Computer Assisted Mechanics and Engineering Sciences, 18: 231–241, 2011.

c 2011 by Institute of Fundamental Technological Research, Polish Academy of Sciences

Copyright

Applying Artificial Neural Networks

for analysis of geotechnical problems

Maria J. Sulewska
Bialystok University of Technology, Faculty of Civil and Environmental Engineering
Wiejska 45A, 15-351 Białystok, Poland
e-mail: m.sulewska@pb.edu.pl

The paper presents a discussion of some applications of Artificial Neural Networks (ANNs) in geo-
engineering using the analysis of the following six geotechnical problems, related mainly to prediction
and classification purposes: 1) prediction of Overconsolidation Ratio (OCR), 2) determination of poten-
tial soil liquefaction, 3) prediction of foundation settlement, 4) evaluation of piles bearing capacity, 5)
prediction of compaction parameters for cohesive soils, 6) compaction control of embankments built of
cohesionless soils. The problems presented are based on the applications of the Multi-Layered Perceptron
(MLP) neural networks.

Keywords: geotechnical problems, artificial neural networks (ANNs), Multi-Layer Perceptron (MLP),
Overconsolidation Ratio (OCR), bearing capacity of piles, settlement of foundation, soil liquefaction,
compaction control.

1. INTRODUCTION

The present study contains a review of selected applications of Artificial Neural Networks (ANNs)
in geotechnical engineering, where they appeared to be useful in the analysis of various problems.
The literature references indicate that ANNs have been successfully used to solve very important
geotechnical problems. From among many problems the following ones were selected: a) 3D mate-
rial characteristics, b) modelling unsaturated soils behaviour under static and dynamic loadings,
c) modelling pressure-deformation curves σ-ε for statically and dynamically loaded soils. In par-
ticular: ANNs have also been applied in: d) analysis of slope stability assuming a circular or wedge
failure mechanism, e) evaluation of the geotechnical parameters on the base of mineral composi-
tion for classification of soils referring to their particle size or structure, f) prediction of hydraulic
conductivity of cohesive soils, g) density and compaction parameters of various soil types or other
properties, for description of suction, collapse potential, and lateral pressure phenomena, h) mod-
elling subsoil-building interaction during mining shocks, i) evaluation of measurement uncertainties,
etc. The list of references related to application of ANNs in geotechnics is included in [14].
The aim of the paper is to discuss the possibilities of ANN applications in geotechnics in the
analysis of six geotechnical problems, solutions related to prediction and classification purposes,
listed in the abstract of this paper. The most representative publications are discussed in [14]. In
the present paper all the neural network support applications were related to the Multi-Layered
Perceptron (MLP), see Haykin [18], of architecture MLP: D-H-M , where: D – number of inputs,
H – number of hidden neurons, M – number of outputs.

2. MEASURES OF ERRORS

The models were evaluated on the basis of the analysis of the following error measures calculated
separately for learning and testing data sets:
232 M.J. Sulewska

• Root Mean Squared Error:

v
u P
u1 X (p) (p)
RMSi = t (ti − yi )2 , (1)
P
i=p

(p) (p)
where ti , yi are target and neurally computed i-th outputs (i = 1, . . . , M ) for p-th pattern
(p = 1, . . . , P );
• Mean Absolute Error:
P 
P 
 (p) (p) 
 ti − y i 
p=1
MAEi = ; (2)
P
• Maximum of Absolute Relative Error:
(p)
MAREi = max(REi ); (3)

• Relative Error can be defined as:

 
 t(p) − y (p) 
(p)  
REi =  i (p) i  100%; (4)
 ti 

• Mean Relative Error:

P
1 X (p)
MREi = REi ; (5)
P p=1

• Regression coefficient Ri (and determination coefficient Ri2 ):

P
P (p) (p)
(ti − ti )(yi − yi)
p=1
Ri = s , (6)
P
P P
P
(p) (p)
(ti − ti )2 (yi − y i )2
p=1 p=1

(p) (p)
where ti , y i are mean values of sets {ti } and {yi }.

