You are on page 1of 11

Multiscale and Multidisciplinary Modeling, Experiments and Design

https://doi.org/10.1007/s41939-023-00148-x

ORIGINAL PAPER

Suitability assessment of the best liquefaction analysis procedure


based on SPT data
Divesh Ranjan Kumar1 · Pijush Samui1 · Avijit Burman1

Received: 4 January 2023 / Accepted: 14 February 2023


© The Author(s), under exclusive licence to Springer Nature Switzerland AG 2023

Abstract
At present, there exist many methods for liquefaction analysis of a soil deposit. Some of them are suitable for only coarse-
grained soils, while a few others can also evaluate the liquefaction potential of both fine- and coarse-grained soils. It is important
to identify the most suitable method for liquefaction analysis. The current study looks at measuring soil’s liquefaction potential
using various methods such as Seed (ASCE J Soil Mech Found Div 97:1249–1273, 1971), Toprak et al. (CPT-and SPT-based
probabilistic assessment of liquefaction potential. In: 7th US–Japan Workshop on Earthquake Resistant Design of Lifeline
Facilities and Countermeasures against Liquefaction, Seattle. Multidisciplinary Center for Earthquake Engineering Research
Buffalo, NY, p 18,1999), and Idriss and Boulanger (Soil Dyn Earthq Eng 26:115–130, 2006). The main objective of this
study is to identify the most suitable method for liquefaction analysis based on factor of safety and Performance Fitness
Error Metrics (PFEMs), Rank analysis, Gini index, etc. The contribution of various independent variables, such as corrected
SPT N values, fine content, maximum horizontal ground acceleration, total vertical stress, total effective stress, magnitude
moment, and depth below ground level towards the evaluated liquefaction potential and probability of liquefaction has been
assessed through the use of Gini Index (GI). The computed probability of liquefaction (PL ) values are compared with the
given liquefaction status of case history data using Performance Fitness Error Metrices. Performance Fitness Error Metrics
for the Idriss and Boulanger (2006) method are found to be higher than those for Toprak et al. (1999) and Seed and Idriss
(ASCE J Soil Mech Found Div 97:1249–1273) methods. Based on these PFEMs, it is found that the Idriss and Boulanger
(2006) methods of liquefaction analysis at any site are more accurate than the other two methods. It is observed that Idriss
and Boulanger (2006) method gives the highest rate of successful prediction percentage of correctly predicted liquefied and
non-liquefied cases.

Keywords Liquefaction analysis · SPT · Cyclic resistance ratio · Cyclic stress ratio

1 Introduction lose all of it, rendering it incapable of supporting structures


above it. The word "liquefaction" is employed because soil
Liquefaction is one of the most hazardous outcomes that take silt will appear to flow like fluid as a result of this process.
place in saturated fine sand deposits due to earthquakes or As soil pore water pressure gradually rises, the cumulative
any external vibration. Soil collapses as a result of liquefac- effective stress decreases. When the total overburden pres-
tion during an earthquake. As a result of liquefaction, soil sure equals the pore pressure at a certain point, soil particles
loses a significant amount of its shear strength and can even are suspended in water, and the effective stress is decreased to
a minimum. Researchers became interested in the devastat-
B Divesh Ranjan Kumar ing occurrence of liquefaction during the Niigata earthquake
ranjandivesh453@gmail.com (1964), which caused numerous structures to tilt. Turkey suf-
Pijush Samui fered miserably in 1999, when significant portions of the city
pijush@nitp.ac.in were destroyed due to widespread liquefaction, resulting in
Avijit Burman the deaths of a large number of people. Following the seismic
avijit@nitp.ac.in incidents in Alaska (1964) and Nîgata in Japan (1964), Seed
and Idriss (1971) created a simplified approach for evaluating
1 Department of Civil Engineering, National Institute of
liquefaction probability based on in-situ tests. They defined a
Technology, Patna, India

