSM design. It’s best to use mixture designs to experiment on
formulations.
100%
90%
80%
70%
60% Plasticizer
Filler
50% Polymer
40% Other
30%
20%
10%
0%
1 2 3 4 5 6 7 8 9 10 11 12 13
Figure 11.3 Sealant experiment done with a mixture design.

MIXTURES ARE LIKE A BOX OF CHOCOLATES

Hollywood’s stereotype of a mixture experimenter is a mad scientist—
typically a bug-eyed fellow with very thick glasses and a few wisps of hair
badly combed over his balding pate—crowing over a smoking beaker of
alarmingly colored chemicals. However, a box of chocolate-covered cher-
ries is a mixture too, as recognized by an astute female food scientist far
from the movie character described above. With our help (paid for in
candy), she set up an experiment that modeled consumer response to
various packing defects for their confection:
1. Upside down, 0%–33%

2. Sideways, 0%–33%
3. Leakers, 0%–33%
4. OK, the remainder to total 100%
The chocolatiers were most concerned by pieces that leaked (component C)

when the sticky filling burst their bubble in transit as seen pictured in
Figure 11.4.
This they knew would create complaints, and it turned out to be so
in the experimental results, which you see evidenced by the downward
track for component C in the trace plot (Figure 11.5).
Figure 11.4 Cherry filling leaking out of a broken chocolate.

(Continued )
C
7 D
A
Overall liking
B B
6
D A
C
5
−0.600−0.400−0.200 0.000 0.200 0.400 0.600

Figure 11.5 Trace plot for the chocolate-covered cherry experiment.
However, the mixture design produced an unexpected result—consumers

liked getting some cherries flipped upside down in any given box—hence
the sharp upward slope to component A. Who knew?
My momma always said, “Life is like a box of chocolates. You never

know what you’re gonna get.”
—Forrest Gump (Paramount Pictures, 1994)
Strategy for Formulation Experimentation

In Chapter 8 we devoted our attention to mixture-screening designs, which
we advise for experiments on 6 or more components. The flowchart shown
in Figure 11.6 refines this recommendation by temporarily setting aside
components you know to be important and screening only those whose
impacts on the formulation remain relatively unknown.
After screening previously unknown components, the strategy calls
for combining the vital few with the important ones you set aside earlier
into an optimization design that models nonlinear blending effects. This
requires a quadratic (second degree or second order) or higher (e.g., special
cubic) design. The last strategic phase puts the predictive model to the test
Phase:
Known Unknown
Screening
components components
Designs:
Simplex screening Trivial
Extreme vertices Screening
many
Vital few
Phase:
Optimization
Estimate
Designs: nonlinear
Simplex lattice ≥ 2nd degree blending effects
Optimal for ≥ 2nd order
Phase:
Verification No
Confirm? Backup
Design:
Confirmation runs Yes
Celebrate!!!
Figure 11.6 Flowchart for the strategy of mixture experimentation.
via confirmation runs; for example, those done in the coffee experiment
detailed in Chapter 10, or more rigorous methods (see Martin Bezener’s
March 2015 webinar on “Practical Strategies for Model Verification” posted at
www.statease.com/training/webinar.html).
PROPERLY SIZING YOUR MIXTURE DESIGN

Due to the collinearity of components that is inherent in mixture designs
(i.e., once all but one of the ingredients is set, the last must make up the
difference to the total amount), power calculations fall short. The rea-
son is to ensure that, for a given variable that can naturally be expected
in your blending process, the difference you wish to see will likely be
detected. Instead, we advise you make use of the sizing tools that are
detailed in the section on “Right-Sizing Designs via Fraction of Design
Space (FDS) Plots” of RSM Simplified, 2nd edition, Chapter 12.
(Continued )
A designer knows he has achieved perfection not when there is

nothing left to add, but when there is nothing left to take away.
—Antoine de Saint-Exupery (Wind, Sand,
and Stars, 1939)
Read more at: https://www.brainyquote.com/quotes/
topics/topic_design.html
This concludes our book on mixture design for optimal formulation. We hope
you put these methods to good use for revising your recipe to the sweetest
spot ever.
References
Adulteration. Encyclopaedia Britannica, 10th ed., 1902. www.1902encyclopedia.

