Explainable Machine-Learning Predictions For Peak

applied
sciences
Article
Explainable Machine-Learning Predictions for Peak
Ground Acceleration
Rui Sun 1,2 , Wanwan Qi 1,2 , Tong Zheng 1,2, * and Jinlei Qi 1,2
1 Key Laboratory of Earthquake Engineering and Engineering Vibration, Institute of Engineering Mechanics,
China Earthquake Administration, Harbin 150080, China; iemsr@163.com (R.S.); iem_qiww@163.com (W.Q.);
qijinlei521@163.com (J.Q.)
2 Key Laboratory of Earthquake Disaster Mitigation, Ministry of Emergency Management,
Harbin 150080, China
* Correspondence: zhengt0928@163.com
Abstract: Peak ground acceleration (PGA) prediction is of great significance in the seismic design
of engineering structures. Machine learning is a new method to predict PGA and does have some
advantages. To establish explainable prediction models of PGA, 3104 sets of uphole and downhole
seismic records collected by the KiK-net in Japan were used. The feature combinations that make
the models perform best were selected through feature selection. The peak bedrock acceleration
(PBA), the predominant frequency (FP), the depth of the soil when the shear wave velocity reaches
800 m/s (D800), and the bedrock shear wave velocity (Bedrock V s ) were used as inputs to predict
the PGA. The XGBoost (eXtreme Gradient Boosting), random forest, and decision tree models were
established, and the prediction results were compared with the numerical simulation results The
influence between the input features and the model prediction results were analyzed with the SHAP
(SHapley Additive exPlanations) value. The results show that the R2 of the training dataset and
testing dataset reach up to 0.945 and 0.915, respectively. On different site classifications and different
PGA intervals, the prediction results of the XGBoost model are better than the random forest model
and the decision tree model. Even if a non-integrated algorithm (decision tree model) is used, its
prediction effect is better than the numerical simulation methods. The SHAP values of the three
machine learning models have the same distribution and densities, and the influence of each feature
Citation: Sun, R.; Qi, W.; Zheng, T.;
on the prediction results is consistent with the existing empirical data, which shows the rationality of
Qi, J. Explainable Machine-Learning the machine learning models and provides reliable support for the prediction results.
Predictions for Peak Ground
Acceleration. Appl. Sci. 2023, 13, 4530. Keywords: machine learning; XGBoost; explainable; SHAP; prediction of PGA
https://doi.org/10.3390/
app13074530
Academic Editors: José A. Peláez,

1. Introduction
Jian Song and Bin Ruan
The strong ground motion caused by earthquake will cause surface ruptures, building
Received: 24 February 2023 collapses, and casualties and thus presents a serious threat to the safety of human life and
Revised: 28 March 2023
property. The rapid estimation of ground motion intensity after an earthquake can be used
Accepted: 31 March 2023
to evaluate the loss after the earthquake. It is important for probabilistic seismic hazard
Published: 3 April 2023
analysis to predict the possible ground motion intensity in the potential source area before
an earthquake. Peak ground acceleration (PGA) can directly reflect the intensity of ground
motion [1,2]. PGA is an intuitive characterization parameter of seismic action characteristics
Copyright: © 2023 by the authors.
in the seismic design of engineering structures, and it plays an important role in earthquake
Licensee MDPI, Basel, Switzerland. early warning systems, seismic risk assessment, ground motion zoning, and the dynamic
This article is an open access article responses of buildings [3–7]. The accurate acquisition and prediction of the PGA is of great
distributed under the terms and significance in guiding post-earthquake rescue and reconstruction, as well as in the design
conditions of the Creative Commons and construction of major projects. With the progress of seismic technology, there are fewer
Attribution (CC BY) license (https:// and fewer casualties caused by building damage in strong earthquakes, but the functional
creativecommons.org/licenses/by/ losses of structures and infrastructures caused by site effects has not been significantly
4.0/). reduced. Therefore, site effects should be the focus of seismic engineering research [8–10].
Appl. Sci. 2023, 13, 4530. https://doi.org/10.3390/app13074530 https://www.mdpi.com/journal/applsci

Appl. Sci. 2023, 13, 4530 2 of 22
For areas with strong motion stations, the PGA can be obtained directly from the
stations. In addition, the PGA can be predicted by ground motion prediction equations
(GMPEs) and numerical simulation. GMPEs [11–14] are essentially predictions of ground
motion parameters for a certain type of site with common characteristics. They have the
characteristics of clear forms and fast calculation speeds, and they are commonly used in
ground motion zoning and probabilistic seismic hazard analysis. However, most GMPEs
have insufficient consideration of site effects, so when they are used for specific sites, the
necessary amendments need to be made according to the site’s characteristics. Numerical
simulation [15–19] includes one-dimensional, two-dimensional, and three-dimensional
analysis methods. The one-dimensional method has a simple constitutive relationship
and can consider the relative high-frequency components of ground motion at a lower
computational cost. Soil seismic response analysis programs based on one-dimensional
numerical simulation method, such as SHAKE [20] and DEEPSOIL [21], are widely used in
the world, but their calculation results on soft soil sites are not ideal.
With the continuous development of artificial intelligence technology, machine learn-
ing has also been widely used in the prediction of ground motion parameters. For example,
Kerh et al. [22] used epicentral distance, focal depth, and magnitude as input parameters to
establish a neural network model to predict the PGA of 10 railway stations along Taiwan’s
high-speed rail system and identify potentially dangerous stations. Arjun and Kumar [23]
established a PGA prediction artificial neural network model suitable for Japan and com-
pared and analyzed the prediction effects of the model with six input features and three
input features. Günaydın et al. [24] compared three artificial neural network methods using
the same data and inputs and developed a PGA prediction model for northern Turkey.
Derras et al. [25] established a fully data-driven PGA prediction model by avoiding any
prior function form specification. However, these machine learning models established by
the existing research mostly use magnitude, epicentral distance, etc., as input parameters,
and make less use of parameters that characterize site characteristics. The influence of
site conditions on seismic activity cannot be ignored, so it is necessary to consider site
conditions in input features. For site effects, Zhu et al. [26] developed an amplification
model using the random forest algorithm, which achieved better performance than the
physics-based modeling (GRA) using detailed 1D velocity profiles. Mori et al. [27] used
machine learning to establish ground motion prediction maps and provided a ground mo-
tion variability consistent with advanced numerical simulations results based on detailed
subsoil models.
It is equally important to understand the reason why the model makes a prediction
as well as the accuracy of the prediction. Simple models (such as linear models, single
decision tree models, etc.) are interpretable. Ensemble learning models or deep learning
models have better prediction performance and higher model complexity, which reduces
the interpretability of the model. At present, some methods have been applied to the
explanation of these models, such as LIME (Local Interpretable Model Agnostic Explana-
tion) [28] and SHAP (SHapley Additive exPlanations) [29–31]. LIME is a local explanation
method that can only be used to evaluate the contributions of features to a single prediction
and cannot be used to explain the model globally. SHAP does not have this limitation.
It can not only explain a single sample locally, but it can also explain all samples glob-
ally. Due to its unique advantages, the SHAP method has been successfully applied to
many scientific fields such as biomedical engineering [32–34], ecological engineering [35],
concrete strength prediction [36,37], and the prediction of rheological properties of self-
compacting concrete [38]. It has also been applied in earthquake engineering: for example,
Chen et al. [39] established an explainable prediction model of ground motion parameters
based on SHAP.
Based on the ground motion data of the KiK-net strong motion network in Japan, the
feature combination with the best performance of the models were selected by feature im-
portance ranking, and the PGA prediction models based on ground motion parameters and
site characteristic parameters were established by XGBoost (eXtreme Gradient Boosting).
Appl. Sci. 2023, 13, 4530 3 of 22
The SHAP was used to explain the machine learning models, and the influence of each
feature on the prediction results was analyzed.
Based on the ground motion data of the KiK-net strong motion network in Japan, this
paper established the machine learning model for predicting PGA. The input parameters
included both ground motion characteristics and site characteristics, and SHAP was used
to explain the model. Firstly, we introduced the data and algorithm principles required
for machine learning model training. Then, two ground motion characteristics and nine
site characteristics were selected for feature selection, and 14 feature importance ranking
methods were used for comprehensive scoring. We finally selected the feature combinations
that made the model perform best, including the peak bedrock acceleration (PBA), the
predominant frequency (FP), the depth of the soil when the shear wave velocity reaches
800 m/s (D800), and the bedrock shear wave velocity (Bedrock V s ), as the input features
for machine learning. Additionally, we compared the prediction results with the measured
records and numerical simulation results, respectively. Finally, the SHAP method was used
to explain the machine learning model and the variation laws of the SHAP value of each
input feature were analyzed.
2. Establishment of the Dataset

The KiK-net strong motion network covers 661 stations throughout Japan, and each of
its stations are equipped with three-component, high-sensitivity acceleration sensors on
the uphole and downhole [40]. The network has extremely rich ground motion records and
obtains site information through drilling for external sharing.
Based on the seismic records of the KiK-net, the seismic records between 1 March 1996
and 1 November 2020 were selected and filtered as follows:
(1) We only selected the horizontal site stations with strong seismometers on the
surface. Many stations in KiK-net are located on a hillside with large slopes rather than a
horizontal site, which is why topography has a certain influence on PGA. The aim of this
study was to predict the PGA of horizontal sites, mainly by taking the one-dimensional
horizontal site as the research object. The influence of terrain was not within the scope of
this study, so it was excluded.
(2) In terms of PGA selection, all records greater than 0.01 g were included in the
dataset as much as possible.
According to the above principles, a total of 40 KiK-net stations and 3104 sets of uphole
and downhole seismic records were collected (the uphole and downhole seismic records
represent the acceleration records of the surface and bedrock, respectively). To test the
applicability of machine learning in different site classification tasks, we divided the sites
into different classes. The site classification methods of different countries can reflect the
extent of soft and hard areas of the site, although their methods are different. For example,
China divides sites into 5 classes, America divides sites into 6 classes, and Japan divides
sites into 3 classes. Considering that the classification in Japan is slightly rough and that the
plasticity index is required as a parameter for class E and class F in the American code, but
these parameters are not included in the station profile, we classified the sites according
to “GB 50011–2010: Code for seismic design of buildings” in China [41]. The classification
criteria are shown in Table 1. Table 2 shows the number of stations corresponding to the
site and the number of seismic records, and the distribution of PGA is shown in Figure 1.
Table 1. Chinese code for classification of sites.
Shear Wave Velocity of Rock Vs or Equivalent Shear Site Class and Overburden Thickness
Wave Velocity of Soil Vse (m/s) I0 I1 II III IV
V s > 800 0
800 ≥ V s > 500 0
500 ≥ V se > 250 <5 ≥5
250 ≥ V se > 150 <3 3–50 >50
V se ≤ 150 <3 3–15 15–80 >80
I0 and I1 are subclasses of site class I.
Vs > 800 0
800 ≥ Vs > 500 0
500 ≥ Vse > 250 <5 ≥5
250 ≥ Vse > 150 <3 3–50 >50
Appl. Sci. 2023, 13, 4530 Vse ≤ 150 <3 3–15 15–80 >80 4 of 22
I0 and I1 are subclasses of site class I.
Table 2. Number
Table of of
2. Number stations and
stations records
and of of
records seismic activity
seismic at at
activity various sites.
various sites.
Site Class
Site Class Number of of
Number Stations
Stations Number of Seismic
Number Records
of Seismic Records
II II 6 6 1524
1524
III III 26 26 1296
1296
IVIV 8 8 284284
Total
Total 40 40 3104
3104
Figure 1. Distribution of PGA.

