Professional Documents
Culture Documents
Abstract
Six artificial neural network (ANN) models are developed to predict various response parameters of kinematic soil pile interaction.
These responses include (1) kinematic response factors for free and fixed head piles in homogenous soil layer to derive foundation input
motion (2) normalized bending moment at fixed head of pile in homogenous soil layer (3) normalized kinematic pile moment at the
interface of two soil layers of sharply different soil stiffnesses. These ANN models represent simple solutions that can be implemented in
a simple calculator capable of matrix operation and bypass the site response analysis and the complex wave diffraction analysis. The data
required for ANN training is generated using beam on dynamic Winkler formulation (BDWF). Fifty percent of the data is used to train
the ANN models while remaining 50% is used to test the ANN models. The trained ANN models show good agreement with BDWF
results.
r 2007 Elsevier Ltd. All rights reserved.
0267-7261/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.soildyn.2006.12.009
ARTICLE IN PRESS
I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905 893
where U g is the ground displacement amplitude and The pile–soil interaction involves seismic energy dissipation
o is the excitation circular frequency. The pile group through hysteretic (material damping) and radiation
effect is not considered as it is negligible in kinematic (geometric damping). The former represents the energy
interaction [9]. dissipation through the soil inelastic behavior, and is
Parametric study on free head piles in homogenous soil related to soil damping ratio, bs . The latter represents the
reveals that the kinematic interaction factors depend on the radiation of energy by waves spreading geometrically away
ratio of E p =E s , L=d, soil Poisson’s ratio n, soil damping from the pile–soil interface. Hence, the distributed dashpot
ratio bs , and rp =rs . Out of these parameters, E p =E s and coefficient is cx ¼ cr þ cm , where cr is the distributed
L=d are the most governing parameters [10]. For a two- radiation dashpot coefficient and cm the distributed
layer soil profile, the parameters that govern the pile material dashpot coefficient. In this study, these coefficients
bending moment at the interface of the two layers are are adopted from Gazetas and Dobry [12,13], i.e.
E p =E s1 , L=d, V s2 =V s1 , and h1 =L [6].
2kx bs
The BDWF approach is adopted in this paper to cm ¼ , (2)
o
generate data for the training of ANN models. In all cases,
soil damping ratio (bs Þ is considered to be 10%, which is 5=4 !
Vc 1=4
compatible with the design level of strain in soil, n is cr ¼ 2drs V s 1þ a0 , (3)
considered to be 0.4 and rp /rs of 1.6. The ranges of Vs
input parameters selected to generate data for the ANN where V s is the shear wave velocity of soil. The apparent
models and the corresponding output range are given in velocity of the compression waves, V c is taken as the
Tables 1–3. The considered ranges cover most of the Lysmer’s analog velocity V La at z42:5d, where
practical situations.
3:4V s
V c ¼ V La ¼ . (4)
pð1 uÞ
4. Beam-on-dynamic-Winkler-foundation (BDWF) model
Near the ground surface (zp2:5d), three-dimensional
The data for the ANN models was generated by effects arising from the stress-free boundary are better
analyzing the pile–soil system using BDWF formulation. reproduced using V c ffi V s .
The pile is connected to free field soil along its length
through continuously distributed interface elements. The 4.2. Solution of differential equation governing kinematic
end of the interface element connected to the free field soil is response of piles
excited by the corresponding free field displacement uff ðz; tÞ.
The steady state pile displacement in response to the
4.1. Interface element bedrock harmonic excitation ug ðtÞ is governed by the
following differential equation:
Each interface element consists of a spring and a dashpot U 0000 4
pp ðzÞ l U pp ðzÞ ¼ aU ff ðzÞ, (5)
arranged in parallel. The spring stiffness (kx ) is adopted
from Kavvadas and Gazetas [11], i.e. kx ¼ dE s , where where
1=8 1=12 1=15
2 E s1 d 4 L h1 V s1 mp o2 Sx Sx
d¼ . (1) l4 ¼ ; a¼ ; Sx ¼ K x þ icx o. (6)
2
1 n EpI p d h2 V s2 EpI p EpI p
Table 1
Ranges of parameters used for data generation for ANNKDFix, Table 3
ANNKDFree, and ANNKRFree Ranges of parameters used for data generation for ANNMNINT1
Table 2
Ranges of parameters used for data generation for ANNMNFIX and ANNMNFIX1 shown in parenthesis
In Eqs. (5) and (6), Sx is complex impedance function, mp The solution of the differential equation is coded in
is the pile mass per unit length, I p is pile second moment of MATLAB as a general case for a two-layer soil system
inertia, and U ff ðzÞ and U pp ðzÞ are displacement amplitudes overlying rigid rock formation. The free field displacements
of free field and pile at depth z. The differential equation are evaluated as a function of the depth and their first,
has the general solution: second, and third derivatives. Eight arbitrary constants are
8 9 determined through four boundary conditions and four
> D1 >
>
> > continuity conditions and then the maximum pile displace-
< D2 >=
lz lz ilz ilz
U pp ðzÞ ¼ ½e e e e þ sU ff ðzÞ, (7) ment at different depths are calculated.