3. EXAMPLES OF ANN APPLICATION IN GEOTECHNICS

3.1. Prediction of Overconsolidation Ratio (OCR)

The ANN was applied in the analysis of results from studies carried out on clays. The Piezocone
Penetration Test (PCPT) was applied in order to model the pressure history in the soil, which
was characterized by means of Overconsolidation Ratio (OCR) [6]. The OCR is the ratio of the
maximum past vertical effective stress in a soil in the past σp′ to the existing effective overburden
stress in a soil σv′ . The OCR is a parameter that characterizes the strength, stress-strain behaviour,
and the compressibility characteristics of cohesive soils. It is determined in a laboratory during
a durable oedometric test using undisturbed soil samples collected in the field.
A set of 195 study results from intact (165 results), and fissured (30 results) clay deposits, in
soil ranging from soft, normally consolidated clays to very stiff, heavily overconsolidated clays, were
collected. 63% of the data were selected to learn the network, while the remaining 37% were used
to test the network. The ANN inputs were composed of variables achieved from the field surveys
applying Piezocone Penetration Test basing on the following quantities:
Applying Artificial Neural Networks for analysis of geotechnical problems 233

1. qt – the corrected cone resistance;

2. σv0 – total vertical overburden pressure;
3. u1 – pore water pressure at the cone tip;
4. u2 – pore water pressure measured just above the cone base;
5. u0 – the hydrostatic pore water pressure.
The only output of the ANN is the OCR value from laboratory studies performed in oedometer.
Three neural networks were formulated: MLP1 – network tested by all the data related to
intact and fractured clays, MLP2 – trained network, using only data for intact clays, and MLP3 –
network using data only for fissured clays. It was found that MLP2 and MLP3 became specialised,
i.e. predicted the OCR values. They were specified for intact and fissured clays better than MLP1
that was trained for all data.
The comparison of the OCR prediction quality (within the set of testing data) by neural as
well as empirical and theoretical models (developed by various authors on the basis of Piezocone
Penetration Testing) is presented in Table 1.

Table 1. Mean OCR values achieved using particular methods [6].

Ratio of OCR value predicted

by a model to OCR value from the test
Model Intact clays Fissured clays
Average Standard deviation Average Standard deviation
MLP1 5-8-1 1.07 1.00 1.37 1.46
MLP2 5-8-1 1.00 0.44 – –
MLP3 5-5-1 – – 1.28 1.09
Tumay et al. [17] 2.00 1.77 1.30 0.96
Sully et al. [16] 1.32 0.97 1.61 1.67
Mayne [7]: OCR = f (u1 ) 2.70 2.49 2.17 2.14
Mayne [7]: OCR = f (u2 ) 1.52 1.32 2.17 2.60
Chen and Mayne [2]: OCR = f (u1 ) 1.04 0.97 0.71 0.66
Chen and Mayne [2]: OCR = f (u2 ) 0.97 0.67 0.80 0.83

It can be seen that for intact clays, MLP2 was the best model. Considering fissured clays on
the basis of Chen and Mayne’s models [2], the OCR values are predicted with some deficiency in
relation to the standards from oedometer test. The MLP3 predicted the OCR values with some
excess compared with oedometric standards.

3.2. Determination of potential soil liquefaction

Liquefaction results from seismic vibrations, mainly in loose and hydrated sands. The phenomenon
is caused by the loss of compressive strength due to the increase of hydraulic pressure within
pores. Prediction of soil liquefaction is difficult because there are many critical factors influencing
liquefaction.
Artificial Neural Networks were applied to evaluate the probability of soil liquefaction [4]. They
aimed at classifying the soil either as liquefied or non-liquefied one. The probability of soil liquefac-
tion was tested on the basis of sand tests applying Standard Penetration Test (SPT) and adopting
234 M.J. Sulewska

the results from other seismic surveys. A set of 85 cases was analysed, including 42 cases represent-
ing liquefied soils, while the phenomenon of soil liquefaction was not recorded in remaining cases.
The networks were learned and tested using 59 and 26 patterns, respectively.
The best neural network consisted of the following eight inputs:

1. σ0 – total vertical stress;

2. σ0′ – effective vertical stress;

3. M – the intensity of earthquake;

4. (N1 )60 − N1 is standardised energy for driving energy in drill rods for 60% of the theoretical
free-fall energy of a SPT hammer;

5. a/g – peak horizontal acceleration at soil surface;

6. τav /σ0′ – equivalent dynamical shear stress;

7. F – percent % of fines content;

8. D50 – mean diameter of particles.

Two optional responses were adopted as the outputs: 1 – when the soil was liquefied, or 0 – when
no soil liquefaction was recorded. Two error classifications resulted from the tests of the network,
which were carried out for the relative error 7.7%. It suggested that neural networks could have
a high success ratio at evaluating the potential soil liquefaction.