com/A/ADU/adulteration.html.
Anderson, M. Mixture design enhances the nectar from an exotic Amazonian fruit.
Stat-Teaser, March 2005. Stat-Ease, Minneapolis, MN.
Anderson, M., and P. Whitcomb. Computer-aided design of experiments for
formulations. Modern Paint and Coatings, June 1997a, pp. 38–39.
Anderson, M., and P. Whitcomb. How statisticians keep things simple. RSM
Simplified. New York: Productivity, Press, 2005.
Anderson, M., and P. Whitcomb. Mixing it up with computer-aided design. Today’s
Chemist at Work, 1997b, p. 34.
Anderson, M., and P. Whitcomb. Mixture DOE uncovers formulations quicker.
Rubber & Plastics News, October 21, 2002, pp. 16–18.
Anderson, M., and P. Whitcomb. DOE Simplified: Practical Tools for Effective
Experimentation, 3rd ed. New York: Productivity Press, 2015.
Anderson, M., and P. Whitcomb. Practical aspects for designing statistically optimal
experiments. Journal of Statistical Science and Application, 2014, 2, 85–92.
Anderson, M., and P. Whitcomb. RSM Simplified: Optimizing Processes Using
Response Surface Methods for Design of Experiments, 2nd ed. New York:
Productivity Press, 2016.
Ansmann, A. et al. Personal care formulations. Chapter 7 of the Handbook of
Detergents: Part D: Formulation, M. S. Showell (Ed.). Boca Raton, FL: CRC
Press, 2005, p. 248.
Bezener, M., and H. Anderson. Brewing the perfect pot of office coffee. Stat-Teaser,
September 2016. Stat-Ease, Minneapolis, MN. www.statease.com/publications/
newsletter/stat-teaser-09-16.
Blackledge, R. Finding the sweet spot of your cricket bat. Live and Breathe Cricket
blog, December 14, 2012. http://liveandbreathecricket.blogspot.com/2012/12/
finding-sweet-spot.html.
Box, G. E. P., and N. R. Draper. Response Surfaces, Mixtures, and Ridge Analyses,
2nd ed. New York: John Wiley & Sons, 2007.
Box, G. E. P., and H. L. Lucas. Design of experiments in nonlinear situations.
Biometrika, 1959, 46, 77–90.
175
176 ◾ References
Cornell, J. Experiments with Mixtures, 3rd ed. New York: John Wiley & Sons, 2002.
Cornell, J., and G. Piepel. Methods for Designing and Analyzing Mixture
Experiments, notes from a short course presented at the Fall Technical
Conference by the Chemical & Process Industries Divisions (CPID) and the
Statistics Division of the American Society for Quality (ASQ), and by the
Section on Physical and Engineering Sciences (SPES) and the Section on
Quality & Productivity (Q&P) of the American Statistical Association (ASA),
October 8, 2008, Phoenix, AZ.
Cox, D. R. A note on polynomial response functions for mixtures. Biometrika,
1971, 58, 155–159.
Crosier, R. B. Mixture experiments: Geometry and pseudocomponents.
Technometrics, 1984, 26(3), 209–216.
Daniel, C., and E. L. Lehmann. Henry Scheffe 1907–1977. The Annals of Statistics,
1979, 7(6), 1149–1161.
Del Vecchio, R. J. Design of Experiments. Hanser/Gardner, Cincinnati, OH, 1997,
pp. 100–101.
DeVries, P. Let Me Count the Ways. Boston, MA: Little Brown, 1965.
Dick, P. K. Valis (Valis Trilogy #1). Santa Ana, CA, 2011. www.amazon.com/
VALIS-Valis-Trilogy-Philip-Dick/dp/0547572417.
Draper, N. R., and I. Guttman. Rationalization of the “alphabetic-optimal” and
“variance plus bias” approaches to experimental design. Technical Report 841,
1988, Department of Statistics, University of Wisconsin.
Draper, N. R., P. Prescott, S. M. Lewis, A. M. Dean, P. W. M. John, and M. G. Tuck.
Mixture designs for four components in orthogonal blocks. Technometrics,
1993, 35(3), 268–276.
Dryden, J. All for Love; Or, The World Well Lost: A Tragedy. 1678. www.bartleby.
com/18/1/.
Goos, P., and A. N. Donev. The D-optimal design of blocked and split-plot
experiments with mixture components, Research Report 0303, Departement
Toegepaste Economische Wetenschappen, Katholieke Universiteit Leuven,
January 2003, p. 1. https://lirias.kuleuven.be/bitstream/123456789/118367/1/
OR_0303.pdf.
Guidance for Industry. Q8(R2) Pharmaceutical development. U.S. FDA, November
2009, p. 9.
Hensley, C. Design of experiments helps reduce time to remove aerospace
coatings. Aerospace Engineering & Manufacturing Technology Update, 2008,
pp. 21–23. http://articles.sae.org/2916/.
Humphrey, J. W., J. P. Oleson, and A. N. Sherwood. Greek and Roman Technology:
A Sourcebook. Routledge, New York, 1998.
Kalichevsky, V. A. The Amazing Petroleum Industry. New York: Reinhold, 1943, p. 7.
Koons, G. F., and M. H. Wilt. Design and analysis of an acrylonitrile—butadiene—
styrene (ABS) pipe compound experiment. Computer Applications in Applied
Polymer Science, Chapter 27. Washington, DC: American Chemical Society,
1982, pp. 439–448.
References ◾ 177
Kowalski, S., J. A. Cornell, and G. G. Vining. A new model and class of designs
for mixture experiments with process variables. Communication in Statistics:
Theory and Methods, 2000, 29(9–10), 2255–2280.
Kowalski, S., J. A. Cornell, and G. G. Vining. Split-plot designs and estimation
methods for mixture experiments with process variables. Technometrics, 2002,
44(1), 72.
Kris-Etherton, P., R. H. Eckel, B. V. Howeard, et al. Lyon Diet Heart Study: Benefits
of a Mediterranean-lifestyle, National Cholesterol Education Program/American
Heart Association Step I dietary pattern on cardiovascular disease. Circulation,
2001, 103, 1823–1825.
Lawson, J., and C. Willden. Mixture experiments in R using mixexp. Journal of
Statistical Software, 2016, 72, 1–20.
Lewis, S. M., A. M. Dean, N. R. Draper, and P. Prescott. Mixture designs for
q components in orthogonal blocks. Journal of the Royal Statistical Society.
Series B (Methodological), 1994, 56(3), 457–467.
McLean, R. A., and V. L. Anderson. Extreme vertices design of mixture
experiments. Technometrics, 1966, 8(3), 449.
Meadows, S. L., C. Gennings, W. H. Carter Jr., and D. S. Bae. Experimental
designs for mixtures of chemicals along fixed ratio rays–classic methodology
for detecting and characterizing departures from additivity, isobolograms.
Environmental Health Perspectives, 2002, 110(Suppl 6), 979.
Myers, R. H., D. C. Montgomery, and C. M. Anderson-Cook. Response Surface
Methodology, Process and Product Optimization Using Designed Experiments,
3rd ed. New York: John Wiley & Sons, 2009.
Piepel, G. F. Measuring component effects in constrained mixture experiments.
Technometrics, 1982, 25, 97–105.
Piepel, G. F. Programs for generating extreme vertices and centroids of linearly
constrained experimental regions. Journal of Quality Technology, 1988, 20(2),
125–139.
Piepel, G. F., and J. A. Cornell. Designs for mixture-amount experiments. Journal
of Quality Technology, 1987, 19(1), 11–28.
Roesler, R. R. How to bake the perfect cake. Paint & Coatings Industry, 2004.
Roudsari, S. F., R. Dhib, and F. Ein-Mozaffari. Mixing effect on emulsion
polymerization in a batch reactor. Polymer Engineering and Science, 2014,
55(4), 945–956.
Ruhlman, M. Ratio: The Simple Codes behind the Craft of Everyday Cooking, 2009,
Scribner, New York, p. ix.
Sahrmann, H. F., G. F. Piepel, and J. A. Cornell. In search of the optimum Harvey
Wallbanger recipe via mixture experiment techniques. American Statistician,
1987, 41, 190–194.
Scheffé, H. Experiments with mixtures. Journal of Royal Statistical Society, 1958,
B20, 344–360.
Scheffe, H. Simplex-centroid design for experiments with mixtures. Journal of the
Royal Statistical Society, Series B (Methodological), 1963, 25(2), 235–263.
178 ◾ References
Simon, H. Administrative Behavior: A Study of Decision-Making Processes in