>
> D3 >>
>
: >
;
D4
5. Artificial neural network models
where
p s ¼ a=ðq4 l4 Þ and q is complex wave number ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
o=V s þ ibs . 5.1. Architecture of ANN models
For a homogenous soil profile, there are four arbitrary
constants to be evaluated through four boundary condi- The ANN is a computational procedure able to acquire,
tions, and there are eight constants for the two-layer soil represent, and compute a mapping from one multivariate
profile, which requires four compatibility equations in space of information to another, given a set of data
addition to four boundary conditions. representing that mapping [14]. One of the most common
engineering applications of ANN is to use inputs to predict
4.3. Boundary conditions certain outputs. About 80% of neural network applications
utilize back-propagation neural network for prediction.
4.3.1. Boundary conditions for piles in homogenous soil Back-propagation ANN has been applied to various
layer problems of civil engineering like predicting concrete shear
The moment at pile base is zero, MðL; tÞ ¼ 0; and capacity [15–18], seismic liquefaction potential [19], friction
displacement at pile base is equal to ground displacement capacity of driven piles [20] and overturning response of
at bedrock, i.e. up ðL; tÞ ¼ ug ðtÞ. The shear force and rigid block under near-fault type excitation [21].
moment at the pile head are zero, i.e. V ð0; tÞ ¼ 0, Mð0; tÞ ¼ A Levenberg–Marquardt back-propagation algorithm
0 for free head piles, whereas for rotationally fixed head was used in this research. It is one of the fastest methods
piles the shear force and rotation at pile head are zero available for training moderate-sized feed-forward neural
i.e. V ð0; tÞ ¼ 0, yð0; tÞ ¼ 0. networks [22]. The architecture of ANNMNFIX1 model
consisting of an input layer of two input neurons, a hidden
4.3.2. Boundary conditions for piles in two-layer soil profile layer of two neurons, and an output layer consisting of one
The boundary conditions for the pile in Fig. 2 suggest output neuron is shown in Fig. 3. The symbols w and b in
that moment and shear at the top and base of the pile are Fig. 3 represent connection and bias weights with sub-
zero, i.e. Mð0; tÞ ¼ 0, V ð0; tÞ ¼ 0, MðL; tÞ ¼ 0, V ðL; tÞ ¼ 0. scripts representing the corresponding neurons between
two layers.
4.4. Compatibility conditions for piles in two-layer soil The first step in the design of a neural network is to select
profile an appropriate architecture. In this study, each ANN
model consists of an input layer, a hidden layer, and an
Four compatibility equations are obtained by satisfying output layer. The input and output layers consist of nodes
continuity of the pile moment, shear, displacement and (neurons) equal in number to the input and output
rotation at the interface of the two soil layers, i.e. parameters, respectively. No processing on the inputs is
done in the input layer. Sandwiched between the input and
Pile rotation amplitude : fpp ðzÞ ¼ U 0pp ðzÞ. output layers is the hidden layer, consisting of neurons
Pile moment amplitude : MðzÞ ¼ E p I p U 00pp ðzÞ. called hidden neurons. The hidden neurons, or sometime
the output neurons, have a nonlinear transfer function.