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

term called factor of safety (F S ) computed as the ratio of the the likelihood of liquefaction, some geotechnical researchers
cyclic resistance ratio (CRR) to the cyclic stress ratio (CSR), created machine learning techniques. Goh (1994) proposed a
and this metric is used all over the world to determine whether neural network based model to forecast and evaluate the like-
or not any soil mass will liquefy at a certain depth below the lihood of liquefaction of saturated, cohesionless soil. Later,
ground. CSR represents the induced stress in the soil mass other researchers in geotechnical fields created a number
caused by the process of shaking or vibration, and CRR is a of machine learning methods, including neural networks,
measure of the resistance offered by the soil mass against col- support vector machine (SVM), genetic programming (GP),
lapse. In the years that followed, this process was altered and least square support vector machine (LSSVM) and stochastic
enhanced, in particular by Seed (1979) (1979), Seed (1982), gradient boosting (SGB) to perform the liquefaction analysis
Seed et al. (1985), and Youd and Idriss (1997a, 2001). This (Pal 2006; Samui and Karthikeyan 2013; Hanna et al. 2007a;
approach’s foundation is the simplification of the hypothesis, Samui et al. 2011; Samui and Hariharan 2015; Xue and Liu
which is done by treating the soil column as a rigid body under 2017). Zhang et al. (2015) proposed a non-parametric, mul-
the assumption that the actual peak shear stress produced at a tivariate adaptive regression spline (MARS) algorithm to
depth is always smaller than what is predicted by the method access the liquefaction potential of sands and sandy soils
of Seed and Idriss (1971). Few researchers studied differ- using the energy concept. Zhang et al. (2021) proposed the
ent experimental techniques to reduce the potential of any extreme learning machine (ELM) to assess the liquefaction
soil’s chances of liquefaction by deep soil mixing method potential of soil deposits using CPT data. The most popu-
(Rahmani et al. 2022). In the past few decades, many geotech- lar ML method in this field, the ANN model, has been used
nical researchers have successfully predicted the probability successfully to forecast liquefaction in the past (Samui and
of liquefaction using reliability analysis, considering model Sitharam 2011; Ramakrishnan et al. 2008). Some of the major
and parameter uncertainties (Haldar and Tang 1979; Juang drawbacks of ANN models include poor generalization capa-
et al. 2006). Liao et al. (1988) did a logistic regression analy- bility, a slow rate of convergence, and model overfitting,
sis on the available case histories to interpret the probability which can affect prediction of results. The ML-based liq-
of liquefaction (PL ). In this analysis, to predict the nominal uefaction evaluation models that are currently available are
probability, a simple regression equation has been proposed, inherently opaque since they prioritize accuracy over explain-
which reflects the uncertainties in the proposed model. ability. These models perform differently when compared
Juang et al. (1999) established a reliability method based to databases from other parts of the world because the cur-
on a Bayesian mapping function to predict the probability of rently existing liquefaction databases are modest in size and
liquefaction (PL ). Their proposed method was extended later contain proportionally more liquefaction events than non-
by Juang et al. (2002) to develop a mapping function between liquefaction events. Later on, other researchers developed
PL and FS . This mapping function was developed using the some new emerging machine learning techniques such as
procedure followed by Youd and Idriss (2001). The Bayesian random forest (RF), extreme gradient boosting (XGBoost),
network (BN) method has been identified as an effective tool and gradient boosting machine (GBM) (Kumar et al. 2022a;
to assist engineers in assessing the probability of earthquake- Kohestani et al. 2015; Zhang et al. 2021; Zou et al. 2022). The
induced liquefaction. Hu (2021) proposed a new approach existing ML-based models are preferable substitutes when
for predicting the liquefaction of gravelly soil using two dealing with large amounts of data and improving the accu-
Bayesian network models. In another work, Hu et al. (2022) racy of the prediction. Each machine learning technique has
proposed a hybrid BN model to predict earthquake-induced its own set of limitations due to parameters and model uncer-
liquefaction based on shear wave velocity (Vs ) (Hu and Liu tainty (Momeni et al. 2015).
2019; Hu 2021; Hu et al. 2022; Pirhadi et al. 2023). Cetin et al. After extensive literature studies on various available
(2004) developed a new set of probabilistic and determinis- methods of liquefaction potential evaluation, authors were
tic relationships for assessment of likelihood of initiation of unable to find any which discusses the suitability of a partic-
liquefaction using the maximum likelihood function within ular method for liquefaction potential evaluation of any soil
a Bayesian framework based on an updated database of deposit. It may be sometimes confusing to identify the best
case histories. Later, Idriss and Boulanger (2010), Boulanger method for achieving the desired response of the soil deposit
and Idriss (2014) developed a deterministic and probabilis- concerning the liquefaction phenomenon. The present work
tic method based on SPT data, which is an advancement of is dedicated towards finding out the suitability of a partic-
stress based method proposed by Seed and Idriss (1971). All ular method for studying the events of liquefaction. The
of the conventional methods and empirical relationships that primary goals of this work are to calculate the probability
have been offered to predict the liquefaction potential have of liquefaction (PL ) based on the SPT dataset and the soil
numerous assumptions and approximations. The liquefaction liquefaction safety factor utilizing a variety of techniques,
assessment problems are fundamentally highly nonlinear. To including the Idriss and Boulanger (2006), Seed and Idriss
get around nonlinearity and other difficulties in predicting (1971), and Toprak et al. (1999) methods. We will undertake

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

Table 1 Statistical descriptive of 


the independent variables Statistics Depth (m) N1, 60 FC (%) σv (kPa) σv (kPa) amax (g) Mw

Minimum 0.800 1.000 0.000 12.100 7.500 0.050 5.900


Q25% 3.500 7.000 15.000 62.925 40.000 0.200 7.400
Median 5.600 11.000 58.500 101.000 61.250 0.400 7.400
Q75% 8.800 18.000 94.000 164.075 96.675 0.400 7.600
Maximum 19.800 50.400 100.000 408.900 233.700 0.700 8.300
Mean 6.765 13.320 54.173 126.547 74.437 0.363 7.398
Std. Dev 4.417 9.051 37.584 87.869 47.296 0.150 0.310
Skewness 1.037 1.192 – 0.121 1.121 1.068 0.392 – 1.912
Kurtosis 0.403 1.103 – 1.054 0.608 0.511 −0.208 5.129

this evaluation using streamlined deterministic and proba-


bilistic methods. To calculate the PL value, the factor of
safety against liquefaction for a different layer of soil is com-
puted using the available SPT pos-liquefaction data. Finally,
a comparative analysis has been performed to evaluate which
method has highest prediction accuracy of probability of liq-
uefaction (PL ). Each section of the paper follows the structure
described below. Section 2 presents the details of the methods
used to calculate the probability of liquefaction. Liquefac-
tion analysis is performed using 834 historic data locations
presented in Sect. 2. Then three methods including, Idriss
and Boulanger (2006), Seed and Idriss (1971), and Toprak
et al. (1999) methods result are compared, and we introduced
fifteen PFEMs, score analysis, Gini index, and criteria for liq-
uefaction and non-liquefaction to evaluate the performance
of utilized methods for soil liquefaction in Sect. 3. The con-
clusions are presented in the final section.
Fig. 1 Distribution of normalized values of all independent variables
implemented to perform the study