Administrative Organization, 4th ed. New York: The Free Press, 1997, p. xxix.
Singh, S. K., J. Dodge, M. J. Durrani, and M. A. Khan. Optimization and
characterization of controlled release pellets coated with an experimental
latex: I. Anionic drug. International Journal of Pharmaceutics, 1995, 125(2),
243–255.
Smith, W. Experimental Design for Formulation. Philadelphia, PA: ASA-SIAM Series
on Statistics and Applied Probability, 2005.
Snee, R. D. Experimental designs for mixture systems with multicomponent
constraints. Communications in Statistics, 1979, 303–306.
Snee, R. D., R. W. Hoerl, and G. Bucci. A statistical engineering approach to
mixture experiments with process variables. Quality Engineering, 2016, 28(3),
263–279.
Snee, R. D., and D. W. Marquardt. Extreme vertices designs for linear mixture
models. Technometrics, 1974, 16(4), 399–408.
Snee, R. D., and D. W. Marquardt. Screening concepts and designs for experiments
with mixtures. Technometrics, 1976, 18(1), 24.
Spagnola, L., J. Klang, M. Gupta, and X. Drujon. New advances in
UV-curable soft-touch coatings. Coatings World, October 8, 2016.
https://www.coatingsworld.com/issues/2016-08-01/view_features/
new-advances-in-uv-curablesoft-touch-coatings/.
Suzuki, J. A History of Mathematics. Upper Saddle River, NJ: Prentice Hall, 2002,
p. 731.
Thomson, J. R., and J. R. Thomson. Taste Senior Editor. Which has
more caffeine, light or dark roast? Here’s the truth. Huffington
Post, June 27, 2016. http://www.huffingtonpost.com/entry/
coffee-roast-caffeine_us_576c8725e4b017b379f57ff8.
Vieira, M. C., and C. L. M. Silva. Optimization of a cupuacu (Theobroma
grandiflorum) nectar formulation. Journal of Food Process Engineering,
2004, 27, 181–196.
Vojnovic, D., B. Campisi, A. Mattei, and L. Favretto. Experimental mixture design
to ameliorate the sensory quality evaluation of extra virgin oils. Chemometrics
and Intelligent Laboratory Systems, 1995, 27, 205–210.
Whitcomb, P. Mixture Design for Optimal Formulations Workshop. Section 1.
Slide 3. Stat-Ease, Minneapolis, MI, 2009.
Whitcomb, P. J., and M. J. Anderson. Using DOE with Tolerance Intervals to Verify
Specifications. 11th Annual Meeting of the European Network for Business
and Industrial Statistics (ENBIS), University of Coimbra, Coimbra, Portugal,
September 2011, pp. 4–8. www.statease.com/pubs/using_DOE_with_
tolerance_intervals.pdf.
About the Software
To make mixture design for optimal formulation easy, this book is

augmented with a fully functional, time-limited, version of a commercially
available computer program from Stat-Ease, Inc.—called Design-Expert®
software. Download this Windows-based computational tool, as well
as enlightening tutorials, from www.statease.com/rsm-simplified-2.html.
There you will also find files of data for most of the exercises in the book.
The datasets are named to be easily cross-referenced with corresponding
material in the book. Also, if you work through all the problems
(the only way to a working knowledge of mixture design), check your
answers by downloading them from the book’s website.
You are encouraged to reproduce the results shown in the book and to
explore further. The Stat-Ease software offers far more detail in statistical
outputs and many more graphics than can be included in this book. You
will find a great deal of information on program features and statistical
background in the help system built into the software.
Technical support for the software can be obtained by contacting:
Stat-Ease, Inc.
2021 East Hennepin Ave, Suite 480
Minneapolis, MN 55413
Telephone: 612-378-9449
Fax: 612-378-2152
E-mail: support@statease.com
Website: www.statease.com
179
Index
Note: Page numbers followed by f and t refer to figures and tables respectively.
A Black and Tan, 39

