You are on page 1of 18

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322117796

Prediction of Bearing Capacity of Stone Columns Placed in Soft Clay Using


ANN Model

Article  in  Geotechnical and Geological Engineering · December 2017


DOI: 10.1007/s10706-017-0436-0

CITATIONS READS

3 948

1 author:

Manita Das
National Institute of Technology, Silchar
6 PUBLICATIONS   12 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Ground improvement technique View project

All content following this page was uploaded by Manita Das on 19 August 2018.

The user has requested enhancement of the downloaded file.


Geotech Geol Eng (2018) 36:1845–1861
https://doi.org/10.1007/s10706-017-0436-0

ORIGINAL PAPER

Prediction of Bearing Capacity of Stone Columns Placed


in Soft Clay Using ANN Model
Manita Das . Ashim Kanti Dey

Received: 16 January 2017 / Accepted: 22 December 2017 / Published online: 29 December 2017
 Springer International Publishing AG, part of Springer Nature 2017

Abstract It is known that the construction over soft input data and the bearing capacity as an output. The
clay is always a great challenge to the geotechnical predicted bearing capacity was compared with the
engineers. The soft clay poses high compressibility laboratory experimental data and Plaxis 2D data. The
and low bearing capacity. It is a common practice to predicted values were also compared with the values
construct piles in the soft clay to transfer the super- obtained from the established theories. ANN predi-
imposed load to the hard strata below. Construction of cated values showed a very good match even better
stone columns is also a technique of ground improve- than any of the established theories. Sensitivity
ment normally applied to the soft clay for increase in analysis showed that the effect of the / value is
bearing capacity and reduction in compressibility. maximum on determination of the bearing capacity.
Many theories are developed to determine the bearing
of a soft soil reinforced with stone columns. However, Keywords Artificial neural network  Bearing
most of the theories are site specific and do not show a capacity of stone columns  Soft clay
very good match with the field observations. Artificial
neural network (ANN) is an analytical tool which can
be used to predict some specific behavior of soil like
bearing capacity, settlement, etc. based on some input 1 Introduction
properties like density, shear strength parameters, void
ratio, etc. In the present study, 90 data were collected Construction over soft clay is always a great challenge
from the previously published literatures to build an to the geotechnical engineers. High compressibility
ANN model. Five parameters, namely, un-drained and low shear strength are the two primary reasons for
cohesion of soft clay (cu), friction angle of stone the same. It is a usual practice to construct piles in this
column material (/), ratio of spacing to diameter of type of soil to transfer the load to the hard stratum
the stone columns (s/d), length of the stone column below. For low to medium rise buildings cost of
(l) and number of the stone columns (n) were taken as construction of piles is very high as compared to the
cost of the structure. Hence, some ground improve-
ment techniques may be adopted to solve the problem.
M. Das (&)  A. K. Dey
Construction of stone columns is one of the most
Civil Engineering Department, N.I.T. Silchar, Silchar,
India commonly used ground improvement techniques in
e-mail: manitta403@gmail.com the soft clay. Stone columns increase the bearing
A. K. Dey capacity (Bouassida et al. 1995) and expedite the
e-mail: ashim_kanti@yahoo.co.in consolidation (Balaam and Booker 1981) of the soft

123
1846 Geotech Geol Eng (2018) 36:1845–1861

clay. However, use of stone columns is limited for affects the bearing capacity and settlement of stone
residential buildings, since settlement cannot be columns. Elshazly et al. (2008b) studied on the
significantly reduced. In order to minimize the settle- reliability of the settlement value for the soft soil
ment it is a pre-requisite to understand the factors reinforced with stone columns. Al-Homoud and
affecting the bearing capacity and settlement of the Degen (2006) discussed on the dynamic properties
stone columns. In the present study, artificial neural of soil reinforced with stone column. They found a
network (ANN) is used to study the different factors new idea to prevent the liquefaction by using marine
affecting the bearing capacity and finally to predict the stone columns.
bearing capacity of the stone columns. Artificial neural network (ANN) is a computational
The improved bearing capacity and expected technique to find out an output from a given number of
settlement of soft clay due to installation of stone inputs. It has the capability of input–output mapping.
columns are available in the Indian Standard code IS In geotechnical engineering, many researchers used
15284-Part 1 (2003) and other published papers. Many ANN to solve various problems (Goh et al. 1998; Kosti
researchers calculated the bearing capacity and settle- and Vasovic 2015). Chik et al. (2014) and Chik and
ment experimentally (Afshar and Ghazavi 2014b; Aljanabi (2014) used ANN to find out the settlement of
Ambily and Gandhi 2004, 2007; Castro 2014; the stone columns. Taha and Firoozi (2012) applied
Mohanty and Samanta 2015). Other researchers con- artificial intelligence system to predict the cohesion
ducted theoretical study or used numerical solution value of clayey soil. In the present study, ANN is used
through available commercial software and observed to predict the bearing capacity of the stone columns.
the behavior of the stone columns (Afshar and For obtaining an ANN model, data were collected
Ghazavi 2014a; Etezad et al. 2015; Frikha et al. from the published literatures. To select the best ANN
2013; Goh et al. 1998; Lee and Pande 1998; Das and model, four different validation technique and four
Deb 2016; Six et al. 2012; Fattah and Majeed 2012; algorithms were used. The results from the best fit
Nassaji and Asakereh 2013; Niroumand et al. 2011; ANN model were compared with the results from the
Borges et al. 2009; Deb et al. 2010; Tabchouche et al. laboratory tests, from Plaxis 2D analyses and from
2017; Castro 2017). Etezad et al. (2015) studied the well established theories. A good match was observed
group effect of the stone columns on the bearing with the laboratory results. An equation to find the
capacity by assuming a general shear failure. They bearing capacity of stone columns was derived from
used a limit equilibrium method for the bearing the best fit ANN model. Finally a sensitivity analysis
capacity determination. Castro (2014) found an ana- was carried out to obtain the most important input
lytical solution for predicting the settlement of the parameter affecting the bearing capacity of the stone
stone columns beneath rigid footings. Basack et al. columns.
(2017) made a numerical model based on the lateral
deformation of stone columns placed in soft clay.
Although the field tests are more accurate, yet 2 Short Description of the ANN Model
importance of analytical studies cannot be ignored.
Sometimes experimental studies become more costly Back-propagation neural network (BPN) was used
and time consuming and show lack in repeatability instead of radial basis function network (RBFN) since
when a number of parameters are involved for a BPN has a superior capability in pattern classification
certain solution. Moreover, the experimental results system (Haykin 1998; Chik and Aljanabi 2014).
have some uncertainties due to some errors such as Around 90 data were collected from the previously
error due to weather forecast, error due to installation published literatures. Four validation techniques,
effects, etc. So the analytical or numerical studies are namely, 90-10 validation (q90-10, ten-cross validation),
also important. Andreou and Papadopoulos (2014) 80-20 validation (q80-20), 70-30 validation (q70-30) and
discussed on the factors, which affect on the settle- 50-50 validation (q50-50) were used. For every model,
ment of stone column reinforced soft soils. Elshazly four algorithms namely, Bayesian Regularization
et al. (2008a) found out the effect of spacing between (BR), Levenberg–Marquardt (LM), Gradient descent
the stone columns due to installation of stone columns, with momentum (GDM), and Broyden-Fletcher-Gold-
because spacing is the most important factor which farb-Shannon (BFGS) were used as shown in Fig. 1.