Figure 1. Distribution of PGA.
3. Establishment of the Machine Learning Model
3. 3.1.
Establishment of the Machine Learning Model
Algorithm Principle
3.1.3.1.1.
Algorithm Principle
Decision Tree Model
3.1.1. Decision
DecisionTree
treeModel
(DT) is a supervised learning method with strong readability and fast
running speed that
Decision tree (DT) canis deal with nonlinear
a supervised relationships
learning method with well. The decision
strong readabilitytreeand
in SciKit-
fast
Learn uses the CART (Classification and Regression Tree) algorithm. Its
running speed that can deal with nonlinear relationships well. The decision tree in SciKit- basic principle is to
formuses
Learn a binary decision
the CART tree functionand
(Classification by Regression
applying a cyclic dichotomy Its
Tree) algorithm. to the training
basic principledataset.
is
to form a binary decision tree function by applying a cyclic dichotomy to the training da-the
The CART decision tree model can be used for both classification and regression. When
target
taset. Thevariables are discrete
CART decision treecategory
model can values, it is for
be used a classification tree, and
both classification theregression.
and principle of
the minority obeying the majority is used to generate the predicted value.
When the target variables are discrete category values, it is a classification tree, and When the target
the
variables are continuous values, it is a regression tree, and the mean
principle of the minority obeying the majority is used to generate the predicted value.of the predicted values
is used
When theas the predicted
target variables value. In this paper,
are continuous the it
values, regression tree is tree,
is a regression used,andandthe
themean
algorithm
of
steps are as follows:
the predicted values is used as the predicted value. In this paper, the regression tree is
(1) and
used, Enter
thethe trainingsteps
algorithm dataset
are and stop calculation conditions;
as follows:
(2) Select the optimal segmentation dimension j and segmentation point s of the training dataset;
(1) Enter the training dataset and stop calculation conditions;
(3) Use the optimal segmentation dimension j and segmentation point s to segment regions;
(2) Select the optimal segmentation dimension j and segmentation point s of the training
(4) For the two divided sub-regions R1 , R2 , the above second and third steps are recur-
dataset;
sively called until the stopping condition is satisfied;
(3) Use the optimal segmentation dimension j and segmentation point s to segment re-
(5) Finally, the input space is divided into M regions R1 , R2 , . . . Rm , and finally, the CART
gions;
regression decision tree is generated.
The above stopping condition is usually one of the following conditions:
(1) The number of samples in the node is less than the predetermined value;
(2) The Gini coefficient (classification)/square error (regression) of the sample dataset is
less than the predetermined value;
(3) No more features.
3.1.2. Random Forest Regression Model

Random forest (RF) is a representative algorithm of the Bagging integration algorithm,
proposed by Breiman [42]. It takes the decision tree as the base evaluator and adds random
attributes in the training process so that the decision tree uses random characteristics
and random thresholds for node division. Randomness is the main idea of the algorithm.
Owing to the random sampling of training samples and attributes, the generalization ability
Appl. Sci. 2023, 13, 4530 5 of 22
of the model is improved, and the prediction accuracy is improved, without a significant
increase in the amount of computation. The main steps of establishing a random forest
regression model are as follows:
(1) From the N samples of the dataset containing M features, n training subsets are
randomly selected by the bootstrap method to construct n regression trees;
(2) During the construction of regression tree, m features (m ≤ M) are randomly
selected on its nodes, the optimal branches are determined by the specified optimality
criteria, and the tree is built recursively until the termination conditions are met;
(3) Gather all of the generated regression trees to form a random forest regression
model, and determine the prediction effect through the evaluation of the prediction indi-
cators of the model. If the prediction effect is poor, continue modeling by adjusting the
parameters of the random forest modeling process until the expected effect is achieved.
3.1.3. XGBoost Regression Model

XGBoost is an ensemble learning algorithm designed by Chen and Guestrin [43], it is
committed to making the lifting tree break through its own computational limit to achieve
the engineering goal of fast operation and superior performance. It uses the gradient
descent framework to improve the weak learner, which is an efficient implementation of
the GB (Gradient Boosting) algorithm. Compared with other GB algorithms, XGBoost can
achieve better prediction performance in a short time.
For a given training set N = {(x1 , y1 ), (x2 , y2 ), . . . , (xn , yn )}, select a sample (xi , yi ), and
xi ∈ Rn , yi ∈ Rn , i = 1, 2, . . . , n. Then, K additivity functions are used to predict the samples,
and the results are as follows:
n
ŷi = ∑ f k ( x i ), (1)
i =1
where xi is the n dimensional eigenvector of the ith sample; yi is the true value of the ith
sample; ŷi is the predicted value of the ith sample; f (x) is a function that maps the feature
to the leaf node weight of the tree structure, which can be expressed as f (x) = ω q (x); ω is
the weight of the leaf node; and q represents a tree structure.
The objective function of the XGBoost regression model is composed of the loss
function and regularization, which ensures the accuracy and generalization ability of the
model. The objective function formula is as follows:
n K
Obj = ∑ l (yi , ŷi ) + ∑ Ω( f k ), (2)
i =1 k =1
where Obj is the objective function, l(yi , ŷi ) is the loss function, and Ω(f ) is used to control
the complexity of the model and prevent overfitting. The model is trained in the following
iterative way:
ŷi (t) = ŷi (t−1) + f t ( xi ), (3)
where ŷi (t) is the predicted value of the ith sample after the ith iteration.
The iterative form of the objective function is
n
Obj(t) = ∑l yi , ŷi (t−1) + f t ( xi ) + Ω( f t ). (4)
i =1
If the loss function is expanded by second-order Taylor expansion, the objective

function can be written in the following form:
n 1
Obj(t) = ∑ l y ,
i iŷ ( t −1)
+ g i f t ( x i ) + h
2 i
f t
2
( x i ) + Ω ( f t ), (5)
i =1
where
( t −1)
gi = ∂ŷt−1 l (yi , ŷi ), (6)
i
Appl. Sci. 2023, 13, 4530 6 of 22
( t −1)
hi = ∂2ŷt−1 l (yi , ŷi ). (7)
i
It can be seen that the objective function of XGBoost mainly depends on the first-
order partial derivative and second-order partial derivative of the loss function, which is
also a great advantage of this method. As long as the first-order partial derivative and
second-order partial derivative are provided, XGBoost can support the use of a user-defined
loss function.
3.2. Feature Selection

