Available online at www.sciencedirect.

com
Construction
and Building

Construction and Building Materials 22 (2008) 330–342

MATERIALS
www.elsevier.com/locate/conbuildmat

Dynamic soil–structure interaction analysis of buildings by

neural networks
a,*
Murat Pala , Naci Caglar b, Muzaffer Elmas b, Abdulkadir Cevik c, Mehmet Saribiyik d

a
Technical Programs Department, Kilis Vocational School, University of Gaziantep, 79000 Kilis, Turkey
b
Civil Engineering Department, University of Sakarya, Sakarya, Turkey
c
Civil Engineering Department, University of Gaziantep, Gaziantep, Turkey
d
Technical Education Faculty, Sakarya University, Sakarya, Turkey

Received 20 October 2005; received in revised form 24 March 2006; accepted 30 August 2006
Available online 20 October 2006

Abstract

In this study, neural networks (NN), which have been widely applied to engineering problems, have been applied to soil–structure
interaction (SSI) problems. In the modelling of the NN, the back propagation algorithm has been used. The training and testing sets
have been developed by using finite element (SAP2000) package program outcomes. Various analyses have been performed for different
height/width ratios of buildings in terms of different soil properties, local site depths, and different shear wave velocities. These effects to
the periods, top stories acceleration and displacement have been presented. It is observed that the modelling and analysis stages are min-
imized, the solution time is quite fast and the NN has shown a satisfactory performance for the solution of SSI problems. The NN results
are compared with SAP2000 outcomes and found to be encouraging.
Ó 2006 Elsevier Ltd. All rights reserved.

Keywords: Neural networks; Finite element method; Dynamic soil structure interaction; Direct method; Boundary conditions

1. Introduction Various studies and contributions have been appeared in

the literature regarding the effects of SSI on the dynamic
Soil–structure interaction (SSI) is an important issue, seismic response of buildings [4]. It was suggested that
especially for stiff and massive structures constructed on the dynamic SSI effects could be neglected for regular flex-
the relatively soft ground, which may alter the dynamic ible buildings on rock or very stiff soil, but they should be
characteristics of the structural response significantly. significantly considered for structures embedded in rela-
Thus, the interaction effects should be accounted in the tively soft soil. Throughout the last three decades, despite
dynamic analysis of all structures, particularly in severe soil comprehensive research has been carried out on analytical
conditions [1,2]. The SSI system has two characteristic dif- or semi-analytical methods and various analytical formula-
ferences from the general structural dynamic system which tions developed to solve complex practical problems SSI
are the unbounded nature and the non-linear characteris- analysis, the effect of non-linear behaviour of the support-
tics of the soil medium. The radiation of energy towards ing soil on the seismic response of structures has not been
infinity, leading to so-called radiation damping, is the most fully covered [5]. The application of such methods is very
prominent characteristic in an unbounded soil, which is limited since solutions are only available for systems with
never experienced in a bounded medium [3]. simple geometry.
The capability of the neural networks (NN) to describe
processes on the bases of available training data has
become subject of several contributions. It appears to offer
*
Corresponding author. Tel./fax: +90 348 814 26 68 69. a means of dealing with many multi-variety problems for
E-mail address: [email protected] (M. Pala). which an exact analytical model does not exist or at least

0950-0618/$ - see front matter Ó 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conbuildmat.2006.08.015
 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342 331

