Design Variable Selection to Simplify

Process Calculations
WOOYOUNG LEE, JAMES H. CHRISTENSEN, and DALE F. RUDD
University of Wisconsin, Madison, Wisconsin

During the analysis of a large process system, the vaJues of certain design variables must
be preassigned to allow the solution of the design equations. A careless selection of these distinguished variables can lead to trouble, magnifying computational problems above what need
be. A method is proposed for determining a set of design variables which could minimize the
computational labor associated with process analysis. This work is part of a larger project on
the computer aided design of process systems.

Process design is frequently burdened by a. need for extensive computations; a burden which may not be inherent
in the problem itself but which may result from our accidental approach to the problem. A careless first step
may lead directly to horrendous computational problems
which could easily be avoided. This paper deals with the
strategy of making this first step, the selection of design
variables. A simple algorithm is developed which enables
the selection of a set of design variables which reduces the
computational labor associated with the process analysis.
The process design equations can be expressed in symbolic form as
fi to) = 0
i = 1, 2, . . . . IV
/ = 1, 2 , . . . . M

(1)

where N < M.
The design equations ft tie together the process variables Vj through material and energy balances, equipment
design equations, and the like. The equations constitute
independent sources of information and may number into
the hundreds for a typical chemical process. The variables
(flow rates, equipment design parameters, temperatures,
and the like) are even more numerous, the excess variables affording degrees of freedom for optimization.
The process analysis and optimization problem involves:
the numerical specification of certain key design variables;
the solution of the design equations for the unspecified
state, variables; the economic analysis of the results; and
the respecification of the design variables to achieve more
nearly the economic goals.
We wish to determine a set of design variables which
simplifies the solution of the design equations. Our attention is now focused on the first two steps.
THE BIPARTITE GRAPH

We represent the system of equations by a special graph

which exposes the structure of the equations compactly
and lucidly, and which leads directly to their dissection.
This useful notation is the bipartite graph of topology (1,
5).
Consider the following system of equations:
h (i> 2, vz) = 0
fa to, 04, 05) = 0
/s (o4, t>5, e) = 0
U to,s, s) = 0

The corresponding graph is shown in Figure 1.

Page 1104

(2)

f-nodes

v-nodes
Fig. 1. The bipartite graph of Equations (2).

The equations form the / nodes, the variables form the

v nodes, and an edge connects the equations to the corresponding variables.
The number of edges associated with a node defines p,
the local degree of the node. The local degree of the /
node /i is p(fi) = 3 and the local degree of the v node
v5 is f>(vs) = 2.
This symbolic notation is the key to what follows; bipartite graphs have certain properties of extreme interest
to the process engineer.
SELECTION OF DESIGN

VARIABLES

Generally, it is necessary to specify the numerical values of certain of the variables in Equation (1) before solution can begin. The number of such distinguished variables is equal to the degrees of freedom of the system,
that being the excess of variables over equations.

F = M-N

(3)

Equation (3) is valid only when the design equations

constitute independent sources of information, and it gives
no hint as to which F variables should be chosen as design variables.
We are faced with a combinatorial problem, for there
can be very many different ways to select F design variables from N total variables, not all of which lead to a
soluble set of equations. A criterion is required for the
selection of design variables.
Necessary Conditions

A necessary and sufficient condition for the solution of

a set of equations is that the Jacobian determinant be
nonzero (2, 4). While of considerable mathematical significance, this condition is of little direct use to the process
engineer, since it is extremely difficult to compute the
Jacobian determinant for a large system. We must look in
directions other than toward the classical mathematics.

A.I.Ch.E. Journal

November,

1966

In a logically consistent set of IV equations there will be

N unspecified variables, and for each equation it will be
possible to denote one variable as the output variable for
that equation. And, no variable will be an output variable
for more than one equation. If such an output assignment
is possible, the equations are said to satisfy the diversity
condition of Hall (3, 5), a necessary but not sufficient condition for the existence of a solution. This condition might
be stated thus: There must be at least K unspecified variables associated with each and every group of K equations, where K = 1, 2,.. . . N.
For example, the bipartite graph in Figure 2 describes
an insoluble set of equations. Hall's condition fails at K =
2 where equations /i and fa have only i associated with
them.

design vori'qbles

equations

state variables
Fig. 4. YZ and Vp, as design variables.

The fact that an output set has been assigned, denoted

by the double-headed arrows, indicates that the diversity
condition is satisfied. In addition, an order of solution is
suggested by the direction of the arrows. The subgraph
can be arranged as in Figure 5.

Fig. 5. The precedence order of equations.