Pile shear amplitude : V ðzÞ ¼ E p I p U 000
pp ðzÞ. The effectiveness of an ANN model to simulate highly
Table 4
ANN models architecture
ANN models Input parameters No. of input No. of hidden neurons/ Output No. of output neurons/
neurons transfer function parameters transfer function
ANNKDFix E p =E s , L=d, o=o1 3 4/logistic sigmoid (logsig) KDFix 1/tangent hyperbolic (tanh)
ANNKDFree E p =E s , L=d, o=o1 3 5/logistic sigmoid (logsig) KDFree 1/tangent hyperbolic (tanh)
ANNKRFree E p =E s , L=d, o=o1 3 5/logistic sigmoid (logsig) KRFree 1/tangent hyperbolic (tanh)
nonlinear problems is attributed to the nonlinear transfer min P the vector containing minimum values of the original
function. The number of hidden neurons plays a crucial input; max P the vector containing maximum values of the
role in the performance of an ANN. In this study, the original input; min T the vector containing the minimum
number of hidden neurons was optimized, using trial and value of the target output; max T the vector containing the
error, to avoid underfitting (i.e large training and testing maximum value of the target output.
errors) and prevent overfitting (i.e. low training error but The scaled data was then used to train the neural
high testing error). network. The data from the output neurons have to be
A subset of the data points is used to train the ANN with postprocessed to convert the data back into unscaled units
the help of a suitable algorithm. In this study, the network to get the actual target value according to
is supplied with 50% of the total data for training, and the
T ¼ 0:5 ðT n þ 1Þðmax T min TÞ þ min T. (10)
remaining 50% is reserved for testing the performance of
the trained ANN. The network adjusts the connection Tables 1–3 give the minimum and maximum values of the
weights and bias such as to reduce the error between the data scaling parameters.
target and the corresponding network output, using the The training continued until the average sum squared
Levenberg–Marquardt back-propagation algorithm. error over all the training patterns was minimized. This
After the training process is completed, the prediction occurred after about 1500 cycles of training. The connec-
performance of the trained network is tested with data not tion and bias weights obtained after ANN training can be
seen before (50% of the total data in this study). The used to estimate the corresponding kinematic response
network predicts the output using the connection weights parameter. The weight and bias matrices obtained after the
and biases established in the training phase. The details of training phase of the ANN models are W 1 is the weight
ANN models design are given in Table 4. matrix representing connection weights between the input
The data generated to train ANN models are first layer and hidden layer neurons; W 2 the weight matrix
randomized and then divided into two sets namely the representing connection weights between the hidden
training data set and the testing data set. The data was so layer and output layer neurons; B1 the bias vector for the
divided to give comparable statistical properties for hidden layer neurons; B2 the bias vector for the output
training and testing data. layer neuron. These weight matrices for ANN models are
given below.
5.2. Scaling of training data Weight and bias matrices for ANNKDFix:
2 3
0:1575 0:6526 0:1622
Preprocessing of the training data is performed so that 6 0:1462 0:6481 0:1342 7
the processed data is scaled in the range of 1 to +1. The 6 7
W1 ¼ 6 7,
training data sets were scaled (preprocessed) according to 4 0:0119 0:1868 3:1880 5
4:0713 0:0556 0:0599
ðP min PÞ
Pn ¼ 2 1, (8) 2 3
ðmax P min PÞ 0:1908
6 0:2222 7
ðT min TÞ 6 7
Tn ¼ 2 1, (9) B1 ¼ 6 7,
ðmax T min TÞ 4 8:8738 5
where P is the matrix of the input vectors; T the matrix of 10:3919
the target output vectors; Pn the matrix of scaled input
vectors; T n the matrix of scaled target output vectors; W 2 ¼ ½ 275:3 285:7 1165:1 860:8 ,
ARTICLE IN PRESS
I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905 897
1.00 1.00
KDFix
KDFix
0.75 0.75
0.50 0.50
0.25 0.25
0.00 0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.00 0.10 0.20 0.30 0.40 0.50
ao ao
Fig. 4. Kinematic displacement factors for fixed head piles (E p =E s ¼ 1000; L=d ¼ 20, 40).
1.50
1.50 BDWF (Ep/Es=10000, L/d=40)
BDWF (Ep/Es=10000, L/d=20) 1.25
ANN (Ep/Es=10000, L/d=40)
1.25 ANN (Ep/Es=10000, L/d=20)
1.00
1.00
KDFix
KDFix
0.75
0.75
0.50 0.50
0.25 0.25
0.00 0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.00 0.10 0.20 0.30 0.40 0.50
ao ao
Fig. 5. Kinematic displacement factors for fixed head piles (E p =E s ¼ 10 000; L=d ¼ 20, 40).
