DOI: 10.1111/mice.

12852

RESEARCH ARTICLE

Deep convolutional generative adversarial networks for the

generation of numerous artificial spectrum-compatible
earthquake accelerograms using a limited number
of ground motion records
Mehrshad Matinfar1 Naser Khaji2 Goodarz Ahmadi3

1 Facultyof Civil and Environmental

Engineering, Tarbiat Modares University, Abstract
Tehran, Iran Deep learning (DL) methodologies have been recently employed to solve
2 Civil and Environmental Engineering
various civil and earthquake engineering problems. Nevertheless, due to the
Program, Graduate School of Advanced
Science and Engineering, Hiroshima
limited number of reliable data in the field of earthquake engineering, it is not
University, Hiroshima, Japan convenient to obtain accurate results using DL. To tackle this challenge, the
3 Department of Mechanical and
generative adversarial network (GAN) approach may be considered a reliable
Aerospace Engineering, Clarkson
possible candidate. GANs have been introduced as an efficient way to train
University, Potsdam, New York, USA
generative models. GANs exhibited their capabilities as well as versatility in
Correspondence the field of image production. For nonlinear dynamic analyses of structures,
Naser Khaji, Civil and Environmental
artificial ground accelerograms that are compatible with a target response spec-
Engineering Program, Graduate School of
Advanced Science and Engineering, trum are usually generated. In this paper, an efficient algorithm is proposed by
Hiroshima University, 1-4-1, Kagamiyama, which numerous artificial spectrum-compatible earthquake accelerograms are
Higashi-Hiroshima, Hiroshima 739–8527,
Japan.
generated using a few ground motion records. For this purpose, a specific well-
Email: [email protected] established generative model, namely, the deep convolutional GAN (DCGAN),
is adopted for the first time and used. It is shown that DCGAN can easily
generate desirable artificial ground accelerograms by having a limited number
of seismic records as input to train the network. To quantitatively demonstrate
the quality of the artificial ground accelerograms generated by the DCGAN,
several computer experiments are presented, among which the robustness and
feasibility of the proposed method are examined by using only four earthquake
accelerograms as the worst scenario. Moreover, the efficiency of the DCGAN is
illustrated by comparing various seismic parameters and the spectral response
of the generated accelerograms with those of the actual accelerograms. The
outcomes illustrate the efficiency and robustness of the presented DCGAN.

1 INTRODUCTION response spectrum is the design spectrum provided by

seismic codes. Time-history dynamic analysis is fre-
For the aseismic design of structural systems, response quently employed to design critical/important structures
spectra may be employed as a convenient tool to evalu- like high-rise buildings, for which several ground accelera-
ate the maximum responses of these systems. A useful tions should be available in conducting nonlinear dynamic

© 2022 Computer-Aided Civil and Infrastructure Engineering.

Comput Aided Civ Inf. 2023;38:225–240. wileyonlinelibrary.com/journal/mice 225

 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
226 MATINFAR et al.

analyses (ASCE 7–16; American Society of Civil Engineers, case in civil and earthquake engineering problems (Rafiei
2017). Consequently, it is necessary to generate accelero- & Adeli, 2017), where extensive progress was achieved
grams that are compatible with a given target design using ML (Rafiei & Adeli, 2018a) and DL (Nabian & Mei-
spectrum. dani, 2018). Cha et al. (2017) used a convolutional neu-
Various methods have been developed regarding the ral network (CNN) to automatically detect the existing
production of artificial spectrum-compatible earthquake cracks in concrete. For structural health monitoring (SHM)
accelerograms. Kanai (1957) and Tajimi (1960) proposed purposes, Azimi and Pekcan (2020) used extremely com-
the well-known stationary-filtered white noise model for pressed data through deep transfer learning. Addition-
the seismic ground motion. Their model has been widely ally, DL approaches were used for SHM by implement-
employed in the random vibration analysis of struc- ing a sustainable strain sensing model (Oh et al., 2017).
tural systems (Ahmadi, 1979). Various methods based Rafiei and Adeli (2018b) proposed a novel unsupervised
on the generalized nonstationary Kanai–Tajimi models DL model for global and local SHM. Gao and Mosalam
(Fan & Ahmadi, 1990; Rofooei et al., 2001), evolution- (2018) achieved some experiments to detect the existing
ary Kanai–Tajimi models (Lin & Yong, 1987), nonsta- damages in structural components by employing a small
tionary random process models (Conte & Peng, 1997; dataset. Gulgec et al. (2020) proposed a DL model for the
Conte et al., 1992; Mobarakeh et al., 2002), and stochas- fatigue assessment of structures through strain estimation
tic methods (Boore, 2003; Yang & Zhou, 2015) have from acceleration data. Cheng et al. (2021) developed a DL
been developed to simulate ground motion time-histories. approach for post-hurricane aerial damage assessment of
Cacciola and Zentner (2012) developed a methodol- buildings. Sajedi and Liang (2021) used Bayesian inference
ogy for generating artificial spectrum-compatible earth- for deep vision SHM models, where uncertainty is quan-
quake time histories using random joint time-frequency tified using Monte Carlo dropout sampling. F. Wang et al.
distributions. (2021) used a memory-augmented CNN to detect the shear
Neural networks have been extensively used in various loading of through-bolts in bridges. Hsieh et al. (2021) used
fields of engineering (Perez-Ramirez et al., 2019), among CNN to classify jointed plain concrete pavement condi-
which are some remarkable methods for the generation of tions automatically.
artificial spectrum-compatible earthquake accelerograms Goodfellow et al. (2014) were the first who introduced
(e.g., Lee & Han, 2002). Ghaboussi and Lin (1998) solved the GAN model and indicated the power of this genera-
this problem using artificial neural networks by consider- tive model on Modified National Institute of Standards and
ing the problem as reverse mapping and improved their Technology, Tenders Electronic Daily, and Canadian Insti-
work with a stochastic neural network approach (Lin & tute for Advanced Research databases. The deep convolu-
Ghaboussi, 2001). Moreover, wavelet transforms (Zhou & tional generative adversarial network (DCGAN) generates
Adeli, 2003) in combination with neural networks or inde- high-resolution images with better quality by combining
pendently have been employed for generating artificial the GAN and convolutional layers (Radford et al., 2016).
accelerograms (Amiri, Rad, & Hazaveh, 2014; Amiri et al., The GAN can also support labeled data by using CGAN
2012; Sirca & Adeli, 2004). Furthermore, wavelet trans- or InfoGAN (X. Chen et al., 2016). Many researchers have
forms in combination with the Hilbert transform have customized GANs for their specific purpose, most of which
been used in modal parameters identification and health are principally related to images or two-dimensional (2D)
monitoring of large-scale structures (Amezquita-Sanchez data (Bousmalis et al., 2017; Denton et al., 2016; Donahue
et al., 2017; Z. Li et al., 2017). In addition, fuzzy systems et al., 2017). A few researchers employed GANs for the aug-
(Amiri, Khorasani, et al., 2014; Heidari & Khorasani, 2012; mentation of available datasets by producing new synthetic
H. Li et al., 2013) and genetic algorithm-based methods images to enhance the applicability of the given models
(Naeim et al., 2004; Shahjouei & Ghodrati Amiri, 2011) (Zhang et al., 2021). Yu et al. (2021) used a GAN to propose
have played crucial roles in the generation of artificial a probabilistic vehicle weight estimation approach.
earthquake time histories. Park et al. (2018) used artifi- Although the GANs have shown their robustness and
cial accelerograms to study aftershocks that occur after an accuracy in various applications, there are few applica-
earthquake. tions of GANs in the field of civil engineering. After the
Machine learning (ML) (Rafiei & Adeli, 2016; Rafiei introduction of the GAN, researchers tried to use this
et al., 2017) and deep learning (DL) (Dong et al., 2021) powerful network in their dedicated field of civil engineer-
methodologies are becoming great milestones in the rapid ing. Gao et al. (2019) extended the PEER Hub ImageNet
developments and applications of artificial intelligence database using a DCGAN to effectively augment structural
research in numerous fields of science and technol- image data. T. Wang et al. (2019) developed a new method
ogy, such as safety engineering (Shen et al., 2021) and to generate seismic events using a GAN model and verified
autonomous vehicle control (Chen et al., 2021). This is the the performance of their method using seismic field data
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 227

