VOL. 13, NO.

8, APRIL 2018 ISSN 1819-6608

ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

APPLICATIONS OF ARTIFICIAL INTELLIGENCE IN

GEOTECHNICAL ENGINEERING
Nabeel S. Juwaied
College of Engineering, Department of Civil Engineering, Al-Nahrain University, Baghdad, Iraq
E-Mail: [email protected]

ABSTRACT
The use of artificial intelligence (AI) techniques to deal with complex geotechnical problems has recently
escalated. Which may be the result of lack of efficiency of traditional methods, or the promising potential of these
techniques to represent such complexity. Artificial intelligence has been applied in most areas of geotechnical engineering.
In this research, these applications were extensively reviewed and discussed. Where they showed great success in most of
them.

Keywords: artificial intelligence artificial, neural networks, bearing capacity of piles, settlement of foundation, soil liquefaction.

1. INTRODUCTION complex behavior of most geotechnical engineering

Civil engineering deals with enormous variety of materials. Using AI has main advantage of over regression
materials, most of civil engineering materials, such as in process modeling, which is capacity in dealing with
concrete, steel, and timber, which show more homogeneity multiple outputs or responses while each regression model
and isotropy, some of these we can considered it almost as is able to deal with only one response (Park, 2011).
manufactured product, like reinforced concrete, with high Additional to that (Park,2011) referred to another major
possibility control of quality during the production. Some advantage for developing NN process models is that they
other important components are completely natural, like do not depend on simplified assumptions such as linear
soil. behavior or production heuristics make it very suitable to
Geotechnical engineering deals with materials represent materials with many deferent properties as soil.
(e.g., soil and rock). the engineering properties of soil and Neural networks have a number of important properties
rock show varied and uncertain behavior due to the for modeling a complex mechanical behavior: good
complex and imprecise physical processes associated with generalization capability, universal function
the formation of these materials (Jaksa ,1995). Many approximation capability, resistance to noisy or missing
complexity and uncertainty related with geotechnical data, and accommodation of multiple nonlinear variables
engineering materials and it properties. Some sources of for unknown interactions.
uncertainty are incoherent soil composition, errors in soil On can notes that starting from 90s, there is
boring, sampling, in-situ and laboratory testing, and increasingly in employed AI as an effective tool in
characterization of the shear strength and stiffness of soils, geotechnical engineering. AI techniques are artificial
and loading effects, time effects, construction effects, neural networks (ANNs), genetic programming (GP),
human error. As an example difficulty in obtaining evolutionary polynomial regression (EPR), support vector
undisturbed samples as in sand and gravel, cause machines, M5 model trees, and K-nearest neighbors
uncertainty of laboratory testing results, and the same (Elshorbagy et al., 2010), among AI techniques, artificial
clayey soil often exhibits uncertain behavior from one neural networks (ANNs) most successful applied in
location to another. geotechnical engineering including.
The mathematical model is a form created in
science. The fundamental principles of physics and 2. APPLICATION OF AI IN GEOTECHNICS
mechanics have been used to create this model, and used
to predict, simulate, and analyze system behavior. The 2.1 Site characterisation
mathematical model will be appropriate when known the The main objective of site characterization is the
basic condition of a system, the measured inaccuracy and prediction of in situ soil properties at any half-space point
uncertainty did not decrease the accuracy of the model at a site based on limited tests. Site characterization is an
(Rahman and Mulla, 2005). Because, the complexity in area concerned with the analysis and interpretation of
geotechnical problems, its exact solution is the probability geotechnical site investigation data. A multi-layer feed
(Djajaputra, 1997; Griffith et al., 2002). forward neural network is used by (Zhou and Wu, 1994) to
Traditional forms of engineering modelling capture the spatial distribution of the granite rock head by
approaches are very simplified to deal with the such training the network with sample data from seismic
complexities. Some promise is appearing in alternative refraction surveys on over 11 km of traverse lines with
approach “Artificial intelligence (AI)”, which has showed shot-point spacing of 50 and 25 meters for geophones
succeeded in the field of geotechnical engineering. placed at 10 and 5 meters, respectively. The neural
Artificial intelligence (AI) has been shown high network model requires information on the coordinates
predictive ability comparing to traditional methods and as (x,y) and surface elevation of a survey point in the input to
a result of that, it became widely usable in modeling the characterize the spatial distribution of rock head elevations

2764
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

characterize the spatial distribution of rock head counter propagation ANN is a promising parameter
elevations. The trained network was tested to estimate the estimation method when integrating multiple data types to
rock head elevations for all locations within the area of enhance prediction accuracy and reduce uncertainty.
investigation by producing a corresponding contour map. Samui (2012) based on large amount of Standard
Results from the neural network model compare very well Penetration Test, used the Support Vector Machine
with similar contour maps generated using kriging (SVM), to develop a three dimensional site
techniques. The main advantage of this neural network characterization model. Also a sensitivity analysis of SVM
based approach is its ability in establishing patterns or parameters ( , C, and ε) has been presented. The results
relationships through training directly on the data without obtainable in this paper clearly highlight that the SVM is a
building any complicated mathematical models and strong tool for site characterization.
making assumptions on spatial variations. Gomes et al. (2016). discussed the prediction of
Najjar and Basheer (1996a); Basheer et al. (1996) spatial patterns in the depth to bedrock (DTB) using high-
discussed application related to ground water resolution topographic data, numerical modeling, and
characterization by used neural networks to map the Bayesian analysis has been introduced the different
variation of permeability for purposes of identifying building blocks of a DTB model to predict the vertical
boundaries of landfill to be constructed on a real site. The extent of the weathered rock underling soil-mantled
neural network, as a simple technique, was found to be hillslopes from a high-resolution topographic map of the
able to logically predict the variation. The sensitivity of soil surface. Two case studies with synthetic and real-
the produced permeability maps to both the quality and world regolith depth data were used to illustrate the
number of observations was also studied to investigate the effectiveness and applicability DTB model and
accuracy of the proposed mapping methodology. methodology. The results demonstrate that the proposed
Concluded that, the use of neural networks as a mapping DTB model with lumped parameters mimics reasonably
tool can help identify the regions within a site where well the observed regolith depth data with root mean
additional subsoil exploration is warranted. square error (RMSE).
Rizzo and Dougherty (1996) applied and tested a
new pattern method on a variety of site characterization 2.2 Soil properties and behaviour
problems, called it “SCANN” (Site characterization using There is an increasing in employment of ANNs in
Artificial Neural Networks), is based on the application of modelling soil properties and its behavior. Poor learning
artificial neural networks. Applications include developing resulting from overtraining the network was studied using
maps of discrete spatially-distributed fields (e.g. log- artificially generated triaxial data. (Penumadu et al.,
hydraulic conductivity fields) given estimates of hydraulic 1994), provide modelling rate-dependent behavior of clay
conductivity from pumping tests and classifying soil soils. Neural network methods have been used for
lithology given soil sample descriptions from driller well estimating the permeability of clay liners by Gribb and
logs. Unlike the kriging methods, SCANN is data-driven Gribb (1994); Agrawal et al., 1994); and Najjar and
and requires no estimation of a covariance function. It uses Basheer, 1996b).
a feed-forward counter propagation training approach to Goh (1995a; 1995c). provide a study
determine a "best estimate" or map of a discrete spatially- demonstrates the feasibility of using neural networks for
distributed field. capturing nonlinear interactions between various
Gangopadhyay et al (1999) developed a parameters in complex civil engineering systems. A
multilayer perceptron using the back-propagation simple back-propagation neural network was used to
algorithm for characterizing the subsurface and geographic model two problems involving engineering correlations
information system (GIS) the integrated approach of ANN between various soil parameters were used to model the
and GIS, is shown to be a powerful tool for characterizing correlation between the relative density and the cone
complex aquifer geometry, and for calculating aquifer resistance from cone penetration test (CPT), for both
parameters for ground-water flow modeling. normally consolidated and over-consolidated sands. Based
Besaw et al (2006) present a subsurface on calibration chamber tests, laboratory data, were used to
characterization methodology that integrates multiple successfully train and test the neural network model. By
types of data using a modified counter propagation used as inputs the relative density and the mean effective
artificial neural network (ANN) to provide parameter stress of soils and the CPT cone resistance as a single
estimates and define groundwater contamination at a output. The model gave 0.97 and 0.91 for the training and
leaking landfill. The results of this research illustrate the testing data, respectively, which it high value of
feasibility of combining principal component analysis correlation coefficients, meaning that the neural network
(used to reduce data dimensionality) with the counter was successful in modelling the non-linear relationship
propagation ANN and traditional geostatistical methods between the CPT cone resistance and the other parameters.
(kriging) to estimate subsurface contamination. This Ellis et al. (1996), used (ANNs) for modeling the
research demonstrates the potential for applying this ANN stress-strain relationship of sands with varying grain size
estimation technique (as an alternative to kriging) to distribution and stress history. A series of undrained
delineate the leachate contaminated groundwater and triaxial compression tests for eight different sands was
evaluate water quality associated with subsurface performed under controlled conditions to develop the
contamination at a full scale site. These results suggest the database and was used for neural network training and