KDFree
1.00
0.80
0.80
0.60 BDWF (Ep/Es=1000, L/d=40)
0.60
0.40 0.40 ANN (Ep/Es=1000, L/d=40)
0.20 0.20
0.00 0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.00 0.10 0.20 0.30 0.40 0.50
ao ao
Fig. 6. Kinematic displacement factors for free head piles (E p =E s ¼ 1000; L=d ¼ 20, 40).
1.60 1.60
1.40 1.40
1.20 1.20
1.00 1.00
KDFree
KDFree
0.80 0.80
0.60 0.60
0.40 BDWF (Ep/Es=10000, L/d=20) 0.40 BDWF (Ep/Es=10000, L/d=40)
0.20 ANN (Ep/Es=10000, L/d=20) 0.20 ANN (Ep/Es=10000, L/d=40)
0.00 0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.00 0.10 0.20 0.30 0.40 0.50
ao ao
Fig. 7. Kinematic displacement factors for free head piles (E p =E s ¼ 10 000; L=d ¼ 20, 40).
ARTICLE IN PRESS
898 I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905
0.30
KRFree
0.25 0.25
0.20 0.20
0.15 0.15
0.10 0.10
0.05 0.05
0.00 0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.00 0.10 0.20 0.30 0.40 0.50
ao ao
Fig. 8. Kinematic rotation factors for free head piles (E p =E s ¼ 1000; L=d ¼ 20, 40).
0.50 0.50
BDWF (Ep/Es=10000, L/d=40)
0.45 BDWF (Ep/Es=10000, L/d=20) 0.45
0.40 0.40 ANN (Ep/Es=10000, L/d=40)
ANN (Ep/Es=10000, L/d=20)
0.35 0.35
0.30
KRFree
0.30
KRFree
0.25 0.25
0.20 0.20
0.15 0.15
0.10 0.10
0.05 0.05
0.00 0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.00 0.10 0.20 0.30 0.40 0.50
ao ao
Fig. 9. Kinematic rotation factors for free head piles (E p =E s ¼ 10 000; L=d ¼ 20, 40).
7000 7000
R2 = 1.0
2
6000 R = 1.0 6000
ANNMNFIX Prediction
ANNMNFIX1 Prediction
5000 5000
MNFIX1(ω1)
MNFIX(ω)
4000 4000
3000
3000
2000
2000
1000
1000
0
0 0 1000 2000 3000 4000 5000 6000 7000
0 1000 2000 3000 4000 5000 6000 7000 MNFIX(ω)
MNFIX1(ω1) BDWF Prediction
BDWF Prediction
Fig. 11. ANNMNFIX model prediction.
Fig. 10. ANNMNFIX1 model prediction.
B2 ¼ ½ 852:3121 .
Table 5
Weight and bias matrices for ANNKDFree: Soil–pile system
2 3 Soil–pile system Soil–pile parameters
0:7951 3:2456 6:9741
6 1:7072 3:6850 14:9470 7 L (m) d (m) V s (m/s) E p ðN=m2 Þ
6 7
6 7
W1 ¼ 6
6 0:8041 3:2436 6:9865 77, SP1 20 1.0 100 30 109
6 7
4 1:2705 2:3714 6:4716 5 SP2 40 1.5 80 200 109
4:9902 2:3568 5:2564 SP3 10 1.0 150 200 109
ARTICLE IN PRESS
I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905 899
0.15
Tabas Earthquake
0.05
0
0 5 10 15 20 25 30 35
-0.05
time (sec)
-0.1
0.04
Friuli Earthquake
0.03
ground acceleration (g)
0.02
0.01
0
0 2 4 6 8 10 12 14 16 18 20
-0.01
time (sec)
-0.02
-0.03
-0.04
0.4
Northridge Earthquake
ground acceleration (g)
0.3
0.2
0.1
0
-0.1 0 1 2 3 4 5 6 7 8 9 10
-0.2 time (sec)
-0.3
-0.4
-0.5
0.5
Loma Prieta Earthquake
0.4
ground acceleration (g)
0.3
0.2
0.1
0
-0.1 0 5 10 15 20 25 30
-0.2 time (sec)
-0.3
-0.4
-0.5
2 3
0:6999 B2 ¼ ½ 335:4918 .