The main factors that influence the site amplification effect include three aspects. One
is the ground motion characteristics, including amplitude, spectral characteristics and
duration. Another one is site characteristics, such as site class, the equivalent shear wave
velocity of the site, the thickness of the overburden layer, and the site’s fundamental period.
The third one is soil layer characteristics, including physical properties, shear wave velocity,
and depth of soil. In this paper, PBA and FP were selected to describe the characteristics
of the input ground motion, and there are many parameters used to describe the site
characteristics. For example, for the description of the equivalent shear wave velocity of
the site, there are not only the equivalent shear wave velocity (V se ) commonly used for site
classification in Chinese code, but also the average shear wave velocity within 20 m and
30 m below the site surface (V s20 and V s30 ). We use these factors as original input features
and perform feature selection to select features that have a greater influence on the model.
The following 11 factors are initial input features:
(1) Peak Bedrock Acceleration (PBA/g).
(2) Predominant Frequency (FP/Hz): Input the predominant frequency corresponding
to the Fourier spectrum of the seismic acceleration time history.
(3) Site Fundamental Period (SFP/s): The calculation formula is given in Equation (8) [44]:
N
4hi
T= ∑ v
, (8)
i =1 i
where T is the basic period of the site; hi and vi are the thickness and the shear wave velocity,
respectively, of the ith soil layer; and N is the total number of calculated soil layers.
(4) Overburden Thickness (OBT/m): This is determined according to the OBT in the
site classification in “GB 50011–2010: Code for seismic design of buildings” [41]. Generally,
the OBT is the distance from the ground to the top of the soil layer with shear wave velocity
greater than 500 m/s and the shear wave velocity of each underlying layer of rock and soil
is not less than 500 m/s.
(5) The depth of the soil when the shear wave velocity reaches 800 m/s (D800/m): For
the section where the shear wave velocity of the soil layer is less than 800 m/s, the buried
depth of the ground motion input surface is taken as D800.
(6) Profile Depth (Depth/m): This is the buried depth of the downhole recorder.
(7) Surface shear wave velocity (Surface V s /(m/s)): This is the shear wave velocity
where the uppermost layer of the site is located.
(8) Bedrock shear wave velocity (Bedrock V s /(m/s)): This is the shear wave velocity
of the bedrock where the input ground motion is located.
(9) Equivalent shear wave velocity (Vse /(m/s)): This is determined according to the
calculation formula of equivalent shear wave velocity in site classification in “GB 50011–2010:
Code for seismic design of buildings” [41]. Its calculation formula is given in Equation (9):
d0
Vse = n , (9)
∑ (di /vsi )
i =1
Appl. Sci. 2023, 13, 4530 7 of 22
where d0 is the smaller value between OBT and 20 m, di and vsi are the thickness and shear
wave velocity of soil layer i within the d0 , respectively, and n is the number of layers of soil
layer within the d0 .
(10) Average shear wave velocity within 20 m below the site surface (V s20 /(m/s)):
According to the previous numerical simulation and measured records, the amplification
effect of the site is mainly reflected in the upper soil layer, and the soil structure near the
surface has a great influence on the amplification effect. When the depth of the borehole
and wave velocity measurement is less than 20 m, it can be calculated according to the
velocity extrapolation formula established by Boore et al. [45]. The site profiles used in this
paper are more than 30 m, which can be calculated directly;
(11) Average shear wave velocity within 30 m below the site surface (V s30 /(m/s)).
The commonly used feature selection methods include the filter method, the embedded
method, and the wrapper method. The filter method is usually used as a preprocessing step.
It is completely independent of any machine learning algorithm and it selects features based
on scores in various statistical tests and various indicators of correlation. The embedded
method is a method that allows the algorithm to decide which features are used; that is,
feature selection and algorithm training are carried out at the same time. When using the
embedded method, we first use some machine learning algorithms for training to obtain
the weight coefficients of each feature and then select features from large to small according
to the weight coefficients. The wrapper method is very similar to the embedded method;
it also relies on the choice of the algorithm itself to complete feature selection. However,
the difference with the embedded method is that we often use an objective function as a
black box to help us select features, rather than input a threshold of an evaluation index or
statistic. The most typical objective function is recursive feature elimination (RFE).
In this study, we used XGBoost, random forest, and decision tree to establish machine
learning models, and their basic learner CART is not sensitive to feature scaling, so we did
not normalize or standardize the input parameters. We used 14 different feature selection
methods to rank the importance of the 11 extracted influencing factors. The methods include
filter methods (F-test and mutual information method), embedded methods (algorithm
model adopts decision tree, extra tree, adaptive boosting, random forest, gradient boosting,
and extreme gradient boosting), and RFE methods (algorithm model adopts decision
tree, extra tree, adaptive boosting, random forest, and extreme gradient boosting, and
linear support vector machine feature weight ranking). According to the order of feature
importance from high to low, we assigned a score of 11 to the most important feature in
each method and a score of 1 to the least important feature. If two features, set as feature
A and feature B, had the same importance ranking, then they were assigned the same
numerical value as the two features, e.g., both are 11, and then a value of 9 is assigned to
the feature C after A and B. Combining the results of each method, the importance score
of each feature is shown in Figure 2. The numbers above the bar chart in Figure 2 are the
sum of the assignment of the same feature by 14 feature importance ranking methods. The
larger the number, the higher the importance of the feature.
According to the feature importance scores from high to low, we started with the two
most important features, and added features to the combination in turn until it included all
features. Finally, the following 10 feature combinations were used as model inputs.
(1) PBA and Bedrock V s ;
(2) PBA, Bedrock V s , and FP;
(3) PBA, Bedrock V s , FP, and D800;
(4) PBA, Bedrock V s , FP, D800, and OBT;
(5) PBA, Bedrock V s , FP, D800, OBT, and V s20 ;
(6) PBA, Bedrock V s , FP, D800, OBT, V s20 , and V se ;
(7) PBA, Bedrock V s , FP, D800, OBT, V s20 , V se , and Depth;
(8) PBA, Bedrock V s , FP, D800, OBT, V s20 , V se , Depth, and V s30 ;
(9) PBA, Bedrock V s , FP, D800, OBT, V s20 , V se , Depth, V s30 , and Surface V s ;
(10) PBA, Bedrock V s , FP, D800, OBT, V s20 , V se , Depth, V s30 , Surface V s , and SFP.
A and feature B, had the same importance ranking, then they were assigned the same
numerical value as the two features, e.g., both are 11, and then a value of 9 is assigned to
the feature C after A and B. Combining the results of each method, the importance score
of each feature is shown in Figure 2. The numbers above the bar chart in Figure 2 are the
Appl. Sci. 2023, 13, 4530 sum of the assignment of the same feature by 14 feature importance ranking methods.8The of 22
larger the number, the higher the importance of the feature.
Importancescore
Figure2.2.Importance
Figure scoreofoffeatures.
features.
The coefficient of determination (R2 ), mean absolute error (MAE), and root mean
According to the feature importance scores from high to low, we started with the two
squared error (RMSE) were used as the evaluation indices of the prediction effect. The
most important features, and added features to the combination in turn until it included
expressions are given by Equations (10)–(12):
all features. Finally, the following 10 feature combinations were used as model inputs.
(1) PBA and Bedrock Vs; n
2
(2) PBA, Bedrock Vs, and FP; ∑ (ŷi − yi )
2 i =1
R = 1− n
(3) PBA, Bedrock Vs, FP, and D800; , (10)
2
(4) PBA, Bedrock Vs, FP, D800, and OBT;∑ (yi − yi )
i =1
(5) PBA, Bedrock Vs, FP, D800, OBT, and Vs20;
(6) PBA, Bedrock Vs, FP, D800, OBT, Vs201, and n Vse;
(7) PBA, Bedrock Vs, FP, D800, OBT, Vs20n, ∑
MAE = i − ŷDepth;
Vse|,yand i |, (11)
i =1
(8) PBA, Bedrock Vs, FP, D800, OBT, Vs s20, Vse, Depth, and Vs30;
(9) PBA, Bedrock Vs, FP, D800, OBT, Vs201, Vnse, Depth, V2 s30, and Surface Vs;
RMSE
(10) PBA, Bedrock Vs, FP, D800, = Vs20,∑
OBT, (yi − ŷi ) V
, s30, Surface Vs, and SFP. (12)
n i=V1se, Depth,
The coefficient of determination (R ), mean absolute error (MAE), and root mean
2
where n error
squared is the (RMSE) wereyiused
sample size, is theas
true
thevalue, ŷi is theindices
evaluation predicted value,
of the and y is effect.
prediction the average
The
of the true values.
expressions are given by Equation (10)–12):
We used SciKit-Learn [46] in machine learning to randomly divide the data into a
training dataset and testing dataset, of which 80% of the seismic records were used as the
training dataset for model training and feature selection, and the remaining 20% were used
as the testing dataset for model performance testing. In the 10-fold cross-validation, the
training dataset was randomly divided into 10 groups, in which 9 groups were used for
training, the remaining group was used for verification; this process was repeated until
all 10 groups had been used as a validation set. The average error of 10 verifications was
calculated, and the hyper-parameters that make the average performance of the model best
in 10 verifications were selected as the best hyper-parameters for the model. Figure 3 shows
the performance of the verification dataset of the model in the optimal hyper-parameter
when different feature combinations are used as model inputs. The number of horizontal
coordinates in Figure 3 is the number of input features, representing the 10 different feature
combinations in Section 3.2. It can be seen from Figure 3 that when the number of features
is 4 (the feature combination is PBA, Bedrock V s , FP, and D800), the R2 value of the model
is the largest, the MAE and RMSE values are the smallest, and the model performance is
the best. The data distribution of the 4 input features, PBA, Bedrock V s , FP, and D800, are
shown in Figure 4.
different feature combinations in Section 3.2. It can be seen from Figure 3 t
number of features is 4 (the feature combination is PBA, Bedrock Vs, FP, and
value of the model is the largest, the MAE and RMSE values are the sma
model performance is the best. The data distribution of the 4 input features, P
Appl. Sci. 2023, 13, 4530 9 of 22
Vs, FP, and D800, are shown in Figure 4.
Appl. Sci. 2023, 13, x FOR PEER REVIEW 10 of 23

Figure 3. The performance of the model in the optimal hyperparameter with different feature
Figure 3. The
combinations as performance
input. of the model in the optimal hyperparameter with differen
binations as input.
Figure4.4.The
Figure Thedata
datadistribution
distributionforfor
(a)(a)
PBA, (b) (b)
PBA, Bedrock V s , (c)
Bedrock Vs,FP,
(c)and
FP, (d)
andD800.
(d) D800.
The correlation matrix of the four input features is shown in Figure 5. This matrix
The correlation
is symmetrical, matrix
with the same of the four shown
correlation input features
above theis main
shown in Figure
diagonal being5. This matrix is
a mirror
symmetrical, with the same correlation shown above the main diagonal
image of the values below the main diagonal. The number in the correlation matrix and being a mirror
image
the ofofthe
color thevalues below
rectangle the main
represent diagonal.
the strength Thecorrelation.
of the number inThe thecorrelation
correlation matrix and
between
the four
the colorinput
of the rectangle
features represent the strength of the correlation. The correlation between
is weak.
the four input features is weak.
Figure 4. The data distribution for (a) PBA, (b) Bedrock Vs, (c) FP, and (d) D800.
The correlation matrix of the four input features is shown in Figure 5. This matrix is
symmetrical, with the same correlation shown above the main diagonal being a mirror
image of the values below the main diagonal. The number in the correlation matrix and
Appl. Sci. 2023, 13, 4530 10 of 22
the color of the rectangle represent the strength of the correlation. The correlation between
the four input features is weak.
Figure 5. Correlation matrix for input four features.