Blending. See Linear blending; Nonlinear
Adhesive experiment, 107–108, 108f blending
Adjusted R-squared statistic, 29 Blocking, 37, 39, 43
Alka Seltzer “rockets,” 17–20, 18t, 19t, 20f on bread flours, 156–160, 157t, 158f, 159t,
Analysis of variance (ANOVA), 9, 15 160f
Antagonism templates for, 155–156
nonlinear blending behavior and, Blue Moon wheat ale. See Beer cocktail
12–13, 12f experiment
with three-components, 51–52, 51f Bread flours blocking experiment
Archimedes, principle of buoyancy of, 4 with modern tools for optimal design,
Augmented simplex lattice (ASL), 60–62, 60t 157–160, 159t, 160f
component quantity and, 93 proportions for, 156, 157t
for extra-virgin olive oil study, 62–64, results of, 156–157
63t–64t trace plot for, 156, 158f
Budweiser American lager. See Beer cocktail
B experiment
Buoyancy, principle of, 4
Beer cocktail experiment
contour plots for overall liking and cost C
in, 44–46, 45f, 46f
cost as a function of, 44, 44f Caffeine, 163
design for, 36–37, 36f Categorical variables, 149–151, 150f, 151f
history of, 38–39 Centroid. See also Simplex centroid
including cost in, 40t, 44 of Alka Seltzer rockets, 19, 19t
linear model for, 43–44, 43f of gold and copper mixture, 6t
measurement for, 37–38, 38t for rheology modifier experiment, 114, 114t
overview of, 35 for simplex screening, 130, 131t, 132
response surface for, 41–42, 41f for tableting case-study, 96, 97t
results of, 39, 40t in trapezoid nonsimplex, 88
Black and Blue Moon beer cocktail. on trilinear graph paper, 32, 32f
See Beer cocktail experiment in U-pseudo coding, 84, 85f
181
182 ◾ Index
Check blends Computer-aided numerical search, for

for Alka Seltzer rockets, 19, 19t extra-virgin olive oil study,
in beer cocktail experiment, 36f, 37, 40t 68, 68f
in extra-virgin olive oil study, 63, Confidence interval (CI)
63t–64t computation of, 123–124
in rheology modifier experiment, 114, 114t for QbD, 119, 124
in shampoo experiment, 90, 91f for tablet formulation experiment, 121,
for simplex lattices, 58, 60–61 122f, 123
for simplex screening, 130, 131t, 132 Constrained space
in solvent study, 49–50, 50t establishment of, 73–74, 74f
in tableting case-study, 96 pseudo coding for, 76–78, 77t, 78f
for trapezoid nonsimplex, 88, 88f rescaling of component levels for, 75–76,
Chemistry, 23 76t, 76f
Chocolate-covered cherry experiment in shampoo experiment, 89–93, 90t, 90f,
170–171, 170f, 171f 91f, 92f
CI. See Confidence interval (CI) upper constraints, 82–86, 82f, 83f,
Coded equations, 42–43 84f, 85f
Coefficients, in beer cocktail experiment, Constraints. See Minimum constraints;
41–42 Multicomponent constraints
Coffee experiment, 161–163, 162t (MCCs)
Combined mixture designs Contour plots
categorical variables, 149–151, 150f, 151f for beer cocktail experiment, 44–46,
design problem for, 152 45f, 46f
KCV models of, 153–154, 154t of desirability, 118, 119f
mixture-amount design, 143–149, 144f, for extra-virgin olive oil study, 69, 69f
146t, 149f for rheology modifier experiment,
nonlinear, 154 116–117, 116f, 117f
split plots in, 160–161 for shampoo experiment, 90–91, 91f
Components, 7–8. See also Four-component for tablet formulation experiment, 121,
mixture; Three-component mixture; 122f–123f
Two-component mixture Controlled-release ibuprofen tablet case
ASL and quantity of, 93 study
in EVD, 134–135 I-optimal selection for, 145–147, 146t
hard-to-change, 160–163, 162t MAD for, 144–145, 144f
KCV models for, 153–154, 154t R-squared and model reduction for,
polynomials for, 5–10, 127–128, 128t 147–148, 148t
for pound cake experiment, 105, 106t sweet spot for, 148–149, 149f
relative levels of, 78, 78f Coordinate exchange, 96
rescaling of levels of, 75–76, 76t, 76f Copper and gold mixture. See Gold and
screening for vital, 127–128, 128t copper mixture
for shampoo experiment, 89 Correlation (r), inverse, 8
trace plot for combining, 108–109, 109f Cost
on trilinear graph paper, 31–32, 32f beer cocktail experiment inclusion of,
Composite material 40t, 44–46, 44f, 45f, 46f
categorical variables in, 149–150, 150f quality correlation with, 35–36
essentials for, 150 in rheology modifier experiment,
optimal recipe for, 150–151, 151f 114–117, 114t, 116f, 117f
Index ◾ 183
Cox direction response surfaces for, 67–68, 67f