6 9:3396 7
6 7
6 7
B1 ¼ 6 0:7018 7
6
7, Weight and bias matrices for ANNKRFree:
6 7 2 3
4 1:5968 5 0:0466 0:6654 8:0697
5:5271 6 1:0610 1:7342 2:6758 7
6 7
6 7
W1 ¼ 6
6 0:8282 1:9736 5:9451 77,
6 7
4 1:1908 1:7099 2:6744 5
W 2 ¼ ½ 338:083 1:6118 336:9977 2:1119 2:2238 , 1:1245 1:7225 2:6786
ARTICLE IN PRESS
900 I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905
35
Friuli earthquake
Tabas earthquake
30
Northridge earthquake
Loma Prieta earthquake
25
Fourier Amplitude (g-sec)
20
15
10
0
0 1 2 3 4 5 6 7 8 9 10
Frequency (Hz)
2 3 2 3
13:0827 1:5515
6 1:2521 7 6 51:2817 7
6 7 6 7
6 7 6 7
B1 ¼ 6
6 0:4956 7,
7 B1 ¼ 6
6 1:5508 7,
7
6 7 6 7
4 1:2072 5 4 2:0229 5
1:2315 51:3134
W 2 ¼ ½ 470:7 765 1:5 755:8 1522:3 , W 2 ¼ ½ 36:4954 168:1853 36:8649 0:4751 168:1417 ,
B2 ¼ ½ 758:05 .
B2 ¼ ½ 1:0615 .
300
SP1 (Tabas Earthquake) BDWF
ANN
200
moment (KN-m)
100
0
0 5 10 15 20 25 30 35
-100
time (sec)
-200
-300
7500 BDWF
SP2 (Tabas Earthquake)
ANN
5000
2500
moment (KN-m)
0
0 5 10 15 20 25 30 35
-2500
-5000
time (sec)
-7500
-10000
1000
SP3 (Tabas Earthquake) BDWF
750 ANN
500
moment (KN-m)
250
0
10 15 20 25 30 35
-250
-500
-750
time (sec)
-1000
Fig. 14. Comparison of bending moment time history of ANNMNFIX with BDWF for Tabas earthquake.
W 2 ¼ ½ 2:5119 0:6068 175:1881 88:4036 210:5579 , or conducted using a calculator capable of performing
simple matrix operations. The input data is first prepro-
cessed according to Eq. (8) to get a scaled input vector Pn .
The scaled target T n for kinematic response factors is
B2 ¼ ½ 56:2951 .
calculated as follows:
5.3. Procedure for estimating kinematic response parameter T n ¼ tanh½W 2 flogsigðW 1 Pn þ B1 Þg þ B2 , (11)
60 BDWF
SP1 (Friuli Earthquake)
ANN5
40
moment (KN-m)
20
0
0 2 4 6 8 10 12 14 16 18 20
-20 time (sec)
-40
-60
1000 BDWF
SP2 (Friuli Earthquake)
750 ANN
500
moment (KN-m)
250
0
0 2 4 6 8 10 12 14 16 18 20
-250
-500 time (sec)
-750
-1000
200
SP3 (Friuli Earthquake) BDWF
150 ANN
100
moment (KN-m)
50
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
-50
time (sec)
-100
-150
-200
Fig. 15. Comparison of bending moment time history of ANNMNFIX with BDWF for Friuli earthquake.
The scaled target T n for kinematic moment is obtained that the kinematic interaction effect is more pronounced in
as follows: fixed head piles than for free head piles (Figs. 6 and 7). The
kinematic displacement factor for free head piles may be
T n ¼ W 2 ftanhðW 1 Pn þ B1 Þg þ B2 . (12) greater than 1.0 for certain frequency ranges (Figs. 6 and
7), and are also accompanied with kinematic rotation
The scaled output T n is then unscaled using Eq. (10) to
factors (Figs. 8 and 9), which should not be neglected in
obtain the corresponding actual target.
the design.
The kinematic interaction factors can be applied to
5.4. ANN model prediction either Fourier spectra or response of free field motion to
derive the foundation input motion.