2 Data collection and statistical analysis


Figure 1 represents the normalized values of probability
The data used in this study were collected from the database distribution of all datasets. As per Fig. 1 and results presented
of SPT-based liquefaction case histories of past earthquakes in Table 1, the N1, 60 and Mw have the highest value of Skew-
from various literature, such as Tiwan (1999), Niigata (1964), ness (1.192 and − 1.912, respectively) and Kurtosis (1.103
Kocaeli (1999) earthquake, etc. Researchers have compiled a and 5.129 respectively) among all datasets and consequently
thorough list of these case histories (Cetin et al. 2004) (Hanna both of them deviate more from the normal distribution than
et al. 2007b). A total of 834 cases were included in this inves- the other variables.
tigation, of which 397 cases were found to have liquefied and
437 cases to have non-liquefied. Gravels to silt mixtures and
pure sand are all present in the dataset. Independent variables 3 Methodology
considered during this study are adjusted SPT blow count
N1, 60 , total stress σvo , depth (z), Fine content (FC), effective 3.1 Deterministic approaches

stress σvo , maximum horizontal ground acceleration amax ,
magnitude moment (Mw ). Table 1 shows the various statisti- When evaluating liquefaction using various available tech-
cal parameters of all independent variables. The terms Q25% niques, the Seed and Idriss (1971) technique has typically
and Q75% represent the corresponding quartile values of the been demonstrated to be a good fit and a workable choice.
related variable. The statistical characteristics of each inde- However, the method was originally developed for purely
pendent variables demonstrate the diversity of datasets at a cohesionless soils. Idriss and Boulanger (Hanna et al. 2007b;
suitable level. Youd and Idriss 1997b) modified the original Seed and Idriss

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

strategy by incorporating the effects of fine content (FC%) In Eq. (3), cyclic stress ratio and (N1 )60CS , the corrected
to represent the liquefaction response of more fine-grained equivalent clean-sand blow count must be calculated using
soils, and this improvement has been successfully imple- the method proposed by Youd and Idriss (2001).
mented to solve the liquefaction problems all over the world When an earthquake of any magnitude occurs, to adjust
for both sands and fine-grained soils alike. Toprak et al. the induced CSR the factor is calculated which is known as,
(1999) used Magnitude Scaling Factor (MSF) previously magnitude scaling factor (MSF). The mathematical equation
suggested by Youd and Idriss (2001) to suggest an equa- used to calculate the MSF is proposed by Youd and Idriss
tion for calculating the probability of liquefaction (PL ). MSF (2001)
is a parameter relating CSR with the earthquake magnitude
(M w ). Furthermore, the expression of stress reduction factor 102.24
(r d ) was also modified by Idriss and Boulanger (2006). In MSF  2.56
, (4)
MW
this paper, the probability of liquefaction has been evaluated
using these three different methodologies to identify the best where MW represent the moment magnitude of earthquake
available method for studying liquefaction problems. recorded on Richter scale.

3.2 Seed and Idriss (1971) method


3.4 Idriss and Boulanger (2006) method
Seed and Idriss (1971) proposed a stress-based equation to
Idriss and Boulanger (2006) proposed a stress-based sim-
estimate the Cyclic Stress Ratio (CSR), which can be evalu-
plified technique to estimate the value of cyclic stress ratio
ated using the following equation.
(CSR). Mathematical equation used for the calculation of
  
σv amax CSR presented in Eq. (5).
CSR  0.65  rd . (1)
σv g   
σv amax
CSR  0.65  rd . (5)
In Eq. 1, CSR represents the Cyclic Stress Ratio, σv is the σv g

vertical total stress, σv is the vertical effective stress,
amax is the maximum horizontal ground acceleration. To 
In Eq. (5), σv represents the vertical total stress, σv
convert the ultimate cyclic shear stress ratio to CSR, apply represents the vertical effective stress, amax represents the
the factor 0.65 and rd is the shear stress reduction factor. maximum horizontal ground acceleration and g represents
Probability of liquefaction is calculated using the logistic the acceleration due to gravity. To convert the ultimate cyclic
regression equation proposed by Juang et al. (2002). shear stress ratio to CSR, apply the factor 0.65 and rd is the
  shear stress reduction factor calculated by method proposed
PL by Idriss and Boulanger (2006). The capacity of soil to with-
Logit (PL )  ln  10.1129
1 − PL stand liquefaction, as calculated with the help of corrected
− 0.25729 (N1 )60CS + 3.4825 ln (CSR7.5 ) . (2) SPT blow N value of each soil layer in terms of cyclic resis-
tance ratio (CRR) using the following Eq. (6).
In Eq. 2, (N1 )60CS is the correction to the clean sand        
(N1 )60cs (N1 )60cs 2 (N1 )60cs 3 (N1 )60cs 4
equivalent SPT penetration resistance value, which is com- CRR  exp + − + − 2.8
14.1 126 23.6 25.4
puted using the equation proposed by Youd and Idriss
(1997b) and CSR is calculated using Eq. (1). (6)

3.3 Toprak et al. (1999) method CRR


FS  . (7)
CSR
To create SPT-based probabilistic liquefaction boundary
curves, Toprak et al. (1999) (Toprak et al. 1999) used logistic The probability of liquefaction (PL ) is calculated as a func-
regression analyses. Based on the updated worldwide lique- tion of factor of safety (FS ) which is the ratio of CRR to CSR
faction database (total no of dataset  440), the probability using Eq. (7). This PL –FS mapping function is unique and
of liquefaction evaluated by means of the logistic regression factor of safety must be calculated by the method of Idriss
Eq. (5). and Boulanger (2006).