Figure 5. Correlation matrix for input four features.
3.3. Model Training
3.3. Model Training
According to the author’s previous research results [47], Qi used six different machine
learning models to predict PGA, including both simple models, such as linear regression
and decision tree, and complex aggregation algorithm models, such as random forest,
Adaptive Boosting (AdaBoost), and XGBoost. Compared with random forest, AdaBoost,
and other models, XGBoost has the best prediction performance. Ma et al. [48] introduced
the results of 8 different machine learning algorithms in the prediction of axial compressive
capacity of CFRP-confined, concrete-filled steel tubular short columns, including linear
regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), and typical
ensemble learning models such as RF, AdaBoost, gradient boosting decision tree (GBDT),
XGBoost, and LightGBM regressor (LGB). The XGBoost algorithm proved to provide the
best accuracy among these methods. Somala et al. [49] explored the application of LR, KNN,
SVM, RF, and XGBoost models in predicting the PGA and PGV (peak ground velocity)
during an earthquake. The best predictions were obtained using the RF and XGBoost
models for the prediction of PGA and PGV, respectively. Mangalathu [50] considered
8 different machine learning methods to obtain the best prediction model for the punching
shear strength of flat slabs, including LR, ridge regression, SVM, DT, KNN, RF, AdaBoost,
and XGBoost. XGBoost shows the best performance. For this reason, we used the XGBoost
model to predict PGA in this paper. In order to comprehensively evaluate the prediction
results of the machine learning models, we choose two other machine learning algorithms
based on a tree structure for PGA prediction, one is RF and the other is DT. The dataset and
input parameters used are the same as the XGBoost model.
In this study, two hyper-parameters are considered to improve the performance of
the DT model: max_depth and max_features. They have the same meaning as the two
parameters in the RF model. We also used 10-fold cross-validation and grid search to obtain
the optimal hyper-parameters of the DT model, and the results of the hyper-parameters
are shown in Figure 6. The optimal DT hyper-parameters’ values are: max_depth: 7;
max_features: 3.
is DT. The dataset and input parameters used are the same as the XGBoost model.
In this study, two hyper-parameters are considered to improve the performance of
the DT model: max_depth and max_features. They have the same meaning as the two
parameters in the RF model. We also used 10-fold cross-validation and grid search to ob-
tain the optimal hyper-parameters of the DT model, and the results of the hyper-parame-
Appl. Sci. 2023, 13, 4530 11 of 7;
ters are shown in Figure 6. The optimal DT hyper-parameters’ values are: max_depth: 22
max_features: 3.
Figure6.
Figure Determinationof
6.Determination ofhyper-parameters
hyper-parameters for
for DT
DT model.
model.
For the
For the RF
RF model,
model, three
three hyper-parameters
hyper-parametersareareconsidered:
considered:the then_estimators
n_estimatorsparameter,
parame-
which is used to specify the number of decision trees in the RF; the max_depth parameter,
ter, which is used to specify the number of decision trees in the RF; the max_depth param-
which
eter, indicates
which the maximum
indicates depthdepth
the maximum of theof
decision tree; and
the decision tree;the max_features
and parameter,
the max_features pa-
which is the maximum number of features considered when constructing the optimal
rameter, which is the maximum number of features considered when constructing the
model of the decision tree. We also used 10-fold cross-validation to obtain the optimal
optimal model of the decision tree. We also used 10-fold cross-validation to obtain the
hyper-parameters of the RF model. The optimization results of the hyper-parameters are
optimal hyper-parameters of the RF model. The optimization results of the hyper-param-
shown in Figure 7, and the optimal RF hyper-parameters’ values are: n_estimators: 67;
eters are shown in Figure 7, and the optimal RF hyper-parameters’ values are: n_estima-
max_depth: 7; max_features: 3.
tors: 67; max_depth: 7; max_features: 3.
Figure 7. Determination of hyper-parameters for RF model.

Figure 7. Determination of hyper-parameters for RF model.
The hyper-parameters of the XGBoost regression model are as follows: the subsample
parameter, used to specify the
The hyper-parameters proportion
of the XGBoostofregression
sampling from
modelthe aresample; the learning_rate
as follows: the subsam-
parameter, used to reduce the step length of each step; the n_estimators
ple parameter, used to specify the proportion of sampling from the sample; the learn- parameter, used
to specify the number of weak learners in the model; the max_depth parameter,
ing_rate parameter, used to reduce the step length of each step; the n_estimators parame- used to
specify the maximum tree depth of the weak learner; the reg_alpha parameter,
ter, used to specify the number of weak learners in the model; the max_depth parameter, used for
the objective
used function
to specify to use L1 tree
the maximum regularization
depth of theandweak
control the regularization
learner; the reg_alpha strength; and
parameter,
the reg_lambda parameter, used for the objective function to use L2 regularization
used for the objective function to use L1 regularization and control the regularization and
control regularization
strength; strength.
and the reg_lambda We used used
parameter, 10-foldforcross-validation to obtain
the objective function to theL2
use optimal
regu-
hyper-parameters of the XGBoost model, and the maximized average R 2 values of the
larization and control regularization strength. We used 10-fold cross-validation to obtain
validation dataset were taken as the optimal hyper-parameters. The optimization results of
the optimal hyper-parameters of the XGBoost model, and the maximized average R2 val-
the hyperparameters are shown in Figure 8. The optimal XGBoost hyper-parameters’ values
ues of the validation dataset were taken as the optimal hyper-parameters. The optimiza-
tion results of the hyperparameters are shown in Figure 8. The optimal XGBoost hyper-
parameters’ values are: subsample: 0.55; learning_rate: 0.08; n_estimators: 97; max_depth:
4; reg_alpha: 0.01; reg_lambda: 0.73.
ter, used to specify the number of weak learners in the model; the max_depth parameter,
used to specify the maximum tree depth of the weak learner; the reg_alpha parameter,
used for the objective function to use L1 regularization and control the regularization
strength; and the reg_lambda parameter, used for the objective function to use L2 regu-
larization and control regularization strength. We used 10-fold cross-validation to obtain
Appl. Sci. 2023, 13, 4530 the optimal hyper-parameters of the XGBoost model, and the maximized average R212val- of 22
ues of the validation dataset were taken as the optimal hyper-parameters. The optimiza-
tion results of the hyperparameters are shown in Figure 8. The optimal XGBoost hyper-
are:parameters’
subsample:values
0.55; are: subsample: 0.08;
learning_rate: 0.55; n_estimators:
learning_rate: 0.08; n_estimators:
97; max_depth: 4;97; max_depth:
reg_alpha: 0.01;
4; reg_alpha:
reg_lambda: 0.73.0.01; reg_lambda: 0.73.
Figure 8. Determination of hyper-parameters for the XGBoost model.
The predictive performance of the training dataset and testing dataset of the three
machine learning models are shown in Table 3. It can be seen from the table that the
XGBoost model has the best prediction performance. In both the training dataset and the
testing dataset, the prediction results of XGBoost are better than RF and DT. In addition the
prediction performance of the non-integrated algorithm (DT) is not as good as that of the
integrated algorithms (RF and XGBoost).
Table 3. Prediction results of three models.
Evaluation XGBoost RF DT
Indices Training Testing Training Testing Training Testing
R2 0.945 0.915 0.921 0.901 0.888 0.787
MAE 0.012 0.014 0.013 0.015 0.013 0.017
RMSE 0.017 0.022 0.020 0.024 0.024 0.034
3.4. Analysis of XGBoost Model Results

The optimized model in Section 3.2 with 4 features (PBA, Bedrock Vs , FP, and D800) was
used to train and test the samples. The residuals (measured value minus predicted value)
of the training set and the testing set were calculated, and the residual plots were plotted,
as shown in Figure 9. To better show the trend of the residuals, we divided the PGA into
15 groups and calculated the average residual of each group, as shown in the black solid line
in Figure 9. The dotted line is the average residual plus or minus one standard deviation.
The PGA groupings and the corresponding number of earthquake records in each group
are shown in Table 4. In order to ensure that the results are statistically significant, only the
average residual value of the number of each group with five or more ground motions were
calculated; that is, only the average value of the training dataset PGA ranging from 0.01 to
1.0 g and the testing dataset PGA ranging from 0.02 to 0.4 g is calculated.
It can be seen from Figure 9 that, except for individual conditions, the residuals of
most prediction results are within the range of ±0.1 g. The average residuals of the training
set and the testing set show a trend from negative to positive. When PGA is in the range of
0.01–0.05 g, the average residuals of the training set and the testing set are less than 0, and
Appl. Sci. 2023, 13, 4530 13 of 22
of 0.01–0.05 g, the average residuals of the training set and the testing set are less than 0,
and the predicted values are generally larger than the observation values. When the PGA
the
is in predicted
the range ofvalues areg,generally
0.05–1.0 the average larger thanare
residuals thepositive,
observation
and thevalues.
predictionWhen the PGA is in
results
theunderestimated.
are range of 0.05–1.0 The g, the average
standard residuals
deviation increasesare positive,
gradually and
with thethe prediction
increase in PGA,results are
underestimated.
and The standard
the residual discreteness deviation
increases increases
gradually. gradually
The main reason with theresult
for this increase in PGA, and
may be
that
the the number
residual of large earthquakes
discreteness increasesisgradually.
small, and the
Thetraining of the model
main reason for this is not suffi-
result may be that
cient.
the number of large earthquakes is small, and the training of the model is not sufficient.
Figure
Figure9.9.The
Theprediction residuals
prediction on the
residuals ontraining and testing
the training dataset for
and testing XGBoost.
dataset for XGBoost.
TableFigure 10 is the
4. Defined PGAstatistical
bins and histogram
associatedof number
the ratioof
of seismic
the prediction
records.and the observation
of the XGBoost model. The horizontal coordinates in Figure 10 are the ratio of the pre-
dicted value to the observed value. It can be seen that the number Number of positiveofand negative
Motions
Bin Number
30% errors in the training datasetPGA(g) Range
are about 40%, and the number of positive and negative
Training Set Testing Set
30% errors in the testing dataset are 45% and 35%, respectively. The training dataset is
1
more balanced than the testing dataset. 0.00–0.01 0 0
2 0.01–0.02 18 2
50 3 0.02–0.03 50 155 18
Training
4 0.03–0.04 Testing 449 70
40 5 0.04–0.05 40 542 122
6 0.05–0.06 370 102
30
7 0.06–0.07 30
245 69
Counts/%
Counts/%
8 0.07–0.08 164 46
20
9 0.08–0.09 20
129 29
10 0.09–0.10 76 25
11 0.10–0.20 283 77
10 10
12 0.20–0.40 78 10
13 0.40–0.60 13 1
0 0
0.5 14 1.0 1.5 0.60–1.002.5
2.0 0.5 1.0 7 1.5 2.0 2.5 2
15 Predictions/Observations >1.00 Predictions/Observations

1 1
Figure 10. Histogram of the ratio of XGBoost for predicted and observed PGA.
Figure 10 is the statistical histogram of the ratio of the prediction and the observation
WeXGBoost
of the collectedmodel.
the results
The of three numerical
horizontal simulation
coordinates methods,
in Figure 10 areDEEPSOIL
the ratio of (DP),
the predicted
SHAKE2000
value to the observed value. It can be seen that the number of positive andpre-
(SHAKE), and SOILRESPONSE (SR) [51], and compared them with the negative 30%
diction
errors results
in the of machine
training learning
dataset models.
are aboutThe
40%,numerical
and thesimulation
number ofresults in this
positive andpaper
negative 30%
are the author’s previous research results [51]. Because there are no nonlinear parameters
errors in the testing dataset are 45% and 35%, respectively. The training dataset is more
in the site data of the station, the generic nonlinear shear modulus reduction and damping
balanced than the testing dataset.
curves were sourced from Darendeli [52]. DP and SHAKE methods are suitable for the
WeIIcollected
site class the III,
and site class results of calculation
and the three numerical simulation
results are methods,
not good on DEEPSOIL
the soft soil site. (DP),
SHAKE2000
The calculation(SHAKE), and
results of the SRSOILRESPONSE (SR) DP
method are better than [51], and
and compared
SHAKE [51]. Itthem with the pre-
is worth
dictionthat
noting results of machine
previous researchlearning models.
results show thatThe
the numerical simulation
linear equivalent method results
entersina this paper
are the author’s previous research results [51]. Because there are no nonlinear parameters
in the site data of the station, the generic nonlinear shear modulus reduction and damping
curves were sourced from Darendeli [52]. DP and SHAKE methods are suitable for the site
class II and site class III, and the calculation results are not good on the soft soil site. The
calculation results of the SR method are better than DP and SHAKE [51]. It is worth noting
that previous research results show that the linear equivalent method enters a strong non-
linear state on the large ground motion input, so the traditional linear equivalent method
cannot simulate the site response well in this condition [15,16].
Figure 10 is the statistical histogram of the ratio of the prediction and the observation
of the XGBoost model. The horizontal coordinates in Figure 10 are the ratio of the pre-
dicted value to the observed value. It can be seen that the number of positive and negative
30% errors in the training dataset are about 40%, and the number of positive and negative
Appl. Sci. 2023, 13, 4530 14 of 22
30% errors in the testing dataset are 45% and 35%, respectively. The training dataset is
more balanced than the testing dataset.
50 50
Training Testing
40 40
30 30
Counts/%
Counts/%
20 20
10 10