5.4.1. ANNKDFix, ANNKDFree, and ANNKRFree prediction
The ANN models were used to predict the kinematic 5.4.2. Prediction performance of ANNM NFIX 1 and
response factors for four cases of pile–soil properties and ANNM NFIX
for a range of frequencies. As shown in Figs. 4–9, the In order to evaluate the capability of the ANN models,
predictions of the ANN models are reasonable from they were presented with new data that was not part of the
engineering point of view. It is evident from the figures training data set, and the corresponding normalized
that as the rigidity of the pile relative to the soil increases, moments calculated. Figs. 10 and 11 show the prediction
the free field motion is strongly altered. Figs. 4 and 5 show of M NFIX1 and M NFIX plotted against the corresponding
ARTICLE IN PRESS
I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905 903
1000
SP1 (Northridge Earthquake)
BDWF
750
ANN5
500
0
0 1 2 3 4 5 6 7 8 9 10
-250 time (sec)
-500
-750
8000
SP2 (Northridge Earthquake)
6000 BDWF
ANN
4000
moment (KN-m)
2000
0
0 2 4 6 8 10
-2000 time (sec)
-4000
-6000
4500
SP3 (Northridge Earthquake)
BDWF
3000 ANN
moment (KN-m)
1500
0
0 1 2 3 4 5 6 7 8 9 10
-1500 time (sec)
-3000
-4500
Fig. 16. Comparison of bending moment time history of ANNMNFIX with BDWF for Northridge earthquake.
normalized moments calculated using the BDWF model. Fig. 12, while Fig. 13 shows Fourier spectra of acceleration
The correlation coefficient ðRÞ for both ANN models was time histories which imply a range of frequencies present in
1.0 for testing and training data set. the earthquake signals. The bending moments at the fixed
To evaluate the performance of ANNMNFIX model for head of the pile is compared with the corresponding
the full time history range of an earthquake signal, three BDWF solution in Figs. 14–17, which show close agree-
soil pile systems are subjected to four earthquake time ment between the two solutions.
histories. The parameters of pile–soil systems considered
are given in Table 5. The four time histories selected are the 5.4.3. ANNM NINT1 prediction
Tabas earthquake (Iran 1978-09-16) station 70 Bosh- To check the ability of the ANNMNINT1 model to predict
rooyeh, Fuiuli earthquake (Italy 1976-05-06) Station: M NINT1 , it was presented with new data that was not part
8002 Barcis, Northridge (1994-01-17) Station: CDMG of the training data set and M NINT1 was calculated. The
24207 Pacoima Dam (downstream), and Loma Prieta normalized moment M NINT1 was then converted into
(1989-10-18) Station: CDMG 47379 Gilroy Array #1 M INT1 using d ¼ 1 m, rp ¼ 1:8 KNs2 =m, and ag ¼
earthquake. The acceleration time histories are shown in 1 m=s2 . Fig. 18 compares M INT1 calculated using the
ARTICLE IN PRESS
904 I. Ahmad et al. / Soil Dynamics and Earthquake Engineering 27 (2007) 892–905
1000
SP1 (Loma Prieta Earthquake)
800
600 BDWF
moment (KN-m) ANN5
400
200
0
0 2 4 6 8 10
-200
time (sec)
-400
-600
15000
SP2 (Loma Prieta Earthquake)
BDWF
10000 ANN
moment (KN-m)
5000
0
0 1 2 3 4 5 6 7 8 9 10
-5000 time (sec)
-10000
-15000
4500 BDWF
SP3 (Loma Prieta Earthquake)
ANN
3000
moment (KN-m)
1500
0 ????
0 1 2 3 4 5 6 7 8 9 10
-1500
time (sec)
-3000
-4500
Fig. 17. Comparison of bending moment time history of ANNMNFIX with BDWF for Loma Prieta earthquake.
5000 BDWF and the ANN models. The agreement between the
two sets is satisfactory, with a correlation coefficient
4000 R2=1.0 of 1.0.
ANNMNINT1 Prediction
6. Conclusions [8] Mylonakis G. Simplified model for seismic pile bending at soil layer
interfaces. Soils Found 2001;41(4):47–58.
[9] Gazetas G, Fan K, Tazoh T, Shimizu K, Kavvadas M, Makris N.
