ARTICLE IN PRESS

Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

www.elsevier.com/locate/soildyn

Assessment of liquefaction triggering using strain energy concept and

ANN model: Capacity Energy
M.H. Baziar, Y. Jafarian
Center of Excellence for Fundamental Studies in Structural Engineering, College of Civil Engineering, Iran University of Science and Technology,
Tehran 16846-13114, Iran
Received 11 January 2007; accepted 28 March 2007

Abstract

In the present study, an artificial neural network (ANN) model was developed to establish a correlation between soils initial
parameters and the strain energy required to trigger liquefaction in sands and silty sands. A relatively large set of data including 284
previously published cyclic triaxial, torsional shear and simple shear test results were employed to develop the model. A subsequent
parametric study was carried out and the trends of the results have been confirmed via some previous laboratory studies. In addition, the
data recorded during some real earthquakes at Wildlife, Lotung and Port Island Kobe sites plus some available centrifuge tests data have
been utilized in order to validate the proposed ANN-based liquefaction energy model. The results clearly demonstrate the capability of
the proposed model and the strain energy concept to assess liquefaction resistance (capacity energy) of soils.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Liquefaction triggering; Energy concept; Neural network; Capacity energy

1. Introduction of Niigata earthquake, concluded that the equivalent shear

stress is 57% rather than the 65% for 20 cycles of loading.
Liquefaction in loose and saturated sand and silty sands Although, the stress-based procedure has been continually
has been the cause of damage and catastrophic failures in revised and extended due to subsequent studies and
civil engineering structures. This phenomenon was best increase in the number of liquefaction case histories (e.g.,
illustrated in Niigata and Alaska, 1964, Loma Prieta, 1989, [5,6]), the uncertainty concerning random loading still
Kobe, 1995 and Chi-Chi, 1999 earthquakes. persists.
Several methods have been developed to determine Dobry et al. [7] demonstrated that the initiation of pore
liquefaction potential of soil deposits. The stress-based water pressure development in sands is exclusively depen-
procedure [1,2] is the most common method practiced by dent upon the shear strain level. The results of this study
engineers for liquefaction assessment. The shear stress level were used to conservatively assess the liquefaction potential
and the number of cycles are the major criteria in this of soils which has been known as the strain-based
approach. In order to correlate earthquake actual motion approach. This approach is based on the hypothesis that
to laboratory harmonic loading conditions, the equivalent pore pressure initiates to develop when shear strain
stress intensity and the number of cycles have to be defined surpasses a threshold shear strain. They showed that the
[1]. Seed et al. [3] selected this equivalent shear stress to be threshold shear strain is around 0.01%, independent of
as 65% of the maximum shear stress induced in the earth sand type, relative density, initial effective confining
structure. However, Ishihara and Yasuda [4], using a series pressure and sample preparation method. Although, the
of cyclic triaxial and torsional shear tests and time histories strain-based approach theoretically seems reasonable, it
has been less popular than the stress-based method. This
Corresponding author. Fax: +98 21 77240451. may be due to the fact that the strain procedure only
E-mail addresses: baziar@iust.ac.ir (M.H. Baziar), predicts the initiation of pore pressure buildup, which is
yjafarianm@iust.ac.ir (Y. Jafarian). essential for liquefaction to occur, but does not imply that

0267-7261/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.soildyn.2007.03.007
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1057

liquefaction will occur. The same uncertainty still exists in artificial neural network (ANN) model has been proposed
this approach on the subject of random loading since they to correlate W to the soil initial parameters. This ANN-
[7] assumed the same equivalent number of earthquake based model of capacity energy can be employed in energy-
loading as the stress-based method. based liquefaction assessment of sand–silt deposits.
During the past three decades, many researchers have
tried to explore a correlation between potential of pore 2. Previous energy-based models
pressure buildup and stress–strain dissipated energy in
liquefiable soils. These studies have established a founda- When an energy-based pore pressure buildup model is
tion for the energy-based liquefaction assessment ap- employed for liquefaction assessment of level ground soil
proaches. The use of energy concept in liquefaction deposits, the only required step in the pore pressure
assessment of soils is a logical step due to the fact that buildup procedure is the final step (ru ¼ 1), which indicates
the parameters used in this approach can be directly related the liquefaction triggering (e.g., [9,16,17]). This is the
to seismological parameters [8]. Pioneers such as Nemat- condition in which the excess pore water pressure reaches
Nasser and Shokooh [9] presented functional relationships, the initial effective confining pressure. A relationship
relating the accumulated strain energy dissipated in a unit between capacity energy (W), soil initial parameters and
volume of a granular soil to the pore pressure buildup. This model calibration parameters could be derived by setting
accumulative dissipated energy per unit volume of the soil ru ¼ 1 at any energy-based pore pressure buildup model.
mass, denoted as dissipated energy density (DW), can be Calibration parameters, implemented in the majority of
determined by calculating the area of the hysteresis these models, should be obtained by curve fitting of
stress–strain loops generated in the soil mass under cyclic experimental data. In other words, they need at least one
loading. set of laboratory cyclic test results before implementing the
The fact that the shear energy required to liquefy a soil model. Although correlation relationships were proposed
deposit is independent of the stress history has become the for calibration parameters in some of these models, their
main advantage of the energy approach. It was shown that accuracy is not satisfactory for all soils and conditions.
there exist a unique relationship between the dissipated Furthermore, many of these models lack a specified
shear energy and the pore water pressure buildup, procedure to determine the calibration parameters.
independent of the shear stress history [10]. In addition, Excluding inherent errors in the mentioned models and
Liang et al. [11], applying random and sinusoidal excita- difficulties regarding determination of calibration para-
tions to the samples of Reid Bedford sand, concluded that meters, none of those laboratory models is guarantied to be
the strain energy required for liquefaction triggering is generalized for various soil types and loading conditions.
independent of applied load pattern (harmonic or ran- They have been developed on the basis of experimental
dom). Therefore, it is not necessary to decompose the time data obtained from limited number of cyclic tests on a
history of shear stress to find an equivalent number of certain soil within small span range of initial conditions
cycles for a chosen average stress or strain level. In other such as initial void ratio and initial effective stress. Such
words, there is a solid and simple linkage between limitations are also noticed in the more recent model [16].
laboratory and field behavior in contrast with stress and This model contains only one calibration parameter and
strain-based approaches [11]. was developed based on the stress controlled data of cyclic
Various studies have been carried out to propose energy- triaxial tests [18] on Yatesville and Monterey sands and
based models relating pore water pressure increment ratio, silty sands. In their database, parameters such as relative
ru, to dissipated strain energy density, DW, loading density (Dr) and FC were extensively varied but initial
parameters such as cyclic stress ratio (CSR) or strain level, effective stress (s0 0) range was very narrow (93.74–
initial parameters of soils such as initial void ratio (e) or 103.1 kPa). Therefore, their proposed model probably
relative density (Dr), initial effective confining pressure would not appropriately work when the initial effective
(s0 0), and some calibration parameters obtained from curve stress is not around 100 kPa.
fitting of experimental data. Based on those energy-based On the other hand, while it has been proven that some
pore pressure buildup models, several liquefaction evalua- parameters such as FC are effective parameters on the
tion procedures have been developed (e.g., [12–15]). The liquefaction behavior of soil deposits, only one group of
objective of current study is to find a generalized relation- researchers [17] incorporated FC as a determinative
ship between the strain energy density required to induce parameter in their model. However, their model is limited
liquefaction in a specified soil (W) and soil initial up to 37% FC and is associated with three calibration
parameters. This level of imparted energy density, denoted parameters, two of which are correlated to the FC
as capacity energy of the soil, indicates whether liquefac- percentage and an approximate range of values is given
tion is triggered in the soil deposit. for the third one. Rationally, applying this model to soils
Employing a comprehensive and reliable database of with more than 37% fines might lead to unacceptable
cyclic triaxial, torsional and simple shear tests on different results.
sands and silty sands with various initial relative densities, Numerous investigations on the use of energy concept in
effective stresses and non-plastic fines contents (FC), an pore pressure buildup and liquefaction of soils have been
 ARTICLE IN PRESS
1058 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