is very difficult and time consuming to develop. In recent difficulty in modelling the soil region arises from the prop-
years, the NN approach has been effectively applied in agating wave characteristics in the soil medium. The main
many engineering applications [6–9] and NN seems to be goal of many related engineering studies is to develop SSI
very promising. models, which are reliable and easy to implement. There
In this study, NN has been applied to SSI problems. The are valuable works in literature, which is focused on to pro-
training and testing sets has been developed using finite ele- pose reliable mathematical models for soil region of the SSI
ment analysis. Various analyses have been performed for problems [12].
different height/width ratios in terms of different soil prop- The main difficult task in SSI analyses is the representa-
erties, local site depths, and number of stories. tion of energy dissipated at the artificial boundaries. In
many cases, the soil region shows a complex and non-linear
2. Methodology of SSI analysis behaviour, which can be discretized using the finite element
method (FEM), the boundary element method (BEM) or
There are two main methods dealing with SSI analysis: hybrid models. Various advantages and disadvantages
Direct Method, and Substructure Method. In the direct valid for these models. For example, large-scale mesh is
method, the response of the soil and structure is deter- required in finite element models to account for the sur-
mined simultaneously by analysing the idealized soil–struc- rounding soil medium which is bounded by the far-field
ture system in a single step [10]. The soil containing the that is represented by artificial boundaries. In numerical
structure is modelled up to the artificial boundary modelling of wave propagation, problems arise as artificial
(Fig. 1). In some conditions where it is impossible to cover boundaries introduce artificial reflections that contaminate
the unbounded soil domain with finite elements with the solution. To compensate this error, special boundary
bounded dimensions the Substructure Method is used conditions are used, which might be referred to as radiation
instead. In the Substructure Method the semi-infinite damping. These special boundary conditions such as trans-
half-space is represented by approximate boundary condi- mitting, non-reflecting and silent boundaries absorb the
tions. Lysmer and Kuhlemeyer [11] presented a special fre- wave energy. The non-reflecting viscous boundaries have
quency independent viscous dashpot boundary on the been widely used for various dynamic SSI problems
interface nodes in all directions. This method uses finite ele- [11,13]. These boundaries must be located far away from
ment modelling of the core region thus provides a suitable the structure or source of dynamic load as they are only
way to solve complex structures such as dams or nuclear competent to transmit plane and cylindrical waves.
reactors. On the other hand a large number of degree-of- In substructure method [14] the SSI problem ids divided
freedoms arises due to the discretization of the soil region into a sets of simpler problems, which are solved indepen-
as it accounts for the non-linearity of near-structure dently, and the results are then superposed to obtain the
domain. Furthermore, this method does not consider the response of the structure (Fig. 2) which leads this method
infinity of the far field. to be simple and computationally more efficient. The sub-
In SSI problems, the ability to predict the coupled structure approach is most often implemented in the fre-
behaviour of the soil and structure is of vital importance. quency domain to account for the frequency dependence
Thus, combined soil and structure models are required of the foundation impedance functions. The substructure
for these problems. However soil models involve compli- method has a significant advantage where changes due to
cated analysis due to their unbounded nature, structure the properties of the structure and the seismic environment
models are well-founded in the literature. Hence the main have not to be repeated once the scattering and impedance
problems have been solved. The basic step in the substruc-
ture approach is to determine the force–displacement char-
acteristics of the soil. This relationship may be in the form
Superstructure

Artifical
Boundary

Fictitious Dashpot Boundary

Fig. 1. Direct method configuration. Fig. 2. Substructure method configuration.

332 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342

of an impedance (stiffness) function, or, inversely, a compli- 21

Training set (One hidden layer)
ance (flexibility) function [10].
Testing set (One hidden layer)
19
Training set (Two hidden layer)
3. Neural network (NN) 17
Testing set (Two hidden layer)

MAPE (%)
NN is a computational tool, which attempts to simulate 15

the architecture and internal operational features of the 13

human brain and nervous system. NN architectures are
11
formed by three or more layers, which includes an input
layer, an output layer and a number of hidden layers in 9
which neurons are connected to each other with modifiable
7
weighted interconnections (Fig. 3). Each neuron has an 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
associated transfer function, which describes how the Number of hidden layer unit
weighted sum of its inputs is converted to the results into
Fig. 5. Selection of appropriate number of hidden layers and number of
an output value. Each hidden or output neuron receives neurons in hidden layers.
a number of weighted input signals from each of the units
of the preceding layer and generates only one output value
(Fig. 4). This NN architecture is commonly referred to as a
fully interconnected feedforward multi-layer perceptron. In
addition, there is also a bias, which is only connected to Bias
Bias

neurons in the hidden and output layers with modifiable

weighted connections. The number of neurons in each layer
may vary depending on the problem.
Period

H/B TSD
x1 x2 xn
SLT DGL

... Input Layer Vs RTSD

ATS

Input Layer

Output
... Hidden Layer

Hidden Layer
Fig. 6. The architecture of the NN model.