2765
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

testing. An attempt has been made to implement artificial based prediction models could satisfactorily be used in
neural networks (ANNs) for modeling the stress-strain obtaining the swelling pressures of the expansive soils.
relationship of sands with varying grain size distribution After that, Ikizler et al (2012) provided, a new estimation
and stress history. A series of undrained triaxial model to predict the pressures is developed using
compression tests for eight different sands was performed experimental data. The data were collected in the
under controlled conditions to develop the database and laboratory using a newly developed device and
was used for neural network training and testing. The experimental setup also. In the experimental setup, a rigid
investigation confirmed that a sequential ANN with steel box was designed to measure transmitted swelling
feedback is more effective than a conventional ANN pressures in lateral and vertical directions. In the
without feedback, to simulate the soil stress-strain estimation model, approaches of artificial neural network
relationship. The study shows that there is potential to (ANN) and adaptive neuro-fuzzy inference systems
develop a general ANN model that accounts for particle (ANFIS) are employed. In the first stage of the study, the
size distribution and stress history effects. Also lateral and vertical swelling pressures were measured with
demonstrated the ability of neural networks to simulate different thicknesses of expanded polystyrene geofoam
unload-reload loops of the soil stress-strain characteristics. placed between one of the vertical walls of the steel box
This study showed that artificial-neural-network-based soil and the expansive soil in the laboratory. Then, ANN and
models can be developed by proper training and learning ANFIS approaches were trained using these results of the
algorithms based on a comprehensive data set, and that tests measured in the laboratory as input for the prediction
useful inferences can be made from such models. of transmitted lateral and vertical swelling pressures.
Cal (1995) based on three factors (plastic index, Results obtained showed that ANN based prediction and
liquid limit and clay content) introduced a model to a ANFIS approaches could satisfactorily be used to estimate
quantitative soil classification by using a neural network. the transmitted lateral and vertical swelling pressures of
Romero and Pamukcu (1996) used (ANN) expansive soils.
successfully to characterize and estimate the shear Tizpa et al (2014). presented artificial neural
modulus of granular materials. network prediction models which relate compaction
Najjar et al. (1996) a huge data base representing characteristics, permeability, and soil shear strength to soil
413 soils retrieved from 45 different projects was used to index properties. A database including a total number of
develop prediction models. Neural network-based models 580 data sets was compiled. The database contains the
and a statistical model were developed. Showed that results of grain size distribution, Atterberg limits,
neural models provide significant improvements in compaction, permeability measured at different levels of
prediction accuracy over statistical models. compaction degree (90-100 %) and consolidated–drained
Penumadu and Jean-Lou (1997) provided a triaxial compression tests. Comparison between the results
model to simulate the behavior of sand and clay soils by of the developed models and experimental data indicated
using (ANN). Ghaboussi and Sidarta (1997) introduced that predictions are within a confidence interval of 95 %.
nested adaptive neural networks; a new type of neural According to the performed sensitivity analysis, Atterbeg
network developed by Ghaboussi, and applied this neural limits and the soil fine content (silt+clay) are the most
network in modeling of the constitutive behavior of important variables in predicting the maximum dry density
geomaterials. Nested adaptive neural networks take and optimum moisture content. Another aspect that is
advantage of the nested structure of the material test data, coherent from the sensitivity analysis is the considerable
and reflect it in the architecture of the neural network. This importance of the compaction degree in the prediction of
new neural network is applied in modeling of the drained the permeability coefficient. However, it can be seen that
and undrained behavior of sand in triaxial tests.Also, Zhu effective friction angle of shearing is highly dependent on
et al. (1998a; 1998b) provide a modelling the shearing the bulk density of soil
behavior of a fine-grained residual soil, dune sand and
Hawaiian volcanic soil used ANNs. Then Penumadu and 2.3 Pile capacity
Zhao (1999) used ANNs to model volume change and the Evaluation pile bearing capacity usually
stress-strain behavior of sand and gravel under drained performed in situ by static load test (SLT). But, procedure
triaxial compression test conditions. of SLT is always costly and time-consuming. Recently
Ikizler et al (2009) developed prediction model of high strain dynamic pile testing (HSDPT) which is
transmitted lateral swelling pressure, and vertical swelling provided by pile driving analyzer (PDA) is used for
pressures on a retaining structure using artificial neural prediction pile axial bearing capacity. The prediction pile
network (ANN) approach. In the first stage of this study, capacity is always a problematic case, many approximate
the lateral and vertical swelling pressures were measured method has been used, but always there is need to a
with different thicknesses of expanded polystyrene (EPS) method has the ability in finding complex nonlinear
geofoam placed between one of the vertical walls of the relationships between different parameters. Therefore,
steel testing box and the expansive soil. Then, artificial many researchers used neural networks in pile capacity.
neural network was trained using these pressures for Goh (1994a; 1995b), neural network was applied
prediction transmitted lateral swelling pressure, and to predict the friction capacity of piles in clays. Goh used
vertical swelling pressures on a retaining structure. Results empirical data of actual field case records to training the
obtained from this study showed that neural network- neural network and to evaluate friction capacity fs on a

2766
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

side surface of timber and steel pipe driven piles. After were compared with the results obtained by the
piles were driven in soft and very soft clays. The data were conventional prediction methods of (Semple and Rigden,
collected. The model input data were: embedded pile 1986) and the method (Burland 1973). The results for
length (4.7-96.0) m; pile diameter (13.5-76.7) cm; vertical the neural network models with three hidden neurons are
component of mean effective stress 'v (19-718 KPa); and summarized in Table1. NN model had the highest
undrained shear strength su (9-335 kPa). The output was coefficient of correlation and lowest error rate for the
skin friction capacity from compression tests of piles fs training and testing data.
(8.0-192.1) KPa. Prediction results of friction capacity fs

Table-1. Summary of correlation coefficients and error rate for friction pile capacity (Goh 1995b).

Coefficient of correlation Error rate (kPa)

Method
Training Testing Training Testing
NN 0.985 0.956 1.016 1.194
Semple and
0.976 0.885 1.318 1.894
Rigden (1986)
method 0.731 0.704 4.824 3.096

The neural network predictions from the training

and testing sets for NNs model are shown in Figure-1. The
results indicate that the neural network was successful in
modelling the non-linear relationship between fs, and the
other parameters. Comparisons were also made between
the measured fs, values and the values predicted using the
method (Burland, 1973) and the method of (Semple &
Rigden, 1986), as plotted in Figure-2 and Figure-3 the
neural network resulting Figure-1 show less scatter in the
data points.
The back propagation approach was successfully
applied to evaluate the friction capacity of driven piles in
clay from actual field records. Also the study pointed out
that the main shortcoming of the neural network
methodology is its inability at present to trace and explain
the step-by-step logic it uses to arrive at the outputs from Figure-2. Skin friction fs - comparison of values
the inputs provided. calculated according to Semple and Ridgen’s
method with standards from measurements.

Figure-1. Skin friction fs - comparison of values predicted Figure-3. Skin friction fs - comparison of values
by NNs with standards from measurements. calculated according to method with standards
from measurements.

Then (Goh, 1995a; 1996b), estimated the ultimate

load capacity of driven piles in cohesionless soils by using
an ANNs and data were drawn from actual case records

2767
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

compiled by(Flaate, 1964) for timber, precast concrete and calibration chamber and the number of blows. The model
steel piles in cohesionless soils. The inputs to the ANN output was the ultimate bearing capacity. In addition, the
model were the hammer weight, the hammer drop, results of in situ pile load tests obtained from a literature
hammer type, the pile length, the pile weight, the pile survey were also used. Five input variables were used
modulus of elasticity the pile cross sectional area, and the representing the penetration depth ratio, the average
pile set. The pile load capacity Qu was the output neuron. standard penetration number along the pile shaft, the
High coefficients of correlations for the training and average standard penetration number near the pile tip, pile
testing data were obtained; Goh summarized it in a Table- set and hammer energy. The results showed that the
2. The results from the testing phase suggest that although maximum error of prediction did not exceed 25%, except
the model was not explicitly trained for these data, the for some bias data. Then based on the average standard
neural network was capable of generalization and penetration, the results were compared with Meyerhof’s
generally gave reasonable predictions. The results indicate equation (Meyerhof 1976). The results showed that the
that the neural network was successful in modeling the neural networks predicted values corresponding the
nonlinear relationship between QU and the other measured values much better than those obtained from
parameters. Meyerhof’s equation. These limited results indicated the
Depending on the connection weights, the feasibility of utilizing neural networks for pile capacity
hammer weight, the pile set, and the hammer type, are that prediction problems.
the more important input factors. As showed in Table-2 Teh et al. (1997), proposed a back-propagation
the study used coefficients of correlation of predicted neural network model for estimating static pile capacity
versus measured results for compared NNs results with from dynamic stress wave data. The training and testing of
that obtained by the following common relationships: The the network were based on a database of 37 precast
Engineering News formula (Wellington 1892), the Hiley reinforced concrete (RC) piles from 21 different sites. The
formula (Hiley 1922) and the Janbu formula (Janbu 1953). CAPWAP (Rausche et al. 1972) -predicted soil parameters
They indicate that the neural network predictions are more were used as the desired output in training. Three different
reliable than the conventional pile driving formulae. Also network models were used to study the ability of the
the study, indicated important point, during training, neural network to predict the desired output to increasing
unrelated input variables are assigned low connection degree of detail. The study showed that the neural network
weights. These variables can then be omitted from the model predicted the total capacity reasonably well. The
model. neural-network-predicted soil resistance along the pile was
also in general agreement with the CAPWAP solution.
Table-2. Summary of regression analysis results of pile The capability of the network to generalize from limited
capacity prediction (Goh 1995a). training examples was verified by its performance against
dynamic test data obtained from non-RC piles.
Coefficient of correlation Abu-Kiefa (1998), the present study introduces
Method
Training data Testing data three general regression neural network models. The first
model, GRNNM1, was developed to estimate the total pile
Neural network 0.96 0.97 capacity. The second model, GRNNM2, was introduced to
Engineering News estimate the tip pile capacity. The last model, GRNNM3,
0.69 0.61
(EN) was utilized to estimate the shaft pile capacity. Data from
Hiley 0.48 0.76 59 good-quality pile load tests in granular soils were
utilized to construct the networks. Pile capacity
Janbu 0.82 0.89
predictions were made using GRNNMI as well as four
other empirical techniques, and they were also compared
Chan et al. (1995), introduced a neural network with actual measurements. These other methods are those
model were the inputs including the pile driving energy, proposed by (Meyerhof, 1976): (Coyle and Castello,
the elastic compression of the soil and pile, and the pile 1981), the American Petroleum Institute (RP2A 1984),
set, which it the same parameters recorded in the and It may be concluded that the GRNNMI is applicable
simplified Hiley formula (Broms and Lim 1988), as an for all different conditions of driven piles in cohesionless
alternative to pile driving formulae. The pile capacity was soils. (Randolph, 1985). Figure-4 shows the relationship
the model output. The study estimated the value of pile between measured and predicted values for each of the
capacity that was used in the training process by using a five methods. This figure indicates that the new method
commercial computer code called CAPWAP (Rausche et provides the best prediction for total pile capacity. It is
al., 1972) or the CASE method (Goble et al., 1975). clear that GRNNMI is also applicable for all conditions of
Lee and Lee (1996), utilized error back driven piles in granular soils, in contrast to the other
propagation neural networks to predict the ultimate procedures, which are limited to certain conditions and
bearing capacity of piles. For the verification of usually produce inconsistent results.
applicability of neural networks, results of model pile load
tests performed by the authors were simulated, were the
model inputs, the penetration depth ratio (i.e. Penetration
depth of pile/pile diameter), the mean normal stress of the

2768
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Figure-4. Comparison of predicted and measured total

Figure-6. Comparison of predicted and measured shaft
pile capacity (Abu-Kiefa 1998).
pile capacity (Abu-Kiefa 1998).
A comparison between the measured tip pile
Goh et al (2005) used a Bayesian neural network
capacities and the predicted values from GRNNM2 for
algorithm to model the relationship between the soil
both the training and the testing sets was showed very high
undrained shear strength, the effective overburden stress,
coefficients of correlation. The tip pile capacity
and the undrained side resistance alpha factor for drilled
predictions using both the previous empirical procedures
shafts. The study integrated the Bayesian framework into
and GRNNM2 were compared with actual measurements.
the backpropagation algorithm to enhances the neural
Figure-5 showed the relationship between measured and
network prediction capabilities, also provides assessment
predicted values for each of the five methods. Clearly, the
of confidence (error bars) associated with the network
GRNNM2 prediction shown in Figure-5 provided the best
predictions. Both the load test results and the parametric
prediction of tip pile capacity. The proposed method is
studies using the trained neural network suggest that the
superior to the empirical ones when compared with actual
effective overburden stress sv8m directly or indirectly has
measurements.
an influence on the a CIUC value for drilled shafts. The
developed neural network model provided good estimates
of the undrained side resistance adhesion factor.
Furthermore, one distinct benefit of this neural network
model is the computation of the error bars on the
predictions of the adhesion factor. These error bars will
aid in giving confidence to the predicted values and the
interpretation of the results.
Nawari et al. (1999), introduced Neural network
paradigms for the design of piles subjected to axial and
lateral loads. N-SPT value and pile dimension were the
inputs of the network for training and testing phases.
During the training phase, the measured axial pile
capacities were compared with the capacities obtained by
BPNN. The artificial neural network based design
Figure-5. Comparison of predicted and measured tip pile approach consists of feed forward back propagation
capacity (Abu-Kiefa 1998). Neural Network, and Generalized Regression Neural
Network. According to the simulation results, the neural
A comparison between the measured shaft pile network approach is feasible and was found to be more
capacities and the predicted ones from GRNNM3 for both accurate than the commonly used techniques for the
the training and the testing sets was showed very high design of pile foundations.
coefficients of correlation were obtained. The shaft pile The GRNN model has the advantage that it is
capacity predictions using the previous empirical unnecessary to define the number of hidden layers or the
procedures as well as GRNNM3 were also compared with number of neurons per layer in advance. Moreover, the
actual measurements. The relationship between measured GRNN provides an adequate approximation of the full-
and predicted values for each of the five methods is shown scale pile test results. Based on the results from this
in Figure-6. From this figure, it is most evident that investigation, it appeared that the proposed neural network
GRNNM3 also provides the best prediction for shaft pile models furnish a pragmatic and a reliable alternative for
capacity. the current analysis and design techniques of axial pile
capacity and laterally loaded piles.