acquired in Oklahoma, where induced earthquake events duces an unlimited number of accelerograms with high
have been reported. Gatti and Clouteau (2020) used a quality. The approach parallels the GAN’s ability to pro-
GAN to present a strategy for blending the outcome of duce images similar but different from the original images.
physics-based numerical simulations with massive but This paper uses this exciting feature of the GAN to produce
poorly labeled experimental databases. Li et al. (2020) artificial accelerograms that match the design spectrum.
proposed a GAN-based model that uses conditional The resulting accelerograms are visually very similar to the
knowledge to generate high-quality seismic waveforms. records in the database but smoothly different from them
The proposed method generates synthetic samples based in terms of peak ground acceleration, peak ground velocity,
on the statistical characteristics of real seismic waveforms. peak ground displacement, and other seismic parameters.
Maeda et al. (2021) applied GAN to generate new databases The paper is organized as follows. Section 2 describes
for the road damage detection problem. The GAN has also the mathematical principles and basic concepts of GANs
been used for travel time issues and generating traffic data and DCGANs. Section 3 proposes the initial accelerogram
(Song et al., 2020). Gao et al. (2021) proposed a balanced database and the required conversion to 2D arrays. The
semi-supervised GAN for damage assessment from a proposed model, including the generator, the discrimina-
low-data imbalanced-class regime. tor, and the training procedure, is demonstrated in Sec-
As stated earlier, for nonlinear dynamic analyses of tion 4. This is followed by the presentation of the numerical
structures, artificial ground accelerograms compatible experiments and the description of the results in Sections 5
with a target response spectrum are usually generated. The and 6. Section 7 draws the conclusions.
design codes commonly set a minimum for the number
of required accelerograms. The use of a higher number
of accelerograms increases the reliability of the structure 2 GAN
against future earthquakes. Therefore, one essential con-
cern of structural engineers is the minimum number of 2.1 Mathematical principles
accelerograms to be selected, namely, 3, 5, 7, 11, 22, or more
(NEHRP Consultants Joint Venture, 2011). For instance, A GAN is a DL generative model that is mainly used to
ASCE 7–16 requires the use of 11 records as the minimum generate high-quality samples (Goodfellow et al., 2014).
number of ground motion acceleration histories. A GAN typically includes a parameterized discriminator
In many tectonic settings of the world, there are regions and generator, which are any neural network or prescribed
with poor historical/instrumental databases, for which functions. In order to train the GAN, the generator and the
very limited recorded data are available about seismic discriminator play a game in some adversarial sense. In
ground motion for structural engineers to rely upon. this game, the generator and discriminator try to maximize
Therefore, the essential need of structural engineers, under the outcomes and minimize the competitor’s gain itera-
such circumstances, is to generate artificial ground motion tively. The generator has no direct access to the considered
time histories that should be compatible with the tar- database and therefore learns to produce high-quality sam-
get response spectrum. Moreover, the generated database ples through trial-and-error to guess the opinion of the dis-
of accelerograms should have the same specifications as criminator. On the other hand, the discriminator is trained
the actual records, including soil type, frequency content, with actual samples and learns to score the data sent to it
duration, and more. This important challenge has received by the generator. In more detail, suppose that the generator
little attention in the literature. Therefore, this paper pro- intends to generate synthetic data to mislead the discrim-
poses a new approach for generating numerous accelero- inator. However, the discriminator struggles to recognize
grams with high quality using the outstanding features of the synthetic and actual data produced by the generator.
GANs. According to a considerable literature review, to the When the generator succeeds to mislead the discriminator,
best knowledge of the authors, there is no research in rela- the discriminator updates itself so that it can
tion to GANs for generating the required number of artifi- Distinguish synthetic data appropriately. Similarly,
cial spectrum-compatible earthquake accelerograms using when the upgraded discriminator recognizes the synthetic
a limited number of ground motion records as a database. data, the generator updates itself again to mislead the
Hence, it is worthwhile to examine the following two discriminator once more. This procedure continues until
inquiries: (a) Can GANs be implemented in earthquake the discriminator cannot upgrade itself to distinguish
engineering to generate the needed strong ground motion the synthetic data. At this point, the probability of cat-
time histories? (b) How can the suitability of the generated egorization of real data versus synthetic data becomes
records be evaluated? The principal contributions of the almost 50% for the discriminator (Goodfellow et al., 2014).
present paper are to provide answers to the above-raised Accordingly, the synthetic data are adequately close to the
questions. This paper presents a new approach that pro- real data. From the mathematical point of view, GAN’s
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
228 MATINFAR et al.