carried out in Case Western Reserve University. Based on percentage of fines content, FC (%), coefficient of uniformity,
cyclic torsional shear tests, performed on Reid Bedford, Cu, mean grain size, D50 (mm), coefficient of curvature, Cc,
LSI-30 sands and Lower San Fernando Dam (LSFD) silty and measured strain energy density required for triggering
sand [15,19–21], and also centrifuge tests [22], several liquefaction (capacity energy), W (J/m3), which is the
relationships between the soil capacity energy (W), strain accumulative area of stress–strain loops up to the liquefaction
amplitude and some of soil initial parameters such as Dr triggering. From Table 1, it can be seen that s0 mean varies
and s0 0 were proposed. All the proposed relationships have between 41.1 and 294 kPa, while Dr and FC are varying
a similar form because they were thoroughly derived by between 44.5–105.1% and 0–100%, respectively. Also, Cu,
multiple linear regression (MLR) method. D50 and Cc respectively, range from 1.57 to 5.88, 0.03 to 0.46
Based on the results of 27 strain controlled cyclic and 0.74 to 1.61. In addition, the table contains data of some
torsional shear tests on Reid Bedford sand, the researchers element tests under random loading (non-harmonic).
at Case Western Reserve University [19] developed one of The database was randomly divided into two separate
these relationships which is presented in the following groups denoted as training and testing sets consisting about
Equation: 70% and 30% of data, respectively. The testing set was
utilized to determine when training should be stopped to
Log W ¼ 2:002 þ 0:00477s0c þ 0:0116Dr with R2 ¼ 0:937, avoid overfitting. In order to obtain a consistent data
(1) division [30], several combinations of the training and
where, W ¼ required strain energy for liquefaction trigger- testing sets were experienced. The above selection was such
ing in unit volume of the soil (J/m3), s0 c ¼ initial effective that the maximum, minimum, mean and standard devia-
confining pressure (kPa); and Dr ¼ initial relative density tion of parameters were consistent in training and testing
of the sample after consolidation (%). Indicating indepen- data sets (Table 2).
dence of dissipated shear energy density in the granular In this study, the necessity for developing an ANN-based
soils from the pattern of loading as an advantage of energy model for soil capacity energy is primarily examined via
approach, they [19] concluded that the amplitude of assessing other models. In order to assess the performance
applied shear strain, for the range of shear strain amplitude of MLR-based models, the model proposed by Figueroa et
used in their testing program, is not very effective in their al. [19] and a new MLR-based model, developed in this
model. Although the coefficient of determination in this study using the current database (Table 1), were evaluated.
model is high (R2 ¼ 0.937), the fact that it was developed Figs. 1 and 2 illustrate the performance of these two models
based on limited number of cyclic tests on clean sand rises for the data presented in Table 1. Similar to Figueroa et al.
doubts on its generality. Hence, it may not appropriately [19] model, the new MLR-based model was developed
work for sand–silt mixtures. Other relationships presented between logarithm of capacity energy and soil parameters.
by Figueroa and his co-workers, for soil capacity energy, All soil parameters cited in Table 1 excluding Cc, which
are similar to this equation with different coefficients. does not increase precision of the model, were incorporated
in the new MLR model. In fact, the new MLR model was
developed using five input parameters as indicated in
3. The ANN model and database
Log W ¼ 2:1028 þ 0:004566s0mean þ 0:005685Dr þ 0:001821FC
ANN, a powerful tool for statistical data manipulation,  0:02868C u þ 2:0214D50 ;
have been used in many complicated geotechnical en-
with R2 ¼ 0:65; MSE ¼ 7:15%; MAE ¼ 21:33%:
gineering problems such as stress–strain modeling of soils
[23], piles bearing capacity [24], settlement of shallow ð2Þ
foundations [25], earthquake induced liquefaction [26–28]
and seismic lateral spreading [29]. According to Figs. 1 and 2, none of the two models can
In order to present a general and well-built ANN-based predict the capacity energy with a good degree of accuracy.
model for various types of soils with various initial states, a Coefficient of determination (R2), mean squared error
wide-range database was collected from previously published (MSE) and mean absolute error (MAE) are obtained to be
cyclic tests. The database contains 217 cyclic triaxial [8], 61 0.36, 21.6% and 38.6%, respectively, for Figueroa et al.
cyclic torsional shear [10,15] and six cyclic simple shear tests [19] model and 0.65, 7.15% and 21.33%, respectively, for
(VELACS project), presented in Table 1. This table presents a the new MLR-based model. The figures suggest that
total of 284 cyclic triaxial, torsional and simple shear element although the new MLR-based model yields a better
tests on Monterey, Yatesville, Reid Bedford, LSFD, LSI-30, performance than the Figueroa et al. [19] model, its
Toyoura and Nevada 40% clean and silty sands. The criteria precision is not still acceptable.
for failure (liquefaction triggering) is initial liquefaction It should be noted that it is not possible to evaluate the
(ru ¼ 1) or double amplitude of strain of 5% (eDA ¼ 5%), other previously described energy-based pore pressure
whichever occurs first. The other cited information in Table 1 buildup models by the present database, since they
consists of soil initial effective mean confining pressure, s0 mean need calibration parameters. In addition, the correlation
(kPa), initial relative density after consolidation, Dr (%), relationships proposed for the calibration parameters in
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1059

Table 1
The database of element tests incorporated in the ANN model

Test # s0 mean (kPa) Dr (%) FC (%) Cu D50 (mm) Cc W (J/m3) Failure mode

Monterey sand and silty sand, cyclic triaxial, [8]

1 99.74 81.1 0 1.63 0.46 1.09 15220 ru ¼ 1
2 99.84 81.1 0 1.63 0.46 1.09 11330 ru ¼ 1
3 100.20 81.1 0 1.63 0.46 1.09 10340 ru ¼ 1
4 100.40 65.3 0 1.63 0.46 1.09 7680 ru ¼ 1
5 99.11 64.2 0 1.63 0.46 1.09 8260 ru ¼ 1
6 99.53 64.2 0 1.63 0.46 1.09 7520 ru ¼ 1
7 100.00 57.4 0 1.63 0.46 1.09 16420 ru ¼ 1
8 99.11 58.4 0 1.63 0.46 1.09 5020 ru ¼ 1
9 99.74 58.4 0 1.63 0.46 1.09 4180 ru ¼ 1
10 98.16 46.3 0 1.63 0.46 1.09 4180 ru ¼ 1
11 100.80 46.8 0 1.63 0.46 1.09 3880 ru ¼ 1
12 97.74 46.8 0 1.63 0.46 1.09 3520 ru ¼ 1
13 100.00 47.4 0 1.63 0.46 1.09 8350 ru ¼ 1
14 99.63 47.4 0 1.63 0.46 1.09 4020 ru ¼ 1
15 99.21 47.9 0 1.63 0.46 1.09 3590 ru ¼ 1
16 100.40 42.6 0 1.63 0.46 1.09 6260 ru ¼ 1
17 99.95 42.6 0 1.63 0.46 1.09 4190 ru ¼ 1
18 98.79 42.9 0 1.63 0.46 1.09 3370 ru ¼ 1
19 99.53 4.7 0 1.63 0.46 1.09 1040 ru ¼ 1
20 99.95 3.7 0 1.63 0.46 1.09 480 ru ¼ 1
21 101.10 82.1 5 1.64 0.44 1.09 12320 ru ¼ 1
22 100.80 79.9 5 1.64 0.44 1.09 12160 ru ¼ 1
23 100.90 81.6 5 1.64 0.44 1.09 11530 ru ¼ 1
24 101.40 64.7 5 1.64 0.44 1.09 10170 ru ¼ 1
25 100.50 64.4 5 1.64 0.44 1.09 7110 ru ¼ 1
26 99.53 63.3 5 1.64 0.44 1.09 7450 ru ¼ 1
27 99.63 53.4 5 1.64 0.44 1.09 18940 ru ¼ 1
28 100.70 54.0 5 1.64 0.44 1.09 9380 ru ¼ 1
29 99.21 54.5 5 1.64 0.44 1.09 5030 ru ¼ 1
30 97.42 55.6 5 1.64 0.44 1.09 4890 ru ¼ 1
31 98.90 49.7 5 1.64 0.44 1.09 20970 ru ¼ 1
32 99.84 49.7 5 1.64 0.44 1.09 5810 ru ¼ 1
33 93.74 53.4 5 1.64 0.44 1.09 6170 ru ¼ 1
34 98.48 81.0 10 1.64 0.42 1.09 22200 ru ¼ 1
35 99.84 79.7 10 1.64 0.42 1.09 17740 ru ¼ 1
36 99.95 80.6 10 1.64 0.42 1.09 14350 ru ¼ 1
37 100.30 62.9 10 1.64 0.42 1.09 11080 ru ¼ 1
38 98.69 62.9 10 1.64 0.42 1.09 8370 ru ¼ 1
39 99.53 52.2 10 1.64 0.42 1.09 7610 ru ¼ 1
40 99.21 51.7 10 1.64 0.42 1.09 5710 ru ¼ 1
41 99.53 51.7 10 1.64 0.42 1.09 5170 ru ¼ 1
42 99.11 53.3 10 1.64 0.42 1.09 3690 ru ¼ 1
43 100.20 54.1 10 1.64 0.42 1.09 3400 ru ¼ 1
44 99.95 55.2 10 1.64 0.42 1.09 2810 ru ¼ 1
45 100.90 81.7 15 1.65 0.40 1.08 12380 ru ¼ 1
46 99.84 82.1 15 1.65 0.40 1.08 11510 ru ¼ 1
47 99.53 81.7 15 1.65 0.40 1.08 15340 ru ¼ 1
48 98.58 62.5 15 1.65 0.40 1.08 6130 ru ¼ 1
49 98.58 62.9 15 1.65 0.40 1.08 4490 ru ¼ 1
50 98.58 10.8 15 1.65 0.40 1.08 2690 ru ¼ 1
51 99.32 11.2 15 1.65 0.40 1.08 2610 ru ¼ 1
52 99.21 10.8 15 1.65 0.40 1.08 2470 ru ¼ 1
53 99.00 96.8 20 1.65 0.37 1.08 27590 ru ¼ 1
54 97.84 97.5 20 1.65 0.37 1.08 21070 ru ¼ 1
55 99.42 96.8 20 1.65 0.37 1.08 34970 e ¼ 5%
56 99.42 60.8 20 1.65 0.37 1.08 10400 ru ¼ 1
57 99.32 61.1 20 1.65 0.37 1.08 5980 ru ¼ 1
58 98.90 60.8 20 1.65 0.37 1.08 4240 ru ¼ 1
59 99.32 60.4 20 1.65 0.37 1.08 4540 ru ¼ 1
60 97.74 2.1 20 1.65 0.37 1.08 2100 ru ¼ 1
61 98.37 1.1 20 1.65 0.37 1.08 2100 ru ¼ 1
62 99.53 97.9 25 1.66 0.35 1.08 19740 ru ¼ 1
 ARTICLE IN PRESS
1060 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