... Output Layer

Table 1
out1 out2 outm Soil and structure parameters
Soil shear wave Soil Structure
Fig. 3. A typical MLP neural network. velocity,
Elastic modulus, Poisson Elastic modulus, Poisson
Vs (m/s) E (kN/m2) ratio, m E (kN/m2) ratio, m
300 47,706.4 0.3 21,000,000 0.2
x1 w1
400 190,825.6 0.3
x2 w2 500 429,357.8 0.3
Out 600 763,302.7 0.3
700 1,192,660.5 0.3
xi
wi
Σw ij xi + b 800 1,717,431.2 0.3
900 2,337,614.6 0.3
wn 1000 3,053,211.0 0.3
Activation 1500 3,864,220.2 0.3
xn
Function

Fig. 4. The structure of an artificial neuron. Fixed – 0.3

 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342 333

In the forward phase, the weighted sum of input compo-

nents is calculated as
X
n
netj ¼ wij xi þ biasj ð1Þ
i¼1

where netj is the weighted sum of the jth neuron for the in-
put received from the preceding layer with n neurons, wij is
the weight between the jth neuron and the ith neuron in the
preceding layer, xi is the output of the ith neuron in the pre-
ceding layer. The output of thejth neuron outj is calculated
with a sigmoid function as follows:

1
outj ¼ f ðnetj Þ ¼ ð2Þ
1 þ eðnetj Þ

The training of the network is achieved by adjusting the

weights and is carried out through a large number of train-
Fig. 7. El-Centro earthquake records, 1940 (NS component,
PGA = 0.32 g). ing sets and training cycles. The goal of the training proce-
dure is to find the optimal set of weights, which would
The most widely used training algorithm for multi- produce the right output for any input in the ideal case.
layered feedforward networks is perhaps the back-propa- Training the weights of the network is iteratively adjusted
gation (BP) algorithm. The BP algorithm basically to capture the relationship between the input and output
involves two phases. One is the forward phase where patterns.
the activations are propagated from the input to the out- As the BP training algorithm with gradient descent and
put layer. The second is the backward phase where the gradient descent with momentum are slow, several adaptive
error between the observed actual value and the desired training algorithms for NN have recently been discovered
nominal value in the output layer is propagated back- such as conjugate gradient algorithm (CG) and scaled con-
wards in order to modify the weights and bias values. jugate gradient algorithm (SCG). In this study, SCG is
Before training a feed work network, the inputs and the used as optimization algorithm, which is all set to standard
outputs of training and testing sets must be initialized. values suggested in Moller [15].

H
T.S.L.

1400 m

Fig. 8. Finite element mesh of the SSI model.

Table 2
Performance of training set and testing set
Training set Testing set Validation set
2 2
MAPE RMS R MAPE RMS R MAPE RMS R2
Period 0.8809 0.0038 0.9997 1.7710 0.0079 0.9996 3.8578 0.0267 0.9975
TSD 10.4828 0.0260 0.9842 12.3319 0.0155 0.9915 12.1495 0.0256 0.9820
DGL 15.3210 0.0269 0.9897 20.7713 0.5215 0.9742 36.6504 0.8023 0.9514
RTSD 5.6549 0.0351 0.9971 13.0546 0.0215 0.9854 16.8529 0.0373 0.9551
ATS 9.8122 0.0351 0.9836 13.6626 0.0408 0.9618 12.3860 0.0429 0.9740
334 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342

Performance of training set Performance of testing set

1.5
1.5

1.2
1.2
Target value

Target value
0.9
0.9

0.6 0.6

0.3 0.3

0 0
0 0.3 0.6 0.9 1.2 1.5 0 0.3 0.6 0.9 1.2 1.5
NN NN
Fig. 9a. The performance of training set for period of soil and structure. Fig. 9b. The performance of testing set for period of soil and structure.

Performance of training set Performance of testing set

100 60

50
80
Target value

40
Target value

60
30

40
20

20 10

0 0
0 20 40 60
0 20 40 60 80 100
NN
NN
Fig. 10b. The performance of testing set for displacement of top storey.
Fig. 10a. The performance of training set for displacement of top storey.