123
Geotech Geol Eng (2018) 36:1845–1861 1847

Fig. 1 Flow chart of proposed ANN model

Thus, altogether 16 algorithms were performed to prediction of bearing capacity of stone column
investigate the performance of ANN in determination and it shows the better performance than any
of bearing capacity of the stone columns. Results from other algorithms as it shows a very good
all these 16 models were compared to observe the agreement between the results (Fig. 4b).
accuracy of their performances. (b) Levenberg–Marquardt (LM) learning algo-
The algorithms used in this study are briefly rithm: LM algorithm works as an iterative
discussed below: technique which reduces the performance func-
tion in each iteration. That is why it is the fastest
(a) Bayesian Regularization (BR) Learning Algo-
method to train the feed-forward neural network
rithm: Bayesian back propagation algorithm
of moderate size as in the present case. In case
was introduced by MacKay (1991) and Neal
of supervised learning, this algorithm gives the
(1992), which is based on the Bayesian statis-
most accurate result. In mathematics and com-
tical approach (Box and Tiao 1973). BR is a
puting techniques, the LM algorithm is also
mathematical phenomenon that transforms a
known as the damped least-squares (DLS)
nonlinear regression to a ‘‘well-posed’’ statisti-
method. This algorithm can also be used in
cal approach in the similar manner of a ridge
case of curve fitting problems. LM algorithm
regression. This algorithm creates a well gen-
contains the minimum of a multivariate func-
eralized network and forms simple weight
tion which is expressed in terms of sum of the
functions. Like most other algorithms, it does
squares of non-linear real valued functions.
not predict the single optimum prediction;
(c) Gradient descent with momentum (GDM)
rather it provides the probability distribution
learning algorithm: It is a first-order iterative
over all the predicted values which are very
optimization algorithm, which is used to find
important. It provides characteristic error infor-
local minima of a function. This algorithm is the
mation of the prediction which arises from the
steepest descent with momentum which ignores
uncertainties developed with interpolating the
the small features existing in error surface.
noisy data. This algorithm contains an objective
GDM algorithm is the most commonly used
function which includes sum of squared weights
algorithm for the optimization in neural net-
and residual sum of squares to reduce the errors.
work. This algorithm can also be used as a
In the present study, BR algorithm was used to
black-box optimization algorithm because in
observe the performance of ANN model for the

123
1848 Geotech Geol Eng (2018) 36:1845–1861

this case, practical explanations of the strengths ðunscaled value  minimum valueÞ
and weaknesses are very difficult. This algo- Scaled value ¼
ðmaximum value  minimum valueÞ
rithm can be used to measure the output error, to
ð1Þ
estimate the error gradient with the adjustment
of the weights in the direction of descending It is obvious from the above equation that the range
gradient. of the scaled value is 0–1. Table 1 shows the various
(d) Broyden–Fletcher–Goldfarb–Shannon (BFGS) data collected from the literature. The statistical
quasi-Newton back-propagation learning algo- dispersion is measured by the coefficient of variation
rithm: The BFGS algorithm is an iterative (CV). The CV is determined by the following equation
optimization method which can be used to SD
solve unconstrained nonlinear problems as in CV ¼  100 ð2Þ
Mean
the present case. This algorithm is the compu-
tationally lowest cost algorithm which is error where SD is the standard deviation.
tolerant. This algorithm gives best results for the
solution of low to moderate size problems. This 3.2 Optimizing Number of Hidden Neurons
algorithm utilises a local quadratic approxima-
tion of the corresponding error function and The ANN problem is normally a function of opti-
update the weights by approximating the mization problem, where the best network parameters
inverse of the Hessian matrix. BFGS algorithm such as weights and biases can be determined by
results good solutions through a small number minimizing the network error. For multilayer percep-
of iterations (Dennis and Schnabel 1996). tron-neural network (MLP-NN), the optimization of
number of hidden neurons is the most important factor.
The use of more number of hidden neurons results
more accurate output. But use of too many neurons
3 Artificial Neural Network (ANN) Formation results lengthy and time consuming outputs and over-
fitting of data. In this study, the optimum number of
The ANN model for the present study was formed hidden neurons was obtained from the minimum value
through following steps: of the mean squared error (MSE). Taking the number
of neurons from 1 to 20, the optimum number of
3.1 Collection of Data hidden neurons was found out as follows:

For the neural network model, total 90 data were • By plotting MSE versus number of hidden neuron
collected from the published papers (Ali et al. 2010; graphs.
Black et al. 2007; Choobbasti et al. 2011; Deb et al. • For the 90-10 validation technique with BR
2011; Fattah and Majeed 2009; Golait et al. 2009; algorithm, the MSE versus number of hidden
Hassen et al. 2010; Malarvizhi and Ilampurthi 2007; neuron graph was plotted as shown in Fig. 1.
Mohanty and Samanta 2015; Murugesan and Rajago- • From the graph, corresponding to minimum MSE
pal 2006, 2008; Nassaji and Asakereh 2013; Poor- value, the number of hidden neuron was taken as
ooshasb and Meyerhof 1997; Shivashankar et al. optimum number of hidden neuron. In Fig. 2, the
2010, 2011; Vekli et al. 2012; Zhang et al. 2013). optimum number of hidden neuron is noticed as
Some data are from experimental study, some are from four.
numerical study and the rest from analytical study. • Likewise, for all other models with all the algo-
40% of the total data were taken from the experimental rithms, the MSE versus number of hidden neuron
studies. In order to reduce the variations in the values graph was plotted separately and optimum number
due to different types of tests, a data is presented by a of hidden neurons was determined.
scaled value given by Eq. (1). • For all the models, the optimum numbers of hidden
neurons are listed in Table 2.