  1
Logit (PL )  ln
PL
 10.4459 − 0.2295 (N1 )60CS PL  . (8)
1 − PL 1 + exp[7.55(FS − 0.95)]
+ 4.0573ln (CSR/MSF) . (3)

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

Table 2 Liquefaction
classification criteria (Kumar Criteria Probability of liquefaction Liquefaction classification Non-liquefaction
et al. 2022b) classification

A PL > 0.85 High –


B PL > 0.65 Intermediate –
C PL > 0.5 Low –
D PL < 0.15 – High
E PL < 0.35 – Intermediate
F PL < 0.5 – Low

Table 3 Result of liquefied and


non-liquefied cases Criteria for PL Toprak method Idriss and Boulanger Seed and Idriss
method method

Count of Rate Count of Rate (%) Count of Rate


successful (%) successful successful (%)
prediction prediction prediction

Total 397 observed liquefied cases


A(PL > 0.85) 276 70 320 81 285 71
B(PL > 0.65) 340 86 352 89 339 85
C(PL > 0.5) 356 89 354 89 356 89
Total 437 observed non-liquefied cases
D(PL < 0.15) 95 22 133 30 93 21
E(PL < 0.35) 154 35 164 38 153 35
F(PL < 0.5) 188 43 196 45 185 42
Total 834 cases
A+D 371 45 453 54 378 44
B+E 494 59 516 62 492 58
C+F 544 65 550 66 541 64

4 Results and discussion occurrences. The SPT-based dataset of 834 cases, of which
397 cases had surface signs of liquefaction and 437 cases
In this study, an effort has been made to examine the most did not, provided a suitable basis for assessing the efficacy
effective technique for predicting soil liquefaction using data of the suggested approaches in this section. The accuracy of
from 834 locations. A mathematical calculation is performed the proposed methods using field data is evaluated using the
on the given case history SPT-based field data for this pur- likelihood of liquefaction computed using the three recom-
pose. The various soil parameters considered in this study mended methods. The accuracy of the proposed approaches
were used to calculate the factor of safety (Fs ) and prob- is evaluated in the current study using three criteria (A–C):
ability of liquefaction (PL ) to state whether the soil will High PL (0.85–1.0), Intermediate PL (0.65–1.0), and Low
liquify or not. A total of 834 case histories were investigated PL (0.65–1.0). (0.5–1.0). Similar to how Criteria (D) defined
in which actual field observations related to the occurrence as High PL (0.15–0), Criteria (E) defined as Intermediate PL
of the liquefaction phenomenon were stated. Utilizing the (0.35–0), and Criteria (F) defined as Low PL apply to non-
three available probabilistic methods mentioned in Sect. 3, liquefied instances (0.5–0). In Table 2, these classification
the probability of liquefaction (PL ) is calculated and value criteria are shown.
of PL greater than 0.5 is defined as liquefied soil and value Results are shown in Table 2, which also classifies the
of PL less than 0.5 is defined as non-liquefied soil. liquefiable soil (i.e., Criteria A–C) and non-liquefiable (i.e.,
Criteria D–F) based on the rate of successful prediction by
4.1 Criteria for liquefied and non-liquefied cases each proposed approach. Table 3 findings indicate that the
Idriss and Boulanger (2006) based criteria have a better suc-
This section compares the success rates of the vari- cess rate for successful prediction (A  81% and B  89%)
ous approaches for predicting liquefied and non-liquefied for the total of 397 observed liquefied cases than Toprak

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

Fig. 2 Illustration of actual and predicted cases for Toprak method Fig. 4 Illustration of actual and predicted cases for Seed and Idriss
method

Predicted
Actual
Liquefied (+) Liquefied (-)

Liquefied (+) TP FN
Liquefied (-) FP TN

Fig. 5 Illustration of confusion matrix (2 × 2) for classification problem

Binary classification using various Performance Fitness


and Error Metrics (PFEMs). To completely assess the effec-
tiveness of the suggested methodologies when categorizing
liquefaction concerns, different PFEMS are employed in this
part. In the binary class scenario, which includes both liq-
Fig. 3 Illustration of actual and predicted cases for Idriss and Boulanger uefied and non-liquefied examples, there are four alternative
method outcomes for a single prediction. True negative (TN) and true
positive (TP) are the proper classifications (TP). A false posi-
(1999) (A  70%andB  86%) and Seed and Idriss (1971) tive (FP) happens when the output is inaccurately expected to
(A  71%andB  85%) methods. Furthermore, compared be negative, and a false negative (FN) occurs when the result
to Toprak (D  22%andE  35%) and Seed and Idriss (D  is incorrectly labeled as negative. The confusion matrix (2
21%andE  35%) approaches, Idriss and Boulanger (A  × 2) shown in Fig. 5 can be used to evaluate these matrixes.
30%andB  38%) based criterion has a greater success PFEMs are employed to clarify how probabilistic predictions
rate of non-liquefiable case prediction. In overall, for all of proposed methods relate to actual output. These mea-
834 cases, it can be concluded that the Idriss and Boulanger surements frequently relate to the error variance between
(A+D  54%, B+E  62%andC+F  66%) based method expected and observed data. We used certain PFEMS in
attained higher successful prediction rate than Toprak (A + this paper that can be used to assess categorization prob-
D  45%, B+E  59%andC+F  65%) and Seed and Idriss lems (Naser and Alavi 2021). The mathematical expression
(A + D  44%, B + E  58%andC + F  64%) methods. of these PFEMS is listed below:
Scatter plots are also shown for liquefied cases and non- TP
liquefied cases for all the proposed methods between the TPR  , (9)
TP + FN
parameters N1, 60 and CSR. Figures 2, 3, and 4 display the
scatter plot of total liquefied and non-liquefied cases drawn FN
FNR  , (10)
for Toprak et al. (1999), Idriss and Boulanger (2006). and FN + TP
Seed and Idriss (1971) method for actual and predicted TP
dataset. PPV  , (11)
TP + FP