Fig, 2, The diversity condition not satisfied.

Armed with the diversity condition we consider the

problem of selecting design variables which lead to a
solvable set of equations. Later we consider the selection
which leads to an easily solvable set of equations.
Assigning an output set. We now assign a direction to
the edges of the bipartite graph, a step which identifies
one of many possible orders of solution.
If a variable Vj is assigned as the output variable of
equation ft, then the edge connecting nodes ft and Vj will
be oriented from ft to Vj and all other edges associated
with ft will be oriented to ft.

This graph shows the precedence order of the system

of equations (6). Starting from the left and preceding to
the right, the set of five equations can be solved one at a
time! There is no need for the simultaneous solution of
equations, a troublesome task at best.
When such precedence order exists, the system does
not possess a loop; the structure is said to be acyclic.
Should a loop appear, the simultaneous solution of those
equations within the loop must be undertaken (7).
The Effects of Design Variable Selection

The appearance of an acyclic structure in Figure 5 was

not accidental; rather, it was the direct result of the selection of t>2 and Vg as design variables. In this section we
demonstrate the effects of design variable selection on
the structure of the remaining computational problem.
For example, consider the following system and that in
Figure 6.
/i (i, t>2, s) = 0
fa (8, 4, u 5 ) = 0
(5)
fa (vs, e, i) = 0

F= 7-5=2
Fig. 3. The pipartite graph of Equation (3).

Design variables are not candidates for the output set,

and their edges must be oriented to the equations. For
example, consider the following equations and those in
Figure 3.

F - 6-3 * 3
Fig. 6. The bipartite graph of Equations (4).

, a) = 0

fa

, 2, s, vs) = 0

fa

, s,

=0

(4)

ft a, 4, vs, 6) = 0
fs (5, 6, 07) = 0
If 2 and 1)3 are selected as the design variables, the
subgraph in Figure 4 might result.

Vol. 12, No. 6

Case 1. Coupling Variables vi, v3, and u5 are Design

Variables: Once numerical values are assigned to these
three variables the bipartite graph reduces to Figure 7.
The equations can be solved one at a time, and the order
of solution is immaterial.

A.I.Ch.E. Journal

Page 1105

SELECTION ALGORITHMS

2^

VMJ

It has been established that the careful selection of a

set of design variables can alter the complexion of process
design calculations. We now busy ourselves with altering
the problem in our favor, toward a structure yielding an
easily solved set of equations. A simple algorithm for design variable selection will be established.

V6,

A Selection Criterion

Fig. 7. Vi, Vs, and Vs as design variables.

Case 2; Noncoupling Variables t>2, U4, and 6 are Design Variables: The bipartite graph becomes, for example,
Figure 8, This unfortunate selection of design variables

Fig. 8. Vs, V,j, and Vg as design variables.

leads directly to a set of equations which must be solved

simultaneously. We must now be troubled with the development of an appropriate iterative scheme of solution,
with all the problems of convergence and the like. In the
bipartite graph the order is not even unique, that is, the
directions may be reversed.
Case 3; Variables Vi, Vt, and ve are Design Variables:
The bipartite graph is now acyclic and the information
flow is unique (Figure 9). Equation /3 can be solved first
for variable v$, which then allows the solution of /a for
variable t>3, which in turn enables the solution of equation
fi. The happy situation of case 1 has appeared again, in
Figure 9.

Generally, the labor of solving a set of design equations

increases with the number of equations that must be
solved simultaneously. We, therefore, define the following
criterion for the selection of design variables:
A best set of design variables results in a structure in
which the largest number of equations that must be solved
simultaneously is minimized. A corollary of this criterion is
that a best set of design variables results in an acyclic
ordering of the design equations, if such an ordering is
possible, Hence, if the difficulty of solving any equation is
independent of the output variable and an acyclic assignment is possible, this criterion may give the minimum
total effort for solving the set of design equations.*
An algorithm is now presented which selects a set of
design variables, yielding an acyclic ordering of the design equations, if such an ordering is possible. The algorithm is based on the following observation:
The graph of an acyclic system must contain at least
one v node Vj with P(VJ) = 1 and one / node ft with
p(fi) = 1. This is so because every directed path in an
acyclic graph must terminate, by definition, at a v node.
Such a path cannot terminate at an / node, since each
/ node has one outgoing edge; therefore, it must terminate
at a v node. Furthermore, since a v node can have only
In a report yet to be published, we examine the validity of this
criterion, and show that in situations where the difficulties of solving
parts of the process design differ radically the concept of difficulty scores
must be introduced.

