Color profile: Generic CMYK printer profile

Composite Default screen

1195

Risk-based liquefaction potential evaluation using

standard penetration tests
C. Hsein Juang, Caroline J. Chen, Tao Jiang, and Ronald D. Andrus

Abstract: In this paper, a new approach is presented for developing a liquefaction limit state function, which defines a
boundary that separates liquefaction from no-liquefaction occurrence. The new approach is developed using a database
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

consisting of 243 field liquefaction performance cases at sites where standard penetration tests (SPT) had been con-
ducted. This database is first used to train and test an artificial neural network for predicting the occurrence of lique-
faction or no liquefaction. The successfully trained neural network is then used to establish a liquefaction limit state
function. Based on the developed limit state function, mapping functions that relate calculated factors of safety to
probability of liquefaction are established. The established mapping functions form a basis for the development of a
risk-based chart for liquefaction potential evaluation.

Key words: probability, risk-based design, liquefaction potential, SPT, artificial neural network.

Résumé : Dans cet article, on présente une nouvelle approche pour développer une fonction d’état limite de
liquéfaction qui définit une frontière séparant l’incidence de liquéfaction de la non liquéfaction. La nouvelle approche
est développée en partant d’une base de données comprenant 243 cas de liquéfaction sur le terrain venant de sites où
des essais de pénétration standard ont été effectués. Cette base de données est utilisée en premier lieu pour préparer et
tester le réseau de neurones artificiels afin de prédire l’apparition de liquéfaction ou la non liquéfaction. Le réseau de
neurones développé avec succès est alors utilisé pour établir une fonction d’état limite de liquéfaction. En partant de la
For personal use only.

fonction d’état limite développée, on établit des fonctions de transformations qui relient les coefficients de sécurité
calculés à la probabilité de liquéfaction. Les fonctions de transformation établies constituent une base pour le
développement d’une charte de risques pour l’évaluation du potentiel de liquéfaction.

Mots clés : probabilité, conception déterministe, potentiel de liquéfaction, SPT, réseau de neurones artificiels.

[Traduit par la Rédaction] Juang et al. 1208

Introduction are essentially performance functions that were established

from field observations of soil performance in earthquakes
Because of the difficulty and the cost in obtaining high- at sites where in situ test data were available. In this study, a
quality undisturbed soil samples, simplified methods for variant of performance function, termed limit state function,
evaluating liquefaction potential of soils based on in situ is considered. The limit state function may be a multiple-
tests are preferred by geotechnical engineers. The simplified variable function that represents a surface in a multiple-
procedure for evaluation of liquefaction potential developed dimension space, herein called limit state surface. The limit
by the late Professor Seed and his coworkers (Seed and state function can also represent a two-dimensional curve,
Idriss 1971, 1982; Seed et al. 1985) based on the standard herein called limit state curve. Both limit state curve and
penetration test (SPT) is the most widely used method in limit state surface can collectively be called limit state
North America and throughout much of the world. A recent boundaries.
update of this simplified procedure is documented in a Na-
The limit state is a condition at which the liquefaction re-
tional Center for Earthquake Engineering Research
sistance of a soil is equal to the applied seismic load. Soil
(NCEER) workshop proceeding edited by Youd and Idriss
liquefaction resistance is commonly expressed in terms of
(1997).
cyclic resistance ratio (CRR), a term adopted in a recent
In the simplified procedure by Seed and his colleagues,
NCEER workshop (Youd and Idriss 1997). The applied seis-
empirically determined curves are used to predict liquefac-
mic load is expressed as cyclic stress ratio (CSR), a widely
tion and no-liquefaction occurrence. These empirical curves
accepted term after Seed and Idriss (1971, 1982). Concep-
tually, CRR is equal to CSR at the limit state, which is an
Received November 15, 1999. Accepted May 19, 2000. analogy to the concept that shear strength is equal to the
Published on the NRC Research Press website on shear stress at failure. In the present study, the CSR at the
December 18, 2000. limit state is herein termed the critical CSR.
C.H. Juang, T. Jiang, and R.D. Andrus. Civil Engineering Development of a limit state curve in the simplified
Department, Clemson University, Clemson, SC 29634-0911, method by Seed and his colleagues, as well as the methods
U.S.A. by other researchers, generally relied on engineering judg-
C.J. Chen. URS Corporation, 2020 East First Street, Suite ment. There are two drawbacks in this approach. First, con-
400, Santa Ana, CA 92705, U.S.A. siderable judgment is required in drawing the limit state

Can. Geotech. J. 37: 1195–1208 (2000) © 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 18, 2000 2:17:51 PM
 Color profile: Generic CMYK printer profile
Composite Default screen

1196 Can. Geotech. J. Vol. 37, 2000

curve, which inevitably induces uncertainties. Although the through regression analysis. The source data and the devel-
limit state curves in the existing simplified methods were opment of the liquefaction indicator function, the searching
generally drawn conservatively, the degree of conservatism procedure, and the limit state function are described in detail
was typically not quantified. It would be useful to quantify in the sections that follow.
the conservatism and develop consistent probabilistic mea-
sures of liquefaction potential. In this regard, the work by Source data
Liao et al. (1988) is a significant first contribution. Second, The SPT-based field liquefaction performance cases cata-
the limit state curve by Seed et al. (1985) was developed for loged by Fear and McRoberts (1995) were used to establish
clean sands and earthquakes with magnitude 7.5. To account the limit state function. Some cases in their database, how-
for different combinations of factors that affect liquefaction ever, were not used. The cases that were not used fall into
resistance, such as soil types, fines contents, and earthquake three categories: (i) inconsistency in the reported data such
magnitudes, different “discrete” limit state curves were em- as total overburden pressure being less than effective over-
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

pirically established by Seed et al. Limited consideration burden pressure, (ii) multiple records with conflicting infor-
was given to possible interactions among these soil and load mation from the same borehole, and (iii) strong
parameters, as they were difficult to assess. disagreement between the field observation and the predic-
Several methods have been proposed to address the issue tion made with the liquefaction assessment chart by Seed et
of the initiation of liquefaction from a more fundamental al. (1985) and updated in Youd and Idriss (1997). It is desir-
point of view. Davis and Berrill (1982) and Figueroa et al. able to not consider data that are questionable, or redundant,
(1994) have developed energy methods for liquefaction as- in ANN learning. In the first category, the cases were simply
sessment. Desai et al. (1998) have developed a fundamental excluded from further consideration. In the second category,
model for liquefaction instability analysis based on the dis- only the lowest SPT blow count was taken from a borehole.
turbed state concept. Although these methods are fundamen- In the third category, the case was excluded from further
tally attractive, their usage in geotechnical practice has yet consideration if the factor of safety against liquefaction
to be established. Recently, Juang et al. (2000) proposed a based on the Seed et al. chart was less than 0.5 while field
new method for developing the limit state function based on observation indicated no liquefaction. A total of three cases
artificial neural network (ANN) analysis of field liquefaction in the first category, 19 cases in the second category, and 10
For personal use only.

performance records. This method has the same advantages cases in the third category were excluded.
as do the existing simplified methods in that it is based on a After screening the database using the above criteria, a re-
large database of field liquefaction performance records. duced database, consisting of 243 cases (150 liquefied cases
However, it uses a mechanical process to define the limit and 93 non-liquefied cases), was adopted for neural network
state function, which significantly reduces the aforemen- learning. Use of this database to develop a liquefaction limit
tioned drawbacks of the existing simplified methods. state function is described in the next section.
In this paper, the method developed by Juang et al. (2000)
is modified and used to develop a SPT-based limit state Development of liquefaction indicator function
function. A mapping function is then developed based on an As illustrated in Fig. 1, the first step in the development of
analysis of all cases in the database, which relates the calcu- a SPT-based limit state function is to establish a liquefaction
lated factor of safety (FS) to the probability of liquefaction indicator (LI) function. The LI function is a classifier that
(PL). The developed PL–FS mapping function is applied to can predict the occurrence of liquefaction and no liquefac-
develop a risk-based design chart. tion. Conceptually, the LI function may take the following
form:
Development of a limit state function: ANN
[1] LI = f ((N1)60, FCI, σ v′, Rp, SL)
approach
ANN methods (Rumelhart et al. 1986) have been shown where
to be effective in classifying field liquefaction performance LI is the liquefaction indicator (LI = 1 if liquefaction is
cases (Agrawal et al. 1995; Goh 1994, 1996; Juang and observed and LI = 0 if no liquefaction is observed);
Chen 1999, 2000; Juang et al. 1999a; Najjar and Hossam (N1)60 is the corrected SPT N value (Liao et al. 1988);
1998). As described in this section, the ANN approach is FCI is the fines content indicator (see Table 1);
taken one step farther to develop a liquefaction limit state σ v′ is the effective overburden stress at critical depth;
function. Rp is the ratio of total stress to effective stress (σv / σ v′);
The development of the limit state function by means of σv is the total overburden stress at critical depth; and
neural network learning is illustrated in Fig. 1. First, a lique- SL is the seismic load parameter (= (amax /g)/MSF, where
faction indicator function is developed through neural net- amax is the peak horizontal ground surface acceleration, g is
work learning of the actual field performance cases. The the acceleration due to gravity, and MSF is the magnitude
liquefaction indicator function, or classifier, is a trained neu- scaling factor that accounts for the effect of earthquake mag-
ral network that can “predict” the occurrence of liquefaction nitude, Mw).
and no liquefaction with high accuracy. Second, a searching Selection of the above parameters for the liquefaction in-
procedure based on the developed classifier is established dicator function is based on the current liquefaction potential
which can be used to search for points on the unknown limit evaluation procedures. The terms (N1)60 and FCI are consid-
state surface. Third, the limit state function, which is collec- ered essential in any SPT-based method for estimating CRR.
tively defined by the generated data points, is approximated The terms σ v′, Rp, and SL are required in the determination

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:23 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

Juang et al. 1197

Fig. 1. Major steps in the development of a limit state function Table 1. Fines content indicator.
by the ANN approach.
Fines content, Fines content
FC (%) indicator, FCI
FC ≤ 5 1
5 < FC < 35 2
FC ≥ 35 3

Pi is the input parameter i; and

fT(θ) is the transfer function = 1/(1 + e– θ ) (Juang et al.
1999a).
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

In this model, the number of input variables (m) is five

and the number of hidden neurons (n) is four. The input vari-
ables are P1 = (N1)60, P2 = FCI, P3 = σ v′, P4 = Rp, and P5 =
SL. For each case in the database, the values of these five in-
put variables can be readily deduced, and the target value of
the output LI is set to either 0 (for the non-liquefied case) or
1 (for the liquefied case). Roughly two thirds (163 cases) of
the entire database of 243 cases are used to “train” the neu-
ral network. The other 80 cases are reserved for testing the
trained network. The outcome of a successful training is a
set of coefficients for eq. [2], called biases and weights in
neural network literature.
In this study, the intended network is trained with a neural
network toolbox published by MathWork, Inc. (Demuth and
Beale 1998). By gradually reducing the root mean square of
For personal use only.

prediction errors during network training to an acceptable

level, the weights and biases are determined. Table 2 shows
the weights and biases of the trained network.
The success rate of the trained network, defined by eq. [2]
and Table 2, in predicting the occurrence of liquefaction –
no liquefaction is 91% using the 163 case records in the
training set. To examine whether the trained neural network
of CSR. Use of the terms σ v′, Rp, and SL for the calculation has generalized the input–output relationship, it is tested
of CSR is equivalent to use of the terms σ v′, σv, amax, and against the 80 case records in the testing subset. The success
MSF. Experience (Juang et al. 2000) has shown that use of rate is 82% when the trained network is tested with the test-
the regrouped parameters yielded better ANN models. In ad- ing subset. Using all 243 case records, the overall success
dition, use of the SL parameter is convenient in the search rate of the trained network in predicting the occurrence of
for points on the limit state surface, as described later in the liquefaction – non-liquefaction is 88%. These results show
paper. that the developed ANN model, represented by eq. [2] and
The fines-content (FC) indicators shown in Table 1 are in- Table 2, serves the intended purpose well, that is, to predict
tended to accommodate the effect of fines content. They are accurately the occurrence of liquefaction – no liquefaction.
consistent with the three classes of soils (FC ≤ 5%, 5 < FC < It is noted that the developed ANN model is applicable in
35%, FC ≥ 35%) considered in the Seed et al. (1985) lique- the following parameter ranges constrained by the database
faction evaluation procedure. used: 1 ≤ (N1)60 ≤ 43; 1 ≤ FCI ≤ 3; 15 ≤ σ v′ ≤ 320 kPa; 1.0 ≤
The liquefaction indicator function (eq. [1]) is generally a Rp ≤ 2.63; 0.07g ≤ amax ≤ 0.84g; 5.3 ≤ Mw ≤ 8; and 0.04 ≤
highly nonlinear, multiple-dimension function. It may be ap- SL ≤ 0.63. Note that FCI is an ordinal variable (i.e., variable
proximated by a three-layer neural network model: with only three discrete outcomes, 1, 2, and 3) based on
nonplastic fines shown previously in Table 1. All other pa-
 n   m    rameters are continuous variables.
 
[2] LI = fT Bo + ∑ Wk fT  BHk + ∑ Wik Pi    Use of the developed limit state function, and the subse-
 k =1   i =1   
  quent mapping function and risk-based chart for evaluating
liquefaction potential, should be limited to cases where the
where input data fall within the above ranges.
Bo is the bias at the output layer (only one neuron in this
layer); Searching for points on the limit state surface
Wk is the weight of the connection between neuron k of The second step in the development of the ANN-based
the hidden layer and the single output layer neuron; limit state function is to search for points on the unknown
BHk is the bias at neuron k of the hidden layer (k = 1, n); limit state surface. Here, the searching technique developed
Wik is the weight of the connection between input variable by Juang et al. (2000) is modified for use in the present
i (i = 1, m) and neuron k of the hidden layer; study. Figure 2 shows the conceptual model of this tech-

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:25 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

1198 Can. Geotech. J. Vol. 37, 2000

Table 2. Connection weights and biases of ANN model 1 (eq. [2)].

Weight Bias
Wik Wk Bk Bo
Input 1 Input 2 Input 3 Input 4 Input 5 Output Hidden Output
(i = 1) (i = 2) (i = 3) (i = 4) (i = 5) neuron layer layer
Hidden neuron 1 2.466 –40.692 10.661 12.792 26.511 –6.713 695.608 –563.634
(k = 1)
Hidden neuron 2 12.796 –158.149 –2.339 96.545 –211.837 106.023 446.531
(k = 2)
Hidden neuron 3 –75.485 31.479 –27.454 –21.290 60.852 9.075 569.745
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

(k = 3)
Hidden neuron 4 34.854 –33.076 –7.576 26.896 2.547 –8.432 –576.735
(k = 4)

nique. Each liquefied case in the collected database is repre- Fig. 2. Conceptual model for searching for points on the limit
sented by a point located above the unknown limit state state boundary.
surface. A search for the limit state is conducted by decreas-
ing the seismic load (path A in Fig. 2). On the other hand,
each non-liquefied case in the database is represented by a
point located below the unknown limit state surface. In this
situation, a search for the limit state is conducted by increas-
ing the seismic load (path B in Fig. 2).
Figure 3 shows a flow chart that implements the above
conceptual model. Following either path A or path B, an it-
For personal use only.

erative process is required. Using path A as an example, a

new data point is formed by decreasing the seismic load
while maintaining a constant soil resistance. With the new
input data, the ANN model (eq. [2]) will produce a new out-
put. Initially, the output LI is likely to remain the same (LI =
1, liquefaction) with a small decrease in the seismic load.
Continuing this process of decreasing the seismic load, how-
ever, the output LI will eventually change to no liquefaction.
The updated seismic load that yields a change of liquefac-
tion indication is the critical load (or the critical CSR) for
the given soil conditions.
Note that in some cases the search for the critical CSR
may not be successful. Using path B as an example, this
may occur when the seismic load has reached the upper
bound of a typical load range and yet the liquefaction indica-
tion remains unchanged. One reason for this is that some
soils are non-liquefiable, as evidenced by the near-infinity
slope at the upper end of the limit state curves in the SPT- SL in the second line of eq. [3]. The advantage of using a
based evaluation chart (Seed et al. 1985) and the cone pene- single load parameter SL in the search for points on the limit
tration test (CPT) based chart (Robertson and Campanella state surface is explained later in the paper. The term CSR7.5
1985). in eq. [3] is the cyclic stress ratio adjusted to the reference
Since CRR is by definition equal to the critical CSR, each earthquake magnitude of 7.5. It is noted that throughout this
successful search results in a new data point consisting of paper the terms CSR and CRR are referred to the reference
four input parameters, (N1)60, FCI, σ v′, and Rp, and one out- earthquake magnitude of 7.5. The parameter rd is the stress
put parameter, CRR. The magnitude of CRR is equal to the reduction factor that provides an approximate correction for
critical CSR calculated as follows (modified from Seed and flexibility of the soil profile. For a depth z of less than 23 m,
Idriss 1971): the term rd can be calculated using the following equation
(Liao et al. 1988; Youd and Idriss 1997):
σ   a 
0.65rd  v   max  [4a] rd = 1.0 – 0.00765z z ≤ 9.15 m
 σ v′   g  9.15 < z ≤ 23 m
[3] CSR7.5 = [4b] rd = 1.174 – 0.0267z
MSF
The parameter MSF is calculated as
= 0.65rd RpSL − 2.56
M 
Note that the parameters amax/g and MSF in the first line [5] MSF =  w 
of eq. [3] are grouped into a single seismic load parameter  7.5 

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:27 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

Juang et al. 1199

Fig. 3. Flow chart illustrating the searching algorithm for the critical CSR (after Juang et al. 2000).
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14
For personal use only.

The above formulation for MSF represents the lower bound (and thus CSR) takes place. On the other hand, a non-
of the range of MSF values recommended in the 1996 liquefied case may become liquefied if a sufficient increase
NCEER workshop (Youd and Idriss 1997). This formulation in SL takes place. With the second line of eq. [3], change in
is considered a good compromise of a wide range of MSF CSR for a given soil (i.e., a set of fixed rd and Rp) is accom-
values that have been proposed in the literature. plished by change in a single parameter, SL, whereas such
It is important to note that within the ranges of the data change in CSR using the first line of eq. [3] will involve two
used in the present study, choice of a particular MSF for- parameters, amax and MSF. The latter approach, proposed
mula is not critical in the developed limit state function. Re- previously by Juang and Chen (2000), is more difficult to
call that CRR is defined as the critical CSR obtained from implement than the searching algorithm depicted in Figs. 2
the search illustrated in Figs. 2 and 3. Thus, an increase (or and 3. The modified searching technique, using the parame-
decrease) in CSR, as a result of using a particular MSF for- ter SL and the second line of eq. [3], is adopted in the pres-
mula, will result in an increase (or decrease) in the corre- ent study.
sponding CRR in the same proportion. In other words, the From the searching algorithm described above, a total of
formulas for MSF and rd, as part of the calculation for CSR, 138 data points on the limit state surface is generated. These
have been implicitly calibrated by field cases in the database data points collectively define the limit state surface. Devel-
during the development of the limit state function, CRR = opment of the limit state function is described in the sections
f(indexes of soil properties). Thus, choice of a particular that follow.
MSF formula is not critical in the developed limit state func-
tion. It is, however, important that a consistent set of equa- Establishing the SPT-based limit state function
tions for CSR and CRR be adopted in the liquefaction Recall that each generated data point on the limit state
analysis using the developed method. surface consists of four input parameters, (N1)60, FCI, σ v′,
As noted earlier, use of the parameter SL and the second and Rp, and one output parameter, CRR. Conceptually, the
line of eq. [3] is preferred for searching points on the limit limit state function is defined as
state surface. As depicted in Figs. 2 and 3, a liquefied case
may become non-liquefied if a sufficient reduction in SL [6a] CRR = f ((N1)60, FCI, σ v′, and Rp)
© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:28 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

1200 Can. Geotech. J. Vol. 37, 2000

Fig. 4. Liquefaction resistance curve defined by generated data. [9a] β = 1.0 for FC ≤ 5%
[9b] β = [0.99 + (FC1.5/1000)] for 5 < FC < 35%
[9c] β = 1.2 for FC ≥ 35%
Although other methods to account for the effect of fines
content are available in the literature (for example, Robert-
son and Wride 1998), choice of a particular method for cal-
culating (N1)60,cs, as in the case of selecting MSF
formulation previously discussed, is not critical in the devel-
oped risk-based method to be presented later in the paper.
This is so because the developed limit state function is even-
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

tually “calibrated” by the field cases in the development of

mapping functions. The actual fines content, instead of FCI,
is used in the calculation of (N1)60,cs for each generated data
point.
The approach described above yields a set of dimension-
reduced, generated data that can be readily plotted. Figure 4
shows those generated data points which collectively define
the limit state function in reduced dimensions. Also shown
in Fig. 4 is the best-fit limit state curve determined by re-
gression analysis of these data (R2 = 0.75):
 (N ) 
[10] CRR7.5 = − 0.485 + 0.289  1 60,cs 
 35 
For personal use only.

0.739
+
[35 − (N1) 60,cs ] 0.1
This limit state function is a highly nonlinear, multidimen- Note that the broken portion of the regression curve
sional function, and a plot of the generated data points to shown in Fig. 4 reflects the fact that only limited data could
show the trend of the interdependent variables is not possi- be obtained in the range of (N1)60,cs > 25, and that wide scat-
ble. Because of the difficulty in establishing empirical equa- ter of the generated data in that range indicates great uncer-
tions to approximate the limit state function collectively tainty. Also note that eq. [10] is valid for (N1)60,cs < 35 and
defined by the generated data through regression analysis, a implies that the limiting (N1)60,cs, the value beyond which
neural network approach has been proposed to approximate liquefaction will not occur, is 35. This limiting value is a re-
this function (for example, Juang et al. 2000). In the present sult of regression analysis of the generated data, although it
study, a simplified limit state function in the form shown be- does not agree with the limiting value of 30 suggested by
low is developed based on the generated data: Seed et al. (1985), which appears to have been set subjec-
[6b] CRR = f ((N1)60,cs) tively. Equation [10] is more conservative than the Seed et
al. method in the range of (N1)60,cs > 30. Nevertheless, both
where (N1)60,cs is the clean-sand equivalence of the overbur- methods predict strong liquefaction resistance for soils with
den stress corrected SPT blow count. This simplified func- (N1)60,cs in this range, and the same conclusion of “no lique-
tion involves only two of the four input variables, namely faction” would be reached using either method. It is cau-
(N1)60 and FCI. The other two variables, σ v′ and Rp, are sim- tioned that both the Seed et al. method and the proposed
ply discarded. Furthermore, the two variables (N1)60 and FCI method (eq. [10]) are based on very limited data in this
are combined into one single variable, (N1)60,cs, so that the range and should be further investigated when more data be-
generated data points, those which are on the limit state sur- come available.
face, can be displayed in a simple two-dimensional plot. The An evaluation of the regression model shown in Fig. 5 in-
following equation, recommended by the NCEER workshop dicates that eq. [10] is a good approximation of the unknown
(Youd and Idriss 1997), is adopted herein for calculating limit state function. Thus, eq. [10] represents the limit state
(N1)60,cs: function that may be used to calculate CRR. It is important
to note that CRR calculated from eq. [10] must be compared
[7] (N1)60,cs = α + β(N1)60
with CSR calculated using eqs. [3]–[5] and [7]–[9] for a
where α and β are coefficients determined from the follow- proper assessment of liquefaction potential. Equation [10] is
ing equations: referred to hereinafter as the simplified limit state function
(LSF).
[8a] α= 0 for FC ≤ 5% Figure 6 shows a plot of eq. [10] for three classes of FC,
[8b] α = exp[1.76 – (190/FC2)] for 5 < FC < 35% as is conventionally presented. To plot these curves, a value
of FC is selected and the corresponding (N1)60 is calculated
[8c] α = 5.0 for FC ≥ 35% using eqs. [7]–[9] for a given (N1)60,cs. The CRR correspond-

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:30 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

Juang et al. 1201

Fig. 5. Evaluation of the regression model. Fig. 7. The simplified LSF models with historic data.
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14
For personal use only.

Fig. 6. Chart for estimating CRR7.5 based on the simplified LSF

model.

Figure 7 shows a plot of the simplified LSF (eq. [10])

along with all 243 field cases used in this study. The simpli-
fied LSF curve does not encompass all liquefied cases. In
fact, the success rate in “classifying” liquefied cases using
this simplified LSF curve is 82%, whereas that in classifying
non-liquefied cases is 71%. As previously mentioned, the
success rates for classifying liquefied and non-liquefied
cases using the trained ANN (eq. [2] and Table 2) are 92%
and 82%, respectively. The decrease in success rates is un-
derstandable, as this LSF is a simplified model considering
fewer variables than does the trained ANN. However, this
simplified model (LSF) is easier to use and more consistent
in its formulation with existing methods.
As defined previously, the limit state function is a perfor-
mance function. It is important that the LSF is defined ob-
jectively and consistently, as is done in this study through an
objective searching procedure, the accuracy of which de-
pends only on the data used. It is not necessary for an LSF
curve to encompass all liquefied cases, particularly when a
risk-based approach to liquefaction potential evaluation is
taken. A subjectively drawn liquefaction boundary curve that
encompasses all or almost all liquefied cases, as is normally
done in the existing simplified methods such that of Seed et
al. (1985), has no advantage over the LSF defined in this
ing to this (N1)60,cs is obtained from eq. [10]. Thus, a plot of study. To validate this viewpoint, a more conservative per-
CRR versus (N1)60 for a selected FC can be made. Figure 6 formance function defined by applying a reduction factor of
shows that eq. [10] conforms to the general expectation of 70% to the reference LSF (eq. [10]) is also investigated. This
the published liquefaction base curves. conservative performance function, shown in Fig. 7 as curve

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:34 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

1202 Can. Geotech. J. Vol. 37, 2000

B, encompasses almost all liquefied cases. Comparison of fL (β)

[11] PL = P( L β ) =
this reduced performance function with the reference LSF, fL (β) + fNL (β)
in the context of “conservatism,” is presented in detail later
in the paper.
where fL(β) and fNL(β) are the probability density functions
of β for liquefied cases (group L cases) and non-liquefied
Summary of the developed limit state function cases (group NL cases), respectively. The obtained probabil-
approach ity, PL, is the conditional probability, P(L|β), given a calcu-
In summary, an approximation to the unknown, multiple- lated β.
dimensional limit state function has been obtained in the
An investigation of the mapping function approach by
form of a simplified LSF (eq. [10]). The liquefaction resis-
Chen and Juang (2000) reveals that practically the same
tance of a soil, in terms of CRR, can be determined using
mapping function may be obtained by a direct interpretation
eq. [10] for a given (N1)60,cs (defined in eqs. [7]–[9]). On the
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

of probability through calculated factor of safety:

other hand, the seismic effect on the soil, in terms of CSR,
can be determined using eqs. [3]–[5]. In a deterministic ap-
f L ( FS)
proach, the factor of safety (FS), defined as FS = CRR/CSR, [12] PL = P( L FS) =
is calculated and liquefaction is said to occur if FS ≤ 1; no f L ( FS ) + f NL ( FS )
liquefaction occurs if FS > 1.
One issue that should be of interest to practicing engi-
where fL(FS) and fNL(FS) are the probability density func-
neers when using the new method proposed herein is the
tions of FS for liquefied cases (group L cases) and non-
choice of a proper FS. Of course, a proper FS to use depends
liquefied cases (group NL cases), respectively. The obtained
on technical and nontechnical factors, such as the impor-
probability, PL, is the conditional probability, P(L|FS), given
tance and economics of the project. Only the technical con-
a calculated FS.
cern as to whether a soil will liquefy under a seismic load is
considered here. Conceptually, a lower FS value is required Use of eq. [12] to establish a PL–FS mapping function re-
for a performance function that has a built-in conservatism, quires only the calculation of the traditional factor of safety,
such as curve B shown in Fig. 7. On the other hand, a higher FS = CRR/CSR, and the estimation of the probability density
For personal use only.

FS value should be used for the reference LSF (eq. [10]) that functions, fL(FS) and fNL(FS). In this study, each field case is
has no artificially built-in conservatism. To provide a guide- analyzed for the factor of safety using the proposed method.
line, the deterministic method to be used must first be cali- The calculated FS values are grouped into two groups, L and
brated. In this study, the proposed method, summarized NL. Group L consists of 150 cases and group NL consists of
above, is calibrated using field cases. This calibration is car- 93 cases. Figure 8 shows histograms of the calculated FS
ried out through the development of a mapping function de- values in both groups and their fitted density functions.
scribed in the section that follows. Based on the obtained probability density functions, the
probability PL is calculated for each field case according to
eq. [12]. The (PL, FS) data pairs and the mapping function
Development of mapping function and risk- obtained through curve fitting are shown in Fig. 9. The map-
based procedure ping function is expressed as
A mapping function that relates factor of safety (FS), cal-
1
culated by a deterministic method such as the proposed [13] PL = B
method summarized earlier in the paper, to probability of F 
liquefaction (PL) forms a basis for risk-based decisions con- 1 +  S
 A
cerning liquefaction potential of a soil. If the mapping func-
tion is established from a sufficiently large field performance
database, the deterministic method is said to have been cali- where A and B are curve-fitting parameters (A = 0.98 and
brated with field observations. This concept of mapping B = 3.6). This mapping function forms the basis for making
function and the procedure to establish it were first proposed risk-based decisions concerning the liquefaction potential of
by Juang et al. (1999b) and refined in Juang et al. (2000). a soil. In other words, a design decision in terms of selection
Their procedure is summarized as follows: (1) assess the un- of a proper factor of safety can be made based on an accept-
certainty in the input parameters that are used in the calcula- able risk (i.e., probability of liquefaction).
tion of CRR and CSR for each case in the liquefaction As with any empirical methods that are established from a
performance database; (2) conduct a reliability analysis to database, the limitation of the data itself carries over to the
calculate the reliability index β for each case in the database; developed method. Because of the possible sampling bias of
(3) obtain a distribution function of β for both groups of liq- the cases reported in the database used, the probability den-
uefied cases and non-liquefied cases in the database; (4) in- sity functions, fL(FS) and fNL(FS), obtained from the database
terpret the probability of liquefaction (PL) for each case are only approximations of their true density functions. This
based on the distributions of β in both liquefied and non- issue has been discussed in detail by Juang et al. (2000). A
liquefied groups; and (5) calculate factor of safety (FS) for recent study by Juang and Jiang (2000), however, shows that
each case and plot PL versus FS, and then obtain a PL–FS the probabilities obtained through the mapping function ap-
mapping function through curve fitting of the (PL, FS) data proach are comparable with those obtained through logistic
pairs. In step 4, the probability of liquefaction is interpreted regression analyses (Liao et al. 1988) of the field perfor-
according to Bayes’ theorem as follows: mance data.

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:35 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

Juang et al. 1203

Fig. 8. Histograms showing the distributions of calculated FS.

Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14
For personal use only.

Risk-based chart for liquefaction potential evaluation sign level of CSR, what is the required soil strength (in
The mapping function represented by eq. [13] calibrates terms of CRR) to be assured of a risk of liquefaction of less
the proposed LSF model (eq. [10]) for the evaluation of liq- than a specified level?
uefaction resistance of soils. With this mapping function, the As an example of using the design chart shown in Fig. 10,
factor of safety against the occurrence of liquefaction calcu- consider a sandy layer at a site with the following average
lated based on the proposed method is mapped to the proba- properties: z = 5.8 m, σ v′ = 84 kPa, σv = 111 kPa, FC = 2%,
bility of liquefaction based on the field cases. Thus, when and (N1)60 = 6. The structure to be built at the site is to be
using the proposed method for evaluating liquefaction resis- designed with Mw = 7 and amax = 0.20g. In this case, CSR
tance, a proper factor of safety can be selected based on an imposed by the design requirements is 0.14 according to the
acceptable level of risk (probability of liquefaction). The second line of eq. [3] and CRR is 0.09 according to eq. [10].
risk-based chart shown in Fig. 10 is created using the above If the acceptable level of risk (probability of liquefaction) is
SPT-based mapping function (eq. [13]). This can be easily specified as 20%, then the minimum required value of CRR
obtained by replacing FS in eq. [13] with the ratio of CRR to is about 0.20 according to Fig. 10. For the purpose of plan-
CSR. This chart may be used to address questions such as ning a ground-improvement project, it would be of interest
(but not limited to) the following: (1) What is the probability to know what (N1)60 value is needed to satisfy the design re-
of liquefaction for a given soil under a given seismic load quirements. Since a CRR of 0.2 is required, the (N1)60 value
(i.e., given calculated values of CSR and CRR)? (2) At a de- must reach at least 16 after the ground improvement according

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Friday, December 08, 2000 3:00:37 PM
 Color profile: Generic CMYK printer profile
Composite Default screen

1204 Can. Geotech. J. Vol. 37, 2000

Fig. 9. Plot of factor of safety versus probability of liquefaction the reference LSF model (eq. [10]). The other is an updated
and the mapping function obtained through curve fitting. version of the Seed et al. (1985) CRR curve reported in
Youd and Idriss (1997) which may be approximated by the
following equation:

a + cx + ex 2 + gx 3
[14] CRR7.5 =
1 + bx + dx 2 + fx 3 + hx 4

where a = 0.048, b = –0.1248, c = –0.004721, d = 0.009578,

e = 0.0006136, f = –0.0003285, g = –0.00001673, h =
0.000003714, and x = (N1)60,cs, the clean-sand equivalence of
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

stress-corrected blow count. Note that this equation is not a

continuous function and is valid for (N1)60,cs < 30. It must
not be used for (N1)60,cs > 30.
The above two CRR models, along with the reference
LSF model, are shown in Fig. 11. The reduced LSF model is
arbitrarily set at 70% of the reference model (eq. [10]). The
Seed et al. (1985) model is on the average about 15% lower
than the reference LSF model in the range of 5 < (N1)60,cs <
25. Note that the value of CRR determined by the Seed et al.
to eq. [10]. Thus, use of the design chart (Fig. 10) and the model approaches infinity as (N1)60,cs approaches 30. Thus,
associated equations enables the engineer to make a risk- comparison of the Seed et al. and reference CRR models is
based decision regarding the liquefaction potential and the meaningless for (N1)60,cs ≥ 30, although both models predict
requirement for a possible ground-improvement project. very strong liquefaction resistance and the same conclusion
of no liquefaction would be reached for the field cases. Re-
Issues of conservatism and mapping functions peating the procedures previously described, the mapping
For personal use only.

As mentioned earlier in the paper, the issue of conserva- functions for the two CRR models (Seed et al. 1985 and the
tism should be, and can be, addressed with the selection of a reduced LSF models) are obtained. They are expressed in
proper factor of safety, and thus use of a limit state boundary the same form as that for the LSF model, and the corre-
that has a built-in conservatism has no advantage over those sponding mapping function parameters A and B (see
without it. To explore this viewpoint, the above procedure eq. [13]) are shown in Table 3. Figure 12 compares the map-
for establishing a mapping function is repeated for two con- ping functions obtained for the three CRR models.
servative CRR models. One is an arbitrarily constructed The values of parameter B for all three mapping functions
model that is defined by applying a reduction factor (RF) to are almost the same, ranging between 3.5 and 3.6. Note that
Fig. 10. Risk-based chart for liquefaction evaluation based on the simplified LSF model.

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Friday, December 08, 2000 3:00:37 PM
 Color profile: Generic CMYK printer profile
Composite Default screen

Juang et al. 1205

Fig. 11. Comparison of various CRR models. Table 3. Mapping function parameters for three CRR models.
CRR model A B
LSF model 0.98 3.6
Reduced LSF model 0.69 3.6
Seed et al. 1985* 0.80 3.5
* The 1997 NCEER version (Youd and Idriss 1997).

Fig. 12. Mapping functions derived from various CRR models.

Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14
For personal use only.

It is interesting to note that the mapping function obtained

by Chen (1999) is almost identical to the one derived for the
LSF model. Chen’s study yields the following results: A =
1.0 and B = 3.5. In Chen’s study, a more conservative MSF
there is practically no difference between a mapping func- formulation (Idriss 1999) than the one adopted in this study
tion curve with B = 3.5 and that with B = 3.6 for a given A (eq. [5]) is used. Although Chen’s model yields greater CSR
value. The values of parameter A indicate the degree of con- values than does the simplified LSF model presented in this
servatism implicitly built in the CRR models (relative to the paper, the corresponding CRR values are proportionally
CSR model). If A = 1, the CRR model is “unbiased” as far greater because of the limit state concept and the searching
as the degree of conservatism is concerned. In other words, technique implemented. This is easily understood, since
the probability of liquefaction corresponding to FS = 1 for CRR is set to be equal to the critical CSR on the limit state
this particular CRR model would be 50%. The value of A = surface. Thus, the calculated factor of safety remains practi-
0.98 for the reference LSF model implies that this CRR cally unchanged. It is not surprising to see that the mapping
model is practically unbiased in this sense. The reduced LSF function obtained in the present study is almost identical to
model is understandably more conservative and is character- the mapping function obtained by Chen.
ized by a probability of 20% at FS = 1 according to the de- Figure 12 can be used to explain the effect of conserva-
rived mapping function (see Fig. 12). The value of A = 0.8 tism on the obtained PL–FS mapping function. With built-in
for the Seed et al. (1985) model implies that the model is bi- conservatism, the required factor of safety to assure a speci-
ased toward being conservative (about 20% reduction from fied risk level decreases. For example, with no built-in con-
an unbiased CRR model). The Seed et al. model is charac- servatism (the reference LSF model), the required FS is
terized by a probability of about 30% at FS = 1 according to equal to 1.43 at a risk level (probability of liquefaction) of
the derived mapping function (see Fig. 12). Previous studies 0.20. With the Seed et al. (1985) model, the required FS de-
by Liao et al. (1988), Youd and Noble (1997), and Toprak et creases to 1.19. With the reduced LSF model, the required
al. (1999) based on logistic regression analyses, a different FS further decreases to 1.01. It is obvious that a more con-
approach from the one presented herein, characterize the servative CRR model requires a smaller factor of safety than
Seed et al. model with probabilities ranging from 20 to 50% does a less conservative CRR model. However, the same
for different ranges of (N1)60,cs values. The characterization level of risk can be assured if a proper factor of safety is se-
of the Seed et al. model through the derived mapping func- lected based on a mapping function that has been calibrated
tion is quite consistent with the previous studies using logis- with a large database, regardless of which CRR model is
tic regression. The mapping function approach is shown to used. Thus, use of a limit state boundary (i.e., CRR model)
be simple and effective in probabilistic characterization of a that has a built-in conservatism has no advantage over use of
CRR model. those without it if a risk-based approach is taken.

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:20:58 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

1206 Can. Geotech. J. Vol. 37, 2000

Fig. 13. Analysis of the 1989 Loma Prieta cases reported by Fig. 14. Plot of PL–FS data points from Loma Prieta cases.
Toprak et al. (1999).
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

Table 4. Classes of liquefaction potential.

Description (likelihood of
Probability Class liquefaction)
0.85 ≤ PL < 1.00 5 Almost certain that it will liquefy
0.65 ≤ PL < 0.85 4 Liquefaction very likely
0.35 ≤ PL < 0.65 3 Liquefaction and non-liquefaction
For personal use only.

equally likely
0.15 ≤ PL < 0.35 2 Liquefaction unlikely
0.0 ≤ PL < 0.15 1 Almost certain that it will not
liquefy

Survey (Toprak et al. 1999) is analyzed. Following the 1989

Loma Prieta, California, earthquake, the U.S. Geological
Survey compiled extensive SPT and CPT data at 25 sites in
the Monterey and San Francisco Bay regions. Sites with and
without evidence of liquefaction were investigated. Forty-
eight (48) cases with SPT measurements were analyzed us-
ing the proposed method. These data are taken directly from
The comparison of mapping functions in the context of Toprak et al. (1999), except that two cases with questionable
conservatism described earlier in the paper is valid for 5 < data are excluded from consideration. The CSR for each
(N1)60,cs < 25, within which most field cases fall. At the case is calculated by eqs. [3]–[5] and the CRR by eq. [10].
lower end ((N1)60,cs < 5) and higher end ((N1)60,cs > 25), the The factor of safety is calculated as FS = CRR/CSR, and the
comparison does not apply, since both the Seed et al. (1985) probability of liquefaction PL is obtained from the risk-
method and the proposed method were established in these based chart (Fig. 10) or the mapping function (eq. [13]). Fig-
ranges with very limited data. As previously mentioned, in ure 13 shows the calculated values of FS and PL for all cases
the Seed et al. method the curve in these extreme ranges was analyzed. Figure 14 shows a plot of PL versus FS for these
established subjectively by engineering judgment, whereas cases along with the developed mapping function.
in the proposed method the curve in these extreme ranges As a way to interpret the calculated probability of lique-
was established objectively by regression analysis. Although faction, five classes of liquefaction potential are defined, as
it is implicitly assumed that the least squares regression shown in Table 4. These definitions of liquefaction potential
equation is correct, the results obtained for the cases outside classes are suitable for describing the likelihood of the oc-
the range of 5 < (N1)60,cs < 25 should be viewed with cau- currence of liquefaction. Figure 13 shows that non-liquefied
tion. cases generally have a probability of less than 35%, a cutoff
point where liquefaction is considered unlikely (class 2).
Case study: the 1989 Loma Prieta, None of the non-liquefied cases has a probability of greater
California, earthquake than 65%, a cutoff point above which liquefaction is very
likely. Thus, all non-liquefied cases are well explained using
To demonstrate the developed PL–FS mapping function the present methods. As for the liquefied cases, the probabil-
and the risk-based procedures, a database of field perfor- ities range from 21 to 95%. Only three cases have a proba-
mance records published recently by the U.S. Geological bility of less than 35%, the cutoff point below which

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:21:02 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

Juang et al. 1207

liquefaction is unlikely. Using the 20–30% probability, com- tion of a proper factor of safety that is calibrated to the
monly accepted for general engineering practice (e.g., Youd probability of liquefaction based on field observations.
and Noble 1997), as a screening criterion, all liquefied cases
are correctly predicted.
Interpretation of liquefaction potential by means of the Acknowledgments
factor of safety is not as simple as that by means of the
The study on which this paper is based was supported by
probability of liquefaction, because the relation between the
the National Science Foundation (NSF) through Grant CMS-
likelihood of liquefaction and the factor of safety is nonlin-
9612116. The cognizant NSF program official for this grant
ear. The interpretation is also complicated by the fact that
was Dr. Clifford Astill. The financial support of the NSF is
different deterministic methods have different degrees of
greatly appreciated. The Journal reviewers are also thanked
conservatism built into them (e.g., the Seed et al. 1985
for their constructive comments that helped improve this pa-
method is found to be approximately 15% more conservative
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

per.
than the simplified LSF model in the range of 5 < (N1)60,cs <
25). Thus, the interpretation of the calculated factor of safety
could be misleading if the concept or experience is extended References
to a different method. The risk-based approach, on the other
hand, always provides a consistent measure and interpreta- Agrawal, G., Chameau, J.L., and Bourdeau, P.L. 1995. Assessing
tion of liquefaction potential. the liquefaction susceptibility at a site based on information
As a final note to the risk-based methods, the participants from penetration testing. In Artificial neural networks for civil
of a recent liquefaction workshop (Youd and Idriss 1997) did engineers — fundamentals and applications. Chapt. 9. American
not approve recommendations of probabilistic analysis for Society of Civil Engineers, New York, pp. 185–214.
liquefaction evaluation because it is “beyond the normal Chen, C.J. 1999. Risk-based liquefaction potential evaluation using
practice of most geotechnical engineers.” They did encour- cone penetration tests and shear wave velocity measurements.
age “continued development of these concepts” (Youd and Ph.D. thesis, Clemson University, Clemson, S.C.
Chen, C.J., and Juang, C.H. 2000. Calibration of SPT- and CPT-
Noble 1997). In the opinion of the writers, however, the
based liquefaction valuation methods. In Innovations and appli-
mapping function and risk-based chart developed in the
For personal use only.

cations in geotechnical site characterization. Edited by P. Mayne

present study (eq. [13] and Fig. 10) make it easy to continue and R. Hryciw. American Society of Civil Engineers,
using the deterministic approach while making design deci- Geotechnical Special Publication, No. 97, pp. 49–64.
sions based on risk considerations. This is the essence of the Davis, R.O., and Berrill, J.B. 1982. Energy dissipation and seismic
presented risked-based method for liquefaction evaluation. liquefaction in sands. Earthquake Engineering and Structural
As an example, if a risk level (in terms of probability) of Dynamics, 10: 59–68.
15% is acceptable, a factor of safety of 1.58 is required us- Demuth, H., and Beale, M. 1998. Neural network toolbox user’s
ing the simplified LSF model. If a typical range of 20–30% guide. The MathWorks, Inc., Natick, Mass.
probabilities is desired, the required factor of safety will be Desai, C.S., Park, I., and Shao, C. 1998. Fundamental yet simpli-
1.24–1.43 using the simplified LSF model. It is cautioned fied model for liquefaction instability. International Journal for
that these values are applicable only to the simplified LSF Numerical and Analytical Methods in Geomechanics, 22: 721–
model (eq. [10]). If the Seed et al. (1985) method is adopted 748.
for the liquefaction evaluation, the corresponding mapping Fear, C.E., and McRoberts, E.C. 1995. Report on liquefaction po-
function defined in Table 3 must be used for risk-based deci- tential and catalogue of case records. Internal Research Report,
sions. Department of Civil Engineering, University of Alberta, Edmon-
ton, Alta.
Figueroa, J.L., Saada, A.S., Liang, L., and Dahisaria, N.M. 1994.
Evaluation of soil liquefaction by energy principles. Journal of
Conclusions Geotechnical Engineering, ASCE, 120(9): 1554–1569.
A simplified, SPT-based limit state function (LSF) for liq- Goh, A.T.C. 1994. Seismic liquefaction potential assessed by neu-
uefaction resistance is developed based on artificial neural ral networks. Journal of Geotechnical Engineering, ASCE,
network analysis of a large database of field performance re- 120(9): 1467–1480.
cords. The large number of field cases with (N1)60,cs between Goh, A.T.C. 1996. Neural network modeling of CPT seismic lique-
5 and 25 allows the LSF to be well established in this range. faction data. Journal of Geotechnical Engineering, ASCE,
122(1): 70–73.
The developed LSF can be used to evaluate CRR based on
Idriss, I.M. 1999. An update of the Seed–Idriss simplified proce-
(N1)60,cs. This CRR model may be used in conjunction with
dure for evaluating liquefaction potential. In A special CD-Rom
the CSR model defined in eq. [3] and the PL–FS mapping volume of the Transportation Research Board Workshop on New
function (eq. [13]) derived in this study. The derived CRR Approaches to Liquefaction Analysis. FHWA-RD-99-165, Fed-
model and the mapping function together form a practical eral Highway Administration, Washington, D.C.
tool for risk-based liquefaction evaluation using SPT. A risk- Juang, C.H., and Chen, C.J. 1999. CPT-based liquefaction evalua-
based design chart (Fig. 10) is created based on the derived tion using artificial neural networks. Journal of Computer-Aided
mapping function. Civil and Infrastructure Engineering, 14: 221–229.
From a risk-based design viewpoint, use of a limit state Juang, C.H., and Chen, C.J. 2000. A rational method for develop-
boundary with an artificially built-in conservatism has no ad- ment of limit state for liquefaction evaluation based on shear
vantages over those without. The degree of conservatism is a wave velocity. International Journal for Numerical and Analyti-
design issue that can and should be addressed with the selec- cal Methods in Geomechanics, 24: 1–27.

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:21:02 AM
 Color profile: Generic CMYK printer profile
Composite Default screen

1208 Can. Geotech. J. Vol. 37, 2000

Juang, C.H., and Jiang, T. 2000. Assessing probabilistic methods uted processing. Edited by D.E. Rumelhart and J.L. McClelland.
for liquefaction potential evaluation. In Soil Dynamics and MIT Press, Cambridge, Mass., Vol. 1, Chap. 8, pp. 318–362.
Liquefaction 2000. American Society of Civil Engineers, Seed, H.B., and Idriss, I.M. 1971. Simplified procedure for evaluat-
Geotechnical Special Publication, No. 107, pp. 148–162. ing soil liquefaction potential. Journal of the Soil Mechanics and
Juang, C.H., Chen, C.J., and Tien, Y.-M. 1999a. Appraising cone Foundation Division, ASCE, 97(SM9): 1249–1273.
penetration test based liquefaction resistance evaluation meth- Seed, H.B., and Idriss, I.M. 1982. Ground motions and soil lique-
ods: artificial neural network approach. Canadian Geotechnical faction during earthquake. Monograph Series, Earthquake Engi-
Journal, 36: 443–454. neering Research Institute, Berkeley, Calif.
Juang, C.H., Rosowsky, D.V., and Tang, W.H. 1999b. A reliability- Seed, H.B., Tokimatsu, K., Harder, L.F., and Chung, R. 1985. In-
based method for assessing liquefaction potential of sandy soils. fluence of SPT procedures in soil liquefaction resistance evalua-
Journal of Geotechnical and Geoenvironmental Engineering, tions. Journal of Geotechnical Engineering, ASCE, 111(GT12):
ASCE, 125(8): 684–689. 1425–1445.
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

Juang, C.H., Chen, C.J., Tang, W.H., and Rosowsky, D.V. 2000. Toprak, S., Holzer, T.L., Bennett, M.J., and Tinsley, J.C., III. 1999.
CPT-based liquefaction analysis, Part 1: Determination of limit CPT- and SPT-based probabilistic assessment of liquefaction po-
state function. Géotechnique, 50(5): 583–592. tential. In Proceedings of the 7th US–Japan Workshop on Earth-
Liao, S.C.C., Veneziano, D., and Whitman, R.V. 1988. Regression quake Resistant Design of Lifeline Facilities and
models for evaluating liquefaction probability. Journal of Countermeasures Against Liquefaction, Seattle, August 1999.
Geotechnical Engineering, ASCE, 114(4): 389–411. Multidisciplinary Center for Earthquake Engineering Re-
Najjar, Y.M., and Hossam, E.A. 1998. CPT-based liquefaction po- search, Buffalo, N.Y., pp. 69–86.
tential assessment: a neuronet approach. In Geotechnical earth- Youd, T.L., and Idriss, I.M. (Editors). 1997. Proceedings of the
quake engineering and soil dynamics III. American Society of NCEER Workshop on Evaluation of Liquefaction Resistance of
Civil Engineers, Geotechnical Special Publication 75, pp. 542– Soils. Technical Report NCEER-97-0022, National Center for
553. Earthquake Engineering Research, State University of New York
Robertson, P.K., and Campanella, R.G. 1985. Liquefaction poten- at Buffalo, Buffalo, N.Y.
tial of sands using the cone penetration test. Journal of Youd, T.L., and Noble, S.K. 1997. Liquefaction criteria based on
Geotechnical Engineering, ASCE, 111(GT3): 384–403. statistical and probabilistic analyses. In Proceedings of the
Robertson, P.K., and Wride, C.E. 1998. Evaluating cyclic liquefac- NCEER Workshop on Evaluation of Liquefaction Resistance of
For personal use only.

tion potential using the cone penetration test. Canadian Soils. Edited by T.L. Youd and I.M. Idriss. Technical Report
Geotechnical Journal, 35: 442–459. NCEER-97-0022, National Center for Earthquake Engineering
Rumelhart, D.E., Hinton, G.E., and Williams, R.J. 1986. Learning Research, State University of New York at Buffalo, Buffalo,
internal representations by error propagation. In Parallel distrib- N.Y., pp. 201–215.

© 2000 NRC Canada

I:\cgj\Cgj37\Cgj06\[Link]
Monday, December 04, 2000 10:21:02 AM
 This article has been cited by:

1. Pradyut K. Muduli, Sarat K. Das. 2013. Evaluation of liquefaction potential of soil based on standard penetration test using multi-
gene genetic programming model. Acta Geophysica . [CrossRef]
2. A. Johari, A.R. Khodaparast. 2013. Modelling of probability liquefaction based on standard penetration tests using the jointly
distributed random variables method. Engineering Geology 158, 1-14. [CrossRef]
3. J. Karthikeyan, Dookie Kim, Bhairevi G. Aiyer, Pijush Samui. 2013. SPT-based liquefaction potential assessment by relevance
vector machine approach. European Journal of Environmental and Civil Engineering 17:4, 248-262. [CrossRef]
4. Silvia García. 2013. Soft Schemes for Earthquake-Geotechnical Dilemmas. International Journal of Geophysics 2013, 1-34.
[CrossRef]
5. Pijush Samui, J. Karthikeyan. 2012. Determination of liquefaction susceptibility of soil: a least square support vector machine
Can. Geotech. J. Downloaded from [Link] by University of Sydney on 01/31/14

approach. International Journal for Numerical and Analytical Methods in Geomechanics n/a-n/a. [CrossRef]
6. C. Kayadelen. 2011. Soil liquefaction modeling by Genetic Expression Programming and Neuro-Fuzzy. Expert Systems with
Applications 38:4, 4080-4087. [CrossRef]
7. Y. Jafarian, A. Sadeghi Abdollahi, R. Vakili, M.H. Baziar. 2010. Probabilistic correlation between laboratory and field liquefaction
potentials using relative state parameter index (ξR). Soil Dynamics and Earthquake Engineering 30:10, 1061-1072. [CrossRef]
8. Ya-Fen Lee, Yun-Yao Chi, C. Hsein Juang, Der-Her Lee. 2010. Annual probability and return period of soil liquefaction in
Yuanlin, Taiwan attributed to Chelungpu Fault and Changhua Fault. Engineering Geology 114:3-4, 343-353. [CrossRef]
9. Chung-Jung Lee, Ta-Kang Hsiung. 2009. Sensitivity analysis on a multilayer perceptron model for recognizing liquefaction cases.
Computers and Geotechnics 36:7, 1157-1163. [CrossRef]
10. Sanjay K. Jha, Kiichi Suzuki. 2009. Liquefaction potential index considering parameter uncertainties. Engineering Geology 107:1-2,
55-60. [CrossRef]
11. Sanjay K. Jha, Kiichi Suzuki. 2009. Reliability analysis of soil liquefaction based on standard penetration test. Computers and
For personal use only.

Geotechnics 36:4, 589-596. [CrossRef]

12. Adel M. Hanna, Derin Ural, Gokhan Saygili. 2007. Neural network model for liquefaction potential in soil deposits using Turkey
and Taiwan earthquake data. Soil Dynamics and Earthquake Engineering 27:6, 521-540. [CrossRef]
13. Bin-Lin Chu, Sung-Chi Hsu, Yi-Ming Chang. 2004. Ground behavior and liquefaction analyses in central Taiwan-Wufeng.
Engineering Geology 71:1-2, 119-139. [CrossRef]
14. Haiming Yuan, Susan Hui Yang, Ronald D Andrus, C Hsein Juang. 2004. Liquefaction-induced ground failure: a study of the
Chi-Chi earthquake cases. Engineering Geology 71:1-2, 141-155. [CrossRef]
15. C. Hsein Juang, Tao Jiang, Ronald D. Andrus. 2002. Assessing Probability-based Methods for Liquefaction Potential Evaluation.
Journal of Geotechnical and Geoenvironmental Engineering 128:7, 580-589. [CrossRef]
16. C. Hsein Juang, Haiming Yuan, Der-Her Lee, Chih-Sheng Ku. 2002. Assessing CPT-based methods for liquefaction evaluation
with emphasis on the cases from the Chi-Chi, Taiwan, earthquake. Soil Dynamics and Earthquake Engineering 22:3, 241-258.
[CrossRef]