Performance of training set Performance of testing set

25 12

10
20

8
Target value

Target value

15

6
10
4

5
2

0 0
0 5 10 15 20 25 0 5 10 15
NN NN
Fig. 11a. The performance of training set for displacement of ground level. Fig. 11b. The performance of testing set for displacement of ground level.
 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342 335

Performance of training set Performance of testing set

100 60

50
80

40
Target value

Target value
60

30

40
20

20
10

0 0
0 20 40 60 80 100 0 10 20 30 40 50 60

NN NN
Fig. 12a. The performance of training set for relative displacement of top Fig. 12b. The performance of testing set for relative displacement of top
storey–ground level. storey–ground level.

Performance of training set Performance of testing set

100 80

80
60
Target value

Target value

60

40
40

20
20

0 0
0 20 40 60 80 100 0 20 40 60 80
NN NN
Fig. 13a. The performance of training set for acceleration of top storey. Fig. 13b. The performance of testing set for acceleration of top storey.

The training of the network is accomplished by adjust- unit to the jth output, in the batched mode variant the des-
ing the weights and is carried out through a large number cent is based on the gradient
of training sets and training iterations. The goal of the  
dE
learning procedure is to find the optimal set of weights, rE ð3Þ
dwij
which in the ideal case would produce the right output
for any input. The output of the network is compared with for total training set
a desired response to produce an error. The performance of dE
the MLP is measured in terms of a desired signal and the Dwij ðnÞ ¼ e  þ a  Dwij ðn  1Þ ð4Þ
dwij
criterion for convergence.
In this study, a computer program has been developed The gradient gives the direction of error E, The parameters
and performed under MATLAB [16,17]. The back-propa- e and a are the learning rate and momentum, respectively.
gation learning algorithm has been used in feed-forward The learning algorithm used in the study is scaled conju-
with one hidden-layer, back propagation algorithm gate gradients algorithm (SCG), developed by Moller
(BPA), as one of the most famous training algorithms for [15], and was designed to avoid the time consuming line
the multi-layer perceptron (MLP), is a gradient descent search, In the conjugate gradient algorithm (CGA), a
technique to minimize the error E for a particular training search is performed along conjugate directions, which pro-
pattern, For adjusting the weight (wij) from the ith input duces generally faster convergence than steepest descent
336 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342

directions [18], A search is made along the conjugate gradi- derivative approximation and parameter for regulating
ent direction to determine the step size, which will minimize the indefiniteness of the Hessian. In order to get a good
the performance function along that line, A line search is quadratic approximation of E, a mechanism to raise
performed to determine the optimum distance to move and lower kn is needed when the Hessian is positive defi-
along the current search direction, Then the next search nite. Details of the SCGA can be found in Moller’s article
direction is determined in such a way that it is conjugated [15].
to previous search direction, The general procedure for In this study, the number of hidden layer and number of
determining the new search direction is to combine the neurons in hidden layers are determined using trial and
new steepest descent direction with the previous search error technique. Some of the architectures with different
direction, An important feature of the CGA is that the number of hidden layer and number of neurons in hidden
minimization performed in one step is not partially undone layers studied here and the relationship of their percentage
by the next, as it is the case with gradient descent methods, error values are illustrated in Fig. 5. One hidden layer with
A significant drawback of CGA is the requirement of a line 11 neurons was used in the architecture because of its min-
search, which is computationally time consuming, Moller imum percentage error values for training, testing sets.
[15] introduced the scaled conjugate gradient algorithm Some of the tried architecture with different number of
(SCG) as a way of avoiding the complicated line search neurons in hidden layer and their percentage error values
procedure of conventional CGA, According to the SCGA, relationship are illustrated in Fig. 5.
the Hessian matrix is approximated by Thus, NN model developed in this research has 3 neu-
rons (variables) in the input layer, 11 neurons (variables)
E0 ðwn þ rn pn Þ  E0 ðwn Þ
Hn ¼ þ kn p n ð5Þ in the hidden layer and five neurons in the output layer
rn
as illustrated in Fig. 6.
where E 0 and Hn are the first and second derivative infor- In this study, the activation function is sigmoid func-
mation of global error function E(wn), The other terms tion, and number of epochs is 8000, by using a step size
wn, pn, rn and kn represent the weights, search direction, scaling mechanism, SCG avoids a time consuming line
parameter controlling the change in weight for second search per learning iteration.