Piepel direction versus, 138–139, 140f special cubic for, 64–65
in pipe experiment, 139–141, 139t, 140f trace plot for, 65–66, 65f, 66f
Cubic equations. See Third-order mixture Extreme programming, 14
model Extreme-vertices algorithm, 99–100, 100t
Cupuacu fruit nectar, 80, 81f Extreme-vertices screening design (EVD)
component elimination in, 137–138
D components in, 134–135
D-optimal criterion for, 135, 136t
Degrees of freedom (df) gradients for, 136–137, 137t
with gold and copper mixtures, 6 linear effects in, 138
replication and, 61 trace plot for, 135, 137f
Depainting, 84
Desirability F
contour plot of, 118, 119f
details on, 118 Factorials, 167
rheology modifier experiment and, Factorial space, 46–47, 47f
114–118, 114t, 116f, 117f Factors, 7–8. See also Components
Detergent experiment FDS. See Fraction of design space (FDS)
multiple-response optimization for, Fillers, 165–166
124–125 Film-canister rocketry, 17–20, 18t, 19t, 20f
pseudo coding for, 76–78, 77t, 78f First-order mixture model
rescaling of component levels for, 75–76, for beer cocktail experiment,
76t, 76f 43–44, 43f
Df. See Degrees of freedom (df) lack-of-fit test for, 26, 27f, 28, 28t, 28f
DOE Simplified, xvi, 1, 15 number of terms in, 53t
D-optimality, 93–94 for rheology modifier experiment,
for extreme-vertices screening design, 114–115
135, 136t shape of, 22
Drug combinations, isobolograms for, 13 for two components, 11
Fisher, R. A., 5
E Flour, 104–105
Flow, in rheology modifier experiment,
Elasticity, of solid rocket fuel, 51–52, 51f 114–117, 114t, 116f, 117f
Equilateral triangle, how to draw, 56 Formulation experiments. See also specific
Eutectic, 9–10 experiments
Eutectic reaction, 9 application of, xvi–xvii
EVD. See Extreme-vertices screening design history of, 2–3
(EVD) proper sizing for, 172–173
Extra-virgin olive oil study strategy for, 171–172, 172f
augmented simplex lattice for, 62–64, Four-component mixture. See also
63t–64t Extra-virgin olive oil study
computer-aided numerical search for, 68, simplex centroid for, 33–34, 34f
68f simplex lattice for, 58–59, 58f
contour plot for, 69, 69f in tetrahedral space, 55–56, 56f
design for, 62–63, 63f with two categorical variables,
guidelines for, 62 151, 151f
184 ◾ Index
Fourth-order mixture modeling Hydrazine, 20

lack-of-fit test for, 28, 28t, 28f Hypergolic, 20
number of terms in, 53t
special quartic, 52, 53t I
for two-component mixture, 24
Fraction, parts compared with, 5 Inverse correlation, in gold and copper
Fraction of design space (FDS), 86 mixture, 8
Full cubic equation. See Third-order I-optimality, 93–94
mixture model for mixture-amount design, 145–147, 146t
for tablet formulation experiment, 120,
120t–121t
G
for tableting case-study, 96
Gallons per 10,000 (gpm), 16–17 Isobolograms, 13
Gasoline blends, 15–16. See also High-octane Isooctane, 129
simplex screening
Gibbs, J., 32 J
Gluten, 104–105
Jelly bean experiment, 48–49, 48f
Gold and copper mixture
with different blends, 5–6, 6t
K
linear blending model of, 4
LOF test for, 26, 27f, 28, 28t, 28f KCV models, 153–154, 154t
response surface for, 7–10, 8f KISS principle, 25
R-squared statistic for, 29–30, 29t
SMSS for, 25, 26t, 27f L
synergism of, 3–4, 4f
gpm. See Gallons per 10,000 (gpm) Lack-of-fit (LOF) test
application of, 26
H for gold and copper mixture, 26, 27f, 28,
28t, 28f
Half-head test, 89 for pound cake experiment, 106t
Hard-to-change (HTC) replication and, 61
coffee experiment, 161–163, 162t for tableting case-study, 96, 97t
process, 164, 164f Lattice, 57
split plot for, 160–161 Least squares regression, for melting point
Harvey Wallbanger experiment of gold and copper mixture, 7
bundt cake, 75 Linear blending, 12–13, 12f
setting minimum constraints in, Linear effects, 138
73–74, 74f Linear equation. See First-order mixture model
High-octane simplex screening Lithium citrate, 81
design for, 128–130, 130f LOF test. See Lack-of-fit (LOF) test
measuring component effect for, Lower constrained space
133–134, 134t establishment of, 73–74, 74f
results for, 130, 131t, 132 pseudo coding for, 76–78, 77t, 78f
trace plot for, 132, 132f rescaling of component levels for, 75–76,
Hill-climbing algorithm, for beer cocktail 76t, 76f
experiment, 45 in shampoo experiment, 89–93, 90t, 90f,
HTC. See Hard-to-change (HTC) 91f, 92f
Index ◾ 185
L-pseudo-component coding, 76–78, 77t, 78f minimum constraints for, 73–74, 74f
inverted simplex in, 82–83, 83f parts and, 167–169, 168f, 169f
paint remover experiment, 84, 84f quintessentials of, 2–3
RSM application to, 167
M split plots for, 161, 164, 164f
mpg. See Miles per gallon (mpg)
MAD. See Mixture-amount design (MAD) Multicomponent constraints (MCCs)
MCCs. See Multicomponent constraints application of, 101
(MCCs) combining components, 108–111,
Mediterranean diet, 70 109f, 110t, 110f
Melting point, eutectic, 9–10 pound cake experiment, 104–107,
Melting point of gold and copper mixture 106t, 107t
different blends and, 5–6, 6t ratio constraints, 107–108, 108f
response surface for, 7–10, 8f simple compared with, 101–102,
synergism of, 3–4, 4f 102f, 103f
Miles per gallon (mpg), 15–17 sources of, 103
Minimum constraints Multiple response optimization, 113
establishment of, 73–74, 74f desirability, 114–118, 114t, 116f, 117f, 119f
pseudo coding for, 76–78, 77t, 78f for detergent experiment, 124–125
rescaling of component levels for, 75–76, quality by design, 119–124, 120t–121t,
76t, 76f 122f–123f
Mixture-amount design (MAD) MX “Peacekeeper” rockets, 20
I-optimal selection for, 145–147, 146t
sweet spot for, 148–149, 149f N
three-component, 143–145, 144f
Mixture-amount experiment Nitrogen tetroxide, 20
categorical variables, 149–151, 150f, 151f Nonlinear blending
example of, 3 analogy for, 14
mixture-amount design, 143–149, 144f, idiom for, 23
146t, 149f mathematical model for, 7–8
Mixture model, 7 second-order equation for, 12–13, 12f
coded equation for, 42–43 third-order, 158–159
first-order. See First-order mixture model Nonlinear combined mixture models, 154
fourth-order. See Fourth-order mixture Nonsimplex
modeling creation of, 87, 88f
second-order. See Second-order mixture extreme vertices design for, screening,
model 134–138, 136t, 137t, 137f
special cubic. See Special cubic
special quartic, 52, 53t O
third-order. See Third-order mixture
model Octane number, 129
Mixtures. See also Four-component mixture; Olive oil study. See Extra-virgin olive oil
Three-component mixture; Two- study
component mixture Optimal design
definitions of, 1–2 for any feasible region, 93
factorial application to, 167 categorical variables in, 149–151,
fillers and, 166 150f, 151f
186 ◾ Index
Optimal design (Continued) PRESS. See Predicted residual sum of