2769
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Park and Cho (2010). (165) data from dynamic artificial neural network model with existing theories. The
piles load test at various sites were selected to develop an results of the model have the highest performance among
ANN model to predict the resistance of the driven pile in the other methods, even though the difference is not too
dynamic load test. The results showed that the ANN big.
model served as a reliable and simple predictive tool to Lee et al (2014), adopted artificial neural
predict the resistance of the driven pile with correlation networks (ANN), to generate the training samples for
coefficient values close to 0.9. training ANN, finite element (FE) analysis was performed
Shahin (2010), utilized artificial neural networks 50 times for 50 different design cases. The trained ANN
(ANNs), to developed two ANN models (one for driven was verified with another FE analysis case and then used
piles and the other for drilled shafts) were developed with as a structural analyzer. The multilayer neural network
the aid of the software package NEUFRAME Version 4.0 (MBPNN) with two hidden layers was used for ANN. The
(Neusciences 2000). The data used to calibrate and framework of MBPNN was defined as the input with the
validate the ANN models are obtained from the literature lateral forces on the jetty structure and the type of piles
and include a series of 80 in situ driven pile load tests and the output with the stress ratio of the piles. Through
reported by (Eslami, 1996) and 94 in situ drilled shaft load the trained MBPNN, the stress ratios of jetty piles were
tests reported by (Alsamman, 1995), as well as CPT obtained under different loading conditions which were
results. Then the author compared predictions from the not included in the training samples. The feasibility of the
ANN models with those obtained from the most MBPNN was verified by comparing the results from FE
commonly used available CPT-based methods: model and the MBPNN. The results from the MBPNN are
very close to the FE analysis results. The RMSE is also
 Eslami and Fellenius (1997) very small regardless of the pile patterns. For the highly
complex jetty pile patterns the results from the MBPNN
 LCPC (Bustamante and Gianeselli 1982) show very good agreement with those from FE analysis.
With the more training samples and the expansion of input
 European method (de Ruiter and Beringen 1979). parameters for jetty structure design, the MBPNN showed
possibility to replace the repetitive and time-consuming
and statistical analyses were carried out to rank and FE analysis. Also the study referred to that only 50 cases
evaluate the performance of the ANN models and CPT had been modeled for this study, the merit of MBPNN
methods. would be clearer as the number of cases increases.
Usually the produced equations of the developed Maizir et al. (2015), development ANN model
ANN models, difficult to use, so the author, translated it for prediction of axial capacity of a driven pile based on
into simple design equations suitable for hand Pile Driving Analyzer (PDA) test data. As many as 300
calculations. The results indicate that the ANN models sets of high quality test data from dynamic load test
were capable of accurately predicting the ultimate capacity performed. Input considered in the modeling are pile
of pile foundations with high coefficients of correlation, r. characteristics (diameter, length as well as compression
For driven piles, the ANN model had a coefficient of and tension capacity), pile set, and hammer characteristics
correlation (r) of 0.96 and 0.85 in the calibration and (ram weight, drop height, and energy transferred). An
validation sets, respectively, whereas the ANN drilled ANN model (named: ANN-HM) was developed using a
shafts model had an r of 0.97 in both the calibration and computerized intelligent system for predicting the total
validation sets. The sensitivity analyses carried out on both pile capacity as well as shaft resistance and end bearing
the driven piles and drilled shafts ANN models indicated capacity for various pile and hammer characteristics. The
that predictions from the ANN models compare well with results show that the ANN-HM serves as a reliable
what one would expect based on available geotechnical prediction tool to predict the resistance of the driven pile
knowledge and experimental results. with coefficient of correlation (R) values close to 0.9 and
Wardani et al. (2013), applied NN model for mean squared error (MSE) less than 1% after 15,000
prediction of ultimate bearing capacity of single pile number of iteration process.
foundation, was named NN_Qult model. The results of Mazaher and Berneti, (2016), applied MLP neural
analysis model were then compared with (Meyerhof, network for prediction of load carrying capacity of open
1976) and (Briaud ,1985) formulas. At the stage of ended metal piles pounded in sandy soils. Four parameters
modeling, data from full-scale pile load test and SPT were of pile length, its diameter, soil elastic modulus and soil
used. The selected input variables are: d (pile diameter), L internal friction angle are used as input and pile load
(length of the pile embedded), the N60 (shaft) value, and carrying capacity is applied as output.
the N60 (tip) value. The study generated design Charts Obtained results revealed that neural network
that are expected to predict the ultimate bearing capacity with minimum error, high speed and learning capacity, has
of a single pile foundation. The design chart is used as a very high efficiency in predicting load carrying capacity of
tool to calculate the ultimate bearing capacity of a single metal piles. Also based on performed sensitivity analysis
pile in sand soil. The results showed that neural networks on the best obtained network, all parameters have an
can be used for prediction of ultimate bearing capacity of increasing effect on load carrying capacity and according
single pile foundation. This is particularly due to the to sensitivity values and their distances from base line, it is
sensitivity analysis results indicated the suitability of concluded that soil internal friction angle, soil elastic

2770
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

modulus, pile diameter and pile length respectively have Schmertmann’s method overestimate the settlements by
maximum effect on load carrying capacity of piles. about 2.18 times and 3.39 times respectively; the neural
network model predictions were very good.
2.4 Settlement of foundations Arnold (1999) used the same features of NN
Two criteria that govern foundation design, model prepared by (Sivakugan et al., 1998) with a larger
bearing capacity and settlement; often settlement number of data cases. His neural network model contained
governing. It is very complex to estimating the settlement 18 hidden layer nodes with correlation coefficients equal
of foundations, uncertain and not yet entirely understood. to 0.954, 0.955 and 0.944 for the training, testing and
So, many researchers encouraged by the previews succeed validation sets, respectively. (Shahin et al., 2004a)
of ANN techniques in complex problems to apply it in mentioned that 18 hidden layer nodes are considered to be
settlement prediction. large for a network with 5 input variables, which may
Goh (1994a), proposed a ANN model for the affect the generalization ability of the model. Also
prediction of settlement of a vertically loaded pile mentioned that the ANN models developed above for
foundation in a homogeneous soil stratum. The inputs settlement prediction of shallow foundations on
were the ratio of the elastic modulus of the pile to the cohesionless soils have been built on either a limited
shear modulus of the soil, pile length, pile load, shear number of data cases (e.g. Sivakugan et al. 1998) or have
modulus of the soil, Poisson’s ratio of the soil and radius suffered from the lack of a comprehensive procedure for
of the pile. The pile settlement was the output. The target testing their robustness and generalization ability (e.g.
output used for training the ANN model was obtained by Arnold 1999).
means of finite element and integral equation analyses by Shahin et al. (2000) attempted to carried out
(Randolph and Wroth, 1978). The results indicated the similar work for predicting the settlement of shallow
successfully of the ANN model to predict the settlement of foundations on cohesionless soils. Using a large database
pile foundations. of actual measured settlements. In his work, 272 data
Sivakugan et al. (1998) attempted use neural records were used for modelling. The input variables
network for predicting settlements of shallow foundations considered to have the weightiest influence on settlement
on granular soils. Five inputs were used to train the neural prediction were, the footing width, the footing length, the
network: average blow count from the standard applied pressure of the footing and the soil
penetration test, net applied pressure, width of foundation, compressibility. Then results of the ANN model were
shape of foundation and depth of foundation. The output compared with three of the most commonly used
was the settlement of the foundation. Settlement of the traditional methods. (Meyerhof, 1965), (Schultze and
foundation was the output. 79 settlement records with Sherif, 1973) and (Schmertmann et al., 1978). The results
necessary foundation and soil data were collected from the of (Sivakugan et al., 1998), were confirmed by this study
literature. Out of the 79 records, 69 were used to train the results, meaning that ANNs showed good ability to predict
neural network and the remaining ten were used to test the the settlement more than traditional methods. Table (3),
network. The predictions are compared with those from showed that the ANNs model produced higher coefficients
two traditional methods, (Terzaghi and Peck, 1967) and of correlation, r, low root mean squared errors, RMSE, and
(Schmertmann, 1970). The results showed that the low mean absolute errors, MAE, compared with the other
traditional method of Terzaghi and Peck and methods.

Table-3. Performance of ANN and traditional methods for the validation set (Shahin et al. 2000).
Meyerhof Schultze & Schmertmann et
Category ANN
(1965) Sherif (1973) al. (1978)
Correlation, r 0.99 0.33 0.86 0.70
RMSE (mm) 3.9 27.0 23.8 45.2
MAE (mm) 2.6 20.8 11.1 29.5

Similar to that, Shahin et al. (2002a; b; 2003a; c) measure of soil compressibility, footing geometry (length
applied a wide-ranging study to predict the settlement of to width of footing) and footing embedment ratio
shallow foundations on cohesionless soils using ANNs. (embedment depth to footing width). The foundation
Depending on a large database of actual measured settlement was the model output. As in Shahin et al.
settlements and multi-layer perceptrons (MLPs) trained (2000) the results between the predicted and measured
with the back-propagation algorithm, Shahin et al. (2002b) settlements obtained by utilizing ANN were compared
introduced an ANN model, the five inputs were the with same three traditional methods, and Table-4 show
footing width, net applied footing load, average blow that the ANN method better than the traditional methods
count obtained from the standard penetration test (SPT) for all performance measures considered.
over the depth of influence of the foundations as a

2771
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Table-4. Performance of ANN and traditional methods for the validation set (Shahin et al. 2002b).
Meyerhof Schultze & Schmertmann et al.
Category ANN
(1965) Sherif (1973) (1978)
Correlation, r 0.905 0.440 0.729 0.798
RMSE (mm) 11.04 25.72 23.55 23.67
MAE (mm) 8.78 16.59 11.81 15.69