123
Geotech Geol Eng (2018) 36:1845–1861 1849

Table 1 Details of collected data


Number of stone Undrained cohesion, s/d ratio Friction angle, / Length, Load carrying capacity
columns, n cu in kPa in degree (l) in mm (q) in kPa

Minimum 7 5 1.5 20 10.5 20


Maximum 14 32 4 48 600 905
Mean 10.7 14.7783 2.431698 36.95755 298.7594 260.0283
SD 6.8 8.713176 0.735956 11.54476 191.9605 208.4899
CV 28.33 58.95925 30.26511 31.2379 64.25253 80.1797

0.0972 repeating this process, ten iterations were completed


until all the subsets were used for training.
Figure 3a–d show the Mean Standard Error (MSE)
0.0972
versus epoch graph of 90-10 (tenfold cross validation)
model for the algorithms BR, LM, GDM, and BFGS
0.0971 algorithms respectively. From the figure, it can be seen
that in all the cases as the number of epoch increases,
MSE

MSE value decreases which confirms that there is no


0.097 chance of over-fitting.
Figure 4a–d show the BPN network output versus
training of 90-10 (tenfold cross validation) model for
0.097
the algorithms BR, LM, GDM, and BFGS respec-
tively. From the figure, it is seen that the BR algorithm
0.0969 shows the best prediction with experimental value
0 2 4 6 8 10 12 14 16 18 20
among all other algorithms. In case of 90-10 validation
No of hidden neuron
model, the BR algorithm shows the best result for four
Fig. 2 Number of neurons versus mean standard error (MSE) numbers of hidden neurons.
graph It is observed from the Fig. 4, that the four
algorithms (BR, LM, GDM, and BFGS) show differ-
3.3 Training and Testing of ANN Model ent results for the same ANN model (90-10 i.e. tenfold
cross validation model). In case of BR algorithm, the
The collected data from the previous literature on correlation value, R, defined as the coefficient of
bearing capacity of stone columns was divided into determination is obtained as R = 0.99781; for LM
training, testing and validation data. With the training algorithm, R = 0.99617 and for GDM and BFGS
data, the network was developed and with the testing algorithms these values are R = 0.98134 and
and validation data, the trained model was verified. R = 0.97926 respectively. Standard error (SE) is
defined as the minimum distance between the data
3.3.1 90-10 Validation Technique (Tenfold Cross points and the fitted line. The Standard Error values for
Validation, q90-10) these algorithms are SE = 1.6124, 1.612, 2.305 and
2.321 respectively. All the prediction values are within
In 90-10 validation model, the whole data were the 95% prediction interval. From R and SE values it
divided into ten equal sections. Firstly, the first nine can be seen that the BR and LM learning algorithms
sets of data were used for training and the remaining show almost similar results but BR shows the best
one set was used for testing. In the second iteration, out performance.
of these ten sets of data, other nine subsets were used
for training and remaining one was used for testing. By

123
1850 Geotech Geol Eng (2018) 36:1845–1861

Table 2 Optimum number ANN models Algorithms R SE No of hidden neurons


of hidden neurons
q90-10 BR 0.99781 1.613 4
LM 0.99617 1.612 11
GDM 0.98134 2.305 18
BFGS 0.97926 2.321 16
q80-20 BR 0.98865 1.914 4
LM 0.9841 2.342 11
GDM 0.96038 2.505 9
BFGS 0.96862 2.621 7
q70-30 BR 0.979 2.114 4
LM 0.973 2.53 7
GDM 0.959 2.625 13
BFGS 0.958 2.632 14
q50-50 BR 0.972 2.414 11
LM 0.964 2.73 9
GDM 0.95335 2.825 16
BFGS 0.94347 3.032 15

3.3.2 80-20 Validation (Fivefold Cross Validation, testing and validation. To know the best performance
q80-20) of this technique, four algorithms were also used.
The correlation value (R), SE and optimum number
In 80-20 validation model, the whole data were of hidden neurons (n) for all the 16 models are listed in
divided into five equal sections. Firstly, the first four Table 2. From the table, it is seen that among all the 16
sets were used for training and remaining one set was models, the BR algorithm of 90-10 model shows the
used for testing and validation. Iteratively the remain- best performance as its R value is the largest and the SE
ing sets were used for training, testing and validation. value is the lowest of all the values from other models.
In this technique, four algorithms were used to know
the performance such as BR, LM, GDM and BFGS.
4 Validation of ANN Model with Experimental
3.3.3 70-30 Validation (Non-cross Validation, q70-30) and Plaxis-2D Results

In 70-30 validation model, 70% of the total collected 4.1 Experimental Investigation
data were used for training and the remaining 30% of
the data were used for testing and validation of the An experiment is an investigating method by which a
ANN model. In this case also, previously defined four hypothesis can be scientifically tested. In an experi-
algorithms were used. The details of all the algorithms ment, by assuming various independent variables, the
can be seen in Table 2. effect of dependent variable can be measured. Nor-
mally the independent variables are assumed as the
3.3.4 50-50 Validation (Two-Cross Validation, q50-50) cause and the dependent variables as the effect.
Moreover, the experimental methods allow the control
In 50-50 validation model, in the first iteration, first of variables. In this study, total ten plate load tests
50% of the total collected data were used for training were conducted and the bearing capacity for each test
and the remaining 50% of the data were used for was determined. All the tests were conducted in a
testing and validation of the ANN model. In second square tank of size 1 m 9 1 m 9 1 m. Locally avail-
iteration, second 50% of the total collected data were able soft clayey soil and stones of size 2–6 mm were
used for training and the first 50% were used for used for the investigation. To construct the model, first
of all, the tank was wrapped with a polythene sheet and

123
Geotech Geol Eng (2018) 36:1845–1861 1851

Best Validation Performance is 0.095966 at epoch 1000 Best Validation Performance is 0.082873 at epoch 2
0
10 Train
0
10 Train Validation
Validation Test

