Sie sind auf Seite 1von 12

INTERNATIONAL JOURNAL FOR NUMERICAL AND ANALYTICAL METHODS IN GEOMECHANICS

Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841


Published online 6 November 2010 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/nag.981

Estimation of compaction parameters of fine-grained soils in terms


of compaction energy using artificial neural networks

Osman Sivrikaya∗, † and Taner Y. Soycan


Nigde University, Civil Engineering Department, Division of Geotechnical Engineering, Nigde 51245, Turkey

SUMMARY
The determination of the compaction parameters such as optimum water content (wopt ) and maximum
dry unit weight (dmax ) requires great efforts by applying the compaction testing procedure which is also
time consuming and needs significant amount of work. Therefore, it seems more reasonable to use the
indirect methods for estimating the compaction parameters. In recent years, the artificial neural network
(ANN) modelling has gained an increasing interest and is also acquiring more popularity in geotechnical
engineering applications. This study deals with the estimation of the compaction parameters for fine-
grained soils based on compaction energy using ANN with the feed-forward back-propagation algorithm.
In this study, the data including the results of the consistency tests, standard and modified Proctor tests, are
collected from the literature and used in the analyses. The optimum structure of a network is determined
for each ANN models. The analyses showed that the ANN models give quite reliable estimations in
comparison with regression methods, thus they can be used as a reliable tool for the prediction of wopt
and dmax . Copyright 䉷 2010 John Wiley & Sons, Ltd.

Received 5 March 2010; Revised 24 August 2010; Accepted 25 August 2010

KEY WORDS: soil compaction; fine-grained soils; index properties; artificial neural network

1. INTRODUCTION

Compacted soils are used in many projects such as highways, railway subgrades, airfield pavements,
earth dams and landfill liners. In the field, fine-grained soils are usually compacted using tampers,
sheepsfoot rollers, rubber-tired rollers and other various equipments. To evaluate compaction in
the field, laboratory compaction parameters are required and determined. In the laboratory, soil
compaction is usually performed using the impact-Proctor compaction apparatus. Compaction
Proctor test is a well-known standard method of compaction and in this test a standard amount
of compactive effort is applied to produce soil density with which site values can be compared.
It is well known that compaction parameters of fine-grained soils are influenced by many factors
such as dry unit weight, water content, compactive effort, soil type and gradation. For a certain
compactive effort, a typical compaction curve that relates the water content (w) of the soil to
its dry unit weight (d ) is usually obtained. The most important point on the compaction curve
is the optimum compaction point in which two important parameters, maximum dry unit weight
(dmax ) and optimum water content (wopt ), are obtained, and they represent compaction behaviour
(Figure 1).

∗ Correspondence to: Osman Sivrikaya, Nigde University, Civil Engineering Department, Division of Geotechnical
Engineering, Nigde, 51245, Turkey.
† E-mail: osivrikaya@nigde.edu.tr, sivrikayaosman@hotmail.com

Copyright 䉷 2010 John Wiley & Sons, Ltd.


ESTIMATION OF COMPACTION PARAMETERS 1831

3
d (kN/m )

dmax
MP

SP

wopt w (%)

Figure 1. Typical compaction curves for different compaction energies.

Many attempts have been made to obtain the optimum water content and maximum dry unit
weight of compacted fine-grained soils. The correlation equations for fine-grained soils relate
the optimum water content and maximum dry unit weight to factors such as soil classification
descriptors, index properties and grain-size distribution [1–4]. The studies showed that wopt can
be more accurately estimated from the plastic limit (wp ) in comparison with the liquid limit and
plasticity index [1–3], and also dmax can be more accurately estimated from wopt in comparison
with index properties [3, 5].
In the recent year, Sivrikaya [3] and Sivrikaya et al. [4] made considerable contribution for
the estimation of the compaction parameters of fine-grained soils. Sivrikaya [3] developed the
correlations using his own standard Proctor (SP) data with index properties of fine-grained soils
as follows:

wopt = 0.92wp , R = 0.99, SE = 3.10% (1)

dmax = 21.84−0.27wopt , R = 0.97, SE = 0.45 kN/m3 (2)

where R is the correlation coefficient and SE is the standard error of regression equation.
Sivrikaya et al. [4] developed the correlations using the literature and his own modified Proctor
compaction data (Equations (3) and (4)), and also the literature and his own standard and modified
Proctor compaction data with index properties of fine-grained soils in terms of compaction energy
(Equations (5) and (6)) as follows:

wopt = 0.69wp , R = 0.99, SE = 2.32% (3)

dmax = 23.87−0.38wopt , R = 0.95, SE = 0.52 kN/m3 (4)


wopt = (1.99−0.165In E)wp (5)
dmax = (14.34−1.195In E)−(0.073In E −0.19)wopt (6)

Index tests are easily performed and are required for cohesive soils in all soil exploration
programs. Therefore, it is useful to estimate the engineering properties of soils by using other
soil parameters that can be easily acquired. In this context, correlations are important to estimate
the engineering properties of soils particularly for the projects with financial limitation, lack of
test equipment or limited time. Correlations are commonly used in the preliminary stage of any
project. However, many of these available in the literature are not clear enough to apply to field
data, which are not well defined, and thus their use generates confusions. Some uncertainties, like
whether the correlation has a statistical meaning, whose test results are used in the correlation
and to which type of soil the correlation is valid for, have considerable effects on the correlation
equations. Therefore, the correlation equations with compaction parameters should be used by
taking these uncertainties into consideration [3].

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
1832 O. SIVRIKAYA AND T. Y. SOYCAN

In regression analysis, a probabilistic model is employed and assumed that it is normally


distributed. However, all conventional regression methods have some limitations. It is also known
that the empirical methods are not applicable to the complex and highly nonlinear problems.
The artificial neural network (ANN) modelling is becoming more popular and has been used
commonly in engineering tasks [6]. The success has been achieved on geotechnical engineering
applications such as the site characterization [7], the soil classification [8], the engineering proper-
ties of clays [9–11], the bearing capacity of piles [12], the liquefaction potential of coarse-grained
soils [13], the compaction and permeability of soils [14, 15] since the ANN modelling has many
advantages containing requirement of less formal statistical training, ability to implicitly detect
complex nonlinear relationships between dependent and independent variables, ability to detect
all possible interactions between predictor variables and the availability of multiple training algo-
rithms. It is a preferable method among the other approaching methods due to having quick results
in the test phase in a short time.
The main objective of the current study is to develop ANN models to predict the maximum dry
unit weight and optimum water content for compacted fine-grained soils using the data compiled
by Sivrikaya [3] and Sivrikaya et al. [4], and to compare the developed ANN models with available
empirical models.

2. MODELLING WITH ARTIFICIAL NEURAL NETWORKS

2.1. Artificial neural networks


The neural networks is a powerful soft computational technique for linear and nonlinear approxi-
mations in many disciplines, inspired by biological cerebral activity called neuroscience [16–19].
ANNs have ability in performing a good amount of generalization from the patterns on which they
are trained. There are different types of neural networks. These networks differ in the architecture
and in the rules of learning. However, many papers on the application of neural networks in engi-
neering applications indicated that a multilayer feed-forward neural network model is the most
commonly used network for its efficient generalization capabilities [17, 20]. The typical multilayer
feed-forward neural networks used are shown in Figure 2. This type of neural network consists
of an input layer, one or more hidden layer(s) and an output layer. Layers consist of a number of
processing units called neurons (Figure 2). The strength of connections between neurons is repre-
sented by numerical values called weights. Each neuron has an activation value that is a function
of the sum of inputs received from other neurons through the weighted connections (Figure 3).
The input layer feeds the network with input from outside. The hidden layers link the input
layer to the output layer, extract and remember useful features from the input data to estimate the

