A New Statistical Correlation between

Shear Wave Velocity and Penetration

Resistance of Soils Using Genetic
Programming
Gholamali Dehghan Nayeri
M. Sc. in Geotechnical Engineering
Department of Civil Engineering, Faculty of Engineering, Yazd
University, Yazd, Iran
e-mail: gdehghannayeri@yahoo.com

Davood Dehghan Nayeri

Undergraduate in Civil Engineering
Islamic Azad University, Central Tehran Branch

Kazem Barkhordari
Assistant Professor
Department of Civil Engineering, Faculty of Engineering, Yazd
University, Yazd, Iran

ABSTRACT
To predict the effects of site conditions in aggravation of ground motion response, accurate
knowledge of dynamic soil characteristics such as shear wave velocity (Vs) is required.
Standard penetration test (SPT) is one of the most common tests for estimation of soil
characteristics. However, if the shear wave velocity of soils can be obtained using SPT
results, then the dynamic characteristics of soils would be in hand in a very economic method.
Since available correlations between Vs and SPT-N are not with enough accuracy to be used
in most sites, the aim of this study was to focus on the representation of more applicable and
accurate relation between Vs and SPT-N. Genetic programming (GP) is one of the most
powerful branches of “soft computing” methods, widely used in geotechnical engineering. A
relatively large data set including 613 pairs of small strain shear wave velocity (Vs) and SPT
blow count (N) measured in geotechnical studies in Iran was collected and utilized to
establish new statistical correlation between two mentioned parameters. The results showed
that the developed correlation, using GP, is more accurate than simple regression correlation
and statistical preciseness of the model is high enough that can be used with the high degree
of reliability.
KEYWORDS: Shear wave velocity; Standard penetration test; Genetic programming

- 2071 -
 Vol. 18 [2013], Bund. K 2072

INTRODUCTION
The importance of site effects in altering the seismic ground response for local hazard studies
has been widely recognized [1, 11]. The evaluation of site effects causing large amplifications of
ground motion, affected by local topography, surface and subsurface geology, lateral
discontinuities, soil types and variation of their properties with depth, has become a requirement
in microzonation studies and building designs. The major factor which controls site effects is
shear wave velocities (Vs) of surficial sediments. With this view point, accurate knowledge of
geometry and near-surface shear wave velocities of sediment deposits are important parameters
for estimation of ground motion amplifications, attenuation mechanisms in seismic wave
transmissions and initial soil stiffness. Shear wave velocity is expressed in terms of the small
strain modulus, Gmax = ρVs2, where ρ is the total soil mass density.
Although averaged shear wave velocities in uppermost layers have been used widely to assess
site conditions, appropriate procedure for estimation of shear wave velocity is still the key issue.
In situ measurements such as cross-hole, downhole or surface techniques, spectral analysis of
surface waves (SASW), cone and seismic cone penetration tests (CPT-SCPT), seismic refraction
microtremor (ReMi), multi-station analysis of surface waves (MASW) and refraction-reflection
seismic methods, are needed to determine the shear wave velocity (Vs). The high cost and noise
levels of field tests and space constrains in populated areas, have limited the applicability of these
methods and driven the alternative indirect methods, such as SPT tests, for estimating Vs values.
Standard penetration test (SPT) is one of the most common and economical in situ tests,
associated with many geotechnical design parameters of soils. A significant number of studies is
available in literature, proposing statistical correlations between shear wave velocity and
penetration resistance of soils [2, 3, 12, 14]. Some of these relations are dependent on material
types (clay, silt or sand), depth, geological age or corrected penetration resistance.
Brandenberg et al. [5] have been represented the shear wave velocity as a statistical function
of SPT-N and vertical stress, using a data-set obtained from various California bridge sites. Their
relations were derived for different soil types: sand, silt and clay. Dikmen [6] suggested new
correlations between N values and Vs for several soil categories (all soils, sand, silt and clay-type
soils). Based on the obtained results, he suggested that if the uncorrected blow counts were taken
into account, the better correlation of Vs estimation could be acquired. Hasancebi and Ulusay [8]
represented new correlation between SPT-N and Vs and compared their relationship with previous
studies in Turkey. They suggested that the number of blow count is a significant parameter in
obtaining Vs and SPT-N correlations, while the influence of soil type can be neglected. Iyisan [9]
studied the variations of Vs measured using in situ methods with variations of SPT-N values. The
effects of soil type, mean grain size, effective overburden pressure and depth on the correlations
were investigated. Jafari et al. [10] proposed new Vs–SPT-N correlations for fine grained soils
based on field geoseismic investigations in south of Tehran.
In general, the correlations between Vs and SPT-N value have considerable dispersals. This
may be due to various methods used to determine the shear wave velocity and SPT-N value and
also other differences such as, geotechnical and geological conditions and so on. Another reason
of this low accuracy is the type of regression analysis to obtain a correlation between Vs and SPT-
N value. In our knowledge, previously published relationships between Vs and SPT-N value
developed using simple regression analysis. As a result, such relationships have a high degree of
errors and inaccuracies. In order to attenuate such errors, a branch of evolutionary computing
methods, namely Genetic Programming (GP), was used in the current study to develop a more
 Vol. 18 [2013], Bund. K 2073