Mean Squared Error (mse)


Best
Mean Squared Error (mse)

Test
Best

-1
-1
10
10

-2
10
-2 10
0 200 400 600 800 1000 0 0.5 1 1.5 2
2 Epochs
1000 Epochs
(a) (b)
Best Validation Performance is 0.064272 at epoch 13 Best Validation Performance is 0.1021 at epoch 2
0 0
10 10 Train
Train
Validation Validation
Test Test
Mean Squared Error (mse)
Mean Squared Error (mse)

Best Best

-1
10
-1 10

-2
-2 10
10
0 0.5 1 1.5 2
0 5 10 15
19 Epochs 2 Epochs
(c) (d)

Fig. 3 Mean Standard Error (MSE) versus epoch graph of 90-10 (tenfold cross validation) model for the algorithms a BR, b LM,
c GDM, and d BFGS algorithms respectively

then 0.9 m depth of the tank was filled with the soft the centre. For experimental set-up, one hydraulic jack
soil in layers with uniform compaction to each layer. of capacity 5 t was used for the loading, two dial
The tank was covered with a thick polythene sheet and gauges were used to measure the settlement and one
kept idle for about a month for self consolidation of the proving ring was used to measure the applied load as
soil. Table 3 shows the properties of the soft clay shown in Fig. 6. For all the tests, one circular footing
which was used for all the experiments. To construct a of diameter 10 cm was placed centrally over the
stone column, a hole was made with the help of an arrangement of stone columns, i.e. over the central
auger of size 50 mm, the stones were inserted and stone column. It was tried to keep the water content
compacted with uniform compaction in layers. As per constant during all the experiments. The experiments
IS 15284-Part 1 (2003), seven stone columns in a were conducted by varying the input parameters such
triangular patter as shown in Fig. 5 are to be installed as the compactness of stone i.e. friction angle of stones
to determine the behavior of a single column placed at (u), spacing to diameter ratio (s/d) and, length of the

123
1852 Geotech Geol Eng (2018) 36:1845–1861

Fig. 4 BPN network output versus training of 90-10 (tenfold cross validation) model for the algorithms a BR, b LM, c GDM, and
d BFGS algorithms respectively

stone column (l). The data from the experiments are engineering. With this user-friendly interface, the
listed in Table 4. models can be efficiently created with a logical
geotechnical workflow. The finite element mesh can
4.2 Plaxis 2D be immediately created with the automatic meshing
procedure.
PLAXIS 2D is a FEM package which is being used for In this study, for determination of the bearing
a two dimensional analysis for the deformation and capacity of stone columns five parameters such as the
stability of different types of structures in geotechnical number of stone columns (n), undrained cohesion of

123
Geotech Geol Eng (2018) 36:1845–1861 1853

Table 3 Properties of clay mesh which was constructed in Plaxis-2D, is presented


Specific gravity 2.51
in Fig. 7. The ultimate bearing capacity of the stone
Bulk unit weight 1.7–1.75 gm/cc
column was obtained from the double tangent method
Liquid limit 61%
applied to the load settlement curve. The results from
the Plaxis-2D analysis are presented in Table 5.
Plastic limit 34.7%
Unified soil classification system CH
4.3 Comparison of Bearing Capacity Values

Since from the ANN study BR algorithm of q90-10


model shows the best performance, hence, all the
experimental and Plaxis-2D data are validated with
this model. The comparisons of all the bearing
capacity values are shown in Table 6. The first column
of Table 6 shows the experimental and Plaxis 2D
results and the second column shows the results of
ANN model. A comparison with the established
theories on bearing capacity obtained from IS
15284-Part 1 (2003), Etezad et al. (2015) and Afshar
and Ghazavi (2014a) is also shown in Table 6 against
column number 3, 4 and 5, respectively.
Fig. 5 Plan view of seven stone column
From the Table 6, it is observed that the ANN
model shows a better match with the experimental and
Plaxis 2D results than any other established method.
The comparison shows that ANN model is reliable in
predicting the bearing capacity of the stone columns. It
Hydraulic jack
also shows that whereas the established theories are
Proving ring
site specific, the ANN model is soil specific.
So clay
Dial gauge
4.4 Performance Evaluation of the Proposed
Stone column
Models
1mx1m tank
The performance of all the prediction models are
evaluated by different statistical parameters such as
Mean Absolute Percentage Error (MAPE), Variance
Absolute Relative Error (VARE), Median Absolute
Error (MEDAE) as mentioned in Eqs. (3), (4), and (5).
Table 7 shows the performance of all the prediction
models with their corresponding algorithms.
"  #
1 Xn 
t  x 
Fig. 6 Experimental set up MAPE ¼   i i   100 ð3Þ
n  ti 
i¼1
soft soil, (cu), the friction angle of stones (u), spacing "     !#
to diameter ratio (s/d) and, length of the stone column 1 Xn t i  xi  ti  xi 2
VARE ¼     
(l) were varied. The triangular arrangement of the n  t   mean t 
i¼1 i i
stone columns was considered for the analysis. The  100
plate size was kept constant as 10 cm for all the
analyses. In the finite element mesh, a 15-noded ð4Þ
triangular element and Mohr–Coulomb failure criteria
MEDAE ¼ medianðti  xi Þ ð5Þ
were considered. One of the figures of the deformed

123
1854 Geotech Geol Eng (2018) 36:1845–1861

Table 4 Experimental data


Number of stone columns, n cu (kPa) s/d ratio / (degree) L (mm) qu (kPa) measured value

7 22 1.5 40 400 206


7 21.5 2 40 400 185
7 21.3 2.5 40 400 165
7 21.5 3 40 400 148
7 21.2 3.5 37.27 400 120
7 21.5 1.5 37.27 500 145
7 21.5 2 37.27 500 134
7 21.5 2.5 35 500 128
7 21.5 3 35 500 118
7 21.5 3.5 35 500 106