X1 Y1
.
.
.
X2 Y2

X3 Y3
.
. .
.
. .
.
. .
Xn Yn

Input Layer Hidden Layer Output Layer

Figure 2. A multilayer ANN configurations.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
ESTIMATION OF COMPACTION PARAMETERS 1833

+ Bias + Bias

X1
Y1
X2
Compare to observed value
(Control the error)
FEEDBACK

Single Neuron
X1 W1
1-e
1+e
ΣXiWi+q Output of
Neuron
X2 W2 tan-sigmoid

Figure 3. The architecture of feed-forward ANN with two inputs and the functional
representation of single neuron.

output of the network. Trial and error should be carried out in order to determine an adequate
number of hidden layers and the number of neurons in each hidden layer [16, 17, 20, 21].
All simulations were carried out in the MATLAB environment in this study. In MATLAB, a
network is trained by using a special learning function and learning rule. In ANNs analyses, some
functions called learning functions are used for initialization, training, adaptation and performance
function. During the training process, a network is continuously updated by a training function
that repeatedly applies the input variables to a network until a desired error criterion is reached.
Adapt functions are employed for the simulation of a network, while the network is updated for
each time step of the input vector before continuing the simulation to the next input. Performance
functions are used to grade the network results [19]. In this study, gradient descent with momentum
and adaptive learning rate (traingdx), gradient descent with momentum weight and bias learning
function (learngdm) and mean-squared error (MSE) were used for training function, adapt function
and performance function, respectively. In the learning stage, network initially starts by randomly
assigning the adjustable weights and threshold values for each connection between the neurons
in accordance with the selected ANNs model. After the weighted inputs are summed and the
threshold values are added, they are passed through a differentiable nonlinear function defined
as a transfer function. The process of evolution of the weights and bias continues until neural
network outputs approach the target values or the iteration threshold is met. An ANN model can be
differently composed in terms of architecture, learning rule and self-organization. The most widely
used ANNs are the feed-forward, multilayer perceptrons trained by back-propagation (FFBP)
algorithms based on the gradient descent method. This algorithm can provide approximating to
any continuous function from one finite-dimensional space to another for any desired degree of
accuracy.

2.2. Back propagation


The back propagation is a learning algorithm most widely used in the ANNs. The main characteristic
of this paradigm is that it works by sending inputs forward and then propagating errors calculated
using a certain error criteria backwards. In this algorithm, the learning procedure based on the
supervised rule is continued by adjusting the weights until the minimal error is obtained. A
feed-forward network with back-propagation configuration is demonstrated in Figure 3. The back-
propagation process of error is performed by two steps; the first step is a feed-forward phase in
which the output from any node is calculated by propagating the input value given from the input
nodes. The second step is a backward phase in which connection weight values are corrected by

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
1834 O. SIVRIKAYA AND T. Y. SOYCAN

using error criteria. An output value for any neuron in the hidden layer is computed as follows:
 n 

Yj = f X i Wi, j +q j , i = 1, . . . , n, j = 1, . . . , m (7)
i=1

where Y j is the output value of any neuron in the hidden layer, X i = {X 1 , X 2 , . . . , X n } is the input
vector, Wij is the weights from the ith neuron to jth neuron, q j is a threshold value (bias value), f
is the transfer function or activation function and n and m are the numbers of neurons in the input
and hidden layers, respectively. The output value of any output neuron is computed as follows:
 
m
Yt = f Yjt Wij +qt , t = 1, . . . , k (8)
j=1

where k is the number of neurons in the output layer.