FIGURE 1 The general architecture of the model used in this study

training is a kind of MinMax game between the generator = 𝔼𝑥∼𝑑𝑎𝑡𝑎 (𝑥) [𝑙𝑜𝑔𝐷 (𝑥)]
and the discriminator. In this game, the generator learns to
make better samples to mislead the discriminator. On the + 𝔼𝑧∼𝑍 (𝑧) [𝑙𝑜𝑔 (1 − 𝐷(𝐺 (𝑧)))] (1)
other hand, the discriminator learns to better understand
the difference between synthetic and real data. This com- in which 𝜃𝐷 and 𝜃𝐺 denote the parameters of the discrim-
petition results in the production of high-quality samples inator and generator functions, respectively. Furthermore,
in various fields where GAN was implemented. This game 𝔼𝛼∼𝛽 [⋅] implies the expectation function of data or ran-
optimization may have an optimal solution with inherent dom variable 𝛼 sampled from the distribution 𝛽 . As men-
difficulties, such as possible nonconvex loss functions, as tioned earlier, the discriminator tries to recognize whether
well as high-dimensional parameter space. Consequently, the fed data are synthetic or real. As a result, the cross-
it is preferable to implement gradient-based algorithms entropy can be simply written in terms of the loss as fol-
to minimize the loss of the generator and discriminator lows:
at the same time. The general form of a typical GAN is 𝑁 𝑁
1 ∑ 1 ∑
illustrated in Figure 1. 𝐷 = − 𝑦𝑖 𝑙𝑜𝑔𝐷 (𝑥𝑖 ) − (1 − 𝑦𝑖 ) 𝑙𝑜𝑔 (1 − 𝐷 (𝑥𝑖 ))
In this figure, x indicates actual data points, and z 𝑁 𝑖=1 𝑁 𝑖=1
denotes a high-dimensional random variable, which is (2)
usually referred to as code. where N is the number of samples. Now, one can label real
Now, let us introduce the probability distributions of real data, 𝑥𝑖 , as 1, while synthetic data, 𝑥̂ 𝑖 , is labeled as 0. To this
data 𝑑𝑎𝑡𝑎 (𝑥), generative model 𝐺 (𝑔), and random vari- end, the label 𝑦𝑖 is assigned to 1 and 0. Hence, Equation (2)
able 𝑍 (𝑧). The discriminator function 𝐷(⋅) and genera- can be simplified as follows:
tor function 𝐺(⋅) are indicated in Figure 1. To calculate the
probability of recognizing the real data by the discrimina- 𝐷 = −𝔼𝑥∼𝑑𝑎𝑡𝑎 (𝑥) [𝑙𝑜𝑔𝐷 (𝑥)] − 𝔼𝑥∼𝐺 (𝑔) [𝑙𝑜𝑔 (1 − 𝐷 (𝑥))]
tor, 𝐷(𝑥) is chosen as a sigmoid function (Goodfellow et al., (3)
2014). Also, (z) generates the synthetic data in accordance When 𝑁 → ∞ in Equation (2), the empirical value of 𝐷
with 𝐺 . As discussed earlier, the generator is trained to asymptotically approaches its theoretical value. In other
produce synthetic samples 𝑥̂ that are statistically very sim- words, when we optimize the discriminator by evaluat-
ilar to real samples, while on the other hand, the discrim- ing its parameters, 𝜃̂𝐷 , we actually minimize the negative
inator is trained to give higher 𝑠 to 𝑥 and lower 𝑠 to 𝑥̂ to amount of the value function, −𝑉(𝐷, 𝐺), which means we
counterplot the generator’s tries (see Figure 1). The value maximize 𝑉(𝐷, 𝐺) as given below:
function is given below (Goodfellow et al., 2014):
𝜃̂𝐷 = 𝑎𝑟𝑔𝑚𝑖n𝜃𝐷 𝐷 = 𝑎𝑟𝑔𝑚𝑎𝑥𝐷 𝑉 (𝐷, 𝐺) (4)
𝑉 (𝐷, 𝐺) = 𝑉 (𝜃𝐷 , 𝜃𝐺 ) = 𝔼𝑥∼𝑑𝑎𝑡𝑎 (𝑥) [𝑙𝑜𝑔𝐷 (𝑥)]
Consequently, the game tries to maximize the discrim-
+ 𝔼𝑥∼𝐺 (𝑔) [𝑙𝑜𝑔 (1 − 𝐷 (𝑥))]
inator’s gain by recognizing the difference of 𝐺 (𝑔) and
𝑑𝑎𝑡𝑎 (𝑥) in full measure.
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 229

As mentioned earlier, the generator produces synthetic

data during the optimization process and tries to produce
synthetic data of a high quality to mislead the discrimi-
nator. As GAN’s training is a kind of MinMax game, the
generator struggles to minimize the discriminator’s gain.
Therefore, we can consider the discriminator’s gain to be
the generator’s loss function as follows:

𝐺 = 𝑚𝑎𝑥 𝑉 (𝐷, 𝐺) (5)

𝐷

for which the parameters are eventually estimated as (a)

𝜃̂𝐺 = 𝑎𝑟𝑔𝑚𝑖n𝜃𝐺 𝐺 = 𝑎𝑟𝑔𝑚𝑖n𝐺 𝑚𝑎𝑥 𝑉 (𝐷, 𝐺) (6)

𝐷

The abovementioned MinMax game attempts to itera-

tively optimize the discriminator and generator loss func-
tions as far as both functions approach some possible
local minima, for which the evaluated parameter sets are
denoted by 𝜃̂𝐷 and 𝜃̂𝐺 .

2.2 DCGAN

In the initial studies of the GANs, a multilayer perceptron

was implemented as a function estimator of 𝐺(⋅) and
𝐷(⋅). As the GANs exhibited some limitations, such as
the impossibility of generating high-resolution images (b)
and instability of training (Bousmalis et al., 2017; X. Chen
et al., 2016), many updated GANs were introduced to F I G U R E 2 Conversion of (a) a typical accelerogram as a
one-dimensional array to (b) a two-dimensional array represented
solve these problems (Denton et al., 2016; Donahue et al.,
by a 32 × 32 matrix
2017; Radford et al., 2016). Radford et al. (2016) proposed
deep convolutional networks for GANs called DCGANs.
The architecture of DCGANs is introduced by employing 5. Dismiss the fully connected layers after the last convo-
a given class of CNNs for both the discriminator and lution layer.
generator. In addition, the architecture is limited to some
constraints as described in the following guidelines: As a result, a particular DCGAN configuration is
proposed in this research. Considering the above
1. Employ a convolutional layer with a stride for down- constraints, the detailed architecture of the genera-
sampling in the discriminator and a fractional con- tor and discriminator is discussed in the following
volutional layer with a stride for upsampling in the sections.
generator.
2. Utilize batch normalization (Ioffe & Szegedy, 2015) in
both the generator and discriminator. 3 DATABASE
3. Apply the Rectified Linear Unit, or leaky ReLU (Xu
et al., 2015) activation function in the discriminator for 3.1 Initial accelerogram database
all layers, except for the output layer, which employs the
sigmoid function. Generating or preparing a database is one of the critical
4. Use the leaky ReLU activation function except for the steps of most DL approaches. In this regard, one of the
output layer of the generator and use the hyperbolic tan- most serious issues is to produce or find suitable databases
gent activation function in the output layer of the gen- for earthquakes with desirable conditions. This is difficult
erator. since the number of appropriate earthquakes that satisfy
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
230 MATINFAR et al.