Table 1 (continued )

Test # s0 mean (kPa) Dr (%) FC (%) Cu D50 (mm) Cc W (J/m3) Failure mode

63 99.21 97.9 25 1.66 0.35 1.08 27650 e ¼ 5%

64 99.11 98.3 25 1.66 0.35 1.08 21420 e ¼ 5%
65 99.95 75.9 25 1.66 0.35 1.08 8800 ru ¼ 1
66 99.21 76.2 25 1.66 0.35 1.08 5320 ru ¼ 1
67 99.63 75.5 25 1.66 0.35 1.08 5160 ru ¼ 1
68 98.60 60.7 25 1.66 0.35 1.08 3650 ru ¼ 1
69 99.21 60.0 25 1.66 0.35 1.08 4230 ru ¼ 1
70 99.42 61.4 25 1.66 0.35 1.08 3620 ru ¼ 1
71 99.11 2.1 25 1.66 0.35 1.08 2200 ru ¼ 1
72 95.95 0.3 25 1.66 0.35 1.08 2290 ru ¼ 1
73 97.42 1.7 25 1.66 0.35 1.08 2260 e ¼ 5%
74 100.20 55.8 35 1.68 0.31 1.08 560 e ¼ 5%
75 99.42 69.2 35 1.68 0.31 1.08 1000 ru ¼ 1
76 98.79 62.5 35 1.68 0.31 1.08 920 e ¼ 5%
77 99.42 67.6 35 1.68 0.31 1.08 570 ru ¼ 1
78 98.79 72.8 35 1.68 0.31 1.08 920 e ¼ 5%
79 97.00 73.9 50 1.72 0.25 1.07 1250 e ¼ 5%
80 98.90 76.9 50 1.72 0.25 1.07 1890 ru ¼ 1
81 98.37 77.3 50 1.72 0.25 1.07 1370 ru ¼ 1
82 98.68 79.6 50 1.72 0.25 1.07 1850 ru ¼ 1
83 98.58 92.0 75 1.87 0.14 1.03 1660 e ¼ 5%
84 100.50 90.5 75 1.87 0.14 1.03 1590 e ¼ 5%
Yatesville sand and silty sand, cyclic triaxial, [8]
85 99.10 68.0 0 2.44 0.17 0.85 2460 ru ¼ 1
86 99.63 70.0 0 2.44 0.17 0.85 2590 ru ¼ 1
87 98.58 68.3 0 2.44 0.17 0.85 2540 ru ¼ 1
88 99.42 27.3 0 2.44 0.17 0.85 490 ru ¼ 1
89 101.30 24.8 0 2.44 0.17 0.85 620 ru ¼ 1
90 98.05 25.6 0 2.44 0.17 0.85 1420 ru ¼ 1
91 99.42 10.4 0 2.44 0.17 0.85 320 ru ¼ 1
92 100.60 32.7 0 2.44 0.17 0.85 300 ru ¼ 1
93 99.74 44.5 0 2.44 0.17 0.85 370 ru ¼ 1
94 98.47 37.2 0 2.44 0.17 0.85 370 ru ¼ 1
95 99.84 32.3 0 2.44 0.17 0.85 530 ru ¼ 1
96 98.26 32.0 0 2.44 0.17 0.85 490 ru ¼ 1
97 99.53 27.5 4 2.45 0.16 0.85 380 ru ¼ 1
98 98.16 22.9 4 2.45 0.16 0.85 670 ru ¼ 1
99 98.16 28.2 4 2.45 0.16 0.85 740 ru ¼ 1
100 100.30 19.6 4 2.45 0.16 0.85 880 ru ¼ 1
101 99.84 52.1 4 2.45 0.16 0.85 2560 ru ¼ 1
102 100.80 52.4 4 2.45 0.16 0.85 1740 ru ¼ 1
103 99.21 52.1 4 2.45 0.16 0.85 2010 ru ¼ 1
104 100.80 29.5 4 2.45 0.16 0.85 1360 ru ¼ 1
105 99.31 30.8 4 2.45 0.16 0.85 1000 ru ¼ 1
106 100.60 29.5 4 2.45 0.16 0.85 1210 ru ¼ 1
107 96.89 10.9 4 2.45 0.16 0.85 430 ru ¼ 1
108 99.53 9.0 4 2.45 0.16 0.85 420 ru ¼ 1
109 96.79 9.6 4 2.45 0.16 0.85 410 ru ¼ 1
110 97.74 36.5 4 2.45 0.16 0.85 310 ru ¼ 1
111 99.84 35.5 4 2.45 0.16 0.85 580 ru ¼ 1
112 98.89 30.2 4 2.45 0.16 0.85 610 ru ¼ 1
113 97.21 34.0 7 2.46 0.16 0.85 1950 ru ¼ 1
114 98.47 34.6 7 2.46 0.16 0.85 2040 ru ¼ 1
115 100.60 33.7 7 2.46 0.16 0.85 1330 ru ¼ 1
116 98.05 33.7 7 2.46 0.16 0.85 1770 ru ¼ 1
117 100.50 27.3 7 2.46 0.16 0.85 2440 ru ¼ 1
118 98.05 29.4 7 2.46 0.16 0.85 1930 ru ¼ 1
119 100.90 30.7 7 2.46 0.16 0.85 1810 ru ¼ 1
120 99.84 9.0 7 2.46 0.16 0.85 490 ru ¼ 1
121 99.74 8.7 7 2.46 0.16 0.85 550 ru ¼ 1
122 96.47 5.5 7 2.46 0.16 0.85 590 ru ¼ 1
123 100.90 20.4 7 2.46 0.16 0.85 510 ru ¼ 1
124 102.50 31.7 7 2.46 0.16 0.85 330 ru ¼ 1
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1061

Table 1 (continued )

Test # s0 mean (kPa) Dr (%) FC (%) Cu D50 (mm) Cc W (J/m3) Failure mode

125 97.63 25.3 7 2.46 0.16 0.85 510 ru ¼ 1

126 98.05 25.3 7 2.46 0.16 0.85 610 ru ¼ 1
127 99.10 45.2 12 2.47 0.15 0.85 2780 ru ¼ 1
128 100.40 42.4 12 2.47 0.15 0.85 3480 ru ¼ 1
129 102.40 43.9 12 2.47 0.15 0.85 3890 ru ¼ 1
130 99.63 54.4 12 2.47 0.15 0.85 2810 ru ¼ 1
131 98.89 55.1 12 2.47 0.15 0.85 1590 e ¼ 5%
132 99.00 56.9 12 2.47 0.15 0.85 1560 ru ¼ 1
133 100.30 25.7 12 2.47 0.15 0.85 2120 ru ¼ 1
134 99.53 24.7 12 2.47 0.15 0.85 1240 ru ¼ 1
135 98.89 27.5 12 2.47 0.15 0.85 1790 ru ¼ 1
136 101.10 3.4 12 2.47 0.15 0.85 780 ru ¼ 1
137 100.60 2.8 12 2.47 0.15 0.85 1650 e ¼ 5%
138 101.00 4.6 12 2.47 0.15 0.85 2140 ru ¼ 1
139 98.68 11.1 12 2.47 0.15 0.85 410 ru ¼ 1
140 100.90 11.8 12 2.47 0.15 0.85 430 ru ¼ 1
141 99.31 9.6 12 2.47 0.15 0.85 610 ru ¼ 1
142 98.89 9.9 12 2.47 0.15 0.85 1050 ru ¼ 1
143 98.89 76.1 17 2.48 0.15 0.85 6630 e ¼ 5%
144 100.60 77.8 17 2.48 0.15 0.85 7170 e ¼ 5%
145 100.30 76.3 17 2.48 0.15 0.85 7080 ru ¼ 1
146 99.00 37.7 17 2.48 0.15 0.85 610 ru ¼ 1
147 98.37 41.0 17 2.48 0.15 0.85 980 ru ¼ 1
148 102.80 38.3 17 2.48 0.15 0.85 2090 ru ¼ 1
149 98.68 25.3 17 2.48 0.15 0.85 1500 ru ¼ 1
150 100.30 26.0 17 2.48 0.15 0.85 1040 ru ¼ 1
151 100.20 27.7 17 2.48 0.15 0.85 1200 ru ¼ 1
152 98.26 18.0 17 2.48 0.15 0.85 1440 ru ¼ 1
153 100.10 18.0 17 2.48 0.15 0.85 1310 ru ¼ 1
154 99.31 17.0 17 2.48 0.15 0.85 1670 ru ¼ 1
155 97.21 10.7 17 2.48 0.15 0.85 640 ru ¼ 1
156 97.32 7.7 17 2.48 0.15 0.85 800 ru ¼ 1
157 97.74 6.7 17 2.48 0.15 0.85 880 e ¼ 5%
158 99.95 40.3 26 2.51 0.13 0.85 1340 ru ¼ 1
159 99.74 42.8 26 2.51 0.13 0.85 2100 ru ¼ 1
160 99.42 43.2 26 2.51 0.13 0.85 2680 ru ¼ 1
161 100.20 26.9 26 2.51 0.13 0.85 1120 ru ¼ 1
162 98.89 28.1 26 2.51 0.13 0.85 1350 ru ¼ 1
163 100.90 28.5 26 2.51 0.13 0.85 1060 ru ¼ 1
164 100.50 24.9 26 2.51 0.13 0.85 1100 ru ¼ 1
165 99.00 31.0 26 2.51 0.13 0.85 1100 ru ¼ 1
166 97.95 27.7 26 2.51 0.13 0.85 1580 ru ¼ 1
167 99.21 10.2 26 2.51 0.13 0.85 580 ru ¼ 1
168 100.90 9.0 26 2.51 0.13 0.85 1240 ru ¼ 1
169 99.74 9.0 26 2.51 0.13 0.85 910 ru ¼ 1
170 102.60 25.1 37 2.56 0.12 0.85 750 ru ¼ 1
171 99.63 25.8 37 2.56 0.12 0.85 1080 ru ¼ 1
172 100.60 24.4 37 2.56 0.12 0.85 2480 ru ¼ 1
173 99.84 4.5 37 2.56 0.12 0.85 490 ru ¼ 1
174 99.42 8.4 37 2.56 0.12 0.85 740 ru ¼ 1
175 99.95 7.7 37 2.56 0.12 0.85 600 ru ¼ 1
176 98.79 1.0 37 2.56 0.12 0.85 400 ru ¼ 1
177 98.47 0.0 37 2.56 0.12 0.85 890 ru ¼ 1
178 99.63 3.8 37 2.56 0.12 0.85 1830 e ¼ 5%
179 95.32 24.7 37 2.56 0.12 0.85 320 ru ¼ 1
180 98.47 11.1 37 2.56 0.12 0.85 580 ru ¼ 1
181 96.68 16.4 37 2.56 0.12 0.85 880 e ¼ 5%
182 100.40 81.1 50 2.63 0.10 0.85 1070 ru ¼ 1
183 100.10 80.0 50 2.63 0.10 0.85 2180 e ¼ 5%
184 100.20 91.2 50 2.63 0.10 0.85 1870 e ¼ 5%
185 99.31 72.2 50 2.63 0.10 0.85 740 ru ¼ 1
186 99.10 69.8 50 2.63 0.10 0.85 1370 ru ¼ 1
187 98.79 71.6 50 2.63 0.10 0.85 1140 ru ¼ 1
188 99.84 63.4 50 2.63 0.10 0.85 650 e ¼ 5%
 ARTICLE IN PRESS
1062 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