0 0
0.5 1.0 1.5 2.0 2.5 0.5 1.0 1.5 2.0 2.5
Predictions/Observations Predictions/Observations
Figure
Figure 10. 10.
strong Histogram
Histogram of the
nonlinear ofratio
statetheon
ratio
of theof XGBoost
XGBoost for motion
forground
large predicted
predicted and observed
and observed
input, PGA. PGA.
so the traditional linear equiva-
lent method cannot simulate the site response well in this condition [15,16].
WeThe comparison
collected the results results of different
of three numericalsite classifications
simulation methods, and different(DP),
DEEPSOIL PGA intervals
The comparison results of different site classifications and different PGA intervals are
are shown
SHAKE2000 in Figure
(SHAKE), and11. It can be seen
SOILRESPONSE from
(SR) [51],Figure 11 that,them
and compared in general, the prediction
with the pre-
shown in Figure 11. It can be seen from Figure 11 that, in general, the prediction results of
diction
resultsresults
of of machine
the machine learning
learning models.
modelsThe are
numerical
better simulation
than that results
of the in this paper simulation
numerical
the machine learning models are better than that of the numerical simulation methods.
aremethods.
the author’s previous
Even if theresearch results [51].algorithm
non-integrated Because there(DT)are no nonlinear parameters results are
Even if the non-integrated algorithm (DT) is used, theisprediction
used, the results
predictionare still better
in still
the site data
better of
thanthe station,
the numerical the generic nonlinear
simulation shear
methods, modulus reduction and
except forconditions
individual damping
conditions (site
curves than
were the numerical
sourced from simulation
Darendeli methods,
[52].For except
DPsite
andclass for
SHAKE individual
methods are suitable (site class IV and
forofthe
class IV and
0.1 gII<and
PGA 0.1 g
< 0.2 < PGA
g).III,
For < 0.2
sitethe g).
class II, the prediction II, the prediction results the machine
site class site class and calculation results areresults
not goodof the machine
on the soft soillearning
site. models
learning
are quitemodels
differentare quitethat different from that of the numerical simulation methods, re-
and
The calculation results offrom the SR method of theare
numerical
better thansimulation methods,
DP and SHAKE and
[51]. the
It is prediction
worth
thesults
prediction results of the machine learning models are obviously better than those
sim-of
noting thatofprevious
the machine learning
research resultsmodels
show arethatobviously
the linearbetter than those
equivalent methodof the numerical
enters a
theulation
numerical simulation
methods. For sitemethods. For site
class III and siteclass
classIV,
III the
andMAE
site class IV, thevalues
and RMSE MAE of and RMSE
several
values of several methods are similar, and the machine learning models are
methods are similar, and the machine learning models are slightly better than the numer- slightly better
than the numerical simulation. For different PGA intervals, the prediction
ical simulation. For different PGA intervals, the prediction results of the machine learning results of the
machine
modelslearning models
are better are better
than those of thethan those of
numerical the numerical
simulation simulation
methods (except methods
for 0.1 g <(except
PGA
for<0.1
0.2gg;< the
PGA < 0.2
MAE g; the
and MAPE MAE and of
values MAPEDT arevalues of DT
slightly are than
larger slightly larger
those of thethan
SR those
and
of the SR and
SHAKE SHAKEThe
methods). methods). Theerror
calculation calculation error simulation
of numerical of numerical simulation
methods methods
increases sig-
increases significantly,
nificantly, while the prediction
while the prediction results of
results of machine machine
learning learning
models models
are more are and
stable, more
the prediction
stable, error increases
and the prediction error slightly
increaseswith the increase
slightly with the in increase
PGA. in PGA.
0.08 0.32 0.08 0.32 0.08 0.32

Site class Ⅱ MAE Site class Ⅲ MAE Site class Ⅳ MAE
0.07 RMSE 0.30 0.07 RMSE 0.30 0.07 RMSE 0.30
MAPE MAPE MAPE
0.06 0.28 0.06 0.28 0.06 0.28
MAE and RMSE
MAE and RMSE
MAE and RMSE
0.05 0.26 0.05 0.26 0.05 0.26

MAPE
MAPE
0.04 0.24 0.04 0.24 0.04 0.24 MAPE
0.03 0.22 0.03 0.22 0.03 0.22
0.02 0.20 0.02 0.20 0.02 0.20
0.01 0.18 0.01 0.18 0.01 0.18

DP SHAKE SR DT RF XGBoost DP SHAKE SR DT RF XGBoost DP SHAKE SR DT RF XGBoost
Prediction Methods Prediction Methods Prediction Methods
0.05 0.40 0.05 0.40 0.40

PGA<0.1 g MAE 0.1 g<PGA<0.2 g MAE PGA>0.2 g MAE
RMSE 0.35 RMSE 0.35 0.35 RMSE
MAPE MAPE MAPE
MAE, RMSE and MAPE
0.04 0.04
0.30 0.30 0.30
MAE and RMSE
MAE and RMSE
0.25 0.25 0.25

MAPE
MAPE
0.03 0.03
0.20 0.20 0.20
0.15 0.15 0.15

0.02 0.02
0.10 0.10 0.10
0.01 0.05 0.01 0.05 0.05

DP SHAKE SR DT RF XGBoost DP SHAKE SR DT RF XGBoost DP SHAKE SR DT RF XGBoost
Prediction Methods Prediction Methods Prediction Methods
Figure 11. Comparison of prediction effects of numerical simulation methods and machine learn-
Figure 11. Comparison of prediction effects of numerical simulation methods and machine learning
ing models on different site classifications and different PGA intervals.
models on different site classifications and different PGA intervals.
4. Explanation of the Prediction Model Based on SHAP
4.1. Principle of the SHAP Method
SHAP is a method based on coalition game theory [53]. It measures the influence of
the feature on the results by calculating its SHAP values. The SHAP value may be positive
or negative, where the positive value improves the prediction results and the negative
Appl. Sci. 2023, 13, 4530 15 of 22
4. Explanation of the Prediction Model Based on SHAP

4.1. Principle of the SHAP Method
SHAP is a method based on coalition game theory [53]. It measures the influence of
the feature on the results by calculating its SHAP values. The SHAP value may be positive
or negative, where the positive value improves the prediction results and the negative
value reduces the prediction results. The greater the role of features in the model, the
greater the absolute value of the SHAP value and the higher the importance of the features.
Suppose the ith sample is xi , the jth feature of the ith sample is xij , and the predicted
value of the model for the sample is yi . The baseline of the model (the mean of target values
for all samples) is ybase and f (xij ) is the SHAP value of the jth feature for the ith sample. The
SHAP value follows the formula below:

yi = ybase + f ( xi1 ) + f ( xi2 ) + . . . + f xij , (13)
and f (xij ) is as follows:
|S|!(| F | − |S| − 1)!

∑

f xij = [v(S ∪ { j}) − v(S)], (14)
S⊆ F \{ j}
| F |!
where F is the full set of all features, S is a subset of features, v(S) is the contribution of
the features included in subset S, and v(S∪{j}) − v(S) is the contribution of the feature j for
the interaction.
The SHAP value f (xj ) of feature j is obtained by accumulating the mean value of
feature j of all samples. The formula is as follows:
M
∑f

f xj = xij . (15)
i =1
4.2. SHAP Global Analysis

To verify the reliability and stability of the SHAP method, we used three machine
learning models for SHAP analysis. Figure 12 shows summary plots of all features for
three machine learning prediction models. The features are sorted according to the global
importance, and the importance gradually decreases from top to bottom. The color of the
points from blue to red represents the value of the features from small to large. Each point
represents the SHAP value of a sample, which represents the contribution of the feature
to a single prediction, and the set of points represents the direction and size of the overall
influence of the feature on the prediction results. Figure 12a shows the summary plot
for the XGBoost model. Among the four features, the PBA is the most important feature
that affects the prediction results of the model, and a larger PBA generally contributes
positively to the prediction results, which is consistent with the existing understanding
that the greater the intensity of seismic input, the larger the PGA. D800 is the second most
important feature, and a smaller D800 contributes positively to the prediction results, while
a larger D800 has a negative SHAP value. The SHAP values of Bedrock V s and FP are
mainly concentrated around 0, and the influence of the two features on the SHAP value
shows a reverse trend. The smaller Bedrock V s is, the smaller the SHAP value, while the
smaller the FP is, the larger the SHAP value. Figure 12b,c are summary plots of the random
forest model and decision tree model, respectively. The feature importance ranking of
the three prediction models is the same, and the change trend of the SHAP value is the
same. The global importance of all input features for the three machine learning models is
shown in Figure 13. It can be seen from Figure 13 that for the three models in this paper,
the importance of the four input features is consistent. PBA is the most important factor,
followed by D800 and Bedrock V s , and FP is the least important.
value shows a reverse trend. The smaller Bedrock Vs is, the smaller the SHAP value, while
the smaller the FP is, the larger the SHAP value. Figure 12b,c are summary plots of the
random forest model and decision tree model, respectively. The feature importance rank-
ing of the three prediction models is the same, and the change trend of the SHAP value is
the same. The global importance of all input features for the three machine learning mod-
els is shown in Figure 13. It can be seen from Figure 13 that for the three models in this
Appl. Sci. 2023, 13, 4530 16 of 22
paper, the importance of the four input features is consistent. PBA is the most important
factor, followed by D800 and Bedrock Vs, and FP is the least important.