FIGURE 3 A typical generator architecture of the model used in this study

TA B L E 1 Typical architecture of the deep convolutional generative adversarial network (DCGAN) presented in this paper
Generator Discriminator
Layer type Output shape Layer type Output shape
L1 Dense 8192 Conv + LR + Dpout 16, 16, 32
L2 Reshape 8, 8, 128 Conv + ZP + BN + LR + Dpout 9, 9, 64
L3 UpSampling + Conv + BN + Activation 16, 16, 128 Conv + BN + LR + Dpout 5, 5, 128
L4 UpSampling 32, 32, 128 Conv + BN + LR + Dpout 5, 5, 256
L5 Conv + BN + Activation 32, 32, 64 Flatten 6400
L6 Conv + Activation 32, 32, 1 Dense 1
Note: Training is performed via the Adam optimization method (Kingma et al., 2015).
Abbreviations: BN, batch normalization; Conv, 2D convolution; Dpout, dropout LR, leaky ReLU; ZP, zero padding.

the desirable conditions required for nonlinear dynamic et al., 2017). Tang et al. (2019) converted SHM raw time
analyses of structures is quite limited. series data to images prior to input, rather than inputting
The following sections show that the proposed DCGAN the raw time series data directly to the neural network.
is capable of generating numerous artificial spectrum- Liao et al. (2021) stored the wavelet transform of one-
compatible earthquake accelerograms using a limited dimensional (1D) time-series vibrations as characteristic
number of ground motion records. For this purpose, we coefficient matrices, as the inputs of deep neural networks.
need to have just a few accelerograms that are compatible Zhang et al. (2021) proposed a translation scheme for con-
with the target spectrum. The few needed accelerograms verting vibration signals to images for bearing fault diagno-
can be produced by well-established methods available in sis of rotating machinery. This extraordinary performance
the literature as reviewed in Section 1 (Introduction). In led us to switch to 2D arrays instead of using GAN with its
this paper, the wavelet algorithm proposed by Hancock ability to generate 1D arrays.
et al. (2006) is used to adjust earthquake accelerograms to In earthquake engineering applications, model input
match a specific target response spectrum. and output accelerograms should be appropriate for struc-
In this paper, we assume that a limited number of suit- tural design purposes. The procedure of converting 1D
able accelerograms for the analysis and design of struc- arrays to 2D arrays is summarized in the following steps:
tures according to the requirements of the target structural
code, such as ASCE 7–16, are available. In general, these 1. Initial step: the duration (𝑇𝑑 ) of the expected gener-
accelerograms can be adapted to any spectrum accord- ated accelerograms as well as the time intervals (or time
ing to the needs of the designer. Moreover, the developed steps) Δ𝑡 are selected properly in association with struc-
DCGAN model has no particular restrictions on the soil tural analysis and design purposes, for example.
condition or the type of spectrum used. 2. First step: The duration of each record of the consid-
ered database is modified according to the duration
selected in the initial step. For real records shorter than
3.2 Conversion to 2D arrays the selected duration, the end portion of the record
is assumed to be zero. Real records longer than the
GANs show outstanding performances for 2D images or selected duration are trimmed by keeping the principal
arrays, for which there are numerous examples in the lit- portion of the record, which includes main excitations
erature (Bousmalis et al., 2017; Isola et al., 2017; Ledig starting from the initial P waves.
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 231

FIGURE 4 A typical discriminator architecture of the model used in this study

3. Second step: Considering the prescribed 𝑇𝑑 and Δ𝑡, the understand the difference between the two images. At the
dimension of the equivalent 1D array is simply deter- end of the training process, the generator should produce
mined as 𝑚 = 𝑇𝑑 ∕Δ𝑡. The least integer number 𝑛 is qualified images. With the help of this network, numerous
then selected so that 𝑚 ≤ 𝑛2 . accelerograms can be created with features similar to those
4. Third step: For each record of the database, the differ- in the database.
ence between these two numbers (i.e., 𝑚 and 𝑛2 ) can be
𝑛2 −𝑚
resolved by adding zero numbers at the beginning
2 4.1 Generator
and the end of the record.
5. Fourth step: All records now constitute 1D arrays of The architecture of the generator is determined according
𝑚(= 𝑛2 ) dimension and can be directly converted to to the duration of the earthquake. In other words, the net-
equivalent 2D arrays of 𝑛2 dimension, by allocating work training process for long-term earthquakes requires
every 𝑛 component of 1D arrays in the associated row larger layer dimensions than for short-term earthquakes.
of 2D arrays. The number of layers is also chosen in accordance with
the complexity of the problem. However, the experience of
To examine the feasibility and accuracy of the DCGAN the user plays a central role in constructing a perfect archi-
to generate desirable accelerograms, the duration and time tecture, while trial-and-error may also lead to acceptable
intervals of 20 and 0.02 s, respectively, are selected as typ- results. A typical generator model used in this study can
ical examples. Consequently, any given accelerogram con- be seen in Figure 3, whose details are shown in Table 1.
20
tains an array of = 1000 numbers. The square 2D array In this typical model, 100 random numbers (or noises)
0.02
(image) closest to a total of 1000 elements has the dimen- are initially generated and applied to Conv2D layers. By
sion of 322 = 1024. The difference between these two inspecting various layers’ numbers, it was revealed that
numbers (i.e., 1000 and 1024) can be resolved by adding four layers of Conv2D result in an appropriate acceleration
12 zero numbers at the beginning and the end of the earth- for learning, creativity, and accuracy in producing desir-
quake record of the database. A typical converted accelero- able accelerometers. As already stated, we use the leaky
gram is shown in Figure 2. ReLU activation function for all the layers.

4 MODEL 4.2 Discriminator

As discussed earlier, the main structure of the model used The discriminator is responsible for scoring its incoming
in this paper is derived from DCGAN (Radford et al., 2016). earthquakes and then determines how well they match
The general architecture of the model used can be seen with what is in the database. Consequently, the end layer
in Figure 1. The earthquakes in the database are con- dimension of this network should be one, which indicates
verted to 2D images, and these images are then provided the layer of the real earthquake. On the other hand, the
directly to the discriminator. Therefore, the discriminator input of the discriminator should be exactly equal to the
network is trained with the provided data. On the other images obtained from the earthquakes in the database,
hand, the generator tries to guess images that are simi- which is, in turn, identical to the output of the generator.
lar to what was taken from the output database by gen- A typical discriminator model used in this study is drawn
erating images so that the discriminator network cannot in Figure 4, whose details are shown in Table 1.
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
232 MATINFAR et al.