Furthermore, Shahin et al. (2002a) attempted to foundations for new cases not used in the development of
simplify the obtained ANN model and translated it into a the model.
tractable and relatively simple formula suitable for hand Nejad and Jaksa (2010) developed an A back-
calculation. propagation neural network model to predicting pile
Based on the same database, inputs and outputs, settlement based on the results of cone penetration test
Shahin et al. (2003c) applied a B-spline neurofuzzy (CPT) data. A database containing 292 case records of
models for settlement prediction of shallow foundations on actual field measurements for settlement of pile was used
cohesionless soils. The model results then compared with for model development and verification. In addition, this
the MLP model developed by Shahin et al. (2002b) in study discusses the choice of input and internal network
terms of prediction accuracy, model parsimony and model parameters which were examined to obtain the optimum
transparency. In terms of prediction accuracy, it was found model. Finally, the predictions obtained by the ANN
that the two models were comparable, although the MLP compares with those given by a number of traditional
model performs slightly better than the neurofuzzy model. methods. The results indicate that back-propagation neural
In terms of model parsimony, it was found that the networks have the ability to predict the settlement of pile
neurofuzzy model was more parsimonious than the MLP with an acceptable degree of accuracy (r=0.956,
model with fewer model inputs and connection weights. In RMSE=1.06 mm) for predicted settlements ranging from
terms of model transparency, it was found that the B- 0.0 to 137.88 mm. The ANNs method has another
spline model was more transparent than the MLP model, advantage over the conventional methods in that once the
as it was able to describe the relationship between model model is trained, the model can be used as an accurate and
inputs and the corresponding output in the form of a set of quick tool for estimating the settlement of piles. In
fuzzy rules. contrast with the conventional methods, the ANNs method
Furthermore, Shahin et al (2005) used The ANN does not need any manual work such as using tables or
model that was developed by (Shahin et al., 2002a) to charts to calculate the settlement.
obtain deterministic settlement predictions of shallow Results of this study indicated that ANNs have a
foundations on granular soils. In this study, the likely number of significant benefits that make them a powerful
distribution of predicted settlements, given the and practical tool for settlement prediction of piles.
uncertainties associated with settlement prediction, is Alkroosh and Nikraz (2011) proposed the
obtained by combining Monte Carlo simulation with a artificial neural network as an alternative to predict the
deterministic ANN model. A set of stochastic design load settlement behavior of pile foundations embedded in
charts, which incorporate the uncertainty associated with sand or mixed soils and subjected to axial loads. Three
the ANN method, is developed. The charts are considered ANN models have been developed, a model for bored
to be useful in the sense that they enable the designer to piles and two other models for driven piles (a model for
make informed decisions regarding the level of risk each of concrete and steel piles). The data used for
associated with predicted settlements and consequently development of the ANN models is collected from the
provide a more realistic indication of what the actual literature and include a series of in-situ piles load tests as
settlement might be. well as cone penetration test (CPT) results. The data of
Rezania and Javadi (2007) presented a new each model is divided into two subsets: Training set for
genetic programming (GP) approach for predicting model calibration and independent validation set for model
settlement of shallow foundations. The GP model was verification. Predictions from the ANN models are
developed and verified using a large database of standard compared with the results of experimental data and with
penetration test (SPT) based case histories that involve predictions of number of currently adopted load-transfer
measured settlements of shallow foundations. The results methods. Statistical analysis is used to verify the
of the developed GP model were compared with those of a performance of the models. The results have revealed that
number of commonly used traditional methods and the developed ANN models have a high capability to
artificial neural network (ANN) based models. It was simulate the nonlinear behavior of soil under loads. The
showed that the GP model is able to learn, with a very models also possess the ability to simulate the strain
high accuracy, the complex relationship between hardening behavior. The graphical and numerical
foundation settlement and its contributing factors, and comparisons between the performance of the developed
render this knowledge in the form of a function. Also the models in training and validation sets indicates that the
authors suggested that the attained function can be used to ANN models are able to forecast the load-settlement
generalize the learning and apply it to predict settlement of behavior accurately. The graphical illustrations of the

2772
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

relationship between load and settlement may indicate that 2.5 Load-settlement response modeling
the models may tend to under-predict the relationship. The We can saw from above, many researchers deal
graphical and numerical comparison between ANN with soil capacity and soil settlement separately, but
models and the other methods demonstrates that the ANN actually soil resistance and settlement are influenced by
models simulate the load-settlement behavior more each other, and the design of foundations should thus
accurately than the load-transfer methods. It can be consider the bearing capacity and settlement inseparably.
concluded that the ANN models are reliable and can be (Shahin, 2014) used a series of full-scale in-situ pile load
applied as an alternative to forecast the load-settlement settlement tests and CPT data collected from the literature
behavior for design practice. to develop a recurrent neural network (RNN) model for
Shahnazari et al (2014) utilized three new simulating the load-settlement response of drilled shafts
evolutionary-based techniques, including evolutionary (or bored piles). The graphical comparison of the load-
polynomial regression (EPR), classical genetic settlement curves between the RNN model and
programming (GP), and gene expression programming experiments showed an excellent agreement and indicates
(GEP) to obtained more accurate predictive settlement that the RNN model can capture the highly non-linear
models. The models were developed using a large load-settlement response of drilled shafts reasonably well.
databank of standard penetration test (SPT)-based case
histories. The values obtained from the new models were 2.6 Liquefaction
compared with other available soft computing models, Earthquakes led to seismic vibrations that cause
including the ANN model developed by (Shahin et al., soil loss of compressive strength as a result of the soil lost
2002a; b) and the GP model developed by (Rezania and its shear strength due to the increase of hydraulic pressure
Javadi, 2007). It should be noted that the comparison within pores. This ground failure phenomenon name
between the soft computing settlement prediction models “Liquefaction”. During earthquakes liquefaction can cause
and traditional methods such as those of Meyerhof (1965), ground failure and very dangerous damage to civil
Schultze and Sherif (1973) and Schmertmann et al. engineering structures. Although, the well-recognized
(1978). were investigated (as mentioned above) in liquefaction mechanism, the prediction of liquefaction
previous studies by Shahin et al. (2002a) and Rezania and potential is very difficult because there are many critical
Javadi (2007). The results showed that the new EPR and factors influencing liquefaction.
GP-based models are able to predict the foundation Many researchers were applied Artificial Neural
settlement on cohesionless soils under the described Networks to evaluate the probability of soil liquefaction.
conditions with coefficient of determination (r2), values Goh (1994b) investigated the feasibility of using neural
higher than 87%. The artificial neural networks (ANNs) networks to model the complex relationship between the
and genetic programming (GP)-based models obtained seismic and soil parameters, and the liquefaction potential.
from the literature, had r2 values of about 85% and 83%, In his study used simple back‐propagation neural‐network
respectively which are higher than 80% for the GEP-based algorithm. The neural networks were trained using actual
model. A subsequent comprehensive parametric study is field records from 13 earthquakes that occurred in Japan,
further carried out to evaluate the sensitivity of the United States and Pan-America during the period 1891-
foundation settlement to the effective input parameters. 1980. The model used eight input variables These
The comparison results proved that the new EPR and GP- variables are: the standard penetration test (SPT) value, the
based models are the most accurate models. In this study, fines content, the mean grain size D50 and the maximum
the feasibility of the EPR, GP and GEP approaches in horizontal acceleration at ground surface with only one
finding solutions for highly nonlinear problems such as output variable. The output values were 1, for sites with
settlement of shallow foundations on granular soils is also extensive or moderate liquefaction, and for no liquefaction
clearly illustrated. The developed models were quite or marginal a value was 0. Then the results of NN model
simple and straightforward and can be used reliably for were compared with the method of (Seed et al., 1985). The
routine design practice. comparison showed that the NN model was success in
It should be noted that, recently many researchers 95% of cases, while (Seed et al., 1985) gave a correct
(as in Shahnazari et al., 2014) used a coefficient of prediction in 84%.
determination (r2) measure for model performance, rather Goh (1996a) used cone penetration test (CPT)
than, r, so as it, (as indicated by (Das and Sivakugan, resistance data to build neural networks model to measure
2011+), r, sometimes may not necessarily indicate better liquefaction potential. This study data was taken from
model performance due to the tendency of the model to records of sand and silty sand deposits, representing five
deviate toward higher or lower values, particularly when earthquakes that happened during the period 1964–1983 in
the data range is very wide and most of the data are Japan, China, United States and Romania. As a
distributed about their mean. Consequently, the coefficient comparison the neural network gave a 94% success rate,
of determination, r2, is used as it can give unbiased which is the same number of error predictions as the
estimate and may be a better measure for model conventional method by (Shibata and Teparaksa, 1988).
performance. Najjar and Ali (1998) developed neural networks
model by using field data sets of various earthquake sites
from around the world to characterized the soil
liquefaction resistance. The study produced a liquefaction

2773
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

potential valuation chart that could be useful tool for SPT-based case histories for free face and gently sloping
geotechnical engineers in prediction of liquefaction. ground conditions. Two other specific models were
More recent Baziar and Ghorbani (2005) presented for more accurate prediction of moderate lateral
developed a back-propagation neural network model using displacements of up to 1.5 m. The attained function can
professional software called STATISTICA, to predict the then be used to generalize the learning to predict
horizontal ground displacement in both ground slope and liquefaction induced lateral spreading for new cases not
free face conditions due to liquefaction-induced lateral used in the construction of the model. The results of the
spreading. A database including 464 case histories from 10 developed GP models are compared with those of a
sites located in Japan and USA (Youd et al, 2002)) was commonly used multi linear regression (MLR) model and
used for development of the model. A feed-forward the advantages of the proposed GP model over the
network with the Levenberg- Marquardt algorithm was conventional method are highlighted.
used in the training stage. A sensitivity analysis was also Hanna et al (2007) presented General Regression
carried out to study the relative importance of the factors, Neural Network (GRNN) model that addresses the
affecting lateral spreading. The sensitivity analysis collective knowledge built in simplified procedure. The
indicates that the two factors of source distance (R) and model utilizes 12 soil and seismic parameters that
mean grain size (D5015) have the most significant effect characterize soil types, material properties, seismic
on the predicted displacements while the moment characteristics, magnitude and nature of loads, and stresses
magnitude of earthquake (M) and ground slope (S) have and strains, strengths, saturation and seismological aspects
respectively moderate and small impact on displacements. of the soil. These parameters are real world parameters
The results obtained in this study indicate that the ANN that can easily be obtained using widely accepted testing
model has ability to predict the lateral spreading with an techniques and empirical formulas. The sensitivity of each
acceptable degree of accuracy (r20.92, RMSE 0.7 m) for of these parameters to liquefaction potential in soil was
displacements ranging from 0.01 to 10.16 m. This examined. In this investigation, 3,895 case records mostly
accuracy shows the superiority of the ANN model over from the cone penetration test (CPT) results produced
multilinear regression and suggests that the model can be from the two major earthquakes that took place in Turkey
applied in engineering practice. and Taiwan in 1999. Data used were randomly divided for
Young and Byung (2006) developed a the development, testing, and validation. Soil liquefaction
backpropagation artificial neural network model to predict decision in terms of seismic demand and seismic capacity
the liquefaction cyclic resistance ratio (CRR) of sands is determined by the stress-based method and strain-based
using data from several laboratory studies involving method, and further tested with the well-known Chinese
undrained cyclic triaxial and cyclic simple shear testing. criteria. The results produced by the proposed GRNN
The model development was based on data obtained from model explore effectively the complex relationship
published experimental research work available. 346 data between the soil and seismic input parameters and further
sets relating CRR with the number of cycles (N) for forecast the liquefaction potential with an overall success
liquefaction triggering. The data are for several types of ratio of 94 percent. Liquefaction decisions were further
sands tested under different initial conditions (269 and 77 validated by the SPT, confirming the viability of the SPT-
data sets from cyclic triaxial and cyclic direct simple shear to-CPT data conversion, which is the main limitation of
tests, respectively). The model was verified using data that most of the simplified methods.
was not used for training as well as a set of independent Kayadelen (2011) developed four ANN models to
data available from laboratory cyclic shear tests on another presenting the potential of Genetic Expression
soil. The observed agreement between the predictions and Programming (GEP) and Adaptive Neuro-Fuzzy (ANFIS)
the measured CRR values indicate that the model using the computing paradigm to forecast the safety factor for
basic physical properties (i.e., Uc, D50, Dr, emax, and emin) is liquefaction of soils. The study used data of 569 set
capable of effectively capturing the liquefaction resistance collected from the literature. Five parameters were used as
of a number of sands under varying initial conditions. The inputs such as standard penetration test ((N1)60),
predicted CRR values are mostly sensitive to the percentage of finest content (FC), effective overburden
variations in relative density thus confirming the ability of stresses (σ'), cyclic stress ratios (CSR) and angle of
the model to mimic the dominant dependence of shearing resistance (Ø'). Then the performance of models
liquefaction susceptibility on soil density already known was comprehensively valuated using several statistical
from field and experimental observations. Although the verification tools. Equations obtained by GEP Model and
ANN methods lack fundamental linkage with the soil GEP Model, gave the high correlation coefficient and low
response from a mechanistic point of view, the results RMSE values (0.18 and 0.20). On the other hand, the
clearly demonstrated that ANN models have a strong ANFIS models produces satisfactory results with R values
potential and are suitable to serve as a quick ranging from 0.97 to 0.98 and RMSE values ranging from
interpolating/extrapolating tool for the relationship 0.27 to 0.21. The results showed that GEP and ANFIS
between cyclic stress ratio (CSR) and N for liquefaction. models are fairly promising approach for the prediction of
Javadi et al (2006) presented genetic the soil liquefaction potential and capable of representing
programming (GP) approach, for determination of the complex relationship between seismic properties of
liquefaction induced lateral spreading. Two general GP soils and their liquefaction potential.
models were trained and validated using a database of