1 24
NN NN
0.8 20
FEM FEM
Period (sec)

16
TSD (cm)

0.6
12
0.4
8
0.2
4

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
a Samples b Samples

18
8 NN NN
15
FEM FEM
RTSD (cm)

6 12
DGL (cm)

9
4
6
2
3

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
c Samples d Samples

70
NN
60
Accel.(cm/sec 2)

FEM
50
40

30
20
10
0
1 2 3 4 5 6 7 8 9 10
e Samples

Fig. 14. The performance of buildings for H/B = 2 in terms of period (a), displacement of top storey (b), displacement of ground level (c), relative
displacement of top storey–ground level (d) and acceleration of top storey (e).
 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342 337

The values of parameters used in this NN model are as were used for training and the 30 samples (chosen ran-
follows: domly) were used as a test set and 30 samples were used
for verification as validation set. The performance of train-
 Number of input layer unit = 3. ing, and testing sets can be seen in Figs. 10a–13b.
 Number of hidden layer = 1. The dynamic SSI analysis of the soil medium has been
 Number of hidden layer units = 11. performed using 10 different soil properties (Table 1)
 Number of output layer units = 5. with different building heights, shear wave velocity and
local site depth. First 10 seconds of 1940 El-Centro
earthquake (Fig. 7) with a time step of 0.02 seconds
4. Numerical application has been used.
The substructure has been modelled using different soil
In this study the application of the NN to dynamic SSI depths of 100–75–50–25 m for each case with correspond-
problems under seismic loads by using direct method has ing constant width of 1400 m. The superstructure has been
been aimed. The training and testing sets have been prepared modelled as 3 bays with a constant width (B) of 9 m
by finite element analysis (FEA) where 2D Dynamic SSI (Fig. 8). The building height (H) has been selected as mul-
analyses have been carried out in time domain. The perfor- tiples of width and storey height.
mance of buildings in terms of period, displacement of top The analyses have been carried out using the following
story (TSD), displacement of ground level (DGL), relative assumptions:
displacement between top storey and ground level (RTSD),
acceleration of top storey (ATS) have been investigated.  1940 El-Centro earthquake acts on the soil–structure
The training and testing sets (Tables 3–5) have been gen- models with the same magnitude.
erated using outcomes of 240 different soil–structure mod-  The damping of the structure and soil is 5%, and 8%,
els including different height/width ratios of buildings in respectively.
terms of different soil properties, soil layer thickness  Columns and beams are rectangular and do not change
(SLT), and different shear wave velocities. 180 samples through the storey height.

1.2 49
NN NN
1 42
FEM FEM
Period (sec)

0.8 35
TSD (cm)

28
0.6
21
0.4
14
0.2 7
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
a Samples b Samples

14 32
NN 28 NN
12
FEM 24 FEM
10
RTSD (cm)
DGL (cm)

20
8
16
6
12
4 8
2 4
0 0
1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
c Samples d Samples

28
NN
24
Accel.(cm/sec 2)

FEM
20
16

12
8
4
0
1 2 3 4 5 6 7 8 9 10
e Samples

Fig. 15. The performance of buildings for H/B = 4 in terms of period (a), displacement of top storey (b), displacement of ground level (c), relative
displacement of top storey–ground level (d) and acceleration of top storey (e).
338 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342

1.8
NN 35 NN
1.5
FEM 30 FEM
Period (sec)

1.2

TSD (cm)
25

0.9 20
15
0.6
10
0.3
5
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
a Samples b Samples
36
8 NN 32 NN
FEM 28 FEM

RTSD (cm)
6 24
DGL (cm)

20
4 16
12
2 8
4
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
c Samples d Samples

36
NN
FEM
Accel.(cm/sec 2)

27

18

0
1 2 3 4 5 6 7 8 9 10
e Samples

Fig. 16. The performance of buildings for H/B = 6 in terms of period (a), displacement of top storey (b), displacement of ground level (c), relative
displacement of top storey–ground level (d) and acceleration of top storey (e).