5 MVR Model Development

To validate the best ANN model presented in this


study, an MVR analysis (multi-variable regression
analysis) was also performed with the same dataset for
the prediction of bearing capacity of the stone
columns. In the present study, to conduct the MVR
model, n, cu, s/d, / and l were considered as the
independent variables and the bearing capacity qu was
considered as the dependent variable. The relation
between the dependent and independent variables is
expressed in form of a standard equation as in Eq. (6)
Y ¼ c þ B1 X1 þ B2 X2 þ B3 X3 þ . . .Bn Xn  e ð6Þ
where c is the constant on Y intercept. X1 , X2 , X3 and
Xn are the independent variables and B1 , B2 , B3 and Bn
are the slopes associated with the above mentioned
Fig. 7 Deformed mesh of Plaxis-2D model independent variables; Y is the dependent variable and
e is the approximated error.
where ti is the measured value of bearing capacity of
5.1 Comparison of ANN Model
the stone columns, and xi is the predicted value of the
with the Statistical Approach of MVR Model
bearing capacity of stone column.
From Table 7, it can be seen that for the model q90-
The BR algorithm with ten-fold cross validation
10 (ten-cross validation), the algorithm BR has the
technique is also validated with the MVR model with
lowest values of MAPE, VARE, and MEDAE. For the
the same set of data. The comparisons of all the results
model q80-20 (five-cross validation), the algorithm BR
are presented in Table 8. The statistical presentation
has the lowest values of all the statistical parameters
from Table 8 shows the superior capability of ANN
(MAPE, VARE, and MEDAE). For the other two
model than MVR model for the prediction of bearing
models (i.e. q70-30 and q50-50) also, the algorithm BR
capacity of stone columns. The regression analysis of
has the lowest values. Kayri (2016) also obtained a
data was performed with F-test and t test at 95%
better performance with BR algorithm than with LM
confidence interval. The statistical results of MVR
algorithm.

123
Geotech Geol Eng (2018) 36:1845–1861 1855

Table 5 Plaxis 2D data


Number of stone columns, n cu (kPa) s/d ratio / (degree) L (mm) qu (kPa) measured value

7 25 1.5 40 400 232


7 25 2 40 400 210
7 25 2.5 40 500 194
7 25 3 40 500 175
7 25 3.5 37.27 500 150
7 22.5 1.5 37.27 600 190
7 21.5 2 37.27 600 165
7 21.5 2.5 35 600 142
7 23 3 35 600 125
12 25 2.5 37.27 400 227
12 23 2.5 37.27 400 199
12 21.5 2.5 37.27 400 175
19 25 2.5 37.27 400 250
19 23 2.5 37.27 400 227
19 21.5 2.5 37.27 400 200

model are presented in Table 9. The t-stat value, neuron k of the hidden layer; wik is the connecting
p value and upper and lower limits associated with the weight between input variable i and neuron k of the
individual coefficients are listed in Table 9. From hidden layer; Xi is the input variable i; and Y is the
Table 9, it is observed that p values associated with the output parameter. In this study, by using this equation,
parameters n and l are very high i.e. the confidence the normalized Y value (in this case 0–1) is found out.
level is very low [(1 - P) \ 95%] which represents In this case an ANN model is performed to find out the
an insignificant results for the MVR model. Again, for bearing capacity of stone column.
the parameters cu, s/d, and /, the P value is very low, The three layered neural network structure of this
which results a significant coefficient in MVR model. ANN model is presented in Fig. 8. The middle layer is
Therefore, it is concluded that the MVR model rejects called hidden layer which receives inputs from the
the significance of the parameters n and l in bearing input layer and sends output to the output layer. The
capacity prediction and hence, it fails to predict connection weights from input to hidden and hidden to
precisely the bearing capacity of stone columns. output are shown in this figure which was determined
from the matlab R2013a programming. The values of
weights and biases of the trained ANN model of BR
6 ANN Model Equation to Determine the Bearing algorithm with four numbers of hidden neurons are
Capacity of Stone Columns (q) Based on Trained presented in Table 10.
Neural Network To obtain the correlation of Y, the following
expressions are written:
Goh et al. (1998) expressed the mathematical equation s
based on the weights of input and output for ANN A1 ¼ 1:235 þ 0:2345n  0:1915cu þ 0:4325
d
model as shown in Eq. (7). þ 0:3125/ þ 0:1190l
( " ! #)
X h X
m ð8Þ
Y ¼ fsig b0 þ wk fsig bhk þ wik Xi ð7Þ s
k¼1 i¼1
A2 ¼ 0:2420 þ 0:3122n þ 0:2321cu þ 0:3128
d
where fsig is the sigmoid transfer function (in this case  0:1280/ þ 0:2235l
logistic); b0 is the bias at the output layer; wk is the
ð9Þ
connecting weight between kth neuron of the hidden
layer and the single output neuron; bhk is the bias at

123
1856 Geotech Geol Eng (2018) 36:1845–1861

Table 6 Comparison of qu (kPa) values obtained from experiments and Plaxis, ANN model and other established theories
qu (kPa) measured value from qu (kPa) predicted qu (kPa) from IS qu (kPa) from Etezad qu (kPa) from Afshar and
experiment and Plaxis 2D value from ANN CODE (2003) et al. (2015) Ghazavi (2014a)

206 215 182.76 262 350


185 182 170 238.05 335
165 161 155.8 212 330
148 148 131.77 179 327
120 116 118 152 194.07
145 152 164.43 213 257
134 133 142 203 245.89
128 131 153.3 196 182.45
118 118 161 190 170.11
106 109 152 182 159.32
232 250 230 306.8 380.44
210 217 218 296 380.06
194 189 211.5 262 356
175 173 199 228.5 329.7
150 150 185 190 228.55
190 185 189.072 233 206.038
165 160 168.5 206 195.89
142 145 148 182 171.45
125 121 134 205.45 179.5
227 232 199 218.16 259
199 205 188 205 240.8
175 172 149 194 217.2
250 247 211 242.5 298.55
227 223 205 232 260.6
200 203 197 223 226.79

s
A3 ¼ 0:790 þ 0:0353n þ 0:5323cu  0:2972 1:214
d B3 ¼ ð14Þ
½1 þ eA3 
þ 0:0491/  0:3232l
ð10Þ 0:2573
B4 ¼ ð15Þ
s ½1 þ eA4 
A4 ¼ 0:2143 þ 0:0544n þ 0:6520cu þ 0:1781
d C1 ¼ 0:3125 þ B1 þ B2 þ B3 þ B4 ð16Þ
þ 0:6563/  0:3724l
ð11Þ 1
Y¼ ð17Þ
½1 þ eC1 
0:2212
B1 ¼ ð12Þ The Y value in Eq. (17) is the normalized value
½1 þ eA1 
which is to be denormalized to obtain the actual
0:3250 bearing capacity of the stone column (q).
B2 ¼ ð13Þ
½1 þ eA2  q ¼ Ymin þ ðYmax  Ymin ÞY ð18Þ