The networks are capable of mapping the nonlinear input–output relationship by means of
the nonlinearity of the transfer functions. There are different transfer functions to establish a
relationship between the input and output variables at each neuron layer. In the literature, the most
commonly used transfer function is the tangent sigmoid given below:
  
1−e X w+q
f (x, w) =    tan-sigmoid (9)
1+e X w+q

A network propagates the inputs or outputs mapping by using connection weights as far as the
lowest possible error can be obtained by using an error criterion. The error is the difference
between the calculated output and the target value. MSE for all input patterns is computed from
the following equation by altering the connection weights:
2
1 N  k
MSE = (Tst −Yst ) (10)
N t=1 s=1
where N is the number of data set pattern, Tst is the desired data value for the sth pattern and
Yst is the calculated output for the sth pattern. The back-propagation algorithm aims at iteratively
minimizing error. According to the gradient descent method, the back-propagation paradigm fixes
the weights by calculating the gradient (t ) for each neuron on the output layer with the following
equations:
t = Yt (1−Yt )(Tt −Yt ) (11)
The error gradient t is then recursively determined for the hidden layers by computing the weighted
sum of the errors at the previous layer

k
 j = Y j (1−Y j ) (t wjt ) (12)
t=1

The change in connection weights is updated by using error gradients (Equation (12)) as shown
below:
*E
wij (r ) = − =  j wi (13)
*wij
wji (r +1) = wji (r )+wji (r ) (14)
The weight change after the tth data presentation is
wji (r ) =  j xi +wji (r −1) (15)
where  is the learning rate,  is the momentum rate and r is the epoc (iteration) number.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
ESTIMATION OF COMPACTION PARAMETERS 1835

wp or
wopt wopt or
γdmax
E

Figure 4. The architecture of feed-forward ANNs with two inputs.

Table I. ANN models used in the determination of compaction parameters.


Data number
Transfer Epoc
Model Inputs Training Testing Structure function number MSE Output

I E, wp 95 95 2-3-1 Tansig-Tansig 1630 0.95×10−5 wopt


II E, wopt 108 107 2-1-1 Tansig-Tansig 1820 0.48×10−5 dmax

3. ANN MODELS FOR ESTIMATING COMPACTION PARAMETERS

The potential of ANNs to estimate the compaction parameters (wopt , dmax ) is investigated by
developing different ANN models. It is previously pointed out that the main factor affecting
compaction parameters is the compaction energy and plastic limit. This paper has, therefore,
taken into consideration two parameters such as compaction energy (E) and plastic limit (wp ) for
estimation of wopt , and two parameters such as compaction energy (E) and optimum water content
(wopt ) for estimation of dmax as input parameters to the ANN models.
To train and test the ANN models, the data were collected from various sources in the literature
and Sivrikaya’s own data [3, 4]. The details about the data used in the analyses can be obtained
from Sivrikaya [3] and Sivrikaya et al. [4]. In the ANN analyses of wopt , 130 data points from
SP compaction tests and 60 data points from modified Proctor (MP) compaction tests were used.
Half of the data was randomly selected for training and the remaining of the data sets was used
for testing. Considering the ANN analyses of dmax , 152 data points from SP compaction tests and
63 data points from MP compaction tests were used. Hundred and eight of them were used for
training the ANN models. The remaining data set was used for testing. Data sets used for testing
were not included in the training stage.
The ANN toolbox of MATLAB was used to perform the necessary computation [19]. In order
to develop the most adequate ANN architecture, the number of neurons in the hidden layer and a
number of multilayer networks with different transfer functions were tried to predict compaction
parameters. Therefore, they were varied until the convergence was achieved in the MSE. In this
study, the feed-forward neural networks trained by back-propagation algorithms were employed,
and the learning rate and momentum rate were adaptive. In Figure 4, the architecture of the Models
formed by using feed-forward ANNs with two inputs is shown. The details regarding the network
models developed in this study are given in Table I.