Table 3 Table 4
Testing set for H/B = 2 Testing set for H/B = 4
Samples Shear wave velocity Soil layer thickness Samples Shear wave velocity Soil layer thickness
Sample 1 400 25 Sample 1 400 25
Sample 2 700 Sample 2 500
Sample 3 900
Sample 3 500 50
Sample 4 500 50 Sample 4 600
Sample 5 700
Sample 5 700 75
Sample 6 400 75 Sample 6 800
Sample 7 800 Sample 7 900
Sample 8 500 100 Sample 8 500 100
Sample 9 800 Sample 9 1000
Sample 10 1000 Sample 10 1500

 The storey heights for each floor are same. the network model for the period, displacement of top
 The soil medium is isotropic and homogeneous. storey, displacement of ground level, relative displacement
of top storey–ground level and acceleration of top storey
is shown in Table 2. The mean absolute percentage error
5. Results and discussions for the training and testing sets is found to be 8.43% and
13.65%, respectively. The amount of error is quite satis-
Training and testing sets of the proposed NN model factory as there are so many uncertainties in SSI analysis
have been compared with the results of FEA. The results due to soil conditions, assumptions in the dynamic analy-
have been presented by graphs including the training and sis and heterogeneous material properties of soil and
testing sets (Figs. 10a–13b). The statistical performance of concrete.
 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342 339

Table 5 In addition, the absolute fraction of variance (R2), and

Testing set for H/B = 6 mean absolute percentage error (MAPE) is calculated
Samples Shear wave velocity Soil layer thickness using Eqs. (7) and (8), respectively
Sample 1 500 25 P 2
!
Sample 2 600 2 j ðt j  outj Þ
R ¼1 P 2
ð7Þ
Sample 3 300 50 j ðoutj Þ
Sample 4 500
 
out  t
Sample 5 900 MAPE ¼  100 ð8Þ
out
Sample 6 400 75
Sample 7 700 where t is the target value, out is the output value, p is the
Sample 8 900 pattern.
Sample 9 400 100 The FEA and NN outcomes have been presented on the
Sample 10 800 same graphs. The performance of buildings for five H/B
ratios in terms of period, displacement of top storey, dis-
placement of ground level, relative displacement of top sto-
rey–ground level, and acceleration of top storey has been
The error incurred during the learning can be expressed given in Figs. 9–13.
as root-mean-squared (RMS) one and is calculated using The results of Figs. 9a–13a indicate that the neural net-
Eq. (3) work is successful in learning the relationship between the
s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 X
1   different input parameters and outputs for period of soil
RMS ¼  tj  outj 2 ð6Þ and structure, displacement of top storey, displacement
p j of ground level, relative displacement of top storey–ground

Validation
1.2 15
NN NN
1 12
FEM FEM
Period (sec)

0.8
TSD (cm)

9
0.6
6
0.4

0.2 3

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
a Samples b Samples
3
12 NN
NN
10 FEM
FEM
RTSD (cm)

2
DGL (cm)

4 1

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
c Samples d Samples

36
32 NN
FEM
Accel.(cm/sec 2)

28
24
20
16
12
8
4
0
1 2 3 4 5 6 7 8 9 10
e Samples

Fig. A.1. The performance of buildings for H/B = 1 in terms of period, displacement of top storey, displacement of ground level, relative displacement of
top storey–ground level and acceleration of top storey.
340 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342