START

Yes
Assign attached v-node as output
of f, and delete both nodes from
grapn.

Assign v. as output of attached

f-node; delete both nodes from
graph.

v. is a design variable;
1 delete it.

Fig. 9. Vj, V4, and Ve as design variables.

We temporarily assume that acyclic graphs are of the

same difficulty equivalence class, and that cases 1 and 3
constitute equivalent selections of design variables. This
is not strictly true, however, since certain equations can
be solved more easily in one direction than the other. For
example, if we have only a table of sin x, it is easier to
solve y = sin x for y in terms of x, than x in terms of y.
The same analogy holds true in process calculations. Later,
this is accounted for by the assignment of difficulty scores
in a report yet to be published.
Page 1106

A.I.Ch.E. Journal

f. is redundant or inconsistent;
1 delete it.
Yes

System is
nonacyclic
STOP

Fig. 10. Flow diagram of algorithm I.

November,

1966

one incoming edge, this v node must have exactly one

edge, if a path terminates at this v node.
If this v node, and the / node of which it is an output,
are deleted from the graph, the resulting subgraph must
also be acyclic; hence, it will also contain at least one v
node and one / node with one edge each.
Finally, we note that any / node connected to only one
v node must necessarily have that v node as its output.
We then have the algorithm for output and design variable
selection shown in Figure 10.

Fig. 13.

Example of Algorithm I

Suppose that we have the system given in Figure 11.

The design equations of the system are also given. We
apply algorithm I to the graph of Figure 12. The system
consists of a mixer, a reactor, and a separator. There are
five streams in the system each described by two variables;
the mass flow rate q and the composition c. The other
variables in the system are the reactor volume V, the reaction temperature T, and the separation ratio S.

Fig. 14.

System:

Separator

Reactor

Feed Mixer

'

'

Product

Fig. 15.

Fig. 11. A simple process system.

Design Equations:
Mixer:
fi qi + q& <72 0
/2 qici + qscs q&z = 0

total mass balance

component balance

Reactor:
/s </2 qa = 0
total mass balance
ft q%C2 tfsCa R(CS, T, V) = 0 component balance
R is the reaction rate relationship
Separator:
/5 c/s qi q& = 0
total mass balance
fe 73^3 ~ 94^4 <7sCs = 0
component balance
/7 S = Cs/Ci
separation ratio
.The variables of interest are qfs and c/s, and V, T, and
S. The number of design equations is IV = 7. The number
of variables is M = 13.
Graph:

Fig. 16. The ordering of Figure 12.

Preferred Design Variables

For purely practical reasons the engineer may prefer to

have close control over certain variables, and may elect
to elevate such preferred variables to the level of a design
variable regardless of the logical structure of the problem.
For example, if quality control restrictions are imposed on
a product concentration, the engineer may wish to compute only designs which meet those restrictions. This can
be achieved by selecting that concentration as a design
variable, the value of which is free to be selected directly.
This preference offers no difficulty. Such preferred variables are assigned prior to the use of algorithm I. If a restriction exists on the flow rate into the reactor, in the pre-

F = 13-7 = 6
Fig. 12. The bipartite graph far the system.

Step 1: Look for nodes with the local degree of 1.

Step 2: Assign c\, V, and S to fa, /4, and /7, respectively.
Step 3: Eliminate these from the graph to get Figure
13.
Repeat steps 1 to 3: Figures 14 and 15 are obtained.
Design variables: T, c2, c4, cs, qa, and </5
Order of calculations: See Figure 16.

Vol. 12, No. 6

A.I.Ch.E. Journal

F -5
Fig. 17. Preferred design variables.

Page 1107

vious example, q2 might be a preferred variable. The selection of (72 as a design variable consumes one degree of
freedom and results in Figure 17, a graph more easily
reducible by algorithm I.
Irreducible Systems

When algorithm I terminated, either an acyclic output

assignment was produced, or p(A) 2 for any node A
in the remaining subgraph. If an acyclic assignment is
possible, then the latter condition is impossible; this can
be proved by contradiction. If the number of / nodes of
this subgraph equals the number of v nodes, our degrees
of freedom have been consumed and we are not free to
select more design variables. A computational scheme for
this subgraph has been already determined and the information flow can be traced out.
Should the degrees of freedom not be fully consumed,
some more design variables are to be selected; however,
an acyclic structure cannot be attained, regardless of our
selection of the remaining design variables. The information flow can be changed acyclic only by tearing some of
state variables. Suppose that if one assigns outputs to k
f nodes and deletes these nodes and their outputs from
the graph; then an acyclic assignment of the remaining
subgraph may be possible by selecting appropriate design variables, where k is always less than the number of
/ nodes of the graph. It is said that k tears are supplied
to the system and that the computation might be performed by a k dimensional iterative scheme. We wisn to
solve the system with the minimum number of tears k",
since the computational burden of iterations increases
with the dimension. Algorithm I cannot possibly find an
acyclic ordering unless there is at least one v node with a
single edge in the subgraph, This observation leads to the
following algorithm.
Algorithm II