Figure
Figure 12. The
12.The
Figure 12. The SHAPSHAP
SHAP summary
summary
summary plot plot
plot of
of all of allfeatures
all input
input input features
features for (a)
for for (a) (b)
(a) XGBoost,
XGBoost, XGBoost,
(b) Random(b)
Random Random
Forest,
Forest, (c)
(c) Forest,
Decision
(c) Tree.
Decision
Decision Tree.Tree.
Figure
Figure Global
13.Global
Figure13.
13. Global importance
importance
importance of of
allall
of all input
input
input features
features
features for XGBoost,
for (a)
for (a) (a) XGBoost,
XGBoost, (b) (b) Random
(b) Random
Random Forest,
Forest,
Forest, (c) (c) Decision Tree.
(c) Decision
Decision
Tree.
Tree.
In order to show the distribution of SHAP values of different models more clearly,
Figure 14 presents the distributions and densities of the features of XGBoost, Random
Forest, and Decision Tree for the entire dataset. It can be seen from Figure 14 that the
Appl. Sci. 2023, 13, 4530 shape and distribution of the four features in the three machine learning models are17 sim-
of 22
ilar, and the means of the SHAP values in different models are basically the same. The
SHAP values of the four features are mainly concentrated in a certain range and each fea-
ture In
hasorder
positive and negative
to show contributions.
the distribution Figure
of SHAP 14a shows
values the density
of different of SHAP
models morevalues
clearly,
Figure 14 presents the distributions and densities of the features of XGBoost,D800
for XGBoost. It shows that the distribution of PBA and FP is concentrated, while and
Random
Bedrock V s have two obvious peaks.
Forest, and Decision Tree for the entire dataset. It can be seen from Figure 14 that the shape
The influence
and distribution of of
theeach
fourfeature
featuresoninPGA
the prediction
three machineresults will not
learning change
models with
are the dif-
similar, and
the means of the SHAP values in different models are basically the same. The SHAPof
ferent machine learning model. Through the global analysis of the SHAP values the
values
three
of machine
the four learning
features models,
are mainly it can be seen
concentrated in athat the same
certain range feature
and each has the same
feature contri-
has positive
bution
and to the contributions.
negative prediction results on different
Figure 14a showsmodels, and the
the density of distribution
SHAP values and
fordensities
XGBoost.ofIt
the SHAP values are basically the same. This is consistent with the objective
shows that the distribution of PBA and FP is concentrated, while D800 and Bedrock V s knowledge
and also
have two confirms the reliability of SHAP in model explanation.
obvious peaks.
Figure 14. Distributions and densities of the SHAP values for (a) XGBoost, (b) Random Forest, (c)
Figure 14. Distributions and densities of the SHAP values for (a) XGBoost, (b) Random Forest,
Decision Tree.
(c) Decision Tree.
In order to deeply understand the influence of each feature on the prediction results,
The influence of each feature on PGA prediction results will not change with the
the XGBoost model is taken as an example to show the dependency plots in Figure 15.
different machine learning model. Through the global analysis of the SHAP values of
Figure 15a shows the dependency plot of the PBA. It can be seen that the PBA is
the three machine learning models, it can be seen that the same feature has the same
positively correlated with its SHAP value. The SHAP value is negative when the PBA is
contribution to the prediction results on different models, and the distribution and densities
less than 0.015 g; that is, it plays a negative role in the prediction of PGA. When the PBA
of the SHAP values are basically the same. This is consistent with the objective knowledge
is greater than 0.015 g, the SHAP value is positive and has a positive effect on the PGA.
and also confirms the reliability of SHAP in model explanation.
When the PBA is less than 0.015 g, the vertical color stratification is not obvious. When the
In order to deeply understand the influence of each feature on the prediction results,
PBA is greater than 0.015 g, the red dots are mainly distributed below the blue ones, indi-
the XGBoost model is taken as an example to show the dependency plots in Figure 15.
cating that on the same input PBA, the larger D800 is, the smaller the SHAP value, the
Figure 15a shows the dependency plot of the PBA. It can be seen that the PBA is
smaller the predicted PGA value.
positively correlated with its SHAP value. The SHAP value is negative when the PBA is
Figure 15b is the dependency plot of D800. It can be seen that when D800 is less than
less than 0.015 g; that is, it plays a negative role in the prediction of PGA. When the PBA is
50 m, the SHAP value is positive and has a positive effect on the PGA, and the smaller the
greater than 0.015 g, the SHAP value is positive and has a positive effect on the PGA. When
PBA, the smaller the SHAP value. When D800 is greater than 50 m, the SHAP value is
the PBA is less than 0.015 g, the vertical color stratification is not obvious. When the PBA is
greater than 0.015 g, the red dots are mainly distributed below the blue ones, indicating
that on the same input PBA, the larger D800 is, the smaller the SHAP value, the smaller the
predicted PGA value.
Figure 15b is the dependency plot of D800. It can be seen that when D800 is less than
50 m, the SHAP value is positive and has a positive effect on the PGA, and the smaller the
PBA, the smaller the SHAP value. When D800 is greater than 50 m, the SHAP value is mainly
negative and has a negative effect on the PGA. That is, the PGA on the deep site is smaller.
indicates that a small FP has little influence on the prediction result of the PGA. When the
FP is greater than 7 Hz, the SHAP value is negative, and the SHAP value shows a down-
ward trend with the increase in FP, indicating that the larger the FP, the smaller the pre-
dicted PGA value. This is consistent with the existing experience that the gap between the
Appl. Sci. 2023, 13, 4530 FP and the predominant frequency of the site increases with the increase in the FP, 18 of 22
which
avoids resonance and reduces the peak ground acceleration.
15. SHAP
Figure 15.
Figure SHAP dependency
dependencyplots
plotsfor
forXGBoost.
XGBoost.
It can be seen
4.3. Visualization andfrom Figureof15c
Analysis SHAP thatValues
when of a SingleVSample
Bedrock s is less than 600 m/s, the SHAP
value is negative and has a negative effect on the PGA, and when Bedrock V s is greater
than In600addition
m/s, thetoSHAP
providing
value global indication,
is positive and hasSHAPa positivecan effect
also visualize
on the PGA.the Except
featurefor
con-
tributions of a single sample as a force plot. Figure 16 is the force
a few sample points, the SHAP value is mainly concentrated around 0, which has little plot of the typical sample
of XGBoost.
influence Theprediction
on the base valueresults.
is the mean of the model prediction results when there is no
inputFigure
feature,
15d and it is 0.07268
shows that there g for XGBoost.
is no obviousThe red arrow
interaction and blue
between thearrow
FP andindicate
the PBA.that
the feature has a positive or negative role in the PGA prediction
When the FP is less than 7 Hz, the SHAP value is mainly concentrated around 0, which results, respectively. The
longer
indicatesthethat
arrow, the greater
a small the contribution
FP has little influence onofthe theprediction
feature. Inresult
Figureof 16,
thethe actual
PGA. WhenPGA
of
thethis
FP issample
greateris than
0.4127g, and
Hz, thethe predicted
SHAP value value is 0.40and
is negative, g. Both the PBA
the SHAP andshows
value FP make
a
downward
positive trend with the
contributions increase
to the in FP,results,
predicted indicating that
while the larger
Bedrock Vs the
andFP, the play
D800 smaller the
negative
predicted
roles PGA
in the value. This
predicted is consistent
results. In terms with the existingthe
of importance, experience
PBA hasthatthe the gap between
longest arrow and
the FP and the predominant frequency of the site increases with
makes the greatest contribution, followed by Bedrock Vs, FP, and D800. The importance the increase in the FP,
which avoids resonance and reduces the peak ground acceleration.
ranking of a single sample is different from that in the summary plot shown in Figure 13
because the summary plot is based on the global overall indication.
4.3. Visualization and Analysis of SHAP Values of a Single Sample
In addition to providing global indication, SHAP can also visualize the feature contri-
butions of a single sample as a force plot. Figure 16 is the force plot of the typical sample
of XGBoost. The base value is the mean of the model prediction results when there is no
input feature, and it is 0.07268 g for XGBoost. The red arrow and blue arrow indicate that
the feature has a positive or negative role in the PGA prediction results, respectively. The
longer the arrow, the greater the contribution of the feature. In Figure 16, the actual PGA of
this sample is 0.412 g, and the predicted value is 0.40 g. Both the PBA and FP make positive
contributions to the predicted results, while Bedrock V s and D800 play negative roles in
the predicted results. In terms of importance, the PBA has the longest arrow and makes the
greatest contribution, followed by Bedrock V s , FP, and D800. The importance ranking of a
single sample is different from that in the summary plot shown in Figure 13 because the
summary plot is based on the global overall indication.
Appl.
Appl.Sci. 2023,13,
Sci.2023, 13,4530
x FOR PEER REVIEW 19 of
20 of22
23
Figure 16.Force
Figure16. Forceplot
plotofofthe
thetypical
typicalsample
samplefor
forXGBoost.
XGBoost.
5. Conclusions
5. Conclusions
Based on 3104 sets of ground motion records from the KiK-net, XGBoost was used to
train theBased PGA onprediction
3104 sets of ground
model. motionfeature
Through recordsselection,
from the four KiK-net, XGBoost
parameters was used to
representing
train the PGA prediction model. Through feature selection,
ground motion characteristics (PBA, FP) and site conditions (D800, Bedrock V s ) were four parameters representing
ground as
selected motion
inputcharacteristics
features of the (PBA, FP) and
machine site conditions
learning model. For (D800, Bedrock V
comparative s) were se-
analysis, a
lected as input features of the machine learning model.
random forest model and a decision tree model were trained to predict the PGA. In order For comparative analysis, a ran-
todom forest
explain themodel
machine and learning
a decision tree model
models were trained
and analyze whether to the
predict the PGA.
prediction In order
results of theto
machine learning models conform to the objective reality and physical laws, the SHAPofwas
explain the machine learning models and analyze whether the prediction results the
machine learning models conform to the objective reality
introduced to explain the model and analyze the characteristics. The main conclusions are and physical laws, the SHAP
aswas introduced to explain the model and analyze the characteristics. The main conclu-
follows:
sions (1)areTheas Rfollows:
2 of XGBoost in the training set and testing set reach up to 0.945 and 0.915,
(1) The R
respectively. In2 of XGBoost
addition, thein relative
the trainingerrorset of and
about testing
80% set reach is
samples upwithin
to 0.945 and 0.915,
±30%. The
respectively.
model has good In prediction
addition, the and relative error ofability.
generalization about 80% samples is within ±30%. The
model (2) has good prediction
On different and generalization
site classification and different ability.
PGA intervals, the prediction effect of
(2) On different site classification
XGBoost model is better than that of the other two machine and different PGA intervals,
learning themodels.
prediction Excepteffect forof
XGBoost model is better than that of the other two machine
individual working conditions, the prediction results of machine learning models, even if it learning models. Except for
isindividual workingalgorithm,
a non-integrated conditions,are thebetter
prediction
than thatresults of machine
of the numerical learning
simulationmodels, even if
methods.
it is a non-integrated
Especially on strong ground algorithm,
motion are(PGA
better>than that
0.2 g), theofnonlinearity
the numerical simulation
of soil is enhanced,methods.
and
Especially
the calculation on errors
strongofground
numerical motion (PGA >methods
simulation 0.2 g), the nonlinearity
increase of soilwhich
significantly, is enhanced,
reflects
andlimitation
the the calculation errors equivalent
of the linear of numerical simulation
method. methodsinincrease
The increase significantly,
the prediction error of which
the
reflects the limitation of the linear equivalent method. The increase in the prediction error
machine learning models is obviously smaller than that of numerical simulation methods.
of the(3) machine
The influence learningof input
models features shown by
is obviously SHAPthan
smaller value onof
that PGA prediction
numerical results
simulation
ismethods.
consistent with the physical law. The PBA is the most important feature that affects the
prediction
(3) The results of theofPGA.
influence inputThe SHAPshown
features valuesby ofSHAP
the PBA increase
value on PGA with the increase
prediction in
results
the PBA; thatwith
is consistent is, thetheprediction
physical law. value Theof PBA
the PGAis theincreases with thefeature
most important increase inaffects
that the PBA. the
D800 is a secondary
prediction results ofimportant
the PGA. feature,
The SHAP when D800ofisthe
values lessPBAthanincrease
50 m, the SHAP
with the values
increase arein
positive, which have a positive impact on
the PBA; that is, the prediction value of the PGA increases with the prediction results of the PGA; when D800
increase in the PBA. is
greater than 50 m, the SHAP values are mainly negative,
D800 is a secondary important feature, when D800 is less than 50 m, the SHAP values are which have a negative impact
on the prediction
positive, which have results of the PGA.
a positive impact TheonSHAP values ofresults
the prediction Bedrock V s PGA;
of the and FP are mainly
when D800 is
concentrated
greater than 50 near m,0,thewhich
SHAP have littleare
values effect
mainlyon the prediction
negative, which results
haveof the PGA.impact on
a negative
The machine
the prediction learning