level and acceleration of top storey, respectively. The model is appropriate and it predicts the dynamic response
results of testing phase in Figs. 9b–13b show that the NN of buildings with soil–structure interaction accurately when
model is capable of generalizing between input variables compared with the results of FEA.
and the output. As results obtained from the ground surface are evalu-
The statistical study employed herein to determine the ated statistically, significant deviations are observed. Dis-
top displacement and acceleration value of structures are placements at the ground surface due to dynamic motion,
found to be close for training validation and test sets. particularly for stiff soils, are too small. This leads to great
From this study, results with 10%, 12% and 12% errors deviations between computed displacements and neural
are found for training, test and validation set, respectively. network results which can be clearly seen in related graphs
The deviation of relative top displacement of the structure (Figs. 11a and 11b).
is found less in learning set when compared with those The effect of soil structure interaction generally ignored
calculated in test and validation sets. However, statistical in seismic design of building frames. The design is generally
values such as 13.0546%, 16.8529% MAPE and 0.9742, carried out based on the results of dynamic analysis consid-
0.9514 R2 are satisfactory for complex systems like soil– ering fixed-base condition. Flexibility of soil causes length-
structure interaction. ening of lateral natural periods due to overall decrease in
The performance of buildings in terms of period, dis- lateral stiffness. Such lengthening may considerably alter
placement of top storey, displacement of ground level, rel- the seismic response of the building frames [19]. This study
ative displacement of top storey–ground level and also identifies the influential parameters, which can regu-
acceleration of top storey were computed and compared late the effect of soil–structure interaction on the change
with NN model and the FEA for H/B = 1 (Fig. A.1), H/ in lateral natural period, displacement of top storey and
B = 2 (Fig. 14), H/B = 3 (Fig. A.2), H/B = 4 (Fig. 15), acceleration of top storey resting on different soil type
H/B = 5 (Fig. A.3), H/B = 6 (Fig. 16). A careful study of and different soil layer thickness The graphs and tables,
the results in Figs. 13–18 leads to observations of excellent which were presented for testing and verification sets, were
agreement between NN model outcomes and FEA results. showed the trend of variation effect of these characteristics
Almost of the all results also prove that the proposed NN with the soil–structure interaction effect.

1
NN 40 NN
0,8 FEM 35 FEM
Period (sec)

30
TSD (cm)

0,6
25
20
0,4
15
0,2 10
5
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
a Samples b Samples

40
8 NN 35 NN
FEM 30 FEM
RTSD (cm)

6
DGL (cm)

25
20
4
15
2 10
5
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
c Samples d Samples

50
NN
Accel.(cm/sec 2)

40 FEM

30

20

10

0
1 2 3 4 5 6 7 8 9 10
e Samples

Fig. A.2. The performance of buildings for H/B = 3 in terms of period (a), displacement of top storey (b), displacement of ground level (c), relative
displacement of top storey–ground level (d) and acceleration of top storey (e).
 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342 341

As related figures (Figs. 14–16) were examined, the problems. Also for buildings with different heights under
acceleration and relative displacement of peak point pre- different soil conditions, the effect of local site depth, soil
dicted by the NN model were observed to within accept- conditions and building height with respect to resonance
able limits. NN model were also predicted the all outputs of the structure is presented clearly.
of the soil–structure system for high H/B ratios and for soft This study has been a pioneer work on the application of
soils with a high accuracy. NN in dynamic soil structure analysis. The results of the
study can be summarized as follows:
6. Conclusion
1. The proposed neural network approach has demon-
The effects of various soil conditions and local site strated to have a high capacity of generalization for
depths on the structure behaviours have been investigated SSI problems.
and in the preparation of training and verification test sets 2. The computation time for SSI analysis is enormous
solutions obtained for different height/width ratios of where as by using NN the computation time can be
buildings in terms of different soil depths and different extremely reduced.
shear wave velocities are used. Various analysis are per- 3. NN can be used as an alternative efficient tool for SSI
formed for different soil properties, local site depths and analysis with respect to computational methods like
number of building stories and the change of results of lat- FEM. NN present many advantages compared to other
eral displacement of the last stories of buildings, accelera- computational methods as the learning process is speci-
tion and periods with respect to soil conditions are fied for each problem type and application of NN is
presented by graphs and the results are discussed. As a quite flexible.
result by using NN in the solution soil structure interaction
problems it is observed that the assumption made during
the modelling and analysis stages are minimized, the solu- Appendix A. Validation set
tion time is quite fast and NN has shown a satisfactory per-
formance for the solution of soil structure interaction See Figs. A.1–A.3 and Tables A.1–A.3.