Table 1 (continued )

Test # s0 mean (kPa) Dr (%) FC (%) Cu D50 (mm) Cc W (J/m3) Failure mode

189 98.68 62.2 50 2.63 0.10 0.85 730 e ¼ 5%

190 96.37 58.6 50 2.63 0.10 0.85 1130 e ¼ 5%
191 99.95 62.0 50 2.63 0.10 0.85 1100 ru ¼ 1
192 99.53 54.4 50 2.63 0.10 0.85 740 e ¼ 5%
193 100.50 69.3 50 2.63 0.10 0.85 900 e ¼ 5%
194 102.50 85.5 75 2.91 0.07 0.85 530 ru ¼ 1
195 97.95 86.6 75 2.91 0.07 0.85 1010 e ¼ 5%
196 103.10 85.8 75 2.91 0.07 0.85 1180 e ¼ 5%
197 96.47 79.7 75 2.91 0.07 0.85 1110 e ¼ 5%
198 97.53 78.0 75 2.91 0.07 0.85 1280 ru ¼ 1
199 96.89 78.2 75 2.91 0.07 0.85 1570 ru ¼ 1
200 99.53 75.1 75 2.91 0.07 0.85 850 e ¼ 5%
201 99.84 76.9 75 2.91 0.07 0.85 1040 e ¼ 5%
202 99.42 74.2 75 2.91 0.07 0.85 1010 e ¼ 5%
203 99.95 105.1 100 4.44 0.03 0.90 3710 ru ¼ 1
204 101.10 104.3 100 4.44 0.03 0.90 3070 ru ¼ 1
205 100.60 103.3 100 4.44 0.03 0.90 5020 ru ¼ 1
206 98.79 97.5 100 4.44 0.03 0.90 8320 ru ¼ 1
207 99.53 95.9 100 4.44 0.03 0.90 2800 ru ¼ 1
208 99.31 98.5 100 4.44 0.03 0.90 2330 ru ¼ 1
209 98.47 85.6 100 4.44 0.03 0.90 840 ru ¼ 1
210 100.90 86.2 100 4.44 0.03 0.90 1840 e ¼ 5%
211 98.47 87.2 100 4.44 0.03 0.90 1800 e ¼ 5%
212 100.30 87.6 100 4.44 0.03 0.90 980 e ¼ 5%
213 99.95 88.4 100 4.44 0.03 0.90 1070 e ¼ 5%
214 100.10 87.2 100 4.44 0.03 0.90 1380 e ¼ 5%
215 99.63 84.3 100 4.44 0.03 0.90 880 ru ¼ 1
216 99.31 84.6 100 4.44 0.03 0.90 1000 e ¼ 5%
217 100.70 84.7 100 4.44 0.03 0.90 2300 e ¼ 5%

Reid Bedford sand, cyclic torsional, [15]

218 41.40 51.0 0 1.67 0.26 0.74 593 ru ¼ 1
219 82.70 54.5 0 1.67 0.26 0.74 1089 ru ¼ 1
220 124.10 52.9 0 1.67 0.26 0.74 1503 ru ¼ 1
221 41.40 59.1 0 1.67 0.26 0.74 738 ru ¼ 1
222 82.70 61.1 0 1.67 0.26 0.74 1338 ru ¼ 1
223 124.10 59.5 0 1.67 0.26 0.74 1813 ru ¼ 1
224 41.40 75.5 0 1.67 0.26 0.74 1338 ru ¼ 1
225 82.70 67.2 0 1.67 0.26 0.74 1758 ru ¼ 1
226 124.10 68.8 0 1.67 0.26 0.74 2303 ru ¼ 1
227 41.40 52.9 0 1.67 0.26 0.74 669 ru ¼ 1
228 82.70 51.5 0 1.67 0.26 0.74 1034 ru ¼ 1
229 124.10 49.7 0 1.67 0.26 0.74 1496 ru ¼ 1
230 41.40 58.6 0 1.67 0.26 0.74 689 ru ¼ 1
231 82.70 59.3 0 1.67 0.26 0.74 1317 ru ¼ 1
232 124.10 63.8 0 1.67 0.26 0.74 1758 ru ¼ 1
233 41.40 67.4 0 1.67 0.26 0.74 758 ru ¼ 1
234 82.70 71.9 0 1.67 0.26 0.74 1606 ru ¼ 1
235 124.10 72.6 0 1.67 0.26 0.74 2386 ru ¼ 1
236 41.40 52.0 0 1.67 0.26 0.74 607 ru ¼ 1
237 82.70 53.2 0 1.67 0.26 0.74 1241 ru ¼ 1
238 124.10 52.8 0 1.67 0.26 0.74 1717 ru ¼ 1
239 41.40 60.7 0 1.67 0.26 0.74 703 ru ¼ 1
240 82.70 61.2 0 1.67 0.26 0.74 1303 ru ¼ 1
241 124.10 60.1 0 1.67 0.26 0.74 2144 ru ¼ 1
242 41.40 68.4 0 1.67 0.26 0.74 1096 ru ¼ 1
243 82.70 69.8 0 1.67 0.26 0.74 2068 ru ¼ 1
244 124.10 70.9 0 1.67 0.26 0.74 2737 ru ¼ 1
Reid Bedford sand-random loading, cyclic torsional shear, [15]
245 41.10 48.6 0 1.67 0.26 0.74 691 ru ¼ 1
246 41.10 57.5 0 1.67 0.26 0.74 851 ru ¼ 1
247 41.10 71.4 0 1.67 0.26 0.74 1094 ru ¼ 1
248 82.70 54.9 0 1.67 0.26 0.74 1150 ru ¼ 1
249 82.70 58.3 0 1.67 0.26 0.74 1538 ru ¼ 1
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1063

Table 1 (continued )

Test # s0 mean (kPa) Dr (%) FC (%) Cu D50 (mm) Cc W (J/m3) Failure mode

250 82.70 67.5 0 1.67 0.26 0.74 1955 ru ¼ 1

251 124.10 51.2 0 1.67 0.26 0.74 1319 ru ¼ 1
252 124.10 59.9 0 1.67 0.26 0.74 1943 ru ¼ 1
253 124.10 69.0 0 1.67 0.26 0.74 2517 ru ¼ 1
LSFD sand-random loading, cyclic torsional shear, [15]
254 41.10 57.2 28 5.88 0.13 1.61 517 ru ¼ 1
255 41.10 63.7 28 5.88 0.13 1.61 525 ru ¼ 1
256 41.10 74.6 28 5.88 0.13 1.61 522 ru ¼ 1
257 41.10 86.9 28 5.88 0.13 1.61 517 ru ¼ 1
258 82.70 60.0 28 5.88 0.13 1.61 797 ru ¼ 1
259 82.70 59.9 28 5.88 0.13 1.61 802 ru ¼ 1
260 82.70 71.2 28 5.88 0.13 1.61 792 ru ¼ 1
261 82.70 79.2 28 5.88 0.13 1.61 811 ru ¼ 1
262 82.70 87.4 28 5.88 0.13 1.61 827 ru ¼ 1
263 124.10 61.1 28 5.88 0.13 1.61 1294 ru ¼ 1
264 124.10 68.9 28 5.88 0.13 1.61 1232 ru ¼ 1
265 124.10 77.2 28 5.88 0.13 1.61 1246 ru ¼ 1
266 124.10 91.7 28 5.88 0.13 1.61 1379 ru ¼ 1
LSI-30 sand, cyclic torsional shear, [15]
267 41.10 49.1 0 2.39 0.39 0.81 839 ru ¼ 1
268 41.10 62.8 0 2.39 0.39 0.81 890 ru ¼ 1
269 41.10 71.9 0 2.39 0.39 0.81 1544 ru ¼ 1
270 82.70 50.1 0 2.39 0.39 0.81 1273 ru ¼ 1
271 82.70 62.3 0 2.39 0.39 0.81 2167 ru ¼ 1
272 82.70 69.9 0 2.39 0.39 0.81 1991 ru ¼ 1
273 124.10 51.7 0 2.39 0.39 0.81 1771 ru ¼ 1
274 124.10 65.5 0 2.39 0.39 0.81 2574 ru ¼ 1
275 124.10 67.6 0 2.39 0.39 0.81 4098 ru ¼ 1