3.3. Prediction of foundation settlement

The problem of foundation settlement computation is very complex and noised by a great uncer-
tainty. Thus, many authors started with the application of ANNs.
A set of 189 measurements of shallow foundation settlements was completed for the settlement
of foundation on non-cohesive soils [13] in order to apply and verify ANNs. 80% and 20% of
measurements were selected for the learning and testing data sets, respectively. The following
interval data were accepted as input variables in particular ranges:

1. width of foundation footing B = 0.8–60.0 m;

2. pressure transferred to the soil by foundation footing q = 18.3–697.0 kPa;

3. mean number of hits per 30 cm of the SPT probe to the depth equivalent to 2B below the
founding level and characterizing the soil compaction N30 = 4–60;

4. footing geometry L/B = 1.0–10.5, where L is foundation footing length;

5. footing embedment Df /B = 0.0–3.4, where Df is founding depth.

The measured settlement Sm = 0.0–121.0 mm was the only network output.

The optimizing process resulted in the MLP: 5-2-1 network. The foundation settlement was
calculated according to Meyerhof’s [8], Schulze and Sherif’s [11], as well as Schmertmann’s et al.
equations [10]. Comparison of results obtained by MLP vs. traditional equations [8, 10, 11] (within
the subset of validation data) revealed quite high predictive quality of the neural network against
the formulas used (Table 2 and Fig. 1).
The results demonstrate the prevailing role of the neural model over the traditional methods. It
was found that the foot foundation settlement can be predicted by ANN with acceptable precision
(MAE = 8.78 mm).
Applying Artificial Neural Networks for analysis of geotechnical problems 235

Table 2. Comparison of settlement prediction methods [13].

Error measures ANN 5-2-1 Meyerhof [8] Schultze and Sherif [11] Schmertmann et al. [10]
2
R 0.819 0.160 0.518 0.637
RMS [mm] 11.04 25.72 23.55 23.67
MAE [mm] 8.78 16.59 11.81 15.69

Fig. 1. Settlement values – measured and predicted according to analyzed methods [13].

3.4. Evaluating bearing capacity of piles

From among numerous solutions of pile bearing capacity evaluation, a practical example of neural
networks application for the analysis of information obtained from empirical data was presented
in [3]. ANNs were applied to evaluate friction capacity fs on a side surface of timber and steel pipe
driven piles. The driven piles were immersed in soft and very soft clays. The following data were
collected:
236 M.J. Sulewska

1. embedded pile length L = 4.7–96.0 m;

2. pile diameter D = 13.5–76.7 cm;

3. vertical component of mean effective stress σv′ = 19–718 kPa;

4. undrained shear strength was determined mainly from unconfined compression test or from Vane
Shear Test su = 9–335 kPa;

5. friction capacity from compression tests of piles fs = 8.0–192.1 kPa.

The data completed of 45 and 20 tests were selected, respectively, for the training and testing
of neural networks.
Prediction results of friction capacity fs for four best models with a variable number of inputs,
three neurons in hidden layer and a single output are shown in Table 3. They are compared with
results fs obtained by means of conventional methods discussed in [1, 12].

Table 3. Comparison of methods for evaluating the fs values [3].

Input variables R RMS

Model
of ANN Learning Test Learning Test
MLP2: 2-3-1 su , σv′ 0.985 0.936 1.049 1.267
MLP2A: 2-3-1 su , L 0.857 0.944 2.909 2.326
MLP3: 3-3-1 su , L, σv′ 0.981 0.956 1.162 1.278
MLP4: 4-3-1 su , L, D, σv′ 0.985 0.956 1.016 1.194
Method acc. to Semple and Rigden’s [12] – 0.976 0.885 1.318 1.894
Method β acc. to Burland’s [9] – 0.731 0.704 4.824 3.096

The comparison of fs values predicted by MLP4 network with the standards from studies is
presented in Fig. 2. The figure illustrates that using MLP4 the bearing capacity of piles fs can
be better predicted than applying conventional methods (Figs. 2–4). This is confirmed by a higher
regression coefficient and lower value of RMS error, as well as insignificant scatter of results around
the y = x diagonal line.

Fig. 2. Skin friction fs – comparison of values predicted by MLP4 with standards from measurements [3].
Applying Artificial Neural Networks for analysis of geotechnical problems 237

Fig. 3. Skin friction fs – comparison of values calculated according to Semple and Ridgen’s method
with standards from measurements [3].

Fig. 4. Skin friction fs – comparison of values calculated according to β method

with standards from measurements [3].