(a)

F I G U R E 6 Response spectra of accelerograms of Figure 5 as

sample outputs of the DCGAN in various iterations of training of
the generator

4.3 Training procedure

(b)
As stated earlier, the MinMax game tries to iteratively opti-
mize the discriminator and generator loss functions as
far as both functions approach possible local minima, for
which the estimated parameter sets are denoted by 𝜃̂𝐷 and
𝜃̂𝐺 . However, it is difficult to detect whether the discrimi-
nator and generator loss functions have reached local min-
ima in practice. Consequently, a prescribed iterations num-
ber is often set. The procedure of training is summarized in
the following steps:
(c)
1. Initial step: The discriminator and generator are initial-
ized.
2. First step: Considering the distribution 𝑑𝑎𝑡𝑎 (𝑥), 𝑁 data
points {𝑥1 , 𝑥2 , … , 𝑥𝑁 } are sampled from real data. The
distribution 𝑑𝑎𝑡𝑎 (𝑥) is approximated empirically. As
seen in the next numerical experiments, DCGAN shows
very good performances even for small values of 𝑁 = 4,
for example.
3. Second step: Assuming the selected distribution 𝑍 (𝑧),
N noise {𝑧1 , 𝑧2 , … , 𝑧𝑁 } is sampled.
(d) 4. Third step: Using gradient descent, the discriminator is
updated to minimize loss 𝐷 as given below:
F I G U R E 5 Sample outputs of the deep convolutional
[ 𝑁
generative adversarial network (DCGAN) at the (a) 10th, (b) 100th,
1 ∑
(c) 1000th, and (d) 10,000th iterations of training of the generator 𝜃̂𝐷 ← 𝜃̂𝐷 − 𝜂 ⋅ ∇𝜃𝐷 − 𝑙𝑜𝑔𝐷 (𝑥𝑖 )
𝑁 𝑖=1
𝑁
]
1 ∑
− 𝑙𝑜𝑔(1 − 𝐷 (𝐺(𝑧𝑖 ))) (7)
𝑁 𝑖=1
In the typical architecture shown in Figure 4, for exam-
ple, the input of the discriminator is a matrix value with
in which η denotes the learning rate.
dimensions of 32 × 32.
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 233

F I G U R E 7 Comparison of (a) three randomly selected accelerograms of the database with (b) those generated by the DCGAN, (c) whose
differences are remarkable

5. Fourth step: Using gradient descent, the generator is

updated to minimize loss 𝐺 as follows:
[ 𝑁
1 ∑
𝜃̂𝐺 ← 𝜃̂𝐺 − 𝜂 ⋅ ∇𝜃𝐺 𝑙𝑜𝑔𝐷 (𝑥𝑖 )
𝑁 𝑖=1

𝑁
]
1 ∑
+ 𝑙𝑜𝑔(1 − 𝐷 (𝐺(𝑧𝑖 ))) (8)
𝑁 𝑖=1

6. The first to the fourth steps are repeated as far as conver-

(a) gence is obtained. To this end, the response spectrum of
the generated accelerogram (𝑎𝑔𝑒𝑛 (𝑇𝑗 )) of each iteration
is compared to the target spectrum (𝑎𝑡𝑎𝑟 (𝑇𝑗 )). The iter-
ations are terminated when
𝑀
∑ 𝑀
∑
| ( ) ( )| ( )
|𝑎𝑔𝑒𝑛 𝑇𝑗 − 𝑎𝑡𝑎𝑟 𝑇𝑗 | ≤ 𝜀 𝑎𝑡𝑎𝑟 𝑇𝑗 (9)
| |
𝑗=1 𝑗=1

where 𝑀 indicates the number of sample periods of the

spectra, and 𝜀 = 5%.

(b) 5 EXPERIMENTS AND RESULTS

F I G U R E 8 (a) A sample artificial earthquake accelerogram In this section, several numerical experiments are per-
generated using the DCGAN, (b) spectral accelerations of the formed to specify how well the DCGAN model can provide
generated accelerogram in comparison with the target design appropriate accelerograms.
spectrum
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
234 MATINFAR et al.

(a)

F I G U R E 1 0 The Arias intensity (𝐼𝐴 ) of generated artificial

time-histories of Figure 9a (dotted line) and the 𝐼𝐴 of input
earthquake ground motions (solid line)

(b)
accuracy of the generator. In other words, we examine
F I G U R E 9 (a) Six sample accelerograms randomly selected whether the neural network is going in the right direc-
from the output dataset of the DCGAN, (b) mean value of generated tion in the training process. Accordingly, the desired artifi-
responses spectra in comparison with the target design spectrum of cial accelerograms as the output of the GAN are presented
ASCE 7–16 at several stages during the training of the generator in
Figure 5. Figure 6 shows the corresponding response spec-
5.1 Visual inspection of the output tra of the accelerograms in comparison with the selected
results design spectrum of ASCE 7–16 for the soil of type C. As is
clearly observed from these figures, the generator struggles
In reviewing the validation approaches used in many of with trial-and-error to produce accelerograms that are sim-
the earlier works in various scientific fields, one of the ilar to the accelerograms of the database. For further clar-
best approaches is visually inspecting the results. The first ification, Figure 7 illustrates the performance of the GAN
experiment is a simple one, in which we look at the train- network after training. This figure shows that the GAN can
ing process visually to investigate the improvement of the generate data that are similar to the accelerograms in the

TA B L E 2 Selected earthquakes from the NGA database

No. Earthquake name Faulting mechanism Moment magnitude
1 Morgan Hill Strike-slip 6.19
2 Kobe, Japan Strike-slip 6.90
3 Hector Mine Strike-slip 7.13
4 Kocaeli, Turkey Strike-slip 7.51
5 Northridge Reverse 6.69
6 Friuli, Italy Reverse 6.50
7 San Fernando Reverse 6.61
8 Loma Prieta Reverse 6.93
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 235

F I G U R E 1 1 The actual and generated ground motions corresponding to the 1995 Kobe earthquake at Kakogawa station, where (a)
acceleration, (b) velocity, and (c) displacement time histories are shown

trum of ASCE 7–16 (5% damping) for ground type C. It is

found from Figure 8 that the spectral acceleration of the
generated accelerogram is compatible with those of the tar-
get design spectrum.