Toyoura sand, cyclic torsional, [10]

276 294.00 43.1 0 1.57 0.19 1.05 6300 ru ¼ 1
277 294.00 44.2 0 1.57 0.19 1.05 5956 ru ¼ 1
278 294.00 50.7 0 1.57 0.19 1.05 5000 ru ¼ 1
Nevada 40% and Fraser River sands, cyclic simple shear, VELACS project
279 80.00 45.9 0 2.11 0.18 1.50 489 e ¼ 5%
280 160.00 42.6 0 2.11 0.18 1.50 1180 e ¼ 5%
281 100.00 80.0 0 1.65 0.26 1.11 8309 e ¼ 5%
282 200.00 81.0 0 1.65 0.26 1.11 23350 e ¼ 5%
283 100.00 40.0 0 1.65 0.26 1.11 520 e ¼ 5%
284 200.00 44.0 0 1.65 0.26 1.11 1400 e ¼ 5%

Table 2
Statistical characteristics of consistent training and testing sets

Parameters Training set Testing set

Max.a Min.b S.D.c Mean Max. Min. S.D. Mean

s0 mean (kPa) 294.0 41.1 27.64 99.39 294.0 41.1 28.50 97.67
Dr (%) 105.1 44.5 33.70 47.78 104.3 36.5 33.14 47.62
FC (%) 100.0 0.0 26.37 20.28 100.0 0.0 26.35 20.81
Cu 5.88 1.57 1.04 2.38 5.88 1.57 0.99 2.39
D50 0.46 0.03 0.13 0.23 0.44 0.03 0.12 0.23
Cc 1.61 0.74 0.20 0.96 1.61 0.74 0.18 0.92
Log W (J/m3) 4.54 2.48 0.45 3.27 4.35 2.49 0.45 3.27
a
Maximum.
b
Minimum.
c
Standard deviation.
 ARTICLE IN PRESS
1064 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

5.0 previously proposed laboratory models which need at least

Figueroa et al., 1994 Model
one cyclic test to find calibration parameters. Therefore, a
All element tests data
more powerful tool is required to evaluate the liquefaction
4.5 R2 = 0.36 capacity of sands and silty sands in terms of strain energy.
MSE = 21.6%
A batch mode feed-forward multilayer perceptron
Log of Measured Capacity Energy

MAE = 38.6%
(MLP) with back-propagation learning rules was used to
4.0 create the desired ANN model using MATLAB 6.5.
A gradient descent momentum was used in the training
process to bypass the probable local minima in error
3.5 measured=estimated surface. Also, an adaptive learning rate was employed to
keep the learning step size as large as possible while the
training is stable [31]. According to a universal approxima-
3.0 tion theorem, demonstrated concurrently by several
researchers (e.g., [32]) for traditional MLP, a single hidden
layer network is sufficient to uniformly approximate any
2.5 continuous and nonlinear function. The model architecture
was built with one hidden layer, a learning rate of 0.05
updated with a coefficient of 1.05 after each epoch and a
2.0
momentum term of 0.9 updated with a coefficient of 0.9
2.0 2.5 3.0 3.5 4.0 4.5 5.0
after each epoch. The input vector is fully connected to the
Log of Estimated Capacity Energy
hidden neurons by a tan-sigmoid transfer function and the
Fig. 1. Assessment of Figueroa et al. [19] capacity energy model using all neurons of hidden layer are fully connected to the output
element tests data in the database. layer via a linear function. Experimental studies were
started with two hidden neurons to reach the optimum
number of hidden neurons and desired precision [33]. Input
5.0 vector contains soil initial parameters and output (the
MLR-Based Model target vector) is logarithm of strain energy density required
all element tests data
R2= 0.65 to trigger liquefaction (capacity energy). In order to obtain
4.5 MSE = 7.15% a more efficient training process, the input and target were
MAE =21.33% standardized to have zero mean and unity standard
Log of Measured Capacity Energy

deviation. Cross-validation or employing another set of

4.0 measured=1.15* estimated
data for more testing can be used to increase the generality
of the models for future predictions,. In this study, 18
liquefaction triggering centrifuge tests data [22] were used
3.5 as another testing set, referred to as validation set here
(Table 4). In fact, several ANN models using element tests
data (Table 1) were constituted for generating the models.
3.0 Among them, the model with better performance (greater
measured=0.85*
estimated coefficient of determination and smaller MSE and MAE)
for validation data set was selected. In other words, the
2.5
ANN models were developed with the best performance
LogW=2.1028+0.004566 σ'mean+0.005685Dr+0.001821FC-
concurrently for training, testing and validation data sets.
0.02868Cu+2.0214D50
Since, the models are only generated based on the element
2.0
tests data and validated by the centrifuge tests, it is
2.0 2.5 3.0 3.5 4.0 4.5 5.0
reasonable to anticipate that the models are generalized
Log of Estimated Capacity Energy
enough to predict real earthquake induced liquefaction
Fig. 2. The MLR-based capacity energy model, developed in current triggering.
study using all element tests data in the database. An extensive trial study was performed to select the most
relevant input parameters for the ANN model. Five
different ANN models were developed using different
some of those models yield very approximate results for the combinations of input parameters in Table 1. Table 3
present database. summarizes the developed ANN models with their selected
input parameters. R2, MSE and MAE values for all data
4. Developing the ANN model sets are also shown in this table.
It can be seen from Table 3 that, except for model #1,
The complexity of the current problem was highlighted performances of the models are generally improved when
by the previous discussions on MLR-based models and input parameters are increased. Since Cc did not improve
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1065

18.53
the performance of the network, input parameters Cu and
Cc

—
D50 seem to be sufficient to represent the characteristics of
grains size distribution. Consequently, model #2 is
13.74

26.60

32.30
suggested to be the best ANN model. This model has 10
D50

—
hidden neurons and the number of epochs in which the
training, testing and validation data concurrently result to
17.66

12.48

17.91
best outputs is found to be 6500. Values of R2, MSE and
Cu

—
MAE produced from this model are 0.9, 2% and 10.4%,
respectively, for all element tests data and 0.9, 0.6% and
Relative importance (%)

22.76

27.48

34.41

28.23

40.20
FC

6.3% for centrifuge tests data.

18.91

23.60

24.52

25.01

31.96
5. Results and discussions
Dr

The ANN model #2, chosen as the most appropriate

28.85

27.84
model, was constituted by five inputs (s0 mean, Dr, FC, Cu
8.40

9.84

8.77
s0

and D50), one hidden layer with 10 neurons and one output
(log W). Figs. 3–5 illustrate the measured and predicted
MAE (%)

capacity energy values for training (R2 ¼ 0.94, MSE ¼

1.33%, MAE ¼ 8.46%), testing (R2 ¼ 0.82, MSE ¼ 3.7,
10.30

15.00
10.40

14.90
11.00

15.70
11.60

10.30
14.60
12.70
10.60
11.70
15.30
6.30
8.00
8.30

8.46

9.30
9.10

9.70

MAE ¼ 14.9) and all element test data (R2 ¼ 0.9,

MSE ¼ 2, MAE ¼ 10.4), respectively.
MSE (%)

Table 4 illustrates validation data set which consists of

several centrifuge ground level liquefaction tests. Dief [22]
0.60
2.30
1.00
1.30
3.80
2.00

1.33
3.70
2.40
1.40
1.60
4.30
2.60
1.30
2.20
3.70
3.10
1.70
2.60
4.30
Performance

performed these tests on Nevada and Reid Bedford clean

sands and LSFD silty sand at Case Western Reserve
0.90
0.90
0.83
0.94
0.81
0.90

0.94
0.82
0.88
0.78
0.92
0.80
0.87
0.76
0.90
0.81
0.85
0.86
0.88
0.78

University. He processed the recorded accelerations and

R2

lateral displacements of the laminar box segments using

lumped mass models and estimated the shear stress–strain
Centrifuge (validation)

Centrifuge (validation)

Centrifuge (validation)

Centrifuge (validation)
Centrifuge(validation)

history at different depths in centrifuge models. The

accumulative areas within stress–strain loops up to
All element tests

All element tests

All element tests

All element tests

All element tests

liquefaction triggering were calculated at various depths

[22] and are shown in Table 4 (W). Since these accumulated
Training

Training

Training

Training

Training
Testing

Testing

Testing

Testing

Testing

5.0
Training Data
10,000

6500

3000

5800

6000

R2= 0.94
10

10

11

10

10
Summary of all ANN models developed for various input parameters

4.5 MSE = 1.33%

MAE = 8.46%
Log of Measured Capacity Energy
Model architecture

Hidden Neurons

Hidden neurons

Hidden neurons

Hidden neurons

Hidden neurons

4.0 measured=1.1*estimated
Epochs

Epochs

Epochs

Epochs

Epochs

3.5

measured=0.9*estimated
s0 mean, Dr, FC, Cu, D50, Cc

3.0
s0 mean, Dr, FC, Cu, D50

s0 mean, Dr, FC, D50

s0 mean, Dr, FC, Cu

2.5
s0 mean, Dr, FC
INPUTS

2.0
2.0 2.5 3.0 3.5 4.0 4.5 5.0
Log of Estimated Capacity Energy
Model #
Table 3

Fig. 3. Predicted capacity energy by ANN model versus measured values

for training data.
1

5
 ARTICLE IN PRESS
1066 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

5.0 Table 4
Testing Data Data of centrifuge tests, validation data set (after [22]; Case Western
R2= 0.82 Reserve University)
4.5 MSE = 3.7%
Test Soil s0 mean Dr FC Cu D50 Cc W (J/
MAE = 14.9%
Log of Measured Capacity Energy