accurate relationship between Vs and SPT-N value based on database obtained from geotechnical
studies in Iran.

GENETIC PROGRAMMING
Nowadays, soft computing methods, such as evolutionary computing methods are widely
used in many geotechnical fields [4, 7]. Genetic Programming (GP) is a branch of evolutionary
computing methods which can suitably be used for pattern recognition purposes. Population of
GP is represented by parse trees, including non-linear individual entities of different sizes and
shapes. Individuals are structures consisting of functions and terminals, which both are selected
out from larger set of functions and terminals. Each function contains basic arithmetic operators
(+, -, ×, etc.) and Boolean logic functions (AND, NOT, etc.) and also user defined functions.
However, terminals contain constant parameters that form the whole individual structure together
with functions. In each generation, these functions and terminals are selected randomly and can
form many individuals with different fitness values. The fitness value strongly depends on the
definition of fitness function, which is the most influential element in GP. In the current study,
the fitness function is root mean squared error. At each step, by selecting the best individuals and
then breeding them together using GP operations (e.g. cross-over and mutation), a new generation
will be created. The procedure could be continued until the predefined fitness value would be
obtained or specified number of generations would be reached. At the end, the best individual
will be proposed as the best model between input(s) and output data. A new GP toolbox for
MATLAB was used in current study [13].

GP BASED CORRELATION BETWEEN VS AND SPT-N

The database used in current study contains 613 pairs of Vs and SPT blow count (SPT-N),
measured in Iran. The Vs data collected in database were measured by various methods including
downhole (a great part of database), seismic refraction, and SASW (Spectral Analysis of Surface
Waves) methods. These measurements were carried out on different types of soils varying from
coarse grained (sandy soils) to fine grained soils (silty and clayey soils).
Initially, a simple regression analysis was performed to ensure that there is a need to increase
the accuracy of the estimation of Vs using SPT-N. For this purpose, the most accepted functional
form of relation between Vs and SPT-N value proposed in the literature, namely Vs = aNb with a
and b, constant parameters, that must be determined using statistical regression, was used as a
basic format of simple regression analysis. The following relationship was obtained using all data
presented in our database:

Vs = 129.4N0.336 (1)
where Vs is the shear wave velocity (m/sec) and N is the uncorrected SPT blow count. To
evaluate the performance of the models presented in current study, three criteria were used;
coefficient of determination (R2), root mean squared error (RMSE), and mean absolute error
(MAE), which are defined as:
 Vol. 18 [2013], Bund. K 2074