results of themodel
PGA.established
The SHAP values in this ofpaper has achieved
Bedrock Vs and FPgood prediction
are mainly con-
performance, and SHAP analysis proves the rationality
centrated near 0, which have little effect on the prediction results of the PGA. of the prediction results, which
provides The amachine
new method learningformodel
PGA prediction.
established Although
in this paper thehas data come from
achieved gooda prediction
Japanese
site, it does not mean that the model in this paper is
performance, and SHAP analysis proves the rationality of the prediction results, only applicable to Japan, as the site
which
amplification effect has certain universality. However, due
provides a new method for PGA prediction. Although the data come from a Japanese site, to the limited amount of data, it
needs
it does tonot
be further
mean that verified.
the modelResearch on paper
in this this topic can be
is only supplemented
applicable to Japan, andasimproved
the site am- in
the following aspects:
plification effect has certain universality. However, due to the limited amount of data, it
needs (1)toUsing global
be further seismicResearch
verified. datasets,onnot thisjust
topicJapanese seismic datasets,
can be supplemented andtoimproved
verify the in
model in this paper
the following aspects: or establish a new model for PGA prediction;
(2) The ground motion data used in this article are all from the station records of the
(1) Using global seismic datasets, not just Japanese seismic datasets, to verify the
horizontal sites, without considering the influence of complex terrain factors, and further
model in this paper or establish a new model for PGA prediction;
research can predict the PGA under different terrain;
(2) The ground motion data used in this article are all from the station records of the
(3) Deep learning algorithms such as neural networks can be used to predict PGA in
horizontal sites, without considering the influence of complex terrain factors, and further
the further research, and the prediction results can be compared with the XGBoost method
research can predict the PGA under different terrain;
to select the best performing model.
(3) Deep learning algorithms such as neural networks can be used to predict PGA in
the further research, and the prediction results can be compared with the XGBoost method
to select the best performing model.
Appl. Sci. 2023, 13, 4530 20 of 22
Author Contributions: Conceptualization, R.S. and T.Z.; methodology, W.Q. and J.Q.; writing—
original draft preparation, W.Q. and R.S.; writing—review and editing, R.S., W.Q., and J.Q.; funding
acquisition, R.S. and T.Z. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by Scientific Research Fund of Institute of Engineering Mechanics,
China Earthquake Administration (Grant No. 2020C04), the Heilongjiang Provincial Natural Science
Foundation Joint Guidance Project of China (Grant No. LH2020E019), and the National Natural
Science Foundation of China (Grant No. 51808515).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Publicly available datasets were analyzed in this study. The seismic
records are from KiK-net (https://www.kyoshin.bosai.go.jp/, accessed on 1 January 2021).
Acknowledgments: The authors acknowledge the National Research Institute for Earth Science and
Disaster Resilience of Japan (KiK-net) for providing the strong ground motion data.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Gutenberg, B.; Richter, C.F. Earthquake magnitude, intensity, energy, and acceleration (Second paper). Bull. Seismol. Soc. Am.
1942, 32, 105–145. [CrossRef]
2. Hershberger, J. A comparison of earthquake accelerations with intensity ratings. Bull. Seismol. Soc. Am. 1956, 46, 317–320.
[CrossRef]
3. Bose, M.; Heaton, T.; Hauksson, E. Rapid Estimation of Earthquake Source and Ground-Motion Parameters for Earthquake
Early Warning Using Data from a Single Three-Component Broadband or Strong-Motion Sensor. Bull. Seismol. Soc. Am. 2012,
102, 738–750. [CrossRef]
4. Cremen, G.; Galasso, C. Earthquake early warning: Recent advances and perspectives. Earth-Sci. Rev. 2020, 205, 103184.
[CrossRef]
5. Güllü, H. Prediction of peak ground acceleration by genetic expression programming and regression: A comparison using
likelihood-based measure. Eng. Geol. 2012, 141–142, 92–113. [CrossRef]
6. Falcone, G.; Vacca, V.; Mori, F.; Naso, G.; Spina, D. Evaluation of building seismic retrofitting costs foundedon experimental data.
The case study of “San Benedetto” School (Norcia, Italy). Ital. J. Geosci. 2021, 140, 365–381. [CrossRef]
7. Vacca, V.; Occhipinti, G.; Mori, F.; Spina, D. The Use of SMAV Model for Computing Fragility Curves. Buildings 2022, 12, 1213.
[CrossRef]
8. Fayjaloun, R.; Negulescu, C.; Roullé, A.; Auclair, S.; Gehl, P.; Faravelli, M. Sensitivity of Earthquake Damage Estimation to the
Input Data (Soil Characterization Maps and Building Exposure): Case Study in the Luchon Valley, France. Geosciences 2021,
11, 249. [CrossRef]
9. Riga, E.; Apostolaki, S.; Karatzetzou, A.; Danciu, L.; Pitilakis, K. The role of modelling of site conditions and amplification in
seismic hazard and risk assessment at urban scale. The case of Thessaloniki, Greece. Ital. J. Geosci. 2022, 141, 198–215. [CrossRef]
10. Sabetta, F.; Fiorentino, G.; Bocchi, F.; Sinibaldi, M.; Falcone, G.; Mendicelli, A. Influence of local site effects on seismic risk maps
and ranking of Italian municipalities. Bull. Earthq. Eng. 2023, 21, 2441–2468. [CrossRef]
11. Du, K.; Ding, B.; Bai, W.; Sun, J.; Bai, J. Quantifying Uncertainties in Ground Motion-Macroseismic Intensity Conversion Equations.
A Probabilistic Relationship for Western China. J. Earthq. Eng. 2022, 26, 1976–2000. [CrossRef]
12. Dhanya, J.; Raghukanth, S.T.G. Ground Motion Prediction Model Using Artificial Neural Network. Pure Appl. Geophys. 2018,
175, 1035–1064. [CrossRef]
13. Boore, D.M.; Atkinson, G.M. Ground-Motion Prediction Equations for the Average Horizontal Component of PGA, PGV, and
5%-Damped PSA at Spectral Periods between 0.01 s and 10.0 s. Earthq. Spectra 2008, 24, 99–138. [CrossRef]
14. Du, K.; Ding, B.; Luo, H.; Sun, J. Relationship between Peak Ground Acceleration, Peak Ground Velocity, and Macroseismic
Intensity in Western China. Bull. Seismol. Soc. Am. 2019, 109, 284–297. [CrossRef]
15. Régnier, J.; Bonilla, L.F.; Bard, P.; Bertrand, E.; Hollender, F.; Kawase, H.; Sicilia, D.; Arduino, P.; Amorosi, A.; Asimaki, D.; et al.
International Benchmark on Numerical Simulations for 1D, Nonlinear Site Response (PRENOLIN): Verification Phase Based on
Canonical Cases. Bull. Seismol. Soc. Am. 2016, 106, 2112–2135. [CrossRef]
16. Régnier, J.; Bonilla, L.; Bard, P.; Bertrand, E.; Hollender, F.; Kawase, H.; Sicilia, D.; Arduino, P.; Amorosi, A.; Asimaki, D.; et al.
PRENOLIN: International Benchmark on 1D Nonlinear Site-Response Analysis—Validation Phase Exercise. Bull. Seismol. Soc.
Am. 2018, 108, 876–900. [CrossRef]
17. Moczo, P.; Kristek, J.; Bard, P.-Y.; Stripajová, S.; Hollender, F.; Chovanová, Z.; Kristeková, M.; Sicilia, D. Key structural parameters
affecting earthquake ground motion in 2D and 3D sedimentary structures. Bull. Earthq. Eng. 2018, 16, 2421–2450. [CrossRef]
18. Gatmiri, B.; Arson, C. Seismic site effects by an optimized 2D BE/FE method II. Quantification of site effects in two-dimensional
sedimentary valleys. Soil Dyn. Earthq. Eng. 2008, 28, 646–661. [CrossRef]
Appl. Sci. 2023, 13, 4530 21 of 22
19. Maufroy, E.; Chaljub, E.; Hollender, F.; Kristek, J.; Moczo, P.; Klin, P.; Priolo, E.; Iwaki, A.; Iwata, T.; Etienne, V.; et al. Earthquake
Ground Motion in the Mygdonian Basin, Greece: The E2VP Verification and Validation of 3D Numerical Simulation up to 4 Hz.
Bull. Seismol. Soc. Am. 2015, 105, 1398–1418. [CrossRef]
20. Schnabel, P.B.; Lysmer, J.; Seed, H.B. SHAKE: A Computer Program For Earthquake Response Analysis of Horizontally Layered Sites;
Earthquake Engineering Research Center, Report No. UBC/EERC-72-12; University of California: Berkeley, CA, USA, 1972.
21. Hashash, Y.M.; Park, D. Non-linear one-dimensional seismic ground motion propagation in the Mississippi embayment. Eng.
Geol. 2001, 62, 185–206. [CrossRef]
22. Kerh, T.; Ting, S.B. Neural network estimation of ground peak acceleration at stations along Taiwan high-speed rail system. Eng.
Appl. Artif. Intell. 2005, 18, 857–866. [CrossRef]
23. Arjun, C.R.; Kumar, A. Artificial Neural Network-Based Estimation of Peak Ground Acceleration. ISET J. Earthq. Technol. 2009,
501, 19–28.
24. Günaydın, K.; Günaydın, A. Peak Ground Acceleration Prediction by Artificial Neural Networks for Northwestern Turkey. Math.
Probl. Eng. 2008, 2008, 919420. [CrossRef]
25. Derras, B.; Bard, P.-Y.; Cotton, F.; Bekkouche, A. Adapting the Neural Network Approach to PGA Prediction: An Example Based
on the KiK-net Data. Bull. Seismol. Soc. Am. 2012, 102, 1446–1461. [CrossRef]
26. Zhu, C.; Cotton, F.; Kawase, H.; Nakano, K. How well can we predict earthquake site response so far? Machine learning vs
physics-based modeling. Earthq. Spectra 2022, 39, 478–504. [CrossRef]
27. Mori, F.; Mendicelli, A.; Falcone, G.; Acunzo, G.; Spacagna, R.L.; Naso, G.; Moscatelli, M. Ground motion prediction maps using
seismic-microzonation data and machine learning. Nat. Hazards Earth Syst. Sci. Discuss. 2021, 22, 947–966. [CrossRef]
28. Ribeiro, M.; Singh, S.; Guestrin, C. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In Proceedings of the
2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, San Diego,
CA, USA, 12–17 June 2016; pp. 97–101. [CrossRef]
29. Lundberg, S.M.; Lee, S.I. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference
on Neural Information Processing Systems, Long Beach, CA, USA, 4–9 December 2017; pp. 4766–4775.
30. Lundberg, S.M.; Lee, S.-I. Consistent feature attribution for tree ensembles. In Proceedings of the 34th International Conference
on Machine Learning, Sydney, Australia, 6–11 August 2017.
31. Lundberg, S.M.; Erion, G.; Chen, H.; DeGrave, A.; Prutkin, J.M.; Nair, B.; Katz, R.; Himmelfarb, J.; Bansal, N.; Lee, S.-I. From local
explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2020, 2, 56–67. [CrossRef]
32. Johnsen, P.V.; Riemer-Sørensen, S.; DeWan, A.T.; Cahill, M.E.; Langaas, M. A new method for exploring gene–gene and gene–
environment interactions in GWAS with tree ensemble methods and SHAP values. BMC Bioinform. 2021, 22, 230. [CrossRef]
33. Lauritsen, S.M.; Kristensen, M.; Olsen, M.V.; Larsen, M.S.; Lauritsen, K.M.; Jørgensen, M.J.; Lange, J.; Thiesson, B. Explainable
artificial intelligence model to predict acute critical illness from electronic health records. Nat. Commun. 2020, 11, 3852. [CrossRef]
34. Lundberg, S.M.; Nair, B.; Vavilala, M.S.; Horibe, M.; Eisses, M.J.; Adams, T.; Liston, D.E.; Low, D.K.-W.; Newman, S.-F.; Kim,
J.; et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat. Biomed. Eng. 2018,
2, 749–760. [CrossRef]
35. Guo, Y.; Ma, W.; Li, J.; Liu, W.; Qi, P.; Ye, Y.; Guo, B.; Zhang, J.; Qu, C. Effects of microplastics on growth, phenanthrene stress, and
lipid accumulation in a diatom, Phaeodactylum tricornutum. Environ. Pollut. 2020, 257, 113628. [CrossRef] [PubMed]
36. Mangalathu, S.; Hwang, S.-H.; Jeon, J.-S. Failure mode and effects analysis of RC members based on machine-learning-based
SHapley Additive exPlanations (SHAP) approach. Eng. Struct. 2020, 219, 110927. [CrossRef]
37. Yan, F.; Song, K.; Liu, Y.; Chen, S.; Chen, J. Predictions and mechanism analyses of the fatigue strength of steel based on machine
learning. J. Mater. Sci. 2020, 55, 15334–15349. [CrossRef]
38. Cakiroglu, C.; Bekdaş, G.; Kim, S.; Geem, Z.W. Explainable Ensemble Learning Models for the Rheological Properties of
Self-Compacting Concrete. Sustainability 2022, 14, 14640. [CrossRef]
39. Chen, M.; Wang, H. Explainable machine learning model for prediction of ground motion parameters with uncertainty quanti-
fication. Chin. J. Geophys. 2022, 65, 3386–3404. (In Chinese) [CrossRef]
40. Okada, Y.; Kasahara, K.; Hori, S.; Obara, K.; Sekiguchi, S.; Fujiwara, H.; Yamamoto, A. Recent progress of seismic observation
networks in Japan—Hi-net, F-net, K-NET and KiK-net—. Earth Planets Space 2004, 56, 15–28. [CrossRef]
41. GB50011-2010; Code for Seismic Design of Buildings. China Architecture & Building Press: Beijing, China, 2010. (In Chinese)
42. Breiman, L. Random Forest. Mach. Learn. 2001, 45, 5–32. [CrossRef]
43. Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting System. In Proceedings of the KDD ’16: 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794.
[CrossRef]
44. Oweis, I.; Urzua, A.; Dobry, R. Simplified Procedures for Estimating the Fundamental Period of a Soil Profile. Bull. Seismol. Soc.
Am. 1975, 66, 51–58.
45. Boore, D.M.; Thompson, E.M.; Cadet, H. Regional Correlations of VS30 and Velocities Averaged Over Depths Less Than and
Greater Than 30 Meters. Bull. Seismol. Soc. Am. 2011, 101, 3046–3059. [CrossRef]
46. Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al.
Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830.
Appl. Sci. 2023, 13, 4530 22 of 22
47. Qi, J.L. Soil Seismic Response Prediction Model Method Based on Machine Learning Algorithms. Master’s Thesis, Institute of
Engineering Mechanics, China Earthquake Administration, Harbin, China, 2021. (In Chinese).
48. Ma, L.; Zhou, C.; Lee, D.; Zhang, J. Prediction of axial compressive capacity of CFRP-confined concrete-filled steel tubular short
columns based on XGBoost algorithm. Eng. Struct. 2022, 260, 114239. [CrossRef]
49. Somala, S.N.; Chanda, S.; Karthikeyan, K.; Mangalathu, S. Explainable Machine learning on New Zealand strong motion for PGV
and PGA. Structures 2021, 34, 4977–4985. [CrossRef]
50. Mangalathu, S.; Shin, H.; Choi, E.; Jeon, J.-S. Explainable machine learning models for punching shear strength estimation of flat
slabs without transverse reinforcement. J. Build. Eng. 2021, 39, 102300. [CrossRef]
51. Sun, R.; Yuan, X. A holistic equivalent linear method for site response analysis. Soil Dyn. Earthq. Eng. 2021, 141, 106476. [CrossRef]
52. Darendeli, M.B. Development of a New Family of Normalized Modulus Reduction and Material Damping Curves. Ph.D. Thesis,
The University of Texas at Austin, Austin, TX, USA, 2001.
53. Shapley, L.S. A Value for N-Person Games. In Contributions to the Theory of Games; Princeton University Press: Princeton, NJ, USA, 1952.
[CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