criteria for, 93–94 squares (PRESS)
extreme vertices, 89–93, 90t, 90f, 91f, 92f Process factors
in formulation experimentation strategy, hard-to-change, 164, 164f
171, 172f KCV models for, 153–154, 154t
nonsimplex constraints in, 87–88, 88f Propagation of error (POE), 125
summary for, 97–98 Proportions, in mixtures, 2
for tableting case-study, 94–95 Pseudo, spelling trick for, 79–80
using software for, 95–96, 97t Pseudo coding
Orthogonal design matrix, 86 with lower constraints, 76–78,
77t, 78f
P Piepel direction in, 139, 140f
for shampoo experiment, 90, 91f
Paint remover experiment, 83 with upper constraints, 82–86, 82f, 83f,
response surface for, 85–86, 85f 84f, 85f
U-pseudo coding for, 83–85, 84f, 85f Pure blends, 6, 6t
VIF for, 86 Pure error
Parsimony, 12, 24–25, 43, 51 replication for, 6
Parts for simplex screening, 130
failures of, 167–169, 168f, 169f P-values, 15
fraction compared with, 5 in beer cocktail experiment, 42
Perturbation plot, trace plot versus, 66–67 lack-of-fit tests and, 28, 28t
Petrochemicals, experiments on, 2 SMSS and, 25
Pharmachemical jargon, 145
Piepel direction Q
Cox direction versus, 138–139, 140f
in pipe experiment, 139–141, 139t, 140f QbD. See Quality by design (QbD)
Pipe experiment, 139–141, 139t, 140f Quadratic equation. See Second-order
Play putty experiment, 98–99 mixture model
POE. See Propagation of error (POE) Quality, cost correlation with, 35–36
Point exchange Quality by design (QbD)
for shampoo experiment, 92–93, 92f confidence interval in, 121, 122f,
for tableting case-study, 96, 97t 123–124
Polynomial degree. See Mixture model contour plots for, 121, 122f–123f
Pound cake experiment, 104–107, definition of, 119
106t, 107t I-optimal design for, 120, 120t–121t
components for, 105, 106t numerical search for, 121, 122f
design of, 104 tolerance interval in, 123–124, 123f
results from, 106–107, 107t Quartic equations. See Fourth-order mixture
trace plot with both flours separate, modeling
108–109, 109f Quintessence, 2–3
trace plot with flours combined, 109–110,
110t, 110f R
Predicted residual sum of squares
(PRESS), 148 r. See Correlation (r), inverse
Predicted R-squared statistic, 29–30 Randomization
PreDOE, 15 in beer cocktail experiment, 37, 39
Index ◾ 187
in extra-virgin olive oil study, 63, S