 (m
i =1
i − pi ) 2
R2 =1− n
(2)
 (m
i =1
i − m) 2

 (m
i =1
i − pi ) 2
RMSE = (3)
n

m
i =1
i − pi
MAE = (4)
n
where m and p are the values of measured Vs in the field and predicted Vs by the proposed
model, respectively, m is the mean of measured values, and n is the number of data presented in
database. The values of R2, RMSE, and MAE for Eq. (1) are 0.25, 207.0, and 147.5, respectively.
It should be noted that dimensions of RMSE and MAE are in m/sec. These values demonstrate
the low performance and poor preciseness of the model developed using simple regression and
therefore, the necessity of developing a new more accurate model with less error than the simple
regression based models were felt.
In order to establish a correlation between Vs and SPT-N with high degree of accuracy, the
GP was implemented. For this purpose, the 80% of all data were set into training sub-set and
remaining part of data, i.e. 20% of all data, were included into testing sub-set, so that the
maximum, minimum, mean, and standard deviation of Vs and SPT-N in two sub-sets were as
equal to each other as possible.
Finally, based on the results obtained from various analysis containing various combinations
of initial parameters of GP, the best of them was selected based on the values of R2, RMSE, and
MAE and is expressed as the following form:

0.0576 N 10.88527 N 1.35276 6.39661

V s = 136.581 + + + +
A A + 2.91052 A + 1.18224 A − 0.4613
(5)
1.4529 2.9866
− −
A + 0.17753 A + 1.03558
where

A = 0.01918 N − 2.38576 (6)

The measured versus predicted values of Vs for testing, training, and all data-sets are shown
in Figures 1-3, respectively. As can be seen from these figures, all data points are located at the
vicinity of the bisector line corresponded to measured Vs equal to predicted Vs with rather
uniform scattering.
In Table 1, values of R2, RMSE, and MAE for testing, training, and all data-sets are
summarized. Corresponded values for simple regression based model are also presented in Table
1 for comparison purposes. It could be observed that the GP based model have higher R2 value
 Vol. 18 [2013], Bund. K 2075

and lower RMSE and MAE values than simple regression based model. In GP based model the
value of R2 was improved about 88% relative to simple regression based model. Furthermore, the
amount of RMSE and MAE were decreased about 15.9% and 14.5%, respectively. Consequently,
it can be mentioned that the GP based model estimated the Vs more accurately than simple
regression based model and this model was capable to reduce the errors in estimating Vs using
SPT-N.
In order to investigate the statistical performance of previously published models between Vs
and SPT-N in current database, some of these models which all were developed using simple
regression analysis were selected. In Figure 4, their performances are summarized on the basis of
R2 and RMSE. It should be mention that, as negative values of R2 indicate the underestimation of
the method, and also for simplicity, all negative values of R2 obtained from previous studies were
set to zero in Figure 4. It can be seen that the model developed using GP in the current study,
performed better than the all previously published models. The GP based model has the highest
R2 and the lowest RMSE among the models illustrated in Figure 4.

Figure 1: Measured versus predicted Vs in training sub-set

CONCLUSION
In the current study, a new application of evolutionary computing methods in the field of
geotechnical earthquake engineering, seldom considered previously, was investigated. Genetic
Programming, as a sub-area of such methods, was implemented to minimize the error presented
in the estimation of shear wave velocity using penetration resistance of soil. Genetic
Programming can appropriately discover the one functional form between input(s) and output that
fits the data with as higher accuracy as possible.
Vol. 18 [2013], Bund. K 2076

Figure 2: Measured versus predicted Vs in testing sub-set

Figure 3: Measured versus predicted Vs in all data-set

 Vol. 18 [2013], Bund. K 2077

Table 1: Statistical performance of GP and simple regression analysis based models

Simple regression
Parameters GP based model analysis based Improvement Ratio (%)
model
training testing all data all data

R2 0.49 0.39 0.47 0.25 88

RMSE (m/sec) 171.1 185.3 174.0 207.0 15.9
MAE (m/sec) 123.7 136.0 126.1 147.5 14.5