4. THE ANN MODEL ANALYSES

As stated in the previous studies, wp and E, wopt and E have very good correlation with wopt and
dmax , respectively [2, 3]. Therefore, to estimate wopt based on compaction energy, wp and E are
taken as input data, and also E and wopt are taken as input data for estimating dmax based on
compaction energy. In this study, two types of ANN models are examined and developed (Table I).
In order to develop the different ANN models, the input parameters were also individually excluded
from the input parameters. The developed ANN models were tested by data sets from the literature
and current study, which were not employed in the training stage. In order to ensure that ANN

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
1836 O. SIVRIKAYA AND T. Y. SOYCAN

40

35 wopt(est) = 0.98wopt(meas) R = 0.94 SE = ±2.7 %

30

Estimated wopt %
25

20

15

10

0
0 5 10 15 20 25 30 35 40
Measured wopt %

Figure 5. The comparison of measured wopt values with estimated wopt values by ANN for training data.

40
wopt(est) = 0.97wopt(meas) R = 0.88 SE = ±3.9 %
35

30
Estimated wopt %

25

20

15

10

0
0 5 10 15 20 25 30 35 40
Measured wopt %

Figure 6. The comparison of measured wopt values with estimated wopt values by
Equation (5) for training data.

approach is preferable to regression method and to evaluate how accurate the results of the
developed ANN models and the previous empirical correlations are, the correlation coefficient (R)
and the standard error (SE) of correlations obtained from ANN and regression methods are used
as statistical verification tools. It is known that ANN approach is preferable as long as R value
reaches 1 or close to 1; and SE is 0 or close to 0.
wopt and dmax values estimated from ANN method and the correlative equations (Equations 5
and 6) were graphically compared with the measured wopt and dmax values in Figures 5–12. As can
be seen, the ANN models were found to be able to learn the relationship between the compaction
parameters and E, wopt and wp for both training and testing stages.
To estimate wopt based on compaction energy, the analyses are made using wp and E, and
thus Model-I is developed. The results obtained from ANN analyses, the measured values
and the results obtained from Equation (5) are shown during training and test stages in
Figures 5– 8. As far as these figures are concerned, Model-I, which has the same R = 0.93 and the
lowest SE = 2.8 values, appears to be the best model in comparison with Equation (5) of R = 0.93

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
ESTIMATION OF COMPACTION PARAMETERS 1837

40

35 wopt(est) = 0.98wopt(meas) R = 0.93 SE = ±2.8 %

30

Estimated wopt %
25

20

15

10

0
0 5 10 15 20 25 30 35 40
Measured wopt %

Figure 7. The comparison of measured wopt values with estimated wopt values by ANN for testing data.

40

35 wopt(est) = 0.98wopt(meas) R = 0.93 SE = ±3.1 %

30
Estimated wopt %

25

20 g

15

10

0
0 5 10 15 20 25 30 35 40
Measured wopt %

Figure 8. The comparison of measured wopt values with estimated wopt


values by Equation (5) for testing data.

and SE = 3.1 for the testing data. In addition, Model-I with the highest R = 0.97 and the lowest
SE = 2.7 values appears to be the best model in comparison with Equation (5) of R = 0.88 and
SE = 3.9 for the training data.
To estimate dmax based on compaction energy, the analyses are made using wp and E, and
thus Model-II is developed. The results obtained from ANN analyses, the measured values and the
results obtained from Equation (6) are shown during training and test stages in Figures 9–12. As far
as these figures are concerned, Model-II, which has the highest R = 0.98 and the lowest SE = 0.5
values, appears to be the best model in comparison with Equation (6) of R = 0.97 and SE =
0.6 for the testing data. For the training data, Model-II with the same R = 0.97 and the lowest
SE = 0.5 values appears to be the best model in comparison with Equation (5) of R = 0.97 and
SE = 0.6.
Table II gives the statistical performance of all the models. It appears that there is a relatively
good agreement between the ANN and correlative studies predictions and the actual data. This can

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
1838 O. SIVRIKAYA AND T. Y. SOYCAN

22

dmax(est) = dmax(meas) R = 0.97 SE = ±0.5 kN/m3


20

18

16

14

12

10
10 12 14 16 18 20 22

Figure 9. The comparison of measured dmax values with estimated dmax


values by ANN for training data.