1. Apply algorithm I to the graph of the system.

2. If algorithm I succeeds in an acyclic ordering, then
fc* = 0; otherwise, let k = min P(VJ) 1, for Vt in the
vi
remaining subgraph.
3. For each combination of k f nodes giving at least
one v node with just one edge upon their deletion: (1)
Apply algorithm I to the resulting subgraph; if no acyclic
assignment is produced, go on to the next combination.
(2) If an acyclic assignment is obtained, and k of the
variables selected as design variables can be assigned as
outputs of the deleted j nodes, then a minimum number
of tears k* = k has been obtained; otherwise, go to the
next combination.
4. If all combinations of k f nodes have been exhausted,
increment k by one and go to step 3.
As an example, consider applying algorithm II to the
simple system given in Figure 18. Obviously, algorithm I
cannot reduce the size of the graph. We then have min
P(VJ) 2; hence, start with k = 1. Deletion of fi yields

Fig. 19.

the following sequence of edge assignments by using

algorithm I, shown in Figure 19.
Now 2 can be assigned as the output of /i and fc* = 1.
The system can then be arranged as in Figure 20 with
design variables V4 and g. In this case, a minimal tear
was obtained at the first try; however, for large k" the
number of combinations to be considered may become excessively large. It is hoped that further investigation may
yield some way to direct the search in order to reduce the
number of cases which must be examined.

Fig. 20.
ORDERING THE PROCESS FLOW SHEET

Algorithms I and II were designed to aid the computer

during the analysis of the design equations. We now extend these ideas to aid the engineer in the analysis of the
process flow sheet, simplifying the process calculations.
Only a slight extension is required to make this transition.
Figure 21 is the process flow sheet for an example process. The blocks represent unit operations, distillation,
compression, filtration, heating, and so forth, and the arrows trace out the direction of flow of material and energy
between the blocks. Notice that the direction of physical
flow does not necessarily indicate the direction of information flow; the specification on the output product from
stage M actually transfers information back into the system (8).Indeed, we take advantage of the nonuniqueness of the connecting arrows to reorder and simplify the
system. In its present form, the specification of the values
of the design variables leads directly to cumbersome iterative recycle calculations, a condition we shall avoid by the
reassignment of new design variables.

Fig. 21. A process flow diagram.

Fig. 18. Application of algorithm II.

Page 1108

A block in the process flow sheet may be thought to

embrace a number of design equations. For example,
should the block represent a multicomponent distillation
tower it may embrace the hundreds of equations required
for an adequate description of that operation. To apply
A.I.Ch.E. Journal

November, 1966

the algorithms in their present form to such problems is

unnecessary.
Notice that once the input variables to a block are
specified the output variables can be solved for; this is
the usual convention in flow sheet preparation. Thus, we
may think of a block as a special equation which can
possess several output variables, equal in number to the
number of output arrows from the block in the flow sheet,
but not necessarily those particular variables. With this
interpretation we are in a position to apply the algorithms
directly.

Figure 22 is the undirected counterpart of Figure 21

with the number of allowed outputs affixed. The variables
which appear in only one block are denoted as outputs in
Figure 23 with the allowed outputs reduced by the appropriate amount. Those blocks with no more allowed output are eliminated, and any new variables which now appear in only one block are then denoted as outputs from
the reduced graph (Figure 24). When there are more
single block variables than allowed outputs for a block,
one has a choice which can only be made rationally by
physical insight into the type of operation occurring in
the block.
Finally, after all outputs have been assigned this information is placed on the original undirected graph. Any
variables without assigned direction are design variables
(Figure 25), Notice how this assignment has eliminated
all recycle calculations and how the blocks now can be
handled one at a time in the order prescribed in Figure
26,

Fig. 22. An undirected flow diagram.

Fig. 25. A reassignment of design variables for

Figure 21.

Fig. 23.

Fig. 26. A precedence ordering.

FINAL COMMENTS

Fig. 24.