# type (kPa) (%) (%) (mm) m3)

4.0 Nevada
measured=1.1*estimated 1 27.8 49.7 0 2.11 0.18 1.50 370
2 28.7 60.7 0 2.11 0.18 1.50 590
3 33.9 58.5 0 2.11 0.18 1.50 600
3.5 4 28.4 64.7 0 2.11 0.18 1.50 785
5 34.4 66.5 0 2.11 0.18 1.50 935
6 34.7 72 0 2.11 0.18 1.50 1085
3.0 7 34.9 74 0 2.11 0.18 1.50 1740
measured=0.9*estimated 8 34.9 74.7 0 2.11 0.18 1.50 1775
Reid Bedford
2.5 9 34 51 0 1.67 0.26 0.74 700
10 34 60 0 1.67 0.26 0.74 900
11 34 65.5 0 1.67 0.26 0.74 1400
12 34 72.5 0 1.67 0.26 0.74 1700
2.0 13 34 81 0 1.67 0.26 0.74 1800
2.0 2.5 3.0 3.5 4.0 4.5 5.0
LSFD
Log of Estimated Capacity Energy 14 31 62.5 28 5.88 0.13 1.61 460
15 31 67 28 5.88 0.13 1.61 475
Fig. 4. Predicted capacity energy by ANN model versus measured values
16 31 72 28 5.88 0.13 1.61 500
for testing data.
17 31 88 28 5.88 0.13 1.61 510
18 31 95 28 5.88 0.13 1.61 520

5.0 W: calCulated energy density up to liquefaction triggering obtained from

All Element Tests Data lumped mass analyses.
R2= 0.9
MSE= 2%
4.5
MAE = 10.4%
3.5
Log of Measured Capacity Energy

Centrifuge Data, Validation set

4.0 measured=1.1*estimated R2= 0.9
MSE = 0.6%
MAE = 6.3%
Log of Measured Capacity Energy

3.3
3.5

measured=1.05*estimated
3.0
measured=0.9*estimated 3.0

2.5

measured=0.95*estimated
2.8
2.0
2.0 2.5 3.0 3.5 4.0 4.5 5.0
Log of Estimated Capacity Energy

Fig. 5. Predicted capacity energy by ANN model versus measured values

2.5
for all element tests data.
2.5 2.8 3.0 3.3 3.5
Log of Estimated Capacity Energy
dissipated energies were calculated up to liquefaction Fig. 6. Predicted capacity energy by ANN model versus measured values
triggering (ru ¼ 1), they represent the capacity of their for validation data set (centrifuge data).
corresponding soils similar to element tests. Fig. 6 depicts
measured versus predicted energy density for the validation
data set (R2 ¼ 0.9, MSE ¼ 0.6% and MAE ¼ 6.3%). determination (R2), MSE and MAE. The results shown
Target statistical parameters to monitor the performance in Figs. 3–6 indicate reasonable good performance because
of the network were chosen to be the coefficient of the predicted values are satisfactorily distributed between
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1067

two lines illustrating 0.9 and 1.1 times of measured values 6. Parametric study
for training and testing data sets and 0.95 and 1.05 times
for centrifuge data set. A parametric study was carried out to find variations of
A simple approach was proposed to interpret connection soil capacity energy (W) by the range of soil parameters
weight matrixes via a parameter denoted as relative values cited in Table 2 with other parameters kept constant
importance [34]. Relative importance values determine the at their average values. Parametric study was performed
contribution of each input parameter in the model. for further verification of the ANN model. Similar to
Incorporating input-hidden and hidden-output connection laboratory studies carried out by pioneers (e.g., [35]) on the
weights in that approach, relative importance values of dependency of liquefaction resistance of sands on initial
input parameters were calculated. Table 5 shows connec- relative density and effective confining stress, the results of
tion weights and relative importance of parameters for the the parametric study for these parameters confirm that the
chosen model (model #2 in Table 3). Also, relative capacity energy of sands continuously increases with
importance values for the other ANN models with various increasing initial relative density or effective confining
input parameters are presented in Table 3. From Table 5, stress with other parameters kept constant. Fig. 7
initial mean effective stress, relative density, FC, coefficient illustrates the results of the parametric study for FC. The
of uniformity and mean grain size represent 9.84%, effect of non-plastic FC on the liquefaction behavior of
23.58%, 27.48%, 12.48% and 26.61% relative importance, sand–silt mixtures is more complex than the effect of other
respectively. According to these results, it is interesting to parameters. Although several studies have indicated that
note that while capacity energy is more sensitive to FC sands deposits with silt content are much more susceptible
compared with relative density (Dr) and initial mean to liquefaction than clean sand (e.g., [36–42]), there is no
effective stress (s0 mean) it has not been directly incorporated clear consensus in the literature about the effect of silt
in majority of the previous conventional models as content increment on the trend of liquefaction resistance of
described before. sands. While some researchers have shown that an increase
Since s0 0 and Dr represent initial density of soils, they in FC decreases the liquefaction resistance, other research-
were categorized into one group referred as intergranular ers have reported the opposite effect.
contact density. Similarly, Cu and D50 are grain size For non-plastic fines, Polito and Martin [43], based on
distribution parameters and have been categorized into a cyclic undrained triaxial testing results, found that
separate group as grain size characteristic. In addition, FC liquefaction resistance for high FC is generally less than
is individually considered as a category which controls that of materials with low FC for a constant soil relative
potential of pore pressure buildup. Total relative impor- density. Also, they showed a contrary trend for constant
tance values for intergranular contact density, FC and sand skeleton void ratio. This reduction of liquefaction
grain size characteristic categories are 33.43%, 27.48% and resistance with increasing FC in constant relative density is
39.1%, respectively. consistent with the results of a study reported by others [44]

Table 5
Connection weights and relative importance of soil capacity ANN model calculated via Garson [34] technique

Parameters s0 mean (kPa) Dr (%) FC (%) Cu D50 (mm) Log (W)

Hidden units Connection weights

1 0.5992 0.3623 3.1017 0.437 0.7231 2.8011

2 0.3863 1.4293 0.3981 0.1848 0.616 0.7663
3 1.5587 0.2395 3.4745 0.1082 1.5657 1.5435
4 0.1655 0.8185 2.3514 0.0597 0.7639 2.0503
5 0.2136 1.0069 0.675 0.0273 1.2183 0.7329
6 0.0117 1.2689 0.0388 0.5561 1.9546 1.1619
7 0.0927 0.7889 0.4934 0.9895 1.4229 2.816
8 0.1238 0.6106 1.0488 0.83 0.855 0.6822
9 0.2219 1.564 0.4182 0.783 0.8276 0.2182
10 1.2814 0.6115 0.8412 0.9503 0.7668 1.4258
Relative 9.84 23.58 27.48 12.48 26.61
importance (%)
Categories Intergranular contact density Fines content Grain size characteristic

Total relative 33.43 27.48 39.10

importance (%)
 ARTICLE IN PRESS
1068 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

who investigated seismic compression behavior of sand and 7. Verification with measured energy demands
silty sands.
Also, other researchers [42,45] have concluded that for The ANN-based capacity energy model proposed in this
non-plastic FC increasing up to 35% and 44%, respec- paper can predict the energy density required to trigger
tively, the liquefaction resistance of sand–silt mixtures with liquefaction in a susceptible deposit. This energy density is
constant global void ratio decreases with respect to clean a characteristic parameter of soil which stands in the class
sand, whereas for FC greater than these values, the of soil resistance parameters against liquefaction and is
liquefaction resistance increases. comparable with cyclic resistance ratio (CRR) in stress-
Fig. 7 shows that the energy-based liquefaction resis- based, and threshold shear strain in strain-based liquefac-
tance of sand–silt mixture slightly increases when FC tion prediction procedures. It has been known as the
increases up to 10%, and then continuously decreases while capacity energy of liquefiable soils. On the other hand, the
its decrement rate declines for larger FC. This trend is part of released energy imparted to a particular site from
similar to the results of cyclic triaxial tests conducted on an earthquake source is known as demand energy.
Yatesville silty sand specimens [43] prepared by moist According to a logical hypothesis, the demand energy
tamping approach adjusted to 30% relative density imparted in a liquefied site was probably greater than
(Fig. 8). Comparison between Figs. 7 and 8 confirms the capacity energy of liquefied soil and enforced the soil to
results of the parametric study for FC and satisfactory liquefy. This hypothesis was used to examine the proposed
performance of the ANN model. ANN-based model in several liquefaction case histories in
which downhole array data were recorded during earth-
quakes and their stress–strain histories were reported. In
addition, data of some liquefaction centrifuge tests results
3000
from C-Core and VELACS projects were employed for
further verification of the ANN model.
2500
A simple procedure, proposed in basic form for shaking-
Capacity Energy (J/m3)