123
Geotech Geol Eng (2018) 36:1845–1861 1857

Table 7 Statistical errors for all the ANN models for different For cu = 30 kPa, applying Skempton’s formula for
algorithms square footing on the ground surface, the net ultimate
ANN models Algorithms MAPE VARE MEDAE bearing capacity of soil is given by
  
q90-10 BR 2.54166 2.537 1.075 Df B
qnu ¼ 5cu 1 þ 0:2 1 þ 0:2
LM 2.5421 2.539 0.995 B L
GDM 3.1612 3.142 1.743 ¼ 5  30  ð1 þ 0:2  0Þð1 þ 0:2Þ
BFGS 3.0625 3.062 1.401 ¼ 180 kN=m2
q80-20 BR 2.342 2.343 1.213
Considering around two and half times improve-
LM 2.35012 2.322 1.003
ment in qnu due to installation of stone columns, the
GDM 3.1111 3.107 1.821
improved net ultimate bearing capacity
BFGS 3.0272 3.018 1.522
qnu ¼ 2:5  180 ¼ 450 kN/m2 :
q70-30 BR 2.67056 2.648 1.402
Assuming a suitable factor of safety, the net safe
LM 2.646 2.607 1.009
bearing capacity, qns ffi 250 kN/m2 (say).
GDM 3.062 3.041 1.915
Therefore, size of the footing can be assumed as
BFGS 2.9711 2.970 1.523
follows:
q50-50 BR 2.630244 2.641 1.431 rffiffiffiffiffiffiffiffi
LM 2.53025 2.510 1.011 800
B¼ ¼ 1:78 ffi 1:8 m:
GDM 3.1121 3.107 1.942 250
BFGS 3.0642 3.047 1.550
Again, for stone column, qns ffi 250 kN/m2 , mini-
mum and maximum values of Y are considered as
Table 8 Comparision of ANN and MVR model Ymin = 200 kPa and Ymax = 260 kPa respectively.
Model R SE MAPE VARE MEDAE Now, to design the group of stone columns, the
following steps are used:
ANN 0.99781 1.613 2.54166 2.537 1.075
MVR 0.99584 1.825 2.98441 2.512 1.135 • From Eq. (18), Y value is determined.
• C1 value is found out from Eq. (17).
• Then by using Eqs. (8)–(16), with the trial and
where Ymax and Ymin are maximum and minimum error method, all the design parameters (n, s/d and
values of Y respectively which are taken from the l) are estimated as presented in Table 11.
dataset.

8 Sensitivity Analysis of the Input Parameters


7 Mechanistic Analysis
of Experimental and Plaxis-2D Data
Let us consider a column taking an axial load of 800
In this study five input parameters were considered to
kN resting on a group of stone columns installed in a
obtain the bearing capacity of the stone columns. All
soft clayey soil with cu = 30 kPa, angle of friction of
the input parameters affect the bearing capacity little
stone materials, / = 42. The stone column group is
more or less. In order to obtain the parameter which
to be designed.

Table 9 Statistical Coefficients SE t-Stat p value Lower 95% Upper 95%


parameters of MVR model
Intercept - 6.86 3.52 - 5.61539 2.48E-07 - 30.07 -13.603
n - 0.53 0.32 - 0.43 0.253 - 0.62 0.26
cu 1.7 0.003 - 1.342 0.000 - 0.35 0.32
s/d 3.1 0.07 14.8 0.000 0.44 0.98
/ 1.33 0.022 11.5 1.01E-28 0.14 0.44
l 0.05 0.14 8.402 0.293 0.03 0.06

123
1858 Geotech Geol Eng (2018) 36:1845–1861

Fig. 8 Structure of
proposed ANN model

Table 10 Connection weights and biases for the trained ANN model for BR algorithm of 90-10 model
Neuron Weights Biases
Input 1 Input 2 Input 3 Input 4 Input 5 Output Hidden layer Output layer

Hidden neuron 1 (i = 1) 0.2345 - 0.1915 0.4325 0.3125 0.1190 0.2212 1.235 0.3125
Hidden neuron 2 (i = 2) 0.3122 0.2321 0.3128 - 0.1280 0.2235 - 0.3250 0.2420
Hidden neuron 3 (i = 3) 0.0353 0.5323 - 0.2972 0.0491 - 0.3232 1.214 0.790
Hidden neuron 4 (i = 4) 0.0544 0.6520 0.1781 0.6563 - 0.3724 0.2573 0.2143

Table 11 Design of group Number of stone columns, n Spacing to diameter ratio, s/d Length of stone columns, l in m
of stone columns arranged
in a triangular pattern to 7 2 2.8
carry a vertical load of 800
7 2.5 2.9
kN
7 3 3.0
12 3 2.9
12 3.5 3.02
12 4 3.14

mostly affects the bearing capacity of the stone performed in matlab R2013a. The outcome of the
columns, a sensitivity analysis of the input parameters analysis is shown in Fig. 9. It is observed that the
was performed. The sensitivity of an input parameter fourth parameter i.e. friction angle, u of the stone
is defined as the change in the output expressed as column material affects the bearing capacity the most
percentage for 1% increase in the input parameter. For among all other parameters. The second parameter
an example, 1% increase in u value of the soil changes affecting the bearing capacity is spacing to diameter
the value of bearing capacity by 5.88% hence, the ratio (s/d) of a stone column and the parameter which
sensitivity of u is 5.88%. This sensitivity analysis was

123
Geotech Geol Eng (2018) 36:1845–1861 1859

6 (b) In case of 80-20 model, the algorithm BR with 7


numbers of hidden neurons performs the best
5 with correlation, R = 0.98865 and SE = 1.914.
(c) For 70-30 model, the algorithm BR shows the
Change in output in %