3.5. Prediction of compaction parameters for cohesive soil

It can be stated that in case of cohesive soils evaluation of compaction parameters, i.e. the Optimum
Moisture Content (OMC) and the Maximum Dry Density (MDD), by their measurements is highly
labour- and time-consuming. Therefore, a simplified approach for evaluating compaction parameters
on the basis of empirical models, was investigated in [9] applying ANNs.
Different neural models were analysed in [9] basing on the Standard Proctor’s Tests to evaluate
the OMC and MDD parameters. The database of laboratory tests for 39 cohesive soils prepared
artificially (described by 13 variables) and data from 85 different natural soils surveys (described
with 5 variables) was completed. Six and twenty patterns from among the collected data were
selected for testing. The best neural models adopted for compaction parameters, corresponding to
the most important input variables, as well as their MARE error evaluation corresponding to testing
data are presented in Table 4. Evaluation of MLP2: 3-1-2 network for natural soils was presented in
Fig. 5. The network exploitation was performed using the new set of 27 data for worldwide natural
soils, which confirmed quite high predictive usefulness of the model (Fig. 6).
238 M.J. Sulewska

Table 4. Selected neural models built for artificially prepared and natural soils [9].

Soils ANN model Inputs Outputs MARE [%]

Artificially prepared MLP1: 6-4-2 C, S, Sd, Gs , LL, PL OMC; MDD 11.49; 2.51
Natural MLP2: 3-1-2 Gs , LL, PL OMC; MDD 6.26; 2.52
Explanations: C is bentonite content (%), S is silt content (limestone particles with diameter ranging from
0.002 to 0.047 mm) (%), Sd is sand content (sand particles with diameter ranging from 0.074 to 4.76 mm)
(%), Gs is specific density solid particles (g/cm3 ), LL is liquefaction limit according to Casagrande (%),
PL is plasticity limit (%), OMC (%), MDD (pcf, where 1 pcf = 1.016 g/cm3 ).

Fig. 5. Comparison of OMC and MDD values tested for 85 natural soils with those predicted
by MLP2 [9], where 1 pcf = 0.016018 g/cm3 .

Fig. 6. Comparison of OMC and MDD values tested and predicted by MLP2
for new set of 27 natural soils [9].

3.6. Compaction control of embankments built of non-cohesive soils

The Light Falling Weight Deflectometer (LFWD) is one of the field instruments for controlling
the compaction of non-cohesive soils built into the surface layers of embankments. It is applied to
Applying Artificial Neural Networks for analysis of geotechnical problems 239

determine the dynamic deformation modulus ED . In order to determine the quality of embank-
ment compaction applying LFWD, earlier prepared curves of the instrument calibration are used
(i.e. dependence between ED vs. soil compaction measures such as: density index ID or degree of
compaction Is ).
Non-linear dependence of ED vs. ID and Is , as well as the bulk density of soil ρ was studied by
means of the data set completed of 36 measurements for medium sand and 45 measurements for
gravel and sandy gravel [5]. In order to solve the problem, neural network MLP: 3-5-1 architecture
was applied and the results were compared with those from simple linear regression approach,
quoted in [15]. The comparison of ED prediction for medium sand, MLP and regression results is
presented in Figs. 7 and 8.

Fig. 7. Dependence ED = f (ID ) for medium sand [5].

Fig. 8. Dependence ED = f (Is ) for medium sand [5].

The evaluation of the quality of particular prediction methods is illustrated in Table 5. It shows
that the precision of ED prediction by means of ANN is superior to the linear regression method.
240 M.J. Sulewska

Table 5. The error measures for Artificial Neural Network and regression method [5, 15].

MLP: 3-5-1 Regression method

Error measures
ED = f (Is , ID , ρ) ED = f (Is ) ED = f (ID ) ED = f (ρ)
MRE [%] 9.094 10.794 10.521 14.474
MARE [%] 46.033 65.380 64.874 109.494
R 0.937 0.923 0.925 0.861

4. CONCLUSIONS

1. All the examples described above prove that Artificial Neural Networks can be efficiently applied
for solving many geotechnical problems.

2. ANNs approach can be recommended particularly for cases when theoretical process modelling
does not yield expected results, or when some difficulties arise in the application of the statistical
standard methods.

3. Contemporary methods of theoretical modelling have their constraints resulting from incomplete
knowledge of the process or phenomenon understanding. This can affect the complexity or defects
of theoretical analytical tools. Some parameters, influences or effects are neglected in theoretical
models, which is a consequence of simplifications and idealisation of the phenomena described. In
turn, data sets for statistical analysis have to be as large as possible and should be characterised
by their high quality. Sets of data and statistical analyses have to meet many assumptions, e.g.
related to variable distribution normality, however, they are not always met.