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

TN negative cases that are mistakenly classified as positive cases,


N PV  , (12)
TN + FN should be zero for perfect prediction accuracy. However, the
FPR score for the Idriss and Boulanger based criteria (0.650)
FP
FPR  , (13) is lower than the Toprak (0.666) and Seed and Idriss val-
FP + TN ues (0.670). The percentage of people who have a positive
FP test result when the actual condition is negative is known as
FDR  , (14)
FP + TP the False Discovery Rate (FDR). The FDR achieved in this
study for Idriss and Boulanger (0.445) is lower than that for
FN
FOR   1 − NPV, (15) the Toprak (0.450) and Seed and Idriss (0.451) methods, indi-
FN + TPN cating that the Idriss and Boulanger method performed better
2TP than the Toprak and Seed and Idriss methods. False Omission
F1  , (16)
2TP + FP + FN Rate (FOR) measures the percentage of people whose test
results were negative but whose true condition was positive.
TP × TN − FP × FN
MCC  √ , In comparison to Toprak (0.223) and Seed and Idriss (0.222)
(TP + FP)(TP + FN)(TN + FP)(TN + PN) approaches, the value obtained for Idriss and Boulanger
(17)
(0.219) based criteria is lower. The accuracy of the proposed
TP + TN methods is measured by the F1 score. The F1 score value pro-
Accuracy  , (18)
FN + TP + FN + TN duced for the Idriss and Boulanger (0.608) criteria is higher
than that for the Toprak (0.601) and Seed and Idriss (0.600)
TP
Sensitivity  , (19) techniques, showing that the Idriss and Boulanger criteria is
FN + TP
more accurate at predicting the likelihood of liquefaction.
TN Additionally, the performance of a binary classifier is
Specificity  , (20)
FP + TN also tested using Balanced Accuracy (BA), miss or false
negative rate (FNR), G (mean)error and the Matthews correla-
G (mean)error  1 − G mean , (21) tion coefficient (MCC). Balanced accuracy and G (mean)error
are defined in Eqs. (23) and (21) respectively. The number

G mean  (Sensitivity × Specificity), (22) of earthquakes that may have been mistakenly labeled as
not being tsunamigenic is shown by the miss rate. When
 
TP TN each class’s performance is noteworthy and expected to be
BA  0.5 × + . (23) high at the same time, Gmean is frequently utilized (Kubat
TP + FN TN + FP
and Matwin 1997; Yuan and Liu 2011). It is the geometric
Here, TP stands for correctly classified positive instances, mean of each instance of each class’s individual accuracy.
TN for correctly classified negative instances, FN for incor- To evaluate the effectiveness of the classifier model, differ-
rectly classified positive instances, and FP for incorrectly ent research have also used Gmean in terms of error rate in
classified negative instances, respectively. The binary classi- addition to F1 score. when the proportion of positive and
fication forecasts and observations have been compared using negative data differs noticeably, Matthews (1975) introduced
PFEMS data shown in Table 4. The percentage of true posi- the Matthews Correlation Coefficient (MCC), which is more
tives that are accurately classified as positives is measured by useful in these situations. The MCC value should be in the
the true positive rate (TPR). Indeterminate outcomes are not range of − 1.0 and + 1.0. On the other hand, a bias in favor
taken into account in this exercise. We found that the TPR val- of the higher side value denotes a more precise prediction. In
ues for Idriss and Boulanger based criteria (0.892) is higher comparison to Toprak and Seed and Idriss based criteria, the
than that for Toprak (0.882) and Seed and Idriss (0.887) tech- MCC value from Idriss and Boulanger (0.285) based criteria
niques. Positive predictive value measures the percentages of demonstrates that the Idriss and Boulanger based criterion
positive observations that result in real positive values (PPV), for liquefaction probability prediction performs better.
and the PPV value produced for Idriss and Boulanger (0.555)
based criteria are again shown to be higher than Toprak 4.2 Score analysis
(0.550) and Seed and Idriss (0.549). The best and worst values
for the Negative Predictive Value (NPV) are one and zero, Score analysis also performed to compare the effectiveness
respectively. It calculates the proportion of false positives of the proposed methods. The score is calculated for each
among observations that appear to be negative. The Idriss and method corresponding to their performance fitness and error
Boulanger based criteria’s NPV value (0.781) is higher than matrices. The range of score value is defined on the basis
the Toprak and Seed and Idriss values (0.777 and 0.778). The of total number of methods we have used in this study i.e.,
False Positive Rate (FPR), which counts the percentage of 1–3 (3  total number of methods used). In this study score