4
best solution with 19 number of hidden neurons
with correlation, R = 0.979 and SE = 2.114.
(d) For 50-50 model also, the algorithm BR also
3
shows the best solution with nine hidden
neurons with correlation, R = 0.972 and
2
SE = 2.414.
(e) Overall performance of the BR algorithm is the
1 best as compared to other algorithms.
(f) ANN model has the superior capability in
0 predicting the bearing capacity of stone col-
del n del cu del s/d del phi del l
Input parameter umns over MVR model.
(g) From the MVR model, it is reported that the
Fig. 9 Sensitivity analysis of each input data parameters cu, s/d, and / are significant in
prediction of bearing capacity of stone columns.
has the least effect on bearing capacity is the length of (h) ANN model shows a better match with the
the stone column. experimental and Plaxis 2D results than any
other established method.
(i) The established theories are site specific; the
9 Conclusions ANN model is soil specific.
(j) For all the models, the algorithm BR has the
In this study, total 16 ANN models were performed lowest value of all the statistical parameters
with four different combinations and four different such as Mean Absolute Percentage Error
learning algorithms to predict the bearing capacity of (MAPE), Variance Absolute Relative Error
stone columns. For this purpose, 90 data were (VARE), and Median Absolute Error
collected from the previous technical literatures with (MEDAE).
different input values. The bearing capacity of stone (k) From the sensitivity analysis, the bearing
column mainly depends on the friction angle (u) of the capacity of stone column is mostly affected by
stone column materials, diameter (d), length (l), the friction angle of the stone column material
spacing between the stone columns (s), numbers of and is least affected by the length (l) of the stone
stone column (n), and undrained cohesion (cu) of the columns.
surrounding soft soil. So all these parameters were (l) A prediction model equation based on the
considered as inputs and bearing capacity of stone weights of the trained ANN model is presented
column was taken as output. To conduct the ANN to predict the bearing capacity of stone column.
model, a three-layer feed-forward back propagation
neural network was used. By optimizing the number of
hidden neurons for every models separately, different
numbers of hidden neurons were used to get the best References
results. Four types of algorithms i.e. BR, LM, GDM,
and BFGS were used in this case. From the analysis, Afshar JN, Ghazavi M (2014a) A simple analytical method for
the following conclusions are drawn: calculation of bearing capacity of stone column. Int J Civ
Eng 12(1). (Transaction B: Geotechnical Engineering)
(a) In case of 90-10 model, the algorithm BR with 4 Afshar JN, Ghazavi M (2014b) Experimental studies on bearing
capacity of geosynthetic reinforced stone columns. Arab J
numbers of hidden neurons gives the best result Sci Eng 39:1559–1571. https://doi.org/10.1007/s13369-
with correlation, R = 0.99781 and SE = 1.612. 013-0709-8

123
1860 Geotech Geol Eng (2018) 36:1845–1861

Al-Homoud AS, Degen WS (2006) Marine stone columns to Deb K, Basudhar PK, Chandra S (2010) Axi-symmetric analysis
prevent earthquake induced soil liquefaction. Geotech of geosynthetic-reinforced granular fill-soft soil system
Geol Eng 24:775–790. https://doi.org/10.1007/s10706- with group of stone columns. Geotech Geol Eng
005-2783-5.G=8 28:177–186. https://doi.org/10.1007/s10706-009-9291-y
Ali K, Shahu JT, Sharma KG (2010) Behaviour of reinforced Deb K, Samadhiya NK, Namdeo JB (2011) Laboratory model
stone columns in soft soils: an experimental study. In: studies on unreinforced and geogrid-reinforced sand bed
Indian geotechnical conference, 2010, GEOtrendz over stone column-improved soft clay. Geotext Geomembr
December 16–18, IGS Mumbai Chapter and IIT Bombay 2011(29):190–196. https://doi.org/10.1016/j.geotexmem.
Ambily AP, Gandhi SR (2004) Experimental and theoretical 2010.06.004
evaluation of stone column in soft clay. In: ICGGE Dennis JE, Schnabel RB (1996) Numerical methods for
Ambily AP, Gandhi SR (2007) Behavior of stone columns based unconstrained optimization and nonlinear equations.
on experimental and fem analysis. J Geotech Geoenviron SIAM 168–174
Eng 133:405–415. https://doi.org/10.1061/(ASCE)1090- Elshazly H, Elkasabgy M, Elleboudy A (2008a) Effect of inter-
0241(2007)133:4(405) column spacing on soil stresses due to vibro-installed stone
Andreou P, Papadopoulos V (2014) Factors affecting the set- columns: interesting findings. Geotech Geol Eng
tlement estimation of stone column reinforced soils. Geo- 26:225–236. https://doi.org/10.1007/s10706-007-9159-y
tech Geol Eng 32:1175–1185. https://doi.org/10.1007/ Elshazly HA, Hafez DH, Mossaad ME (2008b) Reliability of
s10706-014-9788-x conventional settlement evaluation for circular foundations
Balaam NP, Booker JR (1981) Analysis of rigid rafts supported on stone columns. Geotech Geol Eng 26:323–334. https://
by granular piles. Int J Numer Anal Methods Geomech doi.org/10.1007/s10706-007-9169-9
5:379–403 Etezad M, Hanna AM, Ayadat T (2015) Bearing capacity of a
Basack S, Indraratna B, Rujikiatkamjorn C, Siahaan F (2017) group of stone columns in soft soil. Int J Geomech ASCE.
Modeling the stone column behavior in soft ground with ISSN 1532-3641/04014043(15)
special emphasis on lateral deformation. J Geotech Fattah MY, Majeed QG (2009) Behaviour of encased floating
Geoenviron Eng 143(6):04017016 stone columns. Eng Technol J 27(7):2009
Black JA, Sivakumar V, Madhav MR, Hamill GA (2007) Fattah MY, Majeed QG (2012) Finite element analysis of geo-
Reinforced stone columns in weak deposits: laboratory grid encased stone columns. Geotech Geol Eng
model study. J Geotech Geoenviron Eng 133(9):1154– 30:713–726. https://doi.org/10.1007/s10706-011-9488-8
1161. https://doi.org/10.1061/(asce)1090-0241(2007)133: Frikha W, Bouassida M, Canou J (2013) Observed behaviour of
9(1154) laterally expanded stone column in soft soil. Geotech Geol
Borges JL, Domingues TS, Cardoso AS (2009) Embankments on Eng 31:739–752. https://doi.org/10.1007/s10706-013-
soft soil reinforced with stone columns: numerical analysis 9624-8
and proposal of a new design method. Geotech Geol Eng Goh ATC, Kulhawy FH, Chua CG (1998) Bayesian neural
27:667–679. https://doi.org/10.1007/s10706-009-9266-z network analysis of undrained side resistance of drilled
Bouassida M, De Buhan P, Dormieux L (1995) Bearing capacity shafts. J Geotech Geoenviron Eng 131(1):84–93. https://
of a foundation resting on a soil reinforced by a group of doi.org/10.1061/(asce)1090-0241
columns. Ghotechnique 45(1):25–34 Golait YS, Satyanarayana V, Raju SSV (2009) Concept of under
Box GEP, Tiao GC (1973) Bayesian inference in statistical reamed cemented stone columns for soft clay ground
analysis. Addison-Wesley, Reading improvement. IGC 2009, Guntur
Castro J (2014) An analytical solution for the settlement of stone Hassen G, Buhan PD, Abdelkrim M (2010) Finite element
columns beneath rigid footings. Acta Geotech. https://doi. implementation of a homogenized constitutive law for
org/10.1007/s11440-014-0358-4 stone column-reinforced foundation soils, with application
Castro J (2017) Groups of encased stone columns: influence of to the design of structures. Comput Geotech
column length and Arrangement. Geotext Geomembr 37(2010):40–49
45(2017):68–80 Haykin S (1998) Neural networks: a comprehensive foundation.
Chik Z, Aljanabi QA (2014) Intelligent prediction of settlement Prentice Hall, Englewood Cliffs
ratio for soft clay with stone columns using embankment IS 15284-Part 1 (2003) Design and construction for ground
improvement techniques. Neural Comput Appl 25:73–82. improvement-Guidelines. Part-1: Stone columns [CED 43:
https://doi.org/10.1007/s00521-013-1449-0 Soil and foundation engineering], ICS 93.020
Chik Z, Aljanabi QA, Kasa A, Taha MR (2014) Tenfold cross Kayri M (2016) Predictive abilities of bayesian regularization
validation artificial neural network modeling of the set- and Levenberg–Marquardt algorithms in artificial neural
tlement behavior of a stone column under a highway networks: a comparative empirical study on social data.
embankment. Arab J Geosci 7:4877–4887. https://doi.org/ Math Comput Appl 21:20. https://doi.org/10.3390/
10.1007/s12517-013-1128-6 mca21020020
Choobbasti AZ, Zahmatkesh A, Noorzad R (2011) Performance Kosti S, Vasovic D (2015) Prediction model for compressive
of stone columns in soft clay: numerical evaluation. Geo- strength of basic concrete mixture using artificial neural
tech Geol Eng 29:675–684. https://doi.org/10.1007/ networks. Neural Comput Appl 26:1005–1024. https://doi.
s10706-011-9409-x org/10.1007/s00521-014-1763-1
Das AK, Deb K (2016) Modeling of stone column-supported Lee JS, Pande GN (1998) Analysis of stone-column reinforced
embankment under axi-symmetric condition. Geotech Foundations. Int J Numer Anal Meth Geomech
Geol Eng. https://doi.org/10.1007/s10706-016-0136-1 22:1001–1020