2774
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Rezania et al (2011) developed models to the proposed model, which was derived from laboratory
represent a new approach, based on evolutionary data, can be successfully utilized under field conditions.
polynomial regression (EPR), for assessment of Ardakani and Kohestani (215) examined
liquefaction potential and lateral spreading. Separate EPR capability on of AI techniques call C4.5 decision tree for
models were developed and tested for evaluation of the prediction of seismic liquefaction potential of soil
liquefaction potential and lateral spreading using two based on the Cone Penetration Test (CPT) data. The C4.5
different datasets of actual field case histories. A strong model was trained and validated using a database of 109
EPR model was developed for assessment of liquefaction liquefaction and non-liquefaction field case histories for
potential using a large CPT database of field data. Then sandy soils based on CPT results. The database contains
the models compared to those obtained from neural the information about cone resistance (qc), total vertical
network and linear regression based techniques. The stress ( 0), effective vertical stress ( 0 ′), mean grain size
proposed EPR model overcome the shortcoming in the (D50), normalized peak horizontal acceleration at ground
neural network-based model which it incapability to surface (amax), cyclic stress ratio ( / 0 ′) and earthquake
present an explicit relation- ship between input and output magnitude (Mw). The overall classification success rate for
parameters, by generating a structured representation of all the data set is 98%. The results of C4.5 decision tree
the system. All EPR models have showed coefficient of have been compared with the available artificial neural
determination (COD) values greater than 90%, which is a network (ANN) and importance vector machine (RVM)
good indication of the robustness of models. Comparison models. Unlike available ANN and RVM models, the
of the results shows that the results predicted by the proposed model provides easily interpretable tree structure
proposed EPR models provide an improvement over the that can be used by geotechnical engineering professionals
ANN and MLR models. It was shown that in both cases with the help of a spreadsheet to predict the liquefaction
the EPR models are able to learn, with a very high potential of soil for future seismic event without going into
accuracy, the complex relationship between liquefaction the complexities of model development using C4.5
and its contributing factors in the form of a function, and decision tree. This model can be adopted for modeling
can also generalize the learning to provide predictions for different problems in geosciences.
new cases not used in the construction of the model. An
additional advantage of the EPR approach is that there is 2.7 Slope stability
no need to assume a priori the form of the relationship Landslides or slope failures are important because
between the input and the output parameters, and the form they can lead to the loss of life and property. Therefore,
of the relationship, the number and combination of the slope stability analysis is one of the most important
terms and the values of the coefficients can all be evolved problems in geotechnical engineering. As most of
during construction of the model (training). One of the geotechnical problems, slope failures are complex. The
strongest points of the proposed approach is that a part of analysis of slope stability had been developed with the
the data is set aside and used for validation of the trained development in computational geotechnical engineering.
EPR model in order to examine the generalization Limited data and unclearly defined problems
capabilities of the developed model in making predictions often complicate the study of landslides It is very
for unseen cases. important to understanding and evaluation the behavior of
Baziar et al (2014) presented a new predictive the processes that govern of the slopes, in order to
model utilizing genetic programming technique to eliminate or reduce landslide damages. The successful
estimate the amount of strain energy density, which is performance of Artificial Neural Networks (ANNs), in
required for the liquefaction triggering of sand–silt modeling non-linear multivariate problems, Courage many
mixtures. Model developed with a relatively large researchers to using it in studying slope instability. (Ni et
database including various types of cyclic element tests al., 1996) proposed a combining model of artificial neural
(triaxial, torsional, and simple shear) was used, gathered networks with fuzzy sets of slopes stability evaluating.
from previously published studies. This study selected The model inputs were: horizontal profile, vertical profile,
input variables of the model from the available gradient, location, depth of weathering, direction of slopes,
understandings advanced from the previous studies on the soil texture, vegetation, land use, geological origin,
strain energy-based liquefaction potential assessment. maximum daily precipitation and maximum hour
According to the result, the amount of strain energy precipitation. The slope failure potential was the output of
required for liquefaction beginning increases with increase the model. The results of the neural network showed better
in initial effective overburden pressure, relative density, predicting when compared with the analytical model
and mean grain size. The effect of non-plastic fines on results.
strain energy-based liquefaction resistance shows a more More recent (Ferentinou and Sakellariou, 2007)
complex behavior. Therefore, liquefaction resistance applied supervised ANNs using back-propagation learning
increases with increase in fines up to about 10–15% and algorithm for the prediction of slope performance under
then starts to decline for a higher increase in fines content. static and seismic loading. Then applied unsupervised
Further verifications of the model were carried out using ANNs using the efficient visualization techniques offered
the valuable results of some downhole array data as well by self-organizing maps in lithological classification of
as centrifuge model tests. These verifications confirm that unsaturated soils and in classification of dry and wet
slopes according status of stability and failure mechanism.

2775
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

The study finally proposed a coupled model of SOM ANN (1994) has been used. The database consists of case
with interaction matrix in order to rate slope stability studies of 23 dry and 23 wet slopes where 29 are failed
controlling variables. The work concluded that the and 17 are stable slopes. The input data consist of
application of computational intelligence tools on the real- parameters like height of slope H (m), unit weight
world data sets using both supervised and unsupervised (kN/m3), cohesion c (kPa), internal friction angle ϕ ( ̊ ),
methods gave reasonable results. slope angle ( ̊ ) and pore pressure parameters ru. Yang et
Cho (2009) presented a practical procedure that al. (2004) have used the above database to develop
combined a commercial numerical analysis code and prediction model using genetic programming (GP). The
artificial neural networks into the probabilistic analysis of output database consists of qualitative information
slope stability. An ANN technique was adopted to classifying the slope as stable or failed and quantitative
establish a model for the approximation of the limit state information that consists of factor of safety of the slope
function. Training and test data sets for the model were based on limit equilibrium method. The ANN models were
obtained from numerical calculations based on the strength trained using Bayesian regularization (BRNN), differential
reduction method. Then the ANN model connected to a evolution algorithm (DENN) and commonly used
reliability method, in this case the first- and second order Levenberg–Marquardt (LMNN), and the results are
reliability method and the Monte Carlo simulation method, compared with those from support vector machine and
to predict the failure probability. The obtained results also genetic programming available in literature. The results of
showed that the ANN-based response surface can be DENN are found to better than most likelihood method
successfully applied to the problem of probabilistic slope and comparable to that of GP. Applying sensitivity
stability. analysis to identify important input parameters, based on
Chauhana et al (2010) proposed a new approach DENN model, C is found to most important parameter
for landslide susceptibility zonation (LSZ) mapping based followed by ϕ̊, , , H and ru. As these are circular failure
on the ratings of categories of causative factors derived surfaces, C should be the most important parameters
from an ANN model. These ratings depicted the specific followed by ϕ̊, and . Hence, it may be concluded that the
individual influence of each category on landslide results of DENN model better represents the physical
incidents. The study was conducted in a landslide prone problem.
area in the Himalayan Region. Seven causative factors, Peng et al (2014) presented a systematic method
namely: slope, slope aspect, relative relief, lithology, for slope safety evaluation utilizing multi-source
structural features (e.g., thrusts and faults), land use/ land monitoring information. First, a Bayesian network with
cover, and drainage density, were placed in 42 categories continuously distributed variables for a slope involving the
for which ratings were determined. The structural buffer factor of safety, multiple monitoring indexes and their
category 0–500 m, with the highest rating of 376.5m, was influencing parameters (e.g. friction angle and cohesion) is
found to be the most influential among the 42 categories. constructed. Then the prior probabilities for the Bayesian
The evaluation of LSZ mapping through landslide density network are quantified considering model and parameter
analysis, best-fit quadratic curves and receiver operating uncertainties. After that, multi-source monitoring
curves (ROC), clearly demonstrated the efficiency of the information is used to update the probability distributions
proposed approach and provided an accurate of the soil or rock model parameters and the factor of
representation of the actual scenario of landslide safety using Markov chain Monte Carlo simulation. The
occurrences in the study region. study concluded that he method is able to integrate multi-
Gemitzi et al (2011) presented a methodology for source information based on slope stability mechanisms,
landslide susceptibility assessment to delineate landslide and update the soil or rock parameters, the slope factor of
prone areas by using factor analysis and fuzzy safety, and the failure probability with the integrated
membership functions and Geographic Information monitoring information. Hence the evaluation becomes
Systems (GIS). Six conditioning factors were evaluated: more reliable with the support of multiple sources of site-
slope angle, slope aspect, land use, geology, distance to specific information.
faults and topographical elevation. Fuzzy membership
functions were defined for each factor using the landslide 2.8 Earth retaining structures
frequency data. Factor analysis provided weights (i.e., The bracing of excavation is one of the important
importance for landslide occurrences) for each one of the matter in geotechnical engineering, could be temporary
above conditioning factors, indicating the most important during the construction stage, or permanent as in
factors as geology and slope angle. An overlay and index underground structures or that adjacent to soil ramp. Goh
method was adopted to produce the landslide susceptibility et al., 1995d) utilizing a neural network model for initial
map. In this map 96% of the observed landslides are estimates of maximum wall deflections for braced
located in very high and high susceptibility zones, excavations in soft clay. The model used as input
indicating a suitable approach for landslide susceptibility parameters: the height of excavation, the excavation walls
mapping. stiffness, excavation width, soil unit weight soil,
Das et al (2011) discussed the application of two thickness/excavation width ratio, soil undrained shear
different types of ANN models as alternate tools to strength and undrained soil modulus/shear strength ratio.
classify the slope as stable (1) or failed (0) and to predict The output of the model was the maximum wall
the factor of safety. The database available in Sah et al. deflection. The regression analysis the scatter of the