5.3 Features of the generated artificial

response spectra

In order to illustrate the overall performance of the gener-

ated artificial time histories, six sample accelerograms are
randomly selected from the output database, which con-
tains a large number of accelerograms (see Figure 9a). The
mean value of the generated accelerogram response spec-
FIGURE 12 The 𝐼𝐴 of actual and generated ground motions in tra is plotted in Figure 9b. The accuracy of the proposed
Figure 11 DCGAN is observed in this figure by the agreement of the
generated response spectra to the target spectra.
database. This also confirms the remarkable achievement
of the network.
5.4 Goodness of the generated artificial
time-histories
5.2 Comparison of response and design
spectra As Arias (1970) introduced, the Arias intensity (𝐼𝐴 ) is usu-
ally considered a suitable indicator to evaluate the total
The second numerical experiment was performed to com- input energy of a strong ground motion. This parameter
pare the response spectrum obtained from the gener- is associated with the spectrum of the energy produced
ated accelerograms with the target design spectrum. As by the target strong ground motion and is considered a
described earlier, a limited number of ground motion measure for earthquake strength related to its structural
records are employed to generate an unlimited number of damage, for example. Figure 10 compares the 𝐼𝐴 for the
artificial seismic accelerograms that are compatible with produced time histories of Figure 9a with those of earth-
the target spectrum. Figure 8a represents a typical gener- quake ground motions used as the input of the DCGAN
ated artificial accelerogram, whose response spectrum is database. Obviously, the 𝐼𝐴 of the generated time history
drawn in Figure 8b, compared to the target design spec- matches the 𝐼𝐴 of the target time history very well.
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
236 MATINFAR et al.

TA B L E 3 Seismic parameters of actual and generated ground motions for the 1995 Kobe earthquake at Kakogawa station

Seismic Generated Generated Generated

No. parameter Unit Actual record Record 1 Record 2 Record 3
1 Peak ground g 0.176 0.231 0.215 0.263
acceleration
(PGA)
2 Peak ground cm/s 31.3 21.0 27.4 29.5
velocity (PGV)
3 Peak ground cm 72.5 39.0 52.0 73.9
displacement
(PGD)
4 PGV/PGA s 0.18 0.09 0.13 0.11
5 Acceleration root g 0.033 0.032 0.033 0.033
mean square
(RMS)
6 Velocity RMS cm/s 14.9 6.5 5.5 11.3
7 Displacement cm 39.7 16.8 26.9 41.1
RMS
8 Characteristic – 0.037 0.036 0.039 0.038
intensity
9 Cumulative cm/s 851.1 772.8 735.9 738.9
absolute
velocity
10 Acceleration g.s 0.180 0.156 0.188 0.186
spectrum
intensity (SI)
11 Velocity SI cm 83.0 74.2 88.0 70.9
12 Housner intensity cm 78.0 69.2 84.4 64.6
13 Sustained g 0.148 0.174 0.212 0.244
maximum
acceleration
14 Sustained cm/s 26.7 16.4 15.3 22.8
maximum
velocity
15 Effective design g 0.176 0.216 0.207 0.265
acceleration
16 A95 parameter g 0.172 0.227 0.211 0.259
17 Predominant s 0.48 0.20 0.34 0.20
period (Tp)
18 Mean period (Tm) s 0.59 0.46 0.48 0.40
19 Maximum cm/s 35.4 19.0 23.5 20.8
incremental
velocity
20 Number of – 4.58 6.38 5.99 4.63
effective cycles
21 𝐼𝐴 m/s 0.65 0.63 0.69 0.67
22 Time interval of s 23.14 17.88 15.12 15.98
95% of the 𝐼𝐴
(D5–95 )
23 Time at which 45% s 7.42 8.60 7.06 6.58
of the 𝐼𝐴 is
reached (tmid )
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 237

F I G U R E 1 3 The seismic parameters for 50 artificially generated accelerograms of the Kobe earthquake. The actual record of Kakogawa
station is shown for comparison with the generated accelerograms. For each seismic parameter, four values are normalized to the maximum
value

6 ARTIFICIAL GROUND MOTIONS and displacement time histories, where maximum values
FOR A GIVEN ACTUAL DATABASE are significantly different. Furthermore, the seismic
parameters are computed for 50 artificially generated
In this section, an actual database including eight differ- accelerograms randomly selected from the numerous
ent earthquakes is considered. For this purpose, one of the accelerograms generated by the proposed DCGAN for
existing records of each earthquake is randomly selected the Kobe earthquake. The actual record of the Kakogawa
from the NGA database. The selected earthquakes are station is compared to the generated accelerograms in
summarized in Table 2. For the sake of brevity, only one percentile terms (see Figure 13). Obviously, various actual
set of ground motions (i.e., Kobe, Japan) is discussed in seismic parameters are at different percentiles of the
this section to examine the performance of the proposed generated ones, while all generated accelerograms are
DCGAN. compatible with the design spectrum of ASCE 7–16.
The recorded motions correspond to Kakogawa sta-
tion with a source-to-site distance of 22.5 km. Figure 11a
shows three generated accelerograms randomly selected 7 CONCLUSION
from numerous accelerograms generated by the proposed
DCGAN. The wavelet algorithm (Hancock et al., 2006) has The limited number of available accelerograms is a
been used to adjust the actual record to match the design serious challenge to obtain reliable results with DL
spectrum of ASCE 7–16. Figure 11b,c show the correspond- algorithms. To overcome this barrier, a GAN has been
ing velocity and displacement time histories, respectively. presented in this paper to train generative models. Arti-
Figure 12 compares the 𝐼𝐴 for the generated time histo- ficial ground accelerograms that are compatible with a
ries with that of the actual record. In addition, Table 3 lists target response spectrum are usually generated for use in
the set of seismic parameters (Cosenza & Manfredi, 2000; nonlinear dynamic analyses of structures. In this paper,
D’Angela et al., 2021) for each generated ground motion. an efficient algorithm has been proposed, in which the
The seismic parameters were introduced in Elenas and well-established generative model of DCGAN was adopted
Meskouris (2001), Rezaeian and Der Kiureghian (2010), and used. Using this algorithm, a large number of artificial
and Cao and Ronagh (2014). spectrum-compatible accelerograms can be efficiently
As observed from Figures 11 and 12 as well as Table 3, generated using a limited number of available ground
the generated ground motions show rational seismic motion records. The following conclusions were drawn:
parameters. In other words, although the general features
of the generated ground motions are similar to those of 1. DCGAN is a stable GAN model that can generate
the actual records, they are meaningfully different in their reasonable artificial ground accelerograms even when
characteristics. This is specifically the case for the velocity using only four seismic records as input to train the
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
238 MATINFAR et al.