22
dmax(est) = dmax(meas) R = 0.97 SE = ±0.6 kN/m3

20

18

16

14

12

10
10 12 14 16 18 20 22

Figure 10. The comparison of measured dmax values with estimated dmax values by
Equation (6) for training data.

be interpreted from the R values ranging from 0.93 to 0.98 in Table II. SE values of the models
reflect the overall error performances of the models.
The behaviour of models is also evaluated in terms of underestimating or overestimating. This
evaluation is considered in terms of finding out wopt(est) = a ∗wopt(meas) and dmax(est) = b∗ dmax(meas) .
The coefficients a and b represent underestimating if a or b<1 or overestimating if a or b>1.
The a and b coefficient values for ANN and correlations are given in Table III. As can be seen in
Table III, the wopt values obtained by ANN and regression are found to be sightly underestimated
for both training and testing data (a = 0.98). However, the b coefficient obtained for ANN and
regression is found to be 1 for training data and 0.99 for testing data.
When the results in all figures and tables are evaluated, it can be concluded that the ANN models
can be used for the prediction of compaction parameters of compacted fine-grained soils based on
compactive effort. Consequently, the ANN models in both training and test stages give slightly
better results than the regression method (Equations (5) and (6)) in terms of R and SE (Table II).

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
ESTIMATION OF COMPACTION PARAMETERS 1839

22
dmax(est) = 0.99 dmax(meas) R = 0.98 SE = ±0.5 kN/m3

20

18

16

14

12

10
10 12 14 16 18 20 22

Figure 11. The comparison of measured dmax values with estimated


dmax values by ANN for testing data.

22
3
dmax(est) = 0.99 dmax(meas) R = 0.97 SE = ±0.6 kN/m

20

18

16

14

12

10
10 12 14 16 18 20 22

Figure 12. The comparison of measured dmax values with estimated dmax
values by Equation (6) for testing data.

In other words, the ANN approach has given more accurate and precise results in comparison with
regression methods.

5. CONCLUSION

The determination of compaction characteristics (wopt and dmax ) is both time consuming and
requires great effort. Therefore, it is useful and sometimes inevitable to make use of indirect
methods. In this context, this study deals with developing the reliable ANN models between
characteristic compaction parameters and index properties of fine-grained soils.
Several ANN architectures were compiled to establish an interrelationship between the soil index
properties and compaction parameters. The back-propagation algorithm was used in the analyses.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
1840 O. SIVRIKAYA AND T. Y. SOYCAN

Table II. Statistical performance of results obtained from estimated by ANN and correlations.
R SE
Output Method Training data Testing data Training data Testing data
wopt Model-I 0.94 0.93 2.7% 2.8%
Equation (5) 0.88 0.93 3.9% 3.1%
dmax Model-II 0.97 0.98 0.5 kN/m3 0.5 kN/m3
Equation (6) 0.97 0.97 0.6 kN/m3 0.6 kN/m3

Table III. The regression coefficients obtained for wopt and dmax .

a b
Output Method Training data Testing data Training data Testing data
wopt Model-I 0.98 0.98
Equation (5) 0.97 0.98
dmax Model-II 1 0.99
Equation (6) 1 0.99

The two ANN models, which have different input parameters, were trained by data collected from
the literature and then they were tested. Performances of the models were examined in terms
of some statistical criteria (R and SE). The results of the current and the previous empirical
studies indicate that wopt has the best correlation with wp and E, and however dmax has better
correlation with wopt and E. The results by ANN models are also compared with the previous
empirical studies. It is found that the ANN models and the correlation equations give quite reliable
predictions for both wopt and dmax . It is concluded from this study that the results of ANN models
are very encouraging for the cases trained and tested and the use of ANN is a more reliable method
for the prediction of compaction parameters in comparison with the regression methods. Thus,
the proposed ANN models could be used to estimate the compaction parameters based on the
compaction energy of fine-grained soils in preliminary design stage in the case of time limitations.