123
Geotech Geol Eng (2018) 36:1845–1861 1861

MacKay DJC (1991) Bayesian model comparison and backprop Shivashankar R, Dheerendra Babu MR, Nayak S, Manjunath R
nets. Computation and Neural Systems California Institute (2010) Stone columns with vertical circumferential nails:
of Technology 139-14 Pasadena CA 91125 laboratory model study. Geotech Geol Eng 28:695–706.
Malarvizhi SN, Ilampurthi K (2007) Comparative study on the https://doi.org/10.1007/s10706-010-9329-1
behavior of encased stone column and conventional stone Shivashankar R, Babu MRD, Nayak S, Rajathkumar V (2011)
column. Soils Found Jpn Geotech Soc 47(5):873–885 Experimental studies on behaviour of stone columns in
Mohanty P, Samanta M (2015) Experimental and numerical layered soils. Geotech Geol Eng 29:749–757. https://doi.
studies on response of the stone column in layered soil. Int J org/10.1007/s10706-011-9414-0
Geosynth Ground Eng 1:27. https://doi.org/10.1007/ Six V, Mroueh H, Shahrour I, Bouassida M (2012) Numerical
s40891-015-0029-z analysis of elastoplastic behavior of stone column. Found
Murugesan S, Rajagopal K (2006) Geosynthetic-encased stone Geotech Geol Eng 30:813–825. https://doi.org/10.1007/
columns: numerical evaluation. Geotext Geomembr s10706-012-9500-y
24(2006):349–358. https://doi.org/10.1016/j.geotexmem. Tabchouche S, Mellas M, Bouassida M (2017) On settlement
2006.05.001 prediction of soft clay reinforced by a group of stone col-
Murugesan S, Rajagopal K (2008) Performance of encased umns. Innov Infrastruct Solut 2:1. https://doi.org/10.1007/
stone columns and design guidelines for construction on s41062-016-0049-0
soft clay soils. In: Proceedings of the 4th Asian regional Taha MR, Firoozi AA (2012) Estimating the clay cohesion by
conference on geosynthetics, June 17–20, Shanghai means of artificial intelligence technique. J Asian Sci Res
Nassaji F, Asakereh A, (2013) Effect of granular bed on beha- 2(11):651–657
viour of stone column improved ground. Int J Sci Eng Vekli M, Aytekin M, Ikizler SB, Calik U (2012) Experimental
Investig 2(23), December. ISSN: 2251-8843 and numerical investigation of slope stabilization by stone
Neal RM (1992) Connectionist learning of belief networks. Artif columns. Nat Hazards 64:797–820. https://doi.org/10.
Intell 56(1):71–113 1007/s11069-012-0272-8
Niroumand H, Kassim KA, Yah CS (2011) Soil improvement by Zhang L, Zhao M, Shi C, Zhao H (2013) Settlement calculation
reinforced stone columns based on experimental work. of composite foundation reinforced with stone columns. Int
EJGE, vol 16, Bund. L., pp 1477–1499 J Geomech 13(3):248–256. https://doi.org/10.1061/
Poorooshasb HB, Meyerhof G (1997) G (1997) Analysis of (ASCE)GM.1943-5622.0000212
behavior of stone columns and lime columns. Comput
Geotech 20(1):47–70

123

View publication stats

You might also like