Explainable Machine-Learning Predictions For Peak

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Explainable Machine-Learning Predictions For Peak

Uploaded by

Copyright:

Available Formats

applied

Academic Editors: José A. Peláez,

Appl. Sci. 2023, 13, 4530. https://doi.org/10.3390/app13074530 https://www.mdpi.com/journal/applsci

2. Establishment of the Dataset

Table 1. Chinese code for classification of sites.

I0 and I1 are subclasses of site class I.

Figure 1. Distribution of PGA.

3.1.2. Random Forest Regression Model

3.1.3. XGBoost Regression Model

If the loss function is expanded by second-order Taylor expansion, the objective

3.2. Feature Selection

Appl. Sci. 2023, 13, x FOR PEER REVIEW 10 of 23

Figure 5. Correlation matrix for input four features.

Figure 7. Determination of hyper-parameters for RF model.

Figure 8. Determination of hyper-parameters for the XGBoost model.

Table 3. Prediction results of three models.

3.4. Analysis of XGBoost Model Results

Appl. Sci. 2023, 13, 4530 13 of 22

15 Predictions/Observations >1.00 Predictions/Observations

Appl. Sci. 2023, 13, x FOR PEER REVIEW 15 of 23

0.08 0.32 0.08 0.32 0.08 0.32

MAE and RMSE

MAE and RMSE

0.05 0.26 0.05 0.26 0.05 0.26

0.04 0.24 0.04 0.24 0.04 0.24 MAPE

0.03 0.22 0.03 0.22 0.03 0.22

0.02 0.20 0.02 0.20 0.02 0.20

0.01 0.18 0.01 0.18 0.01 0.18

0.05 0.40 0.05 0.40 0.40

MAE and RMSE

0.25 0.25 0.25

0.15 0.15 0.15

0.10 0.10 0.10

0.01 0.05 0.01 0.05 0.05

4. Explanation of the Prediction Model Based on SHAP

and f (xij ) is as follows:

|S|!(| F | − |S| − 1)!

4.2. SHAP Global Analysis

Appl. Sci. 2023, 13, x FOR PEER REVIEW 17 of 23

You might also like