REFERENCES
1. Gurtug Y, Sridharan A. Compaction behaviour and prediction of its characteristics of fine grained soils with
particular reference to compaction energy. Soil and Foundation 2004; 44(5):27–36.
2. Sridharan A, Nagaraj HB. Plastic limit and compaction characteristics of fine-grained soils. Ground Improvement
2005; 9(1):17–22.
3. Sivrikaya O. Models of compacted fine-grained soils used as mineral liner for solid waste. Environmental Geology
2008; 53(7):1585–1595.
4. Sivrikaya O, Togrol E, Kayadelen C. Estimating compaction behaviour of fine-grained soils based on compaction
energy. Canadian Geotechnical Journal 2008; 45:877–887.
5. Nagaraj HB. Prediction of engineering properties of fine-grained soils from their index properties. Ph.D. Thesis,
Faculty of Engineering, Indian Institute of Science, Bangalore, India, 2000.
6. Shahin MA, Jaksa MB, Maier HR. Artificial neural network applications in geotechnical engineering, Australian.
Geomechanics 2001; 36(1):49–62.
7. Zhou Y, Wu X. Use of neural networks in the analysis and interpretation of site investigation data. Computer
and Geotechnics 1994; 16:105–122.
8. Cal V. Neural network-alternative to pile driving formulas. Advances in Engineering Software 1995; 22(2):95–97.
9. Kurup PU, Dudani NK. Neural networks for profiling stress history of clays from PCPT data. ASCE Journal of
Geotechnical and Geoenvironmental Engineering 2002; 128(7):569–579.
10. Kayadelen C. Estimation of effective stress parameter of unsaturated soils by using artificial neural networks.
International Journal for Numerical and Analytical Methods in Geomechanics 2007; 32:1087–1106.
11. Sivrikaya O. Comparison of artificial neural networks models with correlative works on undrained shear strength.
Eurasian Soil Science 2009; 42(13):1487–1496.
12. Lee IM, Lee JH. Prediction of pile bearing capacity using artificial neural networks. Computers and Geotechnics
1996; 18(3):189–200.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag
ESTIMATION OF COMPACTION PARAMETERS 1841

13. Ural D, Saka H. Liquefaction prediction by neural networks. Electronic Journal of Geotechnical Engineering
1998; 3:1089–3032.
14. Basheer IA. Emprical modeling of the compaction curve of cohesive soils. Canadian Geotechnical Journal 2001;
38:29–45.
15. Sinha SK, Wang MC. Artificial neural network prediction models for soil compaction and permeability.
Geotechnical and Geological Engineering 2008; 26:47–64.
16. Rafiq MY, Bugmann G, Easterbrook DJ. Neural network design for engineering application. Computers and
Structures 2001; 79(17):1541–1552.
17. Kartam N, Flood I, Garrett JH. Artificial Neural Networks for Civil Engineers: Fundamentals and Applications.
ASCE: New York, 1997.
18. Flood I, Kartam N. Neural network in civil engineering. I: Principles and understandings. ASCE Journal of
Computing in Civil Engineering 1994; 8(2):131–148.
19. Flood I, Kartam N. Neural network in civil engineering. II: Systems and applications. ASCE Journal of Computing
in Civil Engineering 1994; 8(2):149–162.
20. Sanad A, Saka MP. Prediction of ultimate shear strength of reinforced-concrete deep beams using neural networks.
ASCE Journal of Structural Engineering 2001; 27(7):818–828.
21. Demuth H, Beale M. Neural Network Toolbox for Use with MATLAB. The MathWorks Inc.: Natick, MA, U.S.A.,
2001; 840.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2011; 35:1830–1841
DOI: 10.1002/nag

Das könnte Ihnen auch gefallen