2776
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

predicted neural network deflections, used as a relative to test at tunnel crown level, the mean blow count from
the deflections obtained using the finite element method. standard penetration test at tunnel spring-line level and the
Results showed high coefficients of correlation mean blow count from standard penetration test at tunnel
for the training and testing data of 0.984 and 0.967, inverted arch level. The outputs were the settlement at the
respectively. Furthermore, some actual case records data face passage, the settlement at the invert closing and the
were used to confirm the performance of the trained neural final settlement after stabilization. This study results
network model. The agreement of the neural network indicated low level of accuracy was obtained from the
predicted and measured wall deflections was promising. neural network model. Then study proposed a modular
The study proposed to use the neural network model as a neural network model based on the concept of integrating
time-saving and easy to use, as alternative to the finite multiple neural network modules in one system, in order
element method. to improve the prediction accuracy, with each module
More recent (Goh and. Kulhawy, 2005) being constrained to operate at one specific situation of a
demonstrates the use of an integrated neural network– complicated real world problem. The modular concept
reliability method to assess the risk of serviceability showed an improvement in terms of model convergence
failure through the calculation of the reliability index. In and prediction. (Shi, 2000) developed the models to
this study applied a series of parametric studies using the improve the capability by applying input data
finite element method and then approximating the non- transformation. This extended study indicated that
linear limit state surface (the boundary separating the safe distribution transformation of the input variables reduced
and ‘failure’ domains) through a neural network model, the prediction error by more than 13%.
the reliability index can be determined with the aid of a Haji hassani et al (2011) used a Back propagation
spreadsheet. Although, descriptive examples are presented neural network model A Back propagation neural network
to show how the serviceability performance for braced modelling was used to predict surface settlement due to
excavation problems can be assessed using the reliability NATM tunneling techniques in soft ground. For model
index. training and verification, 24 data sets containing field
Gordon et al (2007) presented an artificial neural measurement data ware used. The inputs were including
network methodology for predicting deflection of the standard penetration test (SPT), moisture content,
diaphragm walls caused by braced excavation in soft to cohesion, tunnel depth, frictional Angle, unit Weight,
medium clays. The ANN model input variables were, Poisson Ratio, elasticity modulus. While the model
excavation depth, system stiffness, excavation width, shear outputs were the surface settlement at several points. To
strength normalized with vertical effective stress, and obtain the optimal ANN architecture, several models were
Young’s modulus normalized with vertical effective stress. investigated and finally a model with one hidden layer
The maximum wall deflection in a braced excavation in with 20 nodes was chosen. The results manifest that the
soft to medium clays was the output of model. The Back propagation neural networks are capable of
database for training and testing the ANN is generated predicting the surface settlement due to NATM tunneling
from hypothetical cases using finite element method. The in soft ground with adequate accuracy.
performance of the developed ANN reveals that the Juwaied and Al-ZwainyA, (2017) utilizing
influence of each input variable on the wall deflection is database of 63 historical cases collected from five projects
consistent with the excavation behaviors generally in Iraq/Baghdad to be the reference to produce of pile
observed in the field. The validation using 12 excavation design equation. The resulting equation was used to find
case histories collected in this study shows that the wall the prediction design load of the pile and compare it with
deflection caused by braced excavation can be accurately the actual. The results showed relatively high correlation
predicted by the developed ANN. coefficient r, 96% between the actual and predicted
values., as well as the results showed high correlation
2.9 Tunnels and underground openings coefficient r,97% and coefficient of determination r2, 94%
Shi et al. (1998) applied neural networks to between the actual and predicted values, for the 7 cases
predict settlements of tunnels. The study used data from not used in the model development.
the 6.5 km Brasilia Tunnel, Brazil, to training and testing
the model using. The model inputs and three settlement 3. CONCLUSIONS
parameters as the model outputs. The input parameters Through the above review and discussion, the
were, the length of excavation from drive start, the depth success of artificial intelligence applications has been
of soil cover above tunnel crown, the area of tunnel demonstrated in geotechnical engineering. Based on the
section, the delay for closing invert, the water level depth, results of the research above, it can be said that artificial
the rate of advance of excavation, the construction intelligence techniques have performed better than or at
method, the mean blow count from standard penetration least approximated conventional methods.

2777
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Table-5. Summary of some applications of AI techniques in geotechnical engineering.

Data collection
No Researchers Techniques Results
methods
Pile driving records Reanalyzed using neural networks
They indicated that the neural network predictions
1 Goh Actual pile Back Propagation
are more reliable than the conventional pile driving
1996 driving records Neural Networks
formulae
Application of an Artificial Neural Network for Analysis of Subsurface Contamination at the Schuyler Falls
Landfill, NY
Applied and tested a new pattern method on a variety
of site characterization problems, called it “SCANN”
(Site characterization using Artificial Neural
2
Rizzo and Networks), Unlike the kriging methods, SCANN is
Artificial Neural
Dougherty Historical Data data-driven and requires no estimation of a
Networks
1996 covariance function. It uses a feed-forward counter
propagation training approach to determine a "best
estimate" or map of a discrete spatially-distributed
field.
Prediction of Pile Bearing Capacity Using Artificial Neural Networks
In situ pile load
Error Back The results showed that the neural networks predicted
3 Lee and Lee tests obtained
Propagation Neural values corresponding the measured values much
1996 from a
Networks better than those obtained from Meyerhof’s equation.
literatures
General regression neural networks for driven piles in cohesionless soils
Concluded that the GRNNM is applicable for all
4 Abu-Kiefa General Regression
Historical Data different conditions of driven piles in cohesionless
1998 Neural Network
soils.
Prediction of Pile Capacity Using Neural Networks
The study showed that the neural network model
predicted the total capacity reasonably well. The
5 Teh et al. Back propagation
Historical Data neural-network-predicted soil resistance along the
1997 neural networks
pile was also in general agreement with the
CAPWAP solution.
Data collection
No Researchers Techniques Results
Methods
Subsurface Characterization Using Artificial Neural Network And Gis
Multilayer
The integrated approach of ANN and GIS, is shown
perceptron using
6 Gangopadhya to be a powerful tool for characterizing complex
Historical Data the back-
y et al., 1999 aquifer geometry, and for calculating aquifer
propagation
parameters for ground-water flow modeling.
algorithm
Artificial intelligence techniques for the design and analysis of deep foundations
Based on the results from this investigation, it
NN, and
appeared that the proposed neural network models
7 Nawari et al., Generalized
Historical Data furnish a pragmatic and a reliable alternative for the
1999 Regression Neural
current analysis and design techniques of axial pile
Network
capacity and laterally loaded piles.
Bayesian Neural Network Analysis of Undrained Side Resistance of Drilled Shafts
The developed neural network model provided good
estimates of the undrained side resistance adhesion
factor. Furthermore, one distinct benefit of this neural
8 Goh et al., Bayesian neural
Historical Data network model is the computation of the error bars on
2005 network algorithm
the predictions of the adhesion factor. These error
bars will aid in giving confidence to the predicted
values and the interpretation of the results.
Undrained Lateral Load Capacity of Piles in Clay Using Artificial Neural Network
9
Das and Historical Data Back propagation The developed ANN model is more efficient

2778
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Basudhar, neural networks compared to empirical models of Hansen and Broms.

2006
Prediction of Friction Capacity of Driven Piles In Clay Using The Support Vector Machine
With the database collected by Goh (1995) the study
10 Samui, shows that SVM has the potential to be a useful and
Data Base SVM
2008 practical tool for prediction of friction capacity of
driven piles in clay.
Modelling Pile Capacity Using Gaussian Process Regression
The GP regression approach works well in predicting
the load-bearing capacity of piles as compared to the
SVM approach. Another conclusion from this study
Actual pile- is that the Pearson VII function kernel performs well
Pal and Gaussian Process
11 driving records in comparison to the radial basis function kernel with
Deswal (GP) Regression
in cohesion-less both GP- and SVM-based approaches to model the
2010 and SVM
soil pile capacity. The results of this study also suggest
that GP regression works well as compared to the
empirical relations in predicting the ultimate pile
capacity.
Prediction of Pile Settlement Using Artificial Neural NetworksBased on Cone Penetration Test Data
The results indicate that back-propagation neural
Nejad and networks have the ability to predict the settlement of
12 Back-propagation
Jaksa Database pile with an acceptable degree of accuracy (r=0.956,
neural network
2010 RMSE=1.06 mm) for predicted settlements ranging
from 0.0 to 137.88 mm.
Intelligent Computing for Modeling Axial Capacity of Pile Foundations

The results indicate that the ANN models were

capable of accurately predicting the ultimate capacity
13 Shahin Artificial Neural
Historical Data of pile foundations, and compare well with what one
2010 Network (ANN)
would expect based on available geotechnical
knowledge and experimental results.

Data collection
No Researchers Techniques Results
Methods
Neural Network Model for Predicting the Resistance of Driven Piles
The results showed that the ANN model served as a
14 data from
Park and Cho Artificial Neural reliable and simple predictive tool to predict the
dynamic piles
2010 Network (ANN) resistance of the driven pile with correlation
load test
coefficient values close to 0.9.
Neural Network Application in Prediction of Axial Bearing Capacity of Driven Piles
The results showed that the neural network models
Harnedi and give a good prediction of axial bearing capacity of
15 Pile Driving Artificial Neural
Kassim piles if both stress wave data and properties of both
Analyzer (PDA) Network (ANN)
2013 driven pile and driving system are considered in the
input data.
Application of Artificial Neural Network for Predicting Shaft and Tip Resistances of Concrete Piles
Founded that a network with five hidden nodes in one
hidden layer yields the best performance.
16 Momeni et al., Pile Driving Artificial Neural Additionally, through a sensitivity analysis, it was
2015 Analyzer (PDA) Network (ANN) founded/ that the pile length and cross sectional area
are the most influential parameters in predicting the
bearing capacity of piles
17 Analysis of Ultimate Bearing Capacity of Single Pile Using the Artificial Neural

2779
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

The results showed that neural networks can be used

Full-Scale Pile for prediction of ultimate bearing capacity of single
Wardani et al., Artificial Neural
Load Test and pile foundation and the model have the highest
2013 Network (ANN)
SPT performance among the other methods, even though
the difference is not too big.
ANN Prediction of Some Geotechnical Properties of Soilfrom their Index Parameters
Comparison between the results of the developed
models and experimental data indicated that
predictions are within a confidence interval of 95 %.
18 Tizpa et. al artificial neural
Database According to the performed sensitivity analysis,
2014 network
Atterbeg limits and the soil fine content (silt+clay)
are the most important variables in predicting the
maximum dry density and optimum moisture content.
Load–settlement modeling of axially loaded steel driven piles using CPT-based recurrent NNs
Founded that the developed RNN model has the
19 ability to reliably predict the load–settlement
Shahin Pile Load Tests, Recurrent neural
response of axially loaded steel driven piles, and
2014a and (CPT) Data network (RNN)
thus, can be used by geotechnical engineers for
routine design practice.
Evolutionary-Based Approaches for Settlement Prediction of Shallow
Foundations on Cohesionless Soils
Polynomial
regression, genetic In this study, the feasibility of the EPR, GP and GEP
20 Shahnazari et.
programming (GP), approaches in finding solutions for highly nonlinear
al Historical Data
& gene expression problems such as settlement of shallow foundations
2014
programming on granular soils is also clearly illustrated
(GEP)
State-of-the-Art Review of Some Artificial Intelligence Applications in Pile Foundations

21 Shahin Artificial AI techniques perform better than, or at least as good

Historical Data
2014b intelligence as, the most traditional methods.