network. The use of additional seismic records as input Arias, A. (1970). A measure of earthquake intensity. In R. J. Hansen
would enhance the performance of DCGAN. (Ed.), Seismic design for nuclear power plants (pp. 438–483). MIT
2. The proposed algorithm is robust enough and not gen- Press.
Azimi, M., & Pekcan, G. (2020). Structural health monitoring
erally restricted to any specific assumptions regarding
using extremely compressed data through deep transfer learning.
the number of seismic records as input to train the net-
Computer-Aided Civil and Infrastructure Engineering, 35(6), 597–
work, the duration of the records, and time steps. 614.
3. Visual inspection of the output results indicated that Boore, D. (2003). Simulation of ground motion using the stochastic
the proposed DCGAN was able to generate numerous method. Pure and Applied Geophysics, 160, 635–676.
spectrum-compatible and realistic-looking accelero- Bousmalis, K., Silberman, N., Dohan, D., Erhan, D., & Krishnan, D.
grams. (2017). Unsupervised pixel-level domain adaptation with genera-
4. An inspection of the generated artificial response spec- tive adversarial networks. Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, Honolulu, HI (pp. 3722–
tra revealed that the response spectra curves obtained
3731).
from the generated accelerograms were in good agree- Cacciola, P., & Zentner, I. (2012). Generation of response spectrum
ment with the target design spectrum. compatible artificial earthquake accelerograms with random joint
5. In addition to visual inspection for evaluating the good- time–frequency distributions. Probabilistic Engineering Mechan-
ness of the generated artificial time histories, vari- ics, 28, 52–58.
ous seismic parameters calculated from the generated Cao, V. V., & Ronagh, H. R. (2014). Correlation between seismic
accelerograms were considered metrics for evaluating parameters of far-fault motions and damage indices of low-rise
reinforced concrete frames. Soil Dynamics and Earthquake Engi-
the proposed DCGAN performance. The results showed
neering, 66, 102–112.
the efficiency and robustness of the proposed DCGAN
Cha, Y. -J., Choi, W., & Büyüköztürk, O. (2017). Deep learning-based
by comparing various seismic parameters of the gener- crack damage detection using convolutional neural networks.
ated time histories with those of the actual earthquakes. Computer-Aided Civil and Infrastructure Engineering, 32(5), 361–
378.
AC K N OW L E D G M E N T Chen, S., Dong, J., Ha, P. Y. J., Li, Y., & Labi, S. (2021). Graph neu-
The authors are thankful to seven anonymous review- ral network and reinforcement learning for multiagent coopera-
ers for their valuable and constructive comments, which tive control of connected autonomous vehicles. Computer-Aided
Civil and Infrastructure Engineering, 36(7), 838–857.
allowed them to significantly improve the quality of the
Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., &
initial manuscript.
Abbeel, P. (2016). InfoGAN: Interpretable representation learning
by information maximizing generative adversarial nets. Proceed-
REFERENCES ings of the 30th International Conference on Advances in Neu-
Ahmadi, G. (1979). Generation of artificial time-histories compatible ral Information Processing Systems, Barcelona, Spain (pp. 2180–
with given response spectra – a review. Solid Mechanics Archives, 2188).
4(3), 207–239. Cheng, C. S., Behzadan, A. H., & Noshadravan, A. (2021). Deep
American Society of Civil Engineers (2017). Minimum design learning for post-hurricane aerial damage assessment of buildings.
loads and associated criteria for buildings and other structures, Computer-Aided Civil and Infrastructure Engineering, 36(6), 695–
ASCE Standard, ASCE/SEI 7–16. American Society of Civil 710.
Engineers. Conte, J., & Peng, B. (1997). Fully nonstationary analytical earthquake
Amezquita-Sanchez, J. P., Park, H. S., & Adeli, H. (2017). A novel ground-motion model. Journal of Engineering Mechanics, 123, 15–
methodology for modal parameters identification of large smart 24.
structures using MUSIC, empirical wavelet transform, and Hilbert Conte, J., Pister, K., & Mahin, S. (1992). Nonstationary ARMA mod-
transform. Engineering Structures, 147, 148–159. eling of seismic motions. Soil Dynamics and Earthquake Engineer-
Amiri, G. G., Abdolahi Rad, A., Aghajari, S., & Khanmohamadi ing, 11, 411–426.
Hazaveh, N. (2012). Generation of near-field artificial ground Cosenza, E., & Manfredi, G. (2000). Damage indices and damage
motions compatible with median-predicted spectra using PSO- measures. Progress in Structural Engineering and Materials, 2(1),
based neural network and wavelet analysis. Computer-Aided Civil 50–59.
and Infrastructure Engineering, 27, 711–730. D’Angela, D., Magliulo, G., & Cosenza, E. (2021). Towards a reliable
Amiri, G. G., Khorasani, M., Aghajari, S., & Tabrizian, Z. (2014). seismic assessment of rocking components. Engineering Struc-
Assessment of ANFIS networks on wavelet packet levels in gen- tures, 230, 111673.
erating artificial accelerograms. Journal of Intelligent & Fuzzy Sys- Denton, E., Gross, S., & Fergus, R. (2016). Semi-supervised learning
tems, 26, 379–391. with context-conditional generative adversarial networks. arXiv
Amiri, G. G., Rad, A. A., & Hazaveh, N. K. (2014). Wavelet-based preprint arXiv:1611.06430.
method for generating nonstationary artificial pulse-like near- Donahue, J., Darrall, T., & Krähenbühl, P. (2017). Adversarial feature
fault ground motions. Computer-Aided Civil and Infrastructure learning. Proceedings of the 5th International Conference on Learn-
Engineering, 29, 758–770. ing Representations, ICLR 2017 - Conference Track Proceedings.
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
MATINFAR et al. 239