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

Table 4 Observed and ideal


values of PFEMs Matrices Methods

Toprak Idriss and Boulanger Seed and Idriss Ideal value

TPR 0.882 0.892 0.887 1.0


FNR 0.105793 0.108312 0.103275 0.0
PPV 0.550 0.555 0.549 1.0
NPV 0.777 0.781 0.778 1.0
FPR 0.666 0.650 0.670 0.0
FDR 0.450 0.445 0.451 0.0
FOR 0.223 0.219 0.222 0.0
F 1 score 0.601 0.608 0.600 1.0
MCC 0.273 0.285 0.272 − 1.0 to + 1.0
Accuracy 0.601 0.608 0.600 1.0
Sensitivity 0.894 0.892 0.897 1.0
Specificity 0.334 0.350 0.330 –
G(mean) error 0.453 0.441 0.456 0.0
BA 0.614151 0.620901 0.613122 1.0

Table 5 Score analysis result rank 1, followed by the Toprak et al. (1999) and Seed and
Idriss (1971) methods.
Methods Toprak et al. Idriss and Seed and
(1999) Boulanger Idriss (1971)
(2006) 4.3 Assessment of relative contribution of different
parameters in calculation of PL using Gini Index
TPR 1 3 2 (GI)
FNR 2 1 3
PPV 2 3 1 The relative significance of independent variables for pre-
NPV 1 3 2 dicting probability of liquefaction of soil deposit is analyzed
FPR 2 3 1 using the Gini index (GI). The Gini index of each indepen-
FDR 2 3 1 dent variables is computed, and the relative importance has
FOR 1 3 2
been judged. A higher value of the Gini index means more
significance of that individual independent variables for the
F 1 score 2 3 1
prediction of probability of liquefaction of soil deposit. The
MCC 2 3 1
mathematical equation used for the calculation of Gini index
Accuracy 2 3 1
is given as follows:
Sensitivity 2 1 3
 n 
Specificity 2 3 1  n 
 
GI   Cum X i−1 Cum PLi − Cum X i Cum PLi−1 ,
G(mean) error 2 3 1  
i2 i2
BA 2 3 1
(24)
Total Score 25 38 21
Rank 2 1 3 where CumX represents the cumulative value of independent
variables, CumPL represent the cumulative value of proba-
bility of liquefaction for n numbers of observed data.
value is defined on the basis of obtained value of PFEMs. The In Fig. 6, the relative importance of each independent vari-
methods having maximum value for any particular PFEMs ables has been shown for all three proposed methods. The
having ideal value 1.0 is awarded as maximum score (i.e., Gini index value for all three methods for different input

3) and methods having maximum value for any particular variables such as z,σvo ,σvo ,N1, 60 ,FC,amax , and Mw are also
PFEMs having ideal value 0.0 is awarded as minimum score presented in Table 6. The result presented in Table 6 con-
(i.e., 0). In this analysis Idriss and Boulanger (2006) method cludes that the effect of fine content is very less for Seed and
has assigned as the higher score value (i.e.,38) presented in Idriss (1971) methods followed by Toprak et al. (1999) and
Table 5, so that Idriss and Boulanger (2006) has assigned as Idriss and Boulanger (2006) method. The Gini index value

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

TPR value of 0.892 for the Idriss and Boulanger-based liq-


uefaction strategy indicates that 89.2% of the rate is indeed
positive. TPR values for the other two techniques, Toprak and
Seed and idriss, however, are 0.882 and 0.887, respectively.
The research also reveals that other performance fitness and
error matrices data fall within acceptable bounds. Score anal-
ysis also establishes that Idriss and Boulanger (2006) method
is superior.
The identification of Idriss and Boulanger (2006) method
as the best method for liquefaction analysis has been done on
the 834 SPT based datasets of liquefaction events. It is nec-
essary to check the procedure of assessing the suitability of a
Fig. 6 Relative importance of independent variables
particular method for liquefaction study with a larger dataset.
It should be noticed that the present paper discusses a com-
prehensive technique for finding out the relative suitability
Table 6 Gini index value for all proposed methods
of a particular method with other methods while studying
Toprak et al. Idriss and Seed and any physical phenomenon. Therefore, this technique can be
(1999) Boulanger Idriss (1971) further applied for comparing any other methods for grading
(2006)
purpose.
Z (m) 241.5266 320.4422 298.1037
Supplementary Information The online version contains supplemen-
(N1 )60 882.7739 974.8947 885.8942 tary material available at https://doi.org/10.1007/s41939-023-00148-x.
FC 80.37069 500.9136 3.495187
Acknowledgements The authors are thankful to colleague at National
σvo (kPa) 4158.41 5644.592 5222.319

Institute of Technology Patna who effectively contributed to this study.
σvo (kPa) 3722.714 4734.248 4415.888
amax (g) 10.60928 14.13179 10.97233 Author contributions This research is being done with the effort of three
authors. First author had carried out the study and wrote the manuscript
Mw 196.2717 261.4381 202.9882 under the guidance of third author. Second author review and editing
the paper. All authors have read and approved the final manuscript.

Funding The author confirms that he is not currently in receipt of any


obtained for all independent variables in the case of Seed research funding relating to the research presented in this manuscript.
and Idriss (1971) method is found to be less as compared to
Toprak et al. (1999) and Idriss and Boulanger (2006) method. Availability of data and materials The datasets used and analyzed dur-
ing the current study will be made available from the corresponding
author on reasonable request.