Data collection
No Researchers Techniques Results
Methods
Artificial Neural Network Model for Prediction of Bearing Capacity of Driven Pile
The results show that the ANN model serves as a
Pile Driving reliable prediction tool to predict the resistance of the
22 Maizir et. al Artificial Neural
Analyzer (PDA) driven pile with coefficient of correlation (R) values
2015 Network
test data close to 0.9 and mean squared error (MSE) less than
1% .
Toward improved prediction of the bedrock depth underneath hillslopes: Bayesian inference of the bottom-up
control hypothesis using high-resolution topographic data
The results demonstrate that the proposed DTB
23 Numerical
Gomes et al. High-resolution model with lumped parameters mimics reasonably
modeling, and
2016 topographic data well the observed regolith depth data with root mean
Bayesian analysis
square error (RMSE).
Determination bearing capacity of driven piles in sandy soils using Artificial Neural Networks
The NN has very high efficiency in predicting load
24 Mazaher and carrying capacity of metal piles, and it is concluded
MLP Neural
Berneti Database that soil internal friction angle, soil elastic modulus,
Network
2016 pile diameter and pile length respectively have
maximum effect on load carrying capacity of piles.

2780
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

REFERENCES Burland J. B. 1973. Shaft friction of piles in clay. Ground

Engineering.
A. Ardakanivand V. R. Kohestani. 2015. Evaluation of
liquefaction potential based on CPT results using C4.5 C. Kayadelen. 2011. Soil liquefaction modeling by
decision tree. Ardakani & Kohestani/ Journal of AI and Genetic Expression Programming and Neuro-Fuzzy.
Data Mining. Expert Systems with Applications 38.

Abu-Kiefa M. A. 1998. General regression neural Cal Y. 1995. Soil classification by neural-network.
networks for driven piles in cohesionless soils. J. Geotech. Advances in Engineering Software.
Geoenv. Engrg. ASCE, Vol. 123.
Chan W.T., Chow Y.K., Liu L.F. 1995. Neural network:
Adel M. Hanna, Derin Ural, Gokhan Saygili. 2007. an alternative to pile driving formulas. Computers and
Evaluation of liquefaction potential of soil deposits using Geotechnics.
artificial neural networks. Engineering Computations:
International Journal for Computer-Aided Engineering and Cho S.E. 2009. Probabilistic stability analyses of slopes
Software. Vol. 24. using the ANN-based response surface, Computers and
Geotechnics. Vol. 36.
Agrawal G., Weeraratne S., and Khilnani K. 1994.
Estimating clay liner and cover permeability using Djajaputra A. A. 1997. Konsep Beban Terfaktor Dalam
computational neural networks. Proc., First Congress on Perancangan Tiang Bor, Proceeding Seminar PILE’97.
Computing in Civil Engrg., Washington, June.
Ellis G.W.; Yao, C; Zha o R. & Penumadu D. 1996.
Ahmadreza Mazaher and Mazaher Berneti. 2016. Stress–strain modeling of sands using artificial neural
Determination bearing capacity of driven piles in sandy networks, J Geotech Eng. Vol. 121.
soils using Artificial Neural Networks (ANNs). IOSR
Journal of Mechanical and Civil Engineering. Elshorbagy A., Corzo G., Srinivasulu S., Solomatine D.P.,
2010. Experimental investigation of the predictive
Akbar A. Javadi Mohammad Rezania Mohaddeseh capabilities of data driven modeling techniques in
Mousavi Nezhad. 2006. Evaluation of liquefaction hydrology-part 1: concepts and methodology. Hydrology
induced lateral displacements using genetic programming. and Earth System Science.
Computers and Geotechnics.
Eslami A. & Fellenius B. H. 1996. Pile capacity by direct
Alkroosh I., Nikraz H. 2011. Simulating pile load- CPT and CPTu methods applied to 102 case histories.
settlement behavior from CPT data using intelligent Canadian Geotechnical Journal.
computing. Central European Journal of Engineering.
Eslami A. 1996. Bearing capacity of piles from cone
Alsamman, O.M., 1995. The Use of CPT for Calculating penetration test data. Ph.D. thesis, University of Ottawa,
Axial Capacity of Drilled Shafts. PhD Thesis. University Ottawa, Ont.
of Illinois-Champaign, Urbana, Illinois.
F. Pooya Nejad; Mark B. Jaksa. 2010. Prediction of Pile
American Petroleum Institute. RP2A: Recommended Settlement Using Artificial Neural Networks Based on
practice for planning, designing and constructing fixed Cone Penetration Test Data. Advances in Analysis,
offshore platfonns, 15th Ed. 1984. Modeling & Design. 2010 ASCE.

Arnold M. A. 1999. Artificial neural networks applied to F. Pooya Nejad; Mark B. Jaksa. 2010. Prediction of Pile
the prediction of settlements of shallow foundations in Settlement Using Artificial Neural Networks Based on
granular soils. MEngSc. thesis, James Cook University, Cone Penetration Test Data. Advances in Analysis,
Queensland. Modelling & Design. 2010 ASCE.

Basheer I. A., Reddi, L. N. and Najjar Y. M. 1996. Site Ferentinou M.D. & Sakellariou M.G. 2007. Computational
characterisation by neuronets: An application to the intelligence tools for the prediction of slope performance,
landfill siting problem. Ground Water. Computers and Geotechnics.

Briaud J. L. & Tucker L. M. 1988. Measured and Flaate K. S. 1964. An investigation of the validity of three
predicted axial capacity of 98 piles. ASCE, Journal of pile drivingformulae in cohesionless material. Norwegian
Geotechnical Engineering. 114(9). Geotech. Inst., Oslo, Norway.

Broms B. B. and Lim P. C. 1988. A simple pile driving Gemitzi Falalakis, Eskioglou Petalas. 2011. Evaluating
formula based on stress-wave measurements. Proc. 3rd Int. Landslide Susceptibility Using Environmental Factors,
Conf. on the Application of Stress-Wave Theory to Piles.

2781
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Fuzzy Membership Functions and Gis. Global Nest Gribb M. M. and Gribb G. W. 1994. Use of neural
Journal. networks for hydraulic conductivity determination in
unsaturated soil. Proc., 2nd Int. Conf. Ground Water
Ghaboussi J. and Sidarta D. E. 1997. New nested adaptive Ecology, J. A. Stanford and H. M. Valett, eds., Bethesda
neural networks (NANN) for constitutive modeling. J. MD: Amer, Water Resources Assoc.
Computers and Geotechnics.
Griffith D.V., G.A. Fenton, N. Manoharan. 2002. Bearing
Goble G. G., Likins G. E. and Rausche F. 1975. Bearing Capacity of Rough Rigid Strip Footing on Cohesive Soil:
capacity of piles from dynamic measurements. Final Probability Study, Journal of Geotechnical and
Report, Dept. of Civil Eng., Case Western University. Geoenvironmental Engineering.

Goh A. T. C. 1994a. Nonlinear modelling in geotechnical Guilherme J. C. Gomes, Jasper A. Vrugt and Eurfıpedes
engineering using neural networks. Australian Civil A. Vargas Jr. 2016. Toward improved prediction of the
Engineering Transactions. bedrock depth underneath hillslopes: Bayesian inference
of the bottom-up control hypothesis using high-resolution
Goh A. T. C. 1994a. Nonlinear modelling in geotechnical topographic data. Water Resources Research.
engineering using neural networks. Australian Civil
Engineering Transactions. H. Shahnazari M.A. Shahin and M.A. Tutunchian1. 2014.
Evolutionary-based approaches or settlement prediction of
Goh A. T. C. 1994b. Seismic liquefaction potential shallow foundations on cohesionless soils. International
assessed by neural network. J. Geotech. & Geoenv. Eng., Journal of Civil Engineering, Vol. 12, No. 1, Transaction
120. B: Geotechnical Engineering, January.

Goh A. T. C. 1995a. Back-propagation neural networks Hajihassani M., Marto A., Nazmi E., Abad S.,
for modeling complex systems. Artificial Intelligence in Shahrbabaki, M. 2011. Prediction of surface settlements
Engineering. induced by NATM tunneling based on artificial neural
networks. Electro. J. Geotech. Eng.
Goh A. T. C. 1995a. Back-propagation neural networks
for modeling complex systems. Artificial Intelligence in Harnedi Maizir, Nurly Gofar and Khairul Anuar Kassim.
Engineering. 2015. Artificial Neural Network Model for Prediction of
Bearing Capacity of Driven Pile. Jurnal Teoretis dan
Goh A. T. C. 1995b. Empirical design in geotechnics Terapan Bidang Rekayasa Sipil.
using neural networks. Geotechnique.
Hiley A. 1922. The efficiency of the hammer blow, and its
Goh A. T. C. 1995c. Modeling soil correlations using effects with reference to piling. Engineering,
neural networks.”J. Computing in Civil Eng., ASCE.
Hyun U. Park. 2011. Study for Application of Artificial
Goh A. T. C. 1996a. Neural-network modeling of CPT Neural Networks in Geotechnical Problems, Artificial
seismic liquefaction data. J. Geotech. Engrg., ASCE Neural Networks - Application, Dr. Chi Leung Patrick Hui
Goh A. T. C. 1996b. Pile driving records reanalyzed using (Ed.).
neural networks. J. Geotech. Eng.
Hyun U. Park & Cho, C.H. 2010. Neural Network Model
Goh A. T. C., Wong K. S. and Broms B. B. 1995d. for Predicting the Resistance of Driven Piles Marine
“Estimation of lateral wall movements in braced Geosources and Geotechnology, In Press.
excavation using neural networks. Canadian Geotech
Jaksa, M. B. 1995. The influence of spatial variability on
Goh A. T. C.; Fred H. Kulhawy, F. ASCE; and C. G. the geotechnical design properties of a stiff, over
Chua. 2005. Bayesian Neural Network Analysis of consolidated clay, PhD. thesis, The University of
Undrained Side Resistance of Drilled Shafts 10.1061/ Adelaide, Adelaide.
(ASCE)1090.
Janbu N. 1953. Une analyse energetique du battage des
Goh A.T., Kulhawy F.H., Chua C.G. 2005. Bayesian pieux a l'aide de parametres sans dimension. Norwegian
neural network analysis of undrained side resistance of Geotech. Inst., Oslo, Norway.
drilled shafts. J. Geotech. Geoenviron. Eng.
Juwaied Nabeel S. and Faiq M.S. Al-Zwainy. 2017. Pile
Gordon T.C. Kung, Evan C.L. Hsiao, Matt Schuster, C. Design Using Multiple Linear Regression Model. ARPN
Hsein Juang. 2007. A neural network approach to Journal of Engineering and Applied Sciences. 12(15).
estimating deflection of diaphragm walls caused by
excavation in clays. Computers and Geotechnics. Kim Young-Su and Kim Byung. 2006. Use of Artificial
Neural Networks in the Prediction.