Dong, S., Yu, T., Farahmand, H., & Mostafavi, A. (2021). A hybrid deep Kanai, K. (1957). Seismic-empirical formula for the seismic character-
learning model for predictive flood warning and situation aware- istics of the ground. Bulletin of the Earthquake Research Institute,
ness using channel network sensors data. Computer-Aided Civil Japan, 35, 309–325.
and Infrastructure Engineering, 36(4), 402–420. Kingma, D. P., & Ba, J. L. (2015). Adam: A method for stochastic opti-
Elenas, A., & Meskouris, K. (2001). Correlation study between seis- mization. Proceedings of the 3rd International Conference on Learn-
mic acceleration parameters and damage indices of structures. ing Representations, ICLR 2015 - Conference Track Proceedings.
Engineering Structures, 23(6), 698–704. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A.,
Fan, F. G., & Ahmadi, G. (1990). Nonstationary Kanai-Tajimi models Acosta, A., Aitken, A., Tejani, A., Totz, J., & Wang, Z. (2017).
for the El Centro 1940 and Mexico City 1985 earthquake. Proba- Photo-realistic single image super-resolution using a generative
bilistic Engineering Mechanics, 5, 171–181. adversarial network. Proceedings of the IEEE Conference on Com-
Gao, Y., Kong, B., & Mosalam, K. M. (2019). Deep leaf-bootstrapping puter Vision and Pattern Recognition, Honolulu, HI (pp. 4681–
generative adversarial network for structural image data 4690).
augmentation. Computer-Aided Civil and Infrastructure Engineer- Lee, S. C., & Han, S. W. (2002). Neural-network-based models for gen-
ing, 34, 755–773. erating artificial earthquakes and response spectra. Computers &
Gao, Y., & Mosalam, K. M. (2018). Deep transfer learning for image- Structures, 80, 1627–1638.
based structural damage recognition. Computer-Aided Civil and Li, H., Yi, W. J., & Yuan, X. X. (2013). Fuzzy-valued intensity mea-
Infrastructure Engineering, 33(9), 748–768. sures for near-fault pulse-like ground motions. Computer-Aided
Gao, Y., Zhai, P., & Mosalam, K. M. (2021). Balanced generative adver- Civil and Infrastructure Engineering, 28, 780–795.
sarial network for damage assessment from low-data imbalanced- Li, Y., Ku, B., Zhang, S., Ahn, J.-K., & Ko, H. (2020). Seismic data aug-
class regime. Computer-Aided Civil and Infrastructure Engineer- mentation based on conditionalgenerative adversarial networks.
ing, 36(9), 1094–1113. Sensors, 20, 6850.
Gatti, F., & Clouteau, D. (2020). Towards blending physics-based Li, Z., Park, H. S., & Adeli, H. (2017). New method for modal
numerical simulations and seismic databases using generative identification and health monitoring of superhighrise building
adversarial network. Computer Methods in Applied Mechanics and structures using discretized synchrosqueezed wavelet and Hilbert
Engineering, 372, 113421. transforms. The Structural Design of Tall and Special Buildings,
Ghaboussi, J., & Lin, C. -C. J. (1998). New method of generating spec- 26(3), e1312.
trum compatible accelerograms using neural networks. Earth- Liao, W., Chen, X., Lu, X., Huang, Y., & Tian, Y. (2021). Deep trans-
quake Engineering and Structural Dynamics, 27, 377–396. fer learning and timefrequency characteristics-based identifica-
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., tion method for structuralseismic response. Frontiers in Built Envi-
Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial ronment, 7(10), 627058.
nets. Advances in Neural Information Processing Systems, 3, 2672– Lin, C. C. J., & Ghaboussi, J. (2001). Generating multiple spec-
2680. trum compatible accelerograms using stochastic neural networks.
Gulgec, N. S., Takáč, M., & Pakzad, S. N. (2020). Structural sensing Earthquake Engineering and Structural Dynamics, 30, 1021–1042.
with deep learning: strain estimation from acceleration data for Lin, Y., & Yong, Y. (1987). Evolutionary Kanai-Tajimi earth-
fatigue assessment. Computer-Aided Civil and Infrastructure Engi- quake models. Journal of Engineering Mechanics, 113, 1119–
neering, 35(12), 1349–1364. 1137.
Hancock, J., Watson-Lamprey, J., Abrahamson, N. A., Bommer, J. Maeda, H., Kashiyama, T., Sekimoto, Y., Seto, T., & Omata, H.
J., Markatis, A., McCoy, E., & Mendis, R. (2006). An improved (2021). Generative adversarial network for road damage detection.
method of matching response spectra of recorded earthquake Computer-Aided Civil and Infrastructure Engineering, 36(1), 47–60.
ground motion using wavelets. Journal of Earthquake Engineering, Mobarakeh, A. A., Rofooei, F. R., & Ahmadi, G. (2002). Simulation of
10, 67–89. earthquake records using time-varying ARMA (2,1) model. Proba-
Heidari, P. S., & Khorasani, M. (2012). Generation of artificial earth- bilistic Engineering Mechanics, 17(1), 15–34.
quake accelerogram compatible with spectrum using the wavelet Nabian, M. A., & Meidani, H. (2018). Deep learning for accelerated
packet transform and nero-fuzzy networks. International Journal seismic reliability analysis of transportation networks. Computer-
of Geological and Environmental Engineering, 6, 566–569. Aided Civil and Infrastructure Engineering, 33(6), 443–458.
Hsieh, Y. A., Yang, Z., & Tsai, Y. C. (2021). Convolutional neural net- Naeim, F., Alimoradi, A., & Pezeshk, S. (2004). Selection and scaling
work for automated classification of jointed plain concrete pave- of ground motion time histories for structural design using genetic
ment conditions. Computer-Aided Civil and Infrastructure Engi- algorithms. Earthquake Spectra, 20, 413–426.
neering, 36(11), 1382–1397. NEHRP Consultants Joint Venture (2011). Selecting and scaling earth-
Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep quake ground motions for performing response history analysis.
network training by reducing internal covariate shift. Proceedings NIST/GCR Report No. 11- 917-15, NIST.
of the 32nd International Conference on Machine Learning, ICML Oh, B. K., Kim, K. J., Kim, Y., Park, H. S., & Adeli, H. (2017). Evolu-
2015, 1, (pp. 448–456). tionary learning based sustainable strain sensing model for struc-
Isola, P., Zhu, J. -Y., Zhou, T., & Efros, A. A. (2017). Image-to-image tural health monitoring of high-rise buildings. Applied Soft Com-
translation with conditional adversarial networks. Proceedings of puting, 58, 576–585.
the IEEE Conference on Computer Vision and Pattern Recognition, Park, S. W., Park, H. S., Oh, B. K., & Choi, S. W. (2018). Fragility assess-
Honolulu, HI (pp. 1125–1134). ment model of building structures using characteristics of artificial
 14678667, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/mice.12852 by Mcgill University, Wiley Online Library on [11/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
240 MATINFAR et al.

aftershock motions. Computer-Aided Civil and Infrastructure Engi- Song, X., Tian, J., Tao, P., Li, H., & Wu, C. (2020). Traffic state estima-
neering, 33, 691–708. tion of bus line with sparse sampled data. IEEE Access, 8, 216127–
Perez-Ramirez, C. A., Amezquita-Sanchez, J. P., Valtierra-Rodriguez, 216140.
M., Adeli, H., Dominguez-Gonzalez, A., Romero-Troncoso, R. J., & Tajimi, H. (1960). A statistical method of determining the maximum
Osornio-Rios, R. A. (2019). Recurrent neural network model with response of a building structure during an earthquake. Proceedings
Bayesian training and