5 Conclusions Declarations

In this study, three different approaches namely, Idriss Conflict of interest The author declares that he has no competing inter-
ests.
and Boulanger (2006), Toprak et al. (1999) and Seed and
Idriss (1971) are used to perform the liquefaction analysis. Ethics approval and consent to participate Not applicable.
SPT based case histories data having various soil parame-
ters was taken from the records of several past earthquakes Consent for publication Not applicable.
as mentioned above to calculation of probability of lique-
faction. In this paper, the six factors namely, depth, SPT
blow count, fine content, total stress, effective stress, maxi-
mum horizontal ground acceleration, and moment magnitude References
are considered for evaluation of seismic liquefaction. The
most effective technique for calculating the probability of Boulanger RW, Idriss IM (2014) CPT and SPT based liquefaction trig-
soil liquefaction has been determined by the use of the gering procedures. Cent Geotech Model 1:134
various of performance fitness and error matrices. In com- Cetin KO, Seed RB, Der Kiureghian A, Tokimatsu K, Harder LF,
Kayen RE, Moss RES (2004) Standard penetration test-based
parison to Toprak and Seed and idriss methods, the Idriss and
probabilistic and deterministic assessment of seismic soil liquefac-
Boulanger-based criteria have a higher rate of effective pre- tion potential. J Geotech Geoenvironmental Eng 130:1314–1340.
diction, according to the results shown in Table 5. The higher https://doi.org/10.1061/(asce)1090-0241(2004)130:12(1314)

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

Goh ATC (1994) Seismic liquefaction potential assessed by neural net- Pirhadi N, Wan X, Lu J, Hu J, Ahmad M, Tahmoorian F (2023) Seismic
works. J Geotech Eng 120:1467–1480. https://doi.org/10.1061/ liquefaction resistance based on strain energy concept considering
(ASCE)0733-9410(1994)120:9(1467) fine content value effect and performance parametric sensitivity
Haldar A, Tang WH (1979) Probabilistic evaluation of liquefaction analysis. C Model Eng Sci 135:733–754
potential. J Geotech Eng Div 105:145–163 Rahmani F, Hosseini SM, Khezri A, Maleki M (2022) Effect of grid-
Hanna AM, Ural D, Saygili G (2007a) Evaluation of liquefaction poten- form deep soil mixing on the liquefaction-induced foundation
tial of soil deposits using artificial neural networks. Eng Comput settlement, using numerical approach. Arab J Geosci 15:1112
24:5–16. https://doi.org/10.1108/02644400710718547 Ramakrishnan D, Singh TN, Purwar N, Barde KS, Gulati A, Gupta
Hanna AM, Ural D, Saygili G (2007b) Neural network model for S (2008) Artificial neural network and liquefaction susceptibility
liquefaction potential in soil deposits using Turkey and Taiwan assessment: a case study using the 2001 Bhuj earthquake data,
earthquake data. Soil Dyn Earthq Eng 27:521–540. https://doi.org/ Gujarat, India. Comput Geosci 12:491–501. https://doi.org/10.
10.1016/j.soildyn.2006.11.001 1007/s10596-008-9088-8
Hu J (2021) A new approach for constructing two Bayesian network Samui P, Hariharan R (2015) A unified classification model for model-
models for predicting the liquefaction of gravelly soil. Comput ing of seismic liquefaction potential of soil based on CPT. J Adv
Geotech 137:104304 Res 6:587–592. https://doi.org/10.1016/j.jare.2014.02.002
Hu J, Liu H (2019) Bayesian network models for probabilistic eval- Samui P, Karthikeyan J (2013) Determination of liquefaction suscepti-
uation of earthquake-induced liquefaction based on CPT and Vs bility of soil: a Least square support vector machine approach. Int
databases. Eng Geol 254:76–88 J Numer Anal Methods Geomech 37:1154–1161. https://doi.org/
Hu J, Wang J, Zhang Z, Liu H (2022) Continuous-discrete hybrid 10.1002/nag.2081
Bayesian network models for predicting earthquake-induced liq- Samui P, Sitharam TG (2011) Machine learning modelling for predict-
uefaction based on the Vs database. Comput Geosci 169:105231 ing soil liquefaction susceptibility. Nat Hazards Earth Syst Sci
Idriss IM, Boulanger RW (2006) Semi-empirical procedures for eval- 11:1–9. https://doi.org/10.5194/nhess-11-1-2011
uating liquefaction potential during earthquakes. Soil Dyn Earthq Samui P, Kim D, Sitharam TG (2011) Support vector machine for eval-
Eng 26:115–130. https://doi.org/10.1016/j.soildyn.2004.11.023 uating seismic-liquefaction potential using shear wave velocity.
Idriss IM, Boulanger RW (2010) Spt-based liquefaction triggering pro- J Appl Geophys 73:8–15. https://doi.org/10.1016/j.jappgeo.2010.
cedures. Rep. UCD/CGM-10. 2, 4–13 10.005
Juang CH, Rosowsky DV, Tang WH (1999) Reliability-based method Seed HB (1979) Soil liquefaction and cyclic mobility evaluation for
for assessing liquefaction. J Geotech Geoenviron Eng 125(8) level ground during earthquakes. ASCE J Geotech Eng Div
Juang CH, Jiang T, Andrus RD (2002) Assessing probability-based 105:201–255. https://doi.org/10.1061/ajgeb6.0000768
methods for liquefaction potential evaluation. J Geotech Geoenv- Seed HB (1982) Ground motions and soil liquefaction during earth-
iron Eng 128:580–589 quakes. Earthquake Engineering Research Insititue, p 13