63t–64t
importance of, 5 Sam Adams Black Lager. See Beer cocktail
restriction of, 155 experiment
Ratio constraints, 107–108, 108f Scheffé, H., 2–3
Reference blend, 65, 65f mixture model by, 7
Relative component levels, 78, 78f response prediction equations from, 11
Replication Scientific method, xvii
in beer cocktail experiment, 36f, 37, 44 Screening
in gold and silver mixtures, 6, 6t component selection for, 127–128, 128t
in pound cake experiment, 106t extreme vertices design for non-simplex,
for pure error, 6 134–138, 136t, 137t, 137f
for simplex lattices, 60–61 in formulation experimentation strategy,
for simplex screening, 130, 131t 171, 172f
in solvent study, 50, 50t measuring component effect for,
in tableting case-study, 96, 97t 133–134, 134t
Rescaling of component levels, 75–76, 76t, 76f simplex, 128–133, 130f, 131t, 132f
for shampoo experiment, 90, 90t, 90f for vital components, 127
Response surface Second-order mixture model
for beer cocktail experiment, 41–42, 41f for beer cocktail experiment, 43–44, 43f
for cost as a function of beer cocktail derivation of, 7
composition, 44, 44f design for, 14
for extra-virgin olive oil study, 67–68, 67f graphical depiction of, 12f
for melting point of gold and copper KCV models for, 153–154, 154t
mixture, 7–10, 8f lack-of-fit test for, 28, 28t, 28f
for paint remover experiment, 85–86, 85f linear and nonlinear blending in,
for simplex lattices, 62 12–13, 12f
Response surface modeling (RSM) number of terms in, 53t
mixture application of, 167 response prediction with, 11
Scheffé’s polynomial compared with, 7 for rheology modifier experiment, 114–115
three-component, 22 shape of, 22
trace plot versus perturbation plot, 66–67 simplex lattice for, 57–59, 58f
Rheological properties of paint, 115 Sequential model sum of squares (SMSS)
Rheology modifier experiment application of, 25–26
blends and test runs for, 114, 114t for gold and copper mixture, 25, 26t, 27f
contour plots for, 116–117, 116f, 117f 7-Up, 81
desirability for, 117–118, 119f Shampoo experiment
specifications for, 115 algorithm for finding vertices in, 99–100,
Rockets, Alka Seltzer, 17–20, 18t, 19t, 20f 100t
RSM. See Response surface modeling components for, 89
(RSM) contour plot for, 90–91, 91f
RSM Simplified, xvi, 1, 15, 153 point exchange for, 92–93, 92f
R-squared statistic pseudo coding for, 90, 91f
adjusted, 29 rescaling of component levels for, 90,
for gold and copper mixture, 29–30, 29t 90t, 90f
model reduction for, 147–148, 148t shape for, 89–90, 90t, 90f
predicted, 29–30 specifications for, 98
188 ◾ Index
Silicone super ball, 99 sequential model sum of squares, 25–26,

Simplex, 57. See also Nonsimplex 26t, 27f
Simplex centroid U-pseudo coding and, 86
application of, 35 Sweet spot
for beer cocktail experiment, 36–37, 36f for controlled-release ibuprofen tablet
of constrained region, 74, 74f case study, 148–149, 149f
for four-component mixture, 33–34, 34f in multiple response optimization, 113
inverted, 82–83, 82f Synergism
pseudo coding for, 76–78, 77t, 78f of gold and copper mixture, 3–4, 4f
rescaling of component levels in, 75–76, nonlinear blending behavior and, 7,
76t, 76f 12–13, 12f
for Teeny-Beany Experiment, 48–49, 48f with three-components, 51–52, 51f
for three-component mixture, 33, 33f
Simplex lattice T
augmented. See Augmented simplex
lattice (ASL) Tablet formulation experiment
construction of, 57 contour plots for, 121, 122f–123f
for four-component mixture, 58–59, 58f goal for, 119–120
number of design points in, 59 I-optimal design for, 120, 120t–121t
response surface of, 62 numerical search for, 121, 122f
for three-component mixture, 57–58, 58f Tableting case-study, 94–95
for two-component mixture, 57 using software for, 95–96, 97t
Simplex screening, 128–133, 130f, Taylor polynomial series, 11
131t, 132f Teeny-Beany Experiment, 48–49, 48f
SMSS. See Sequential model sum of squares Ternary diagram
(SMSS) of adhesive experiment, 108, 108f
Snee example, 101–102, 102f, 103f with constrained components, 73–76,
Solid rocket fuel, 20 74f, 76f
elasticity of, 51–52, 51f of L-pseudo coding, 77–78, 78f
Solvent study, 49–50, 50t of shampoo experiment, 90, 90f
Special cubic, 51–52, 51f, 53t of Snee example, 102–103, 102f, 103f
for extra-virgin olive oil study, 64–65 of upper component constraints, 82–83,
simplex lattice for, 59 82f, 83f
Special quartic, 52, 53t of U-pseudo coding, 84–85, 85f
Split plots Tetrahedral space
coffee experiment, 161–163, 162t construction of, 55–56, 56f
in combined designs, 160–161 simplex lattices of, 58–59, 58f
for HTC process, 164, 164f Third-order mixture model.
for mixture, 164, 164f See also Extra-virgin olive oil study
mixture application only of, 161 lack-of-fit test for, 28, 28t, 28f
Stainless steel, 33, 33f number of terms in, 53t
Statistical procedures. See also Lack-of-fit shape of, 22
(LOF) test simplex lattice for, 57–59, 58f
blocking, 37, 39 special cubic, 50–52, 51f
for modeling mixture data, 12 special quartic for, 52, 53t
R-squared statistic, 29–30, 29t, for three-component mixture,
147–148, 148t 50–52, 51f
Index ◾ 189
for two-component mixtures, 21–23, first-order modeling of, 11