2782
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

L.E. Besawa, D.M. Rizzob and P.J. Mouserb. 2006. Ni S.H.; Lu P.C. & Juang C.H. 1996. A fuzzy neural
Application of an Artificial Neural Network for Analysis network approach to evaluation of slope failure potential,
of Subsurface Contamination at the Schuyler Falls Journal of Microcomputers in Civil Engineering.
Landfill, NY.
Parichehr Tizpa, Reza Jamshidi Chenari, Mehran
Lee I. M. and Lee J. H. 1996. Prediction of pile bearing Karimpour Fard, Sandro Lemos achado. 2014. ANN
capacity using artificial neural networks. Computers and prediction of some geotechnical properties of soil from
Geotechnics. their index parameters. Arab J Geosci.

M. Peng, X.Y. Li, D.Q. Li, S.H. Jiang, L.M. Zhang. 2014. Penumadu D. and Jean-Lou C. 1997. Geomaterial
Slope safety evaluation by integrating multi-source modeling using artificial neural networks. Artificial Neural
monitoring information. Structural Safety. Networks for Civil Engineers: Fundamentals and
Applications, ASCE.
M.H. Baziar, A. Ghorbani. 2005. Evaluation of lateral
spreading using artificial neural networks. Soil Dynamics Penumadu D. and Jean-Lou C. 1997. Geomaterial
and Earthquake Engineering. modeling using artificial neural networks. Artificial Neural
Networks for Civil Engineers: Fundamentals and
Meyerhof G. G. 1965. Shallow foundations. J. Soil Mech. Applications, ASCE.
& Found. Div.
Penumadu D. and Zhao R. 1999. Triaxial compression
Meyerhof G. G. 1965. Shallow foundations. J. Soil Mech. behavior of sand and gravel using artificial neural
& Found. Div. networks (ANN). J. Computers and Geotechnics.

Meyerhof G. G. 1976. Bearing capacity and settlement of Penumadu D., Jin-Nan L., Chameau J.-L. and Arumugam
pile foundations. J. Geotech. Eng. S. 1994. Rate dependent behavior of clays using neural.

Mohammad H.Baziar, Yase Jafarian, Habib Shahnazari, Pijush Samui. 2012. Three-Dimensional Site
Vahid Movahed, Mohammad Amin Tutunchian. 2014. Characterization Model of Bangalore Using Support
Prediction of strain energy-based liquefaction resistance of Vector Machine. International Scholarly Research
sand-silt mixtures: An evolutionary approach. Computers Network ISRN Soil Science.
& Geosciences.
Rahman M.S. and M. Mulla. 2005. Fuzzy Neural Network
Mohammad Rezania, Asaad Faramarzi, AkbarA. Javadi. Models for Geotechnical Problems, NCDOT Research
2011. An evolutionary based approach for assessment of Project, USA.
earthquake-induced soil liquefaction and lateral Randolph M. F. 1985. Capacity of piles driven into dense
displacement. Engineering Applications of Artificial sand. Rep. Soils TR 171, Engrg. Dept., Cambridge
Intelligence. University, Cambridge.

Najjar Y. M. and Ali H. E. 1998. CPT-based liquefaction Randolph M. F. and Wroth C. P. 1978. Analysis of
potential assessment: A neuronet approach. Geotechnical deformation of vertically loaded piles. J. Geotech. Eng.
Special Publication, ASCE.
Rausche F., Moses F. and Goble G. G. 1972. Soil
Najjar Y. M. and Basheer I. A. 1996a. Utilizing resistance predictions from pile dynamics. J. Soil Mech.
computational neural networks for evaluating the and Found. Div.
permeability of compacted clay liners. Geotechnical and
Geological Engineering. Rausche F., Moses F. and Goble G. G. 1972. Soil
resistance predictions from pile dynamics. J. Soil Mech.
Najjar Y. M. and Basheer I. A. 1996b. Utilizing and Found. Div.
computational neural networks for evaluating the
permeability of compacted clay liners. Geotechnical and Rezania M, Javadi A.A. 2007. A new genetic
Geological Engineering. programming model for predicting settlement of shallow
foundations. Canadian Geotechnical Journal.
Najjar Y. M., Basheer I. A. and Naouss W. A. 1996b. On
the identification of compaction characteristics by Rezania M, Javadi A.A. 2007. A new genetic
neuronets. J. Computers and Geotechnics. programming model for predicting settlement of shallow
foundations, Canadian Geotechnical Journal.
Nawari N. O., Liang R. and Nusairat J. 1999. Artificial
intelligence techniques for the design and analysis of deep Rizzo D. M. and Dougherty D. E. 1994. Application of
foundations. Elec. J. Geotech. Eng. artificial neural networks for site characterization using

2783
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

hard and soft information. Proc., 10th Int. Conf. Shahin M. A., Jaksa M. B. and Maier H. R. 2002a.
Computational Methods in Water Resources. Artificial neural network-based settlement prediction
formula for shallow foundations on granular soils.
Romero S. and Pamukcu S. 1996. Characterization of Australian Geomechanics.
granular meterial by low strain dynamic excitation and
ANN. Geotechnical Special Publication, ASTM-ASCE. Shahin M. A., Jaksa M. B. and Maier H. R. 2003a.
Neurofuzzy networks applied to settlement of shallow
S. Banu Ikizler, Mustafa Vekli, Emrah Dogan, Mustafa foundations on granular soils. 9th Int. Conf. on
Aytekin, Fikret Kocabas. 2012. Prediction of swelling Applications of Statistics and Probability in Civil Eng.
pressures of expansive soils using soft computing
methods. Neural Comput & Applic. Shahin M. A., Maier H. R. and Jaksa M. B. 2002b.
Predicting settlement of shallow foundations using neural
S. Banu Ikizler, Mustafa Aytekin, Mustafa Vekli, Fikret networks. J. Geotech. & Geoenv. Eng.
Kocabas. 2009. Prediction of swelling pressures of
expansive soils using artificial neural networks. Advances Shahin M. A., Maier H. R. and Jaksa M. B. 2002b.
in Engineering Software. Predicting settlement of shallow foundations using neural
networks. J. Geotech. & Geoenv. Eng.
Sarat Kumar Das, Rajani Kanta Biswal, N. Sivakugan,
Bitanjaya Das. 2011. Classification of slopes and Shahin M. A., Maier H. R. and Jaksa M. B. 2003c. Neural
prediction of factor of safety using differential evolution and neurofuzzy techniques applied to modelling settlement
neural networks Environ Earth Sci. of shallow foundations on granular soils. Int. Congress on
Modelling and Simulation.
Schmertmann J. H. 1970. Static cone to compute static
settlement over sand. J. Soil Mech. & Found. Shahin M. A., Maier H. R. and Jaksa M. B. 2003c.
“Neural and neurofuzzy techniques applied to modelling
Schmertmann J. H., Hartman J. P. and Brown P. B. 1978. settlement of shallow foundations on granular soils. Int.
Improved strain influence factor diagrams. J.Geotech. Congress on Modelling and Simulation.
Eng.
Shahin M.A.; Jaksa M.B. & Maier H.R. 2005. Stochastic
Schultze E and Sherif G. 1973. Prediction of settlements simulation of settlement prediction of shallow foundations
from evaluated settlement observations for sand. Proc. 8th based on a deterministic artificial neural network.
Int. Conf. Soil Mechan. & Found. Eng.
Shi J. J. 2000. Reducing prediciton error by transforming
Schultze E. and Sherif, G. 1973. Prediction of settlements input data for neural networks. J. Computing in Civil
from evaluated settlement observations for sand. Proc. 8th Engrg., ASCE.
Int. Conf. Soil Mechan. & Found. Eng.
Shi J., Ortigao J. A. R. and Bai J. 1998. Modular neural
Seed H.D., Tokimatsu K., Harder L. F. and Chung R. M, networks for predicting settlement during tunneling.
1985. Influence of SPT procedure in soil liquefaction J.Geotech. & Geoenv. Engrg., ASCE.
resistance evaluation Journal of Geotechnical Engineering.
ASCE. Shibata T. and Teparaksa W. 1988. Evaluation of
liquefaction potentials of soils using cone penetration
Semple R. M. and Rigden W. J. 1986. Shaft capacity of tests. Soils and Foundations.
driven pipe piles in clay. Ground Engineering.
Shivani Chauhana, Mukta Sharmab, M.K. Aroraa, N.K.
Shahin M. 2014a. Load-settlement modelling of axially 2010. Guptac Landslide Susceptibility Zonation through
loaded drilled shafts using CPT-based recurrent neural ratings derived from Artificial Neural Network.
networks. International Journal of Geomechanics. International Journal of Applied Earth Observation and
Geoformation.
Shahin M. A 2010. Intelligent computing for modeling
axial capacity of pile foundations. Canadian Geotechnical Sivakugan N., Eckersley J. D. and Li H. 1998. Settlement
Journal. predictions using neural networks. Australian Civil
Engineering Transactions.
Shahin M. A., Jaksa M. B. and Maier H. R. 2000.
Predicting the settlement of shallow foundations on Sivakugan N., Eckersley J. D. and Li H. 1998. Settlement
cohesionless soils using back-propagation neural predictions using neural networks. Australian Civil
networks. Research Report No. R 167, The University of Engineering Transactions.
Adelaide,
Subhrendu Gangopadhyay, Tirtha Raj Gautam, and Ashim
Das Gupta. 1999. Subsurface Characterization Using

2784
 VOL. 13, NO. 8, APRIL 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Artificial Neural Network and Gis. Journal of Computing

in Civil Engineering.

Teh C. I., Wong K. S., Goh A. T. C. and Jaritngam S.

1997. Prediction of pile capacity using neural networks. J.
Computing in Civil Eng.

Terzaghi K. and Peck R. 1967. Soil mechanics in

engineering practice, John Wiley & Sons, New York.

Wardani S.P.R., Surjandari N.S., Jajaputra A.A. 2013.

Analysis of Ultimate Bearing Capacity of Single Pile
Using the Artificial Neural. Proceedings of the 18th
International Conference on Soil Mechanics and
Geotechnical Engineering, Paris 2013.

Wellington A. M. 1892. The iron wharf at Fort Monroe,

Va. ASCE Transactions.

Yongjei Lee, Sungchil Lee, and Hun-Kyun Bae. 2014.

Design of Jetty Piles Using Artificial Neural Networks.
Hindawi Publishing Corporation the Scientific World
Journal Volume.

Youd TL, Hansen CM, Bartlett SF. 2002. Revised

multilinear regression equations for prediction of lateral
spread displacement. J Geotech Geoenviron Eng; ASCE.
128.

Zhou Y. and Wu X. 1994. Use of neural networks in the

analysis and interpretation of site investigation data.
Computer and Geotechnics.

Zhu J. H., Zaman M. M. and Anderson S. A. 1998a.

Modeling of soil behavior with a recurrent neural network.
Canadian Geotech. J.

Zhu J. H., Zaman M. M. and Anderson S. A. 1998b.

Modelling of shearing behaviour of a residual soil with
recurrent neural network. Int. J. Numerical and Analytical
Methods in Geomechanics.

2785