table studies [46], was developed and used to study the

2000
downhole earthquake response at the Lotung, Wildlife and
Port Island, Kobe earthquake sites [47–49]. Using this
1500 technique, the recorded accelerations at ground surface and
downhole arrays were employed to assess average shear
1000 stress and strain history at several earthquake sites. These
sites, at which the average stress–strain loops were
500 estimated, included Wildlife site during 1987 Superstition
Hills earthquake (Mw ¼ 6.7) Port Island, Kobe site during
0 1995 Hyogoken-Nanbo earthquake (Mw ¼ 6.9), and Lo-
0 20 40 60 80 100
tung, Taiwan 1986 LSST 16 earthquake. In addition, using
Silt Content (%)
the above-mentioned technique to assess average shear
Fig. 7. Variations of capacity energy predicted by ANN model versus FC, stress–strain history, Zeghal et al. [50] proposed average
other parameters were kept constant at their average values. shear stress–strain loop of a fully liquefied portion of a
centrifuge test at depth 2.6 m conducted during VELACS
project (Davis Model 6), employing the recorded lateral
accelerations. Fig. 9 illustrates the average shear stress–
strain loops proposed by above-mentioned researchers for
Wildlife, Kobe and Lotung downhole array sites, and
Davis model 6 centrifuge test of VELACS project.
One may find the centrifuge tests data from C-Core
project, which was a cooperative research project between
the University of British Columbia, C-CORE Incorpora-
tion, Memorial University of Newfoundland and industrial
partners, at project website. Using the above-mentioned
technique [48], average shear stress and strain loops at fully
liquefied portions of the modeled dykes and slopes in this
project were developed.
Fig. 10 shows the measured pore pressure buildup at
transducer P3 which was installed at the depth of 2 m in
Fig. 8. Effect of silt content on liquefaction behavior of Yatesville silty C-Core centrifuge model CT4. This transducer is located
sand specimens prepared by moist tamping adjusted to 30% relative between accelerometers A2 and A3. The figure illustrates
density (after [43]). that liquefaction occurred at this location and after about
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1069

Fig. 9. Measured average shear stress–strain histories at (a) Wildlife downhole array site (after [48]); (b) Port Island, Kobe site (after [49]); (c) Davis model
6 centrifuge test (after [50]); (d) Lotung site at various depths (after [47]).

12 s the excess pore pressure reaches the initial effective The areas of shear stress–strain loops illustrated in
overburden stress. Using acceleration histories recorded by Fig. 9, and also C-Core project centrifuge data estimated in
A2 and A3 and the above technique [48], average stress- this paper, were accumulatively calculated as real demand
strain loop between A2 and A3 was estimated and the energy imparted in the corresponding soil systems. The
accumulated dissipated demand energy was calculated. calculated demand energy values were compared with the
Fig. 11 depicts the average accumulative dissipated corresponding capacity energy predicted by the proposed
demand energy which was obtained by calculating the ANN-based model (Fig. 12). Fig. 12 shows that the ANN-
area of hysteresis shear stress–strain loop versus dynamic based capacity model predicted larger amount of capacity
loading time in this zone. Also, average capacity energy energy values than their corresponding demand energy
predicted by the ANN model is plotted in the figure. It can values for non-liquefied cases. The reverse is true for the
be seen that the demand energy surpasses capacity energy liquefied case histories.
after 16 s and indicates the occurrence of liquefaction in It is interesting to note that Lotung downhole array site
this zone. did not experience liquefaction in the 1986 LSST 16
 ARTICLE IN PRESS
1070 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

70 earthquake. As can be seen in Fig. 12, the values of

Excess Pore Water Presssure (kPa)

50
Pore Pressure Buildup at P3 (depth =2m) demand energy imparted to three various depths of Lotung
between A2 and A3 accelerometers site are smaller than their corresponding capacity values
30 simulated by the ANN model. Other spots in this figure
10 denote liquefied cases that have fallen on upper portion of
initial effective stress
chart. It means that in these cases, the proposed ANN
-10
model yields reasonable results since the demand energy
-30 values have been greater than the energy-based liquefaction
capacity of corresponding soil systems. Also, the spots
-50 C-Core Centrifuge Test CT4 corresponding to the Wildlife site is located near the
-70 boundary line (demand ¼ capacity) tending to upper
0 5 10 15 20 25 30 35 40 45
portion of the chart. It shows that the ANN model
Time (sec)
predicted approximately an onset liquefaction condition at
Fig. 10. Pore Pressure Buildup at P3 (depth ¼ 2 m) between A2 and A3 this site. Dobry et al. [51] reported liquefaction triggering
accelerometers (C-Core Centrifuge Test CT4). at the depth of 2.9 m in this soil deposit as the measured
pore water pressure buildup at this elevation partially
reached the initial effective overburden stress (Fig. 13).
This indicates an onset liquefaction similar to what is
1000
Average energy demand at a liquefied zone between
Dissipated Energy Density (J/m3)

A2(x=11m, z=18m) and A3(x=11, z=23.5) accelerometers

750

500 Soil Capacity Energy =576

(J/m3)
predicted by the ANN model
C-Core Project
250
Centrifuge Test CT4:
Loose sand layer with drainage dyke
under A2475 earthquake motion
0
0 5 10 15 20 25 30 35 40 45
Time (sec)

Fig. 11. Average energy demand at a liquefied zone between A2 (x ¼ 11,

z ¼ 18 m) and A3 (x ¼ 11, z ¼ 23.5) accelerometers (C-core centrifuge test Fig. 13. Pore water pressure buildup measured by transducer P5 (at 2.9 m
CT4). depth) during Superstition Hill 1987 earthquake, Wildlife site (after [51]).

10000

Wildlife Site
Demand Energy (J/m3)

1000
Port Island, Kobe Site

VELACS UCD Centrifuge Model 6

Demand=Capacity
C-CORE Centrifuge Tests
100 Lotung, Taiwan Site

Solid Spots: Liquefied cases

Hollow Spots: Non-Liquefied cases

10
10 100 1000 10000
Capacity Energy (J/m3)

Fig. 12. Comparison between actual demand energy and predicted capacity energy by ANN model in centrifuge test and actual liquefaction case histories.
 ARTICLE IN PRESS
M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072 1071

Table 6
Comparison between performances of ANN model and other statistical models

Models Number of data Performance

R2 MSE(%) MAE(%)

Figueroa et al. [19]—MLR 27 0.94 0.24 0.39

Figueroa et al. [19]—MLR 284 0.36 21.6 38.6
MLR-based model (this study) 284 0.65 7.15 21.33
ANN-based model (this study) 284 element tests data 0.9 2 10.4
18 centrifuge tests data (validation set) 0.9 0.6 6.3

predicted by the ANN model. Consequently, the capacity density, while it has not been directly incorporated in
energy values predicted by the ANN model show a good majority of the previous models. Table 6 presents the
agreement with the observed demand energy values for superior performance of proposed ANN model in compar-
earthquakes and centrifuge liquefaction case histories. ison with previous models. It can be seen that, although
Figueroa et al. [19] model yields accurate results for data of
8. Summary and conclusion 27 tests, it cannot successfully work for current database.
A parametric study was also conducted in this study using
In this paper, a large database, presenting laboratory the proposed ANN model and the sensitivity of capacity
cyclic data of clean and silty sands, and also data of several energy due to variation of soil initial parameters was
centrifuge liquefaction tests were utilized to develop an evaluated. Results of parametric study illustrate that
ANN model to predict amount of strain energy required up capacity energy of sand–silt mixture continuously increases
to liquefaction triggering. This model yields liquefaction due to increasing s0 0 and Dr while other parameters were
resistance of sand–silt mixtures in terms of strain energy kept constant. Also, it slightly increases when FC increases
and has been known as capacity energy. The developed up to 10%, then decreases continuously while its decrement
model takes into account a hypothesis that for granular rate declines for greater FC. Results of the parametric study,
material, the accumulated area of shear stress–strain loops obtained in this research, were confirmed with the results of
is directly related to their potential of pore pressure experimental studies presented by other researchers.
buildup. In order to assess regression based models, For more verification, some downhole array data of real
Figueroa et al. [19] model and also a new MLR-based earthquake case histories and some liquefaction centrifuge
model developed in this study have been evaluated using tests (VELACS and C-Core projects) were employed to
current database. Unacceptable performance of these compare the actual values of earthquake induced energy
models illustrated necessity for developing an ANN-based imparted to soil mass (demand energy) with their
model. corresponding capacity energy predicted by proposed
Due to a trial study performed to determine suitable input ANN model. The comparison shows the acceptable
parameters, five individual ANN models were constituted. performance of ANN model in predicting capacity energy
Consequently, a model with five input parameters (s0 0, Dr, values of a soil system to experience liquefaction triggering.
FC, Cu and D50) was selected due to its good degree of Comparisons between actual demand energy and ANN-
accuracy and compatibility with basic understanding of based capacity energy of these liquefaction cases have
liquefaction phenomenon. Initial mean effective stress, confirmed the hypothesis that for liquefied soil deposit the
relative density, fines content (FC), coefficient of uniformity imparted demand energy is larger than in situ capacity
and mean grain size compose input parameters of this energy of soil while an opposite correlation governs on
model. In order to develop a more generalized ANN model, non-liquefied cases.
both element and centrifuge tests data were employed as
testing and validation data sets. The ANN model has
yielded coefficients of determination (R2) equal to 0.9 for all References
element tests data and also centrifuge tests data set which
was not incorporated in training process. [1] Seed HB, Idriss IM. Simplified procedure for evaluating soil
liquefaction potential. J Soil Mech Found Div., ASCE 1971;
A sensitivity analysis was carried out [34] to evaluate 97(SM8):1249–74.
contribution of each input parameter in the model. Input [2] Whitman RV. Resistance of soil to liquefaction and settlement. Soils
parameters of the model were categorized into three groups Found 1971;11(4):59–68.
named as intergranular contact density (s0 0 and Dr), FC [3] Seed HB, Idriss IM, Makdisi F, Banerjee N. Representation of
and grain size characteristic (Cu and D50). Total relative irregular stress time histories by equivalent uniform stress series in
liquefaction analyses. Report No. UCB/EERC-75/29, Earthquake
importance values for these three categories have been Engineering Research Centre, U.C. Berkeley, 1975.
33.43%, 27.48% and 39.1%, respectively. According to [4] Ishihara K, Yasuda S. Sand liquefaction in hollow cylinder torsion
this analysis, FC is more important factor than relative under irregular excitation. Soils and Found 1975;15(1):45–59.
 ARTICLE IN PRESS
1072 M.H. Baziar, Y. Jafarian / Soil Dynamics and Earthquake Engineering 27 (2007) 1056–1072

[5] National Research Council. Liquefaction of soils during earthquakes. [26] Young-Su K, Byung-Tak K. Use of artificial neural networks in the
National research Council Report CETS-EE-001, National Academy prediction of liquefaction resistance of sands. J Geotech Geoenviron
Press, Washington, DC, 1985. Eng ASCE 2006;132(11):1502–4.
[6] NCEER. In: Proceedings of the NCEER workshop on evaluation of [27] Goh ATC. Seismic liquefaction potential assessed by neural net-
liquefaction resistance of soils. National Center for Earthquake works. J Geotech Eng ASCE 1994;120(9):1467–80.
Engineering Research. Technical Report No. NCEER970022, [28] Baziar MH, Nilipour N. Evaluation of liquefaction potential using
State University of New York at Buffalo, New York, December 31, neural-networks and CPT results. J Soil Dyn Earthquake Eng
1997. 2003;23:631–6.
[7] Dobry R, Ladd RS, Yokel FY, Chung RM, Powell D. Prediction of [29] Baziar MH, Ghorbani A. Evaluation of lateral spreading using
pore water pressure buildup and liquefaction of sands during artificial neural networks. J Soil Dyn Earthquake Eng 2005;25:1–9.
earthquakes by the cyclic strain method. Building science series, [30] Masters T. Practical neural network recipes in C++. San Diego:
Vol. 138. Washington, DC: National Bureau of Standards, US Academic press; 1993.
Department of Commerce, US Governmental Printing Office; 1982. [31] Hagan MT, Demuth HB, Beale MH. Neural network design. Boston,
[8] Green RA. Energy-based evaluation and remediation of liquefiable MA: PWS Publishing; 1996.
soils. PhD dissertation, Virginia Polytechnic Institute and State [32] Cybenko G. Approximation by superpositions of a sigmoidal
University, Blacksburg, VA; 2001. function. Urbana: University of Illinois; 1989.
[9] Nemat-Nasser S, Shokooh A. A unified approach to densification [33] Haykin S. Neural networks. New York: Macmillan College Publish-
and liquefaction of coheshionless sand in cyclic shearing. Can ing Company Inc.; 1994.
Geotech J 1979;16(4):659–78. [34] Garson GD. Interpreting neural-network connection weights. AI
[10] Towhata I, Ishihara K. Shear work and pore water pressure in Expert 1991;6(7):47–51.
undrained shear. Soils and Foundations 1985;25(3):73–84. [35] Lee KL, Seed HB. Cyclic stress conditions causing liquefaction of
[11] Liang L, Figueroa JL, Saada AS. Liquefaction under random sand. J Soil Mech Found Div. ASCE 1967;93(SM1):47–70.
loading: a unit energy approach. J Geotech Eng ASCE 1995; [36] Zlatovic S, Ishihara K. On the influence of non-plastic fines on
121(GT11):776–81. residual strength. In: Proceedings of the first international conference
[12] Mostaghel M, Habibaghi K. Cyclic liquefaction strength of sands. on earthquake geotechnical engineering, Tokyo, November 14–16,
Earthquake Eng Struct Dyn 1979;7:213–33. 1995.
[13] Berrill JB, Davis RO. Energy dissipation and seismic liquefaction of [37] Lade PV, Yamamuro JA. Effects of non-plastic fines on static
sands: revised model. Soils Found 1985;25(2):106–18. liquefaction of sands. Can Geotech J 1997;34:918–28.
[14] Law KT, Cao YL, He GN. An energy approach for assessing seismic [38] Thevanayagam S, Ravishankar K, Mohan S. Effects of fines on
liquefaction potential. Can Geotech J 1990;27:320–9. monotonic undrained shear strength of sandy soils. Geotech Test J
[15] Liang L. Development of an energy method for evaluating the 1997;20(4):394–406.
liquefaction potential of a soil deposit. PhD dissertation, Department [39] Thevanayagam S. Effect of fines and confining stress on undrained
of Civil Engineering, Case Western Reserve University, Cleveland, shear strength of silty sands. J Geotech Geoenviron Eng ASCE
OH; 1995. 1998;124(6):479–91.
[16] Green RA, Mitchell JK, Polito CP. Energy-based excess pore [40] Baziar MH, Dobry R. Residual strength and large-deformation
pressure generation model for cohesionless soils. In: Proceedings of potential of loose silty sands. J Geotech Eng ASCE 1995;121(12):
the John Booker memorial symposium, Sydney, NSW, Australia, 896–906.
November 2000: p. 16–17. [41] Amini F, Qi GZ. Liquefaction testing of stratified silty sands.
[17] Wang G, Takemura J, Kuwano J. Evaluation of excess pore water J Geotech Geoenviron Eng ASCE 2000;126(3):208–17.
pressures of intermediate soils due to cyclic loading by energy [42] Naeini SA, Baziar MH. Effect of fines content on steady-state
method. In: Yuan, editor. Proceedings of the conference of computer strength of mixed and layered samples of a sand. J Soil Dyn
methods and advances in geomechanics. Rotterdam, Netherlands: Earthquake Eng 2004;24:181–7.
A.A. Balkema; 1997. p. 2215–20. [43] Polito CP, Martin II JR. Effects of non-plastic fines on the
[18] Polito CP. The effects of non-plastic and plastic fines on the liquefaction resistance of sands. J Geotech Geoenviron Eng ASCE
liquefaction of sandy soils. PhD thesis, Virginia Polytechnic Institute 2001;127(5):408–15.
and State University, Blacksburg, VA, 1999. [44] Stewart JP, Whang DH, Moyneur M, Duku P. Seismic compression
[19] Figueroa JL, Saada AS, Liang L, Dahisaria MN. Evaluation of soil of as-compacted fill soils with variables levels of fines content and
liquefaction by energy principles. J Geotech Eng, ASCE 1994;120(9): fines plasticity. CUREE publication no. EDA-05, University of
1554–69. California, Los Angeles, 2004.
[20] Kusky PJ. Influence of loading rate on the unit energy required for [45] Xenaki VC, Athanasopoulos GA. Liquefaction resistance of sand–silt
liquefaction. MS thesis, Department of Civil Engineering, Case mixtures: an experimental investigation of the effect of fines. J Soil
Western Reserve University, Cleveland, OH; 1996. Dyn Earthquake Eng 2003;23:183–94.
[21] Rokoff MD. The influence of grain-size characteristics in determining [46] Koga Y, Matsuo O. Shaking table tests of embankments testing on
the liquefaction potential of a soil deposit by the energy method. MS liquefiable sandy ground. Soils Found 1990;30(4):162–74.
thesis, Department of Civil Engineering, Case Western Reserve [47] Zeghal M, Elgamal A-W, Tang HT, Stepp JC. Lotung downhole
University, Cleveland, OH, 1999. array II: Evaluation of site nonlinear properties. J Geotech Eng
[22] Dief HM. Evaluating the liquefaction potential of soils by the energy ASCE 1995;121(4):363–78.
method in the centrifuge. PhD Dissertation, Department of Civil [48] Zeghal M, Elgamal A-W. Analysis of site liquefaction using
Engineering, Case Western Reserve University, Cleveland, OH, 2000. earthquake records. J Geotech Eng ASCE 1994;120(6):996–1017.
[23] Ellis GW, Yao C, Zhao R, Penumado D. Stress–Strain modeling of [49] Elgamal A-W, Zeghal M, Parra E. Liquefaction of reclaimed island in
sands using artificial neural networks. J Geotech Geoenviron Eng Kobe, Japan. J. Geotech Eng ASCE 1996;122(1):39–49.
ASCE 1995;121(5):429–35. [50] Zeghal M, Elgamal A-W, Zeng X, Arulmoli K. Mechanism of
[24] Keifa MA. General regression neural networks for driven piles in liquefaction response in sand–silt dynamic centrifuge tests. J Soil Dyn
cohesionless soils. J Geotech Geoenviron Eng ASCE 1998;124(12): Earthquake Eng 1999;18:71–85.
1177–85. [51] Dobry R, Elgamal A-W, Baziar MH, Vucetic M. Pore pressure and
[25] Shahin MA, Maier HR, Jaksa MB. Predicting settlement of shallow acceleration response of Wildlife site during the 1987 earthquake. In:
foundations using neural networks. J Geotech Geoenviron Eng O’Rourke TD and Hamada M, editors. The second US–Japan
ASCE 2002;128(9):785–93. workshop on soil liquefaction, 1989; p. 145–60.