21f, 23f fourth-order modeling of, 24
Third-order nonlinear blending, 158–159 second-order modeling of, 11–14, 12f
Three-component mixture. See also Beer simplex lattice for, 57
cocktail experiment third-order modeling of, 21–23, 21f, 23f
mixture-amount design, 143–145, 144f
simplex centroid for, 33, 33f
simplex lattice for, 57–58, 58f U
stainless steel, 33, 33f Upper constrained space, 82–86, 82f, 83f,
Teeny-Beany Experiment, 48–49, 48f 84f, 85f
third-order modeling of, 50–52, 51f U-pseudo coding
trilinear graph paper for, 31–32, 32f equation for, 83
TI. See Tolerance interval (TI) for paint remover experiment, 83–86,
Titan missile, 20 84f, 85f
Tolerance interval (TI) statistical properties gained by, 86
computation of, 123–124
for QbD, 119, 124
for tablet formulation experiment, 123, 123f V
Trace plot
for bread flours blocking experiment, Variance inflation factor (VIF), 86
156, 158f Vertices
for component combining, 108–109, 109f algorithm for finding, 99–100, 100t
Cox versus Piepel direction, 138–139, 140f in Alka Seltzer rockets, 19, 19t
for EVD, 135, 137f in beer cocktail experiment, 36f, 37,
for extra-virgin olive oil study, 65–66, 40t, 61
65f, 66f in extra-virgin olive oil study, 63, 63t
for high-octane simplex screening, in gold and copper mixture, 6t, 61
132, 132f for rheology modifier experiment,
perturbation plot versus, 66–67 114, 114t
in pipe experiment, 139–141, 140f in shampoo experiment, 89–93, 90t, 90f,
for pound cake experiment with both 91f, 92f
flours separate, 108–109, 109f in simplex centroid, 33–34
for pound cake experiment with flours in simplex lattice, 57–59, 58f
combined, 109–110, 110t, 110f for simplex screening, 130, 131t, 132
production of, 65–66, 66f for tableting case-study, 96, 97t
variations in drawing, 133 in tetrahedral space, 55–56, 56f
Trapezoid nonsimplex, 88, 88f in trapezoid nonsimplex, 88
Trial of the Pyx, 10–11 on trilinear graph paper, 31, 32f
Trilinear graph paper. See also Ternary in U-pseudo coding, 84, 85f
diagram View foils, 45–46
application of, 31–32, 32f VIF. See Variance inflation factor (VIF)
factorial space compared with, 46–47, 47f Viscosity, in rheology modifier experiment,
invention of, 32 114–117, 114t, 116f, 117f
stainless steel on, 33, 33f
Tunnel of Fudge, 75 W
Two-component mixture. See also Gold
and copper mixture Weight fraction, 5

Formulation Simplified

Uploaded by

Copyright:

Available Formats

Formulation Simplified

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Formulation Simplified

Uploaded by

Copyright:

Available Formats

Formulation Simplified

A PRODUCTIVITY PRESS BOOK

© 2018 by Mark J. Anderson, Patrick J. Whitcomb, Martin A. Bezener

Productivity Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-1380-5604-6 (Hardback)

and the Productivity Press site at

1 Getting Your Toe into Mixtures ...................................................1

Practice Problems ......................................................................................48

Ratio Constraints ..................................................................................... 107

References ...................................................................................... 175

“experimental design can lead to feelings of pure joy, especially when it

Pat, Martin, and I are indebted to several contributors to the development

Mark J. Anderson, PE, CQE, MBA, is a principal and general manager

Patrick J. Whitcomb, MS, is the founding principal of Stat-Ease, Inc.

Martin A. Bezener, PhD, is a principal and statistician with Stat-Ease,

There are many paths to enlightenment. Be sure to take one with

A MODEL FOR PREDICTING WHAT FORMULAS

Flowchart of recommended readings

Getting Your Toe

Simplicity is the ultimate sophistication.

◾ “Mixtures are combinations of ingredients (components) that together

◾ “What makes a mixture?

The first definition by Cornell and Piepel provides a practical focus on

THE QUINTESSENTIALS OF MIXTURES

With mixtures, the property studied depends on the proportions of

All That Glitters Is Not Gold

PURE BLENDS REPLICATED TO PROVIDE PURE ERROR

Another important element of this experimental design is the replication

Generating a Beautiful Response Surface—Like

Melt point = 1043 x1 + 1072 x 2 − 536 x1x 2

THE JARGON OF DESIGN OF EXPERIMENTS

in the Cuban Missile Crisis (dating ourselves here!)—the proper descriptor

Two component mix

A: Gold (wt %) 0.0 25.0 50.0 75.0 100.0

OTHER DEPRESSING COMBINATIONS OF MATERIALS

A FISHY WAY TO BLEND 50/50

TRIAL OF THE Pyx

Details on Modeling the Performance

First order ( linear ): y = β1x1 + β2 x 2

Second order (quadratic): y = β1x1 + β1x1 + β12 x1x 2

HOW TO ORDER UP JUST THE RIGHT

The mentally disturbed do not employ the Principle of Scientific

—Philip K. Dick (“Valis,” p. 16, Mariner Books, 2011)

0 0.25 0.5 0.75 1 X1

Figure 1.3 Graphical depiction of second-order mixture-model.

When experimenting with mixtures, it pays to design an experiment that

AN ANALOGY FOR NONLINEAR BLENDING

A LIFE PRESERVER FOR THOSE WHO BELIEVE

Consider these questions:

INVERSE TRANSFORMATION PUTS

Figure 1.4 Apparatus for film-canister rocketry.

HEADS UP—DO NOT PICK PRANKSTERS AS

He then set up a two-component mixture design that provided the extreme

Table 1.2 Results from Film-Canister Rocket Experiment

BLASTING OFF FROM TUCSON, ARIZONA

Appendix 1A: Cubic Equations for Mixture

y = β1x1 + β2 x 2 − δ12 x1x 2 ( x1 − x 2 )

0 0.25 0.5 0.75 1 X1