NN 35 NN
1,6
FEM 30 FEM
Period (sec)

TSD (cm)

1,2 25
20
0,8 15
10
0,4
5
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
a Samples b Samples

6 36
NN 32 NN
5 28 FEM
FEM
RTSD (cm)

4 24
DGL (cm)

20
3 16
2 12
8
1
4
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
c Samples d Samples
28
NN
24
Accel.(cm/sec 2)

FEM
20
16
12
8
4
0
1 2 3 4 5 6 7 8 9 10
e Samples

Fig. A.3. The performance of buildings for H/B = 5 in terms of period (a), displacement of top storey (b), displacement of ground level (c), relative
displacement of top storey–ground level (d) and acceleration of top storey (e).
342 M. Pala et al. / Construction and Building Materials 22 (2008) 330–342

Table A.1 References

Testing set for H/B = 1
Samples Shear wave velocity Soil layer thickness [1] Mihailo DT, Maria IT, Tzong-Ying H. Full-scale experimental
studies of soil–structure interaction. In: Proceedings of the 2nd US–
Sample 1 400 25
Japan workshop on soil–structure interaction, Tsukuba City, Japan,
Sample 2 700
March 6–8, 2001.
Sample 3 800 50 [2] John PW, Chongmin S. Some cornerstones of dynamic soil–structure
Sample 4 600 interaction. Eng Struct 2002;24:13–28.
[3] Spyrakosa CC, Chaojin X. Seismic soil–structure interaction of
Sample 5 700 75
massive flexible strip-foundations embedded in layered soils by hybrid
Sample 6 500
BEM–FEM. Soil Dyn Earthquake Eng 2003;23:383–9.
Sample 7 300
[4] Ben Jamaa S, Shiojiri H. A method for three dimensional interaction
Sample 8 900 100 analysis of pile-soil system in time domain. Trans Jpn Soc Comput
Sample 9 400 Eng Sci 2000.
Sample 10 700 [5] Dutta SC, Roy R. A critical review on idealization and modeling for
interaction among soil–foundations structure system. Comput Struct
2002;80:1579–94.
Table A.2 [6] Bishop CM. Neural networks for pattern recognition. Oxford,
Testing set for H/B = 3 England: Oxford University Press; 1995.
Samples Shear wave velocity Soil layer thickness [7] Fausett Laurene V. Fundamentals of neural networks: architectures,
algorithms, and applications. Englewood Cliffs, NJ: Prentice-Hall;
Sample 1 400 25
1994.
Sample 2 900
[8] Haykin S. Neural networks: a comprehensive foundation. NY, USA:
Sample 3 300 50 Prentice Hall; 1999.
Sample 4 600 [9] Kulkarni AD. Artificial neural networks for image understanding.
NY, USA: Van Nosrand Reinhold; 1994.
Sample 5 500 75
[10] Jaya KP, Meher Prasad A. Embedded foundation in layered soil
Sample 6 700
under dynamic excitations. Soil Dyn Earthquake Eng 2002;22:
Sample 7 1000
485–98.
Sample 8 600 100 [11] Lysmer J, Kuhlemeyer RL. Finite element model for infinite media.
Sample 9 500 J Eng Mech Div ASCE 1969;95:859–77.
Sample 10 Fixed [12] Wolf JP, Song C. Finite-element modelling of unbounded media.
England: Wiley; 1996.
[13] White W, Valliappan S, Lee IK. Unified boundary for finite dynamic
Table A.3 models. J Eng Mech ASCE 1977;103(5):949–64.
Testing set for H/B = 5 [14] Kausel E, Roesset JM. Dynamic stiffness of circular footings. J Eng
Mech Div ASCE 1975;101:771–85.
Samples Shear wave velocity Soil layer thickness
[15] Moller AF. A scaled conjugate gradient algorithm for fast supervised
Sample 1 300 25 learning. Neural Networks 1993;6:525–33.
Sample 2 800 [16] Oztas A, Pala M, Ozbay E, Kanca E, Caglar N, Bhatti MA.
Sample 3 1000 Predicting the compressive strength and slump of high strength
concrete using neural network. J Construct Build Mater
Sample 4 300 50
2006;20:769–75.
Sample 5 600
[17] Pala M. New formulation for distortional buckling stress. J Construct
Sample 6 Fixed Steel Res 2006;62:716–22.
[18] Hagan MT, Demuth HB, Beale MH. Neural network design. Boston,
Sample 7 400 75
MA: PWS Publishing; 1996.
Sample 8 600
[19] Bhattacharya K, Dutta SC, Dasgupta S. Effect of soil-flexibilityon
Sample 9 700 100 dynamic behaviour of building frames on raft foundation. J Sound
Sample 10 1500 Vib 2004;274:111–35.