The ANN models developed in this study can be used to
Seismic pile-group—structure interaction. Piles under dynamic loads,
estimate various kinematic response parameters including Geotechnical special publication no. 34, 1992. p. 56–93.
the kinematic response factors for: free and fixed head piles [10] Pender MJ. Aseimic pile foundation design and analysis. Bulletin of
embedded in homogenous soil; the fixed head moment of the New Zealand national society for earthquake engineering 1993;
piles in homogenous soils; and pile bending moment at the 26(1):49–159.
interface of two-layer soil deposit of sharply contrasting [11] Kavvadas M, Gazetas G. Kinematic seismic response and bending of
free-head piles in layered soil. Geotechnique 1993;43(2):207–22.
stiffnesses. The prediction of kinematic response para- [12] Gazetas G, Dobry R. Horizontal response of piles in layered soils.
meters by ANN models was found to be in close agreement J Geotech Eng Div Am Soc Civ Eng 1984;110(1):20–40.
with the BDWF formulation. The ANN-based solutions [13] Gazetas G, Dobry R. Simple radiation damping model for piles and
offer simple scheme of calculation and avoid site response footings. Horizontal response of piles in layered soils. J Geotech Eng
and wave diffraction analyses to find kinematic response Div Am Soc Civ Eng 1984;110(6):937–56.
[14] Garrett Jr JH. Where and why artificial neural networks are
parameters and can easily be implemented in a simple applicable in civil engineering. J Comput Civ Eng ASCE 1994;8(2):
calculator capable of carrying out matrix operation. 129–30.
[15] Ahmad I, Khan AN, El Naggar MH. Neural network for evaluating
shear capacity of concrete in RC beams. Third international
References conference on construction materials: performance, innovations and
structural implications, Vancouver, BC, Canada, August 22–24,
[1] Ahmad I, Akhtar NK. Kinematic seismic response of piles— 2005.
importance and modeling. GeoCongress 2006: geotechnical engineer- [16] Bohigas AC, Mari AR. Shear design procedure for reinforced normal
ing in the information technology age 2006. and high-strength concrete beams using artificial neural networks.
[2] Mizuno H. Pile damage during earthquakes in Japan. In: Nogami T, Part-I beams without stirrups. Eng Struct 2004;26(7):917–26.
editor. Dynamic response of pile foundations. New York: American [17] Seleemah AA. A neural network model for predicting maximum
Society of Civil Engineers; 1987. p. 53–78. shear capacity of concrete beams without transverse reinforcement.
[3] Margason E. Pile bending during earthquakes. Lecture, 6 March Can J Civil Eng 2005;32(4):644–57.
1975; ASCE-UC/Berkeley seminar on design construction and [18] El-Chabib H, Nehdi M, Said A. Predicting shear capacity of NSC
performance of deep foundations (unpublished). and HSC slender beams without stirrups using artificial intelligence.
[4] Dobry R, O’Rourke MJ. In: Flores-Berrones R, Whitman RV, Comput Concr 2005;2(1).
editors. Discussion on seismic response of end-bearing piles, vol. 109. [19] Najjar YM, Ali HE. CPT-based liquefaction potential assessment: a
ASCE, New York. J Geotech Eng Div 1983. neuronet approach. Geotechnical earthquake engineering and soil
[5] NEHERP. Recommended provisions for seismic regulations for new dynamics III, Geotechnical special publication no. 75, vol. 2, 1998.
buildings and other structures. Parts 1 and 2, Building Seismic Safety p. 552–3.
Council, Washington, DC; 1987. [20] Goh ATC. Empirical design in geotechnics using neural networks.
[6] Nikolauo A, Mylonakis G, Gazetas G. Kinematic bending moments Geotechnique 1995;45(4):709–14.
in seismically stressed piles. NCEER technical report, Buffalo, NY; [21] Gerolymos N, Apostolou M, Gazetas G. Neural network analysis of
1995. overturning response under near-fault type excitation. Earthquake
[7] Nikolaou A, Gazetas G. Seismic design procedure for kinematically Eng Eng Vib 2005;4(2):213–28.
stressed piles. In: Seco e Pinto, editor. Seismic behavior of ground [22] Hagan MT, Demuth HB, Beale M. Neural network design. Boston:
and geotechnical structures, Balkema, Rotterdam; 1997. PWS; 1996.