Vol. 12, No. 6

We have demonstrated a systematic approach to the

selection of design variables based on the criterion that
acyclicity is desirable. System acyclicity will be obtained
by use of these algorithms if it is obtainable merely by
selecting the proper design variables. However, there are
several points which require further comment.
It was implied in this report that information flow reversals can be made at any equation or any system component without difficulty. This is not true. It is well known
that difficulties are often encountered should one attempt
to compute certain variables, such as the kind of plates to
use in a distillation tower, and that it is better to specify
such variables and compute the consequences. The direction of information flow in the system determines the ease
of component computation. Thus, in certain situations it
may pay to tolerate cyclic structure in the system to ease
the computation of a troublesome component.
However, before one can begin to understand the implications of partial acyclicity and the role that component
difficulty scores may play in systems analysis, the limited
case discussed in this paper must be presented. Hence, in
this report the problem of systems analysis with completely

A.I.Ch.E. Journal

Page 1109

unhindered information flow reversals and equal component difficulty plays an important role as an introduction
to the more complete problem.
In subsequent reports we will discuss the ordering of
recycle calculations, the effects of direction dependent
component computational difficulties, and the implications
of partial acyclicity. These topics, along with the methods
presented in this first report, may lead to a more orderly
approach to complex systems calculations.
Obviously, the bipartite graph is only useful as a
method of developing the algorithms, and in a practical
problem the ordering of the equations and the selection
of design variables would be performed without the use
of the graph, applying the concept of list processing to a
list of design equations in which the occurrence of each
variable has been noted.
ACKNOWLEDGMENT

The authors wish to thank Donald Engelstad for valuable

suggestions and discussions. This work was supported in part
by a grant from the National Science Foundation.
NOTATION
Ci
F
/i
k

=
=

composition of stream i
degrees of freedom
z'th function
number of tears

kf
= minimum number of tears
M
= total number of variables
N
= total number of equations
qi
= mass flow rate of stream i
R(c, T,V) = reaction rate relationship
S
= separation ratio
T
= reaction temperature
V
= reactor volume
Vj
= /th variable
p(A) = local degree of node A
LITERATURE CITED

1. Berge, Claude, "The Theory of Graphs and Its Applications," Wiley, New York (1962).
2. Buck, R. C., "Advanced Calculus," McGraw-Hill, New
York (1956).
3. Hall, P., ]. London Math. Soc., 10, 26 (1934).
4. Hildebrand, F. B., "Advanced Calculus for Applications,"
Prentice-Hall, Englewood Cliffs, N. J. (1962).
5. Ore, O., "Graphs and Their Uses," Random House, New
York (1963).
6. Rudd, D. F., and D. V. Steward, paper presented at
A.I.Ch.E Las Vegas meeting (Sept., 1964).
7. Steward, D. V., SJAM Rev., 4, 321 (1962).
8. Rudd, D. F., and C. C. Watson, "Strategy in Process Engineering," Wiley, New York, to be published.
Manuscript received January 28, 2966; revilision received April 28,
1966; paper accepted May 3, 1966.

Rate Data and Derivatives

NOEL DE NEVERS
University of Utah, Salt Lake City, Utah

In most rate process problems, the original experimental data must be differentiated to obtain the derivatives, which are the information actually sought. For first derivatives all the
commonly used methods give about the same result. However, the common methods of calculating derivatives can give very different second derivatives for the same set of starting data.
Thus, in problems in which the desired result is a second derivative, the answer obtained is a
strong function of the method used to calculate second derivatives. This is illustrated by
several examples.

Churchill (1) has pointed out one important similarity

between heat, mass, and momentum transfer and reaction
kinetics. In all of these fields, the engineer is interested in
determining rates (heat, mass, momentum transfer, and
chemical reaction rates). These rates are generally obtained by differentiating integral data. For example, the
momentum transfer rate to a particle (force) is proportional to the acceleration. The most common way (14)
to obtain such accelerations is to differentiate positiontime (integral) data twice to obtain d?x/dt2. Similarly,
chemical reaction rates are determined by differentiating
composition-time (integral) data and heat transfer rates
by differentiating temperature-time or temperature-distance data.
Page 1110

In some cases these rates can be obtained from steady

state experiments, for example, steady state heat transfer,
steady state mass transfer, chemical reactions at constant
composition (continuous stirred-tank reactors), or from
force measurements on supported bodies. Even these data
are often differentiated to obtain more interesting information; for example, steady state chemical reaction data are
differentiated with respect to space velocity to elucidate
kinetics (16).
Thus, in all these fields the desired derivatives are obtained from integral data. Given a table of x-y data, there
are several widely known ways to find dy/dx and d2y/dx2.
As will be shown here, all give practically the same values
of dy/dx, but they can give very different values of

A.I.Ch.E. Journal

November, 1966