Figure 4: Performance evaluation of some previously published models with models

developed in the current study
Using a relatively large database contains pairs of Vs and SPT-N values, a GP based model
developed which obtained an estimation of Vs using uncorrected SPT blow count. It was
observed that the R2, RMSE, and MAE of GP based model correlation between Vs and SPT-N
were considerably improved relative to relationship obtained using simple regression analysis.
The value of R2 increased about 88% in GP based model while the values of RMSE and MAE
decreased 15.9% and 14.5%, respectively.
The performance evaluation of previously published models in current database revealed that
the GP based model was with higher preciseness than all other published models. It should be
mention that the use of soft computing methods, such as evolutionary computing and artificial
neural network, can considerably improve the accuracy of Vs estimation using penetration
resistance of soils. Therefore, with regard to important effects of Vs on dynamic behavior of soils
and site response analysis, the results of this study help geotechnical engineers to estimate the Vs
where it is necessary with more accuracy than before.
Vol. 18 [2013], Bund. K 2078

REFERENCES
1. Ashford, S. A., Sitar, N., Lysmer, J. and Deng, N. (1997) “Topographic Effects on
the Seismic Response of Steep Slopes, ” Bulletin of the Seismological Society of
America, 7(3), 701-709.
2. Athanasopoulos, G. A. (1995) “Empirical Correlations of Vs-NSPT for Soils of
Greece: A Comparative Study of Reliability” Proc. 7th Int. Conf. on Soil Dyn.
Earthquake Eng., 19-36.
3. Baziar, M. H., Fallah, H., Razeghi, H. R. and M. M. Khorasani (1998) “The Relation
of Shear Wave Velocity and SPT for Soils in Iran,” Proc. 11th European Conf.
Earthquake Eng., Paris, France.
4. Baziar, M. H., Jafarian, Y., Shahnazari, H., Movahed, V. and Amin Tutunchian, M.
(2011) “Prediction of Strain Energy-based Liquefaction Resistance of Sand–Silt
Mixtures: An Evolutionary Approach, ” Computers & Geosciences, 37, 1883-1893.
5. Brandenberg, S. J., Bellana, N. and Th. Shantz (2010) “Shear Wave Velocity as
Function of Standard Penetration Test Resistance and Vertical Effective Stress at
California Bridge Sites, ” Soil Dyn. Earthquake Eng., 30, 1026-1035.
6. Dikmen, Ü (2009) “Statistical Correlations of Shear Wave Velocity and Penetration
Resistance for Soils,” J. Geophys. Eng., 6, 61-72.
7. Ebrahimian, B., Movahed, V. and Yousefnia Pasha, A. (2012) “Evaluation of
Undrained Shear Strength of Marine Clay using Cone Penetration Resistance at
South Pars Field in Iran, ” Ocean Engineering, 54, 182-195.
8. Hasancebi, N. and R. Ulusay (2007) “Empirical Correlations between Shear Wave
Velocity and Penetration Resistance for Ground Shaking Assessments, ” Bull. Eng.
Geol. Environ., 66, 203-213.
9. Iyisan, R (1996) “Correlations between Shear Wave Velocity and In-Situ Penetration
Test Results, ” Digest, 96, 371-374.
10. Jafari, M. K., Shafiee, A. and A. Razmkhah (2002) “Dynamic Properties of Fine
Grained Soils in South of Tehran, ” JSEE, 4, No.1.
11. Nazari, A., Baziar, M. H. and Shahnazari, H (2010) “Seismic Effects of Two-
Dimensional Semi-Sine Shaped Hills on Ground Motion Response, ” Electronic
Journal of Geotechnical Engineering EJGE, 15/L,
http://www.ejge.com/2010/Ppr10.091/Ppr10.091r.pdf.
12. Ohsaki, Y. and R. Iwasaki (1973) “On Dynamic Shear Moduli and Poisson’s Ratios
of Soil Deposits, ” Soils Founds., 13(4), 61-73.
13. Searson, D (2009) “GPTIPS: Genetic Programming & Symbolic Regression for
MATLAB, ” http://gptips.sourceforge.net.
14. Seed, H. B. and I. M. Idriss (1981) “Evaluation of Liquefaction Potential of Sand
Deposits based on Observation of Performance in Previous Earthquakes, ” ASCE
National Convention (MO), 81-544.

© 2013 ejge