4. ANNs are empirical models that have to adopt not so strong assumptions as standard statistical
methods. New and original solutions obtained by means of ANNs applications are constantly
growing. However, an insufficient number of cases for training and testing the ANN can be
a remarkable limitation for ANNs applications.

5. The number of regression type problems in geotechnics, successfully analysed by ANNs, is con-
stantly increasing. In the paper, only the Multi-Layered Perceptron type neural networks are
discussed. There are many problems for which other standard ANNs can be successfully applied.
The application of fuzzy-neural networks and the Bayessian inference methods seem to be very
promising in the future analysis of geotechnical problems, cf. supplementary paper [19].

REFERENCES

[1] J.B. Burland. Shaft friction of piles in clay – A simple fundamental approach. Ground Engineering, 6: 1–15,
1973.
[2] B.S.-Y. Chen, P.W. Mayne. Profiling the OCR of Clays by Piezocone Tests. Report No. CEEGEO-94-1, Georgia
Institute of Technology, Atlanta, 1994.
[3] A.T.C. Goh. Empirical design in geotechnics using Neural Networks. Géotechnique, 45: 709–714, 1995.
[4] A.T.C Goh. A back propagation approach for predicting seismic liquefaction potential in soils. In: Neural Net-
works, IEEE World Congress on Computational Intelligence, 5: 3322–3325, Orlando, USA 1994.
[5] K. Kądziołka, Z. Waszczyszyn, M.J. Sulewska. Determination of dynamic deformation modulus for non-cohesive
soils by Artificial Neural Networks [in Polish]. In: Proc. the 48-th Conf. of Polish Academy of Science, 3: 249–256,
Krynica 2002.
[6] P.U. Kurup, N.K. Dudani. Neural Networks for profiling stress history of clays from PCPT data. Journal of
Geotechnical and Geoenvironmental Engineering, 128: 569–579, 2002.
[7] P.W. Mayne. Determination of OCR in clays by PCPT using cavity expansion and critical state concepts. Soils
Foundations, 31: 65–76, 1991.
[8] G.G. Meyerhof. Shallow foundations. Journal of the Soil Mechanics and Foundation Division, 91(SM2): 21–31,
1965.
Applying Artificial Neural Networks for analysis of geotechnical problems 241

[9] Y.M. Najjar, I.A. Basheer, W.A. Naouss. On the identification of compaction characteristics by Neuronets.
Computers and Geotechnics, 18: 167–187, 1996.
[10] J.H. Schmertmann, J.P. Hartman, P.B. Brown. Improved strain influence factor diagrams. Journal of Geotech-
nical Engineering, 104: 1131–1135, 1978.
[11] E. Schultze, G. Sherif. Prediction of settlements from evaluated settlement observations for sand. In: Proc. of
the 8-th International Conference Soil Mechanics and Foundation Engineering, 1: 225–230, Moscow, 1973.
[12] R.M. Semple, W.J. Ridgen. Shaft capacity of driven pipe piles in clay. Ground Engineering, 19: 11–17, 1986.
[13] M.A. Shahin, H.R. Maier, M.B. Jaksa. Predicting settlement of shallow foundations using Neural Networks.
Journal of Geotechnical and Geoenvironmental Engineering, 128: 110–120, 2002.
[14] M.J. Sulewska. Artificial neural networks in the evaluation of non-cohesive soil compaction parameters [in Polish].
Committee Civil Engineering of the Polish Academy of Sciences, Warsaw-Bialystok, 2009.
[15] M.J. Sulewska. Modulus of deformation for non-cohesive soil determined applying dynamic method [in Polish].
PhD Research, Faculty of Civil and Environmental Engineering, Bialystok University of Technology, 1993.
[16] J.P. Sully, R.G. Campanella, P.K. Robertson. Overconsolidation ratio of clays from penetration pore pressures.
Journal of Geotechnical Engineering Division, 114: 209–216, 1988.
[17] M.T. Tumay, P.U. Kurup, G.Z. Voyiadjis. Profiling OCR and Ko from piezocone penetration tests. In: Proc.
of the International Symposium on Cone Penetration Testing, 2: 337–342, Swedish Geotechnical Society, SGF
Report No. 3, Linkoping, Sweden, 1995.
[18] S. Haykin. Neural Networks – Comprehensible Foundations, 2nd edition. Upper Saddle River, NJ: Prentice-Hall,
1999.
[19] M. Kłos, M.J. Sulewska, Z. Waszczyszyn. Neural identification of compaction characteristics for granular soils,
Computer Assisted Mech. Eng. Sci., 18(4): 265–273, 2011.