Juang CH, Fang SY, Khor EH (2006) First-order reliability method Seed HB, Idriss IM (1971) Simplified procedure for evaluating soil liq-
for probabilistic liquefaction triggering analysis using CPT. J uefaction potential. ASCE J Soil Mech Found Div 97:1249–1273.
Geotech Geoenviron Eng 132:337–350. https://doi.org/10.1061/ https://doi.org/10.1061/jsfeaq.0001662
(asce)1090-0241(2006)132:3(337) Seed H, Tokimatsu K, Harder LF, Chung RM (1985) Influ-
Kohestani VR, Hassanlourad M, Ardakani A (2015) Evaluation of ence of SPT procedures in soil liquefaction resistance evalu-
liquefaction potential based on CPT data using random forest. ations. J Geotech Eng 111:1425–1445. https://doi.org/10.1061/
Nat Hazards 79:1079–1089. https://doi.org/10.1007/s11069-015- (ASCE)0733-9410(1985)111:12(1425)
1893-5 Toprak S, Holzer TL, Bennett MJ, Tinsley JJ (1999) CPT-and SPT-
Kubat M, Matwin S (1997) Addressing the curse of imbalanced training based probabilistic assessment of liquefaction potential. In: 7th
sets: one-sided selection. In: Icml, Nashville, USA, p 179 US–Japan Workshop on Earthquake Resistant Design of Life-
Kumar DR, Samui P, Burman A (2022a) Prediction of probabil- line Facilities and Countermeasures against Liquefaction, Seattle.
ity of liquefaction using soft computing techniques. J Inst Eng Multidisciplinary Center for Earthquake Engineering Research
(India): Ser A 103(4):1195–1208. https://doi.org/10.1007/s40030- Buffalo, NY, p 18
022-00683-9 Xue X, Liu E (2017) Seismic liquefaction potential assessed by neu-
Kumar DR, Samui P, Burman A (2022b) Determination of best cri- ral networks. Environ Earth Sci 76:1–15. https://doi.org/10.1007/
teria for evaluation of liquefaction potential of soil. Transp Infra s12665-017-6523-y
Geotechnol 1–20. https://doi.org/10.1007/s40515-022-00268-w Youd TL, Idriss IM (1997a) Liquefaction criteria based on statistical
Liao SSC, Veneziano D, Whitman RV (1988) Regression models for and probabilistic analyses. In: Technical Report NCEER-97-0022,
evaluating liquefaction probability. J Geotech Eng 114:389–411. p 281
https://doi.org/10.1061/(ASCE)0733-9410(1988)114:4(389) Youd TL, Idriss IM (1997b) Proceeding of the NCEER workshop on
Matthews BW (1975) Comparison of the predicted and observed sec- evaluation of liquefaction resistance of soils. In: Proceeding of the
ondary structure of T4 phage lysozyme. BBA - Protein Struct NCEER workshop on evaluation of liquefaction resistance of soils,
405(2):442–451. https://doi.org/10.1016/0005-2795(75)90109-9. p 276
Momeni E, Jahed Armaghani D, Hajihassani M, Mohd Amin MF (2015) Youd TL, Idriss IM (2001) Liquefaction resistance of soils: sum-
Prediction of uniaxial compressive strength of rock samples using mary report from the 1996 NCEER and 1998 NCEER/NSF
hybrid particle swarm optimization-based artificial neural net- Workshops on evaluation of liquefaction resistance of soils. J
works. Meas J Int Meas Confed 60:50–63. https://doi.org/10.1016/ Geotech Geoenviron Eng 127:297–313. https://doi.org/10.1061/
j.measurement.2014.09.075 (asce)1090-0241(2001)127:4(297)
Naser MZ, Alavi AH (2021) Error metrics and performance fitness Yuan B, Liu W (2011) A measure oriented training scheme for imbal-
indicators for artificial intelligence and machine learning in engi- anced classification problems. In: Pacific–Asia Conference on
neering and sciences. Archit Struct Constr. https://doi.org/10.1007/ Knowledge Discovery and Data Mining. Springer, pp 293–303
s44150-021-00015-8 Zhang W, Goh ATC, Zhang Y, Chen Y, Xiao Y (2015) Assessment of
Pal M (2006) Support vector machines-based modelling of seis- soil liquefaction based on capacity energy concept and multivariate
mic liquefaction potential. Int J Numer Anal Methods Geomech adaptive regression splines. Eng Geol 188:29–37. https://doi.org/
30:983–996. https://doi.org/10.1002/nag.509 10.1016/j.enggeo.2015.01.009

123
Multiscale and Multidisciplinary Modeling, Experiments and Design

Zhang W, Wu C, Zhong H, Li Y, Wang L (2021) Prediction of undrained Publisher’s Note Springer Nature remains neutral with regard to juris-
shear strength using extreme gradient boosting and random forest dictional claims in published maps and institutional affiliations.
based on Bayesian optimization. Geosci Front 12:469–477. https://
doi.org/10.1016/j.gsf.2020.03.007 Springer Nature or its licensor (e.g. a society or other partner) holds
Zhang Yg, Qiu J, Zhang Y, Wei Y (2021) The adoption of ELM to exclusive rights to this article under a publishing agreement with the
the prediction of soil liquefaction based on CPT. Nat Hazards author(s) or other rightsholder(s); author self-archiving of the accepted
107:539–549. https://doi.org/10.1007/s11069-021-04594-z manuscript version of this article is solely governed by the terms of such
Zou M, Jiang W-G, Qin Q-H, Liu Y-C, Li M-L (2022) Optimized publishing agreement and applicable law.
XGBoost model with small dataset for predicting relative den-
sity of Ti-6Al-4V parts manufactured by selective laser melting.
Materials (basel). 15:5298. https://doi.org/10.3390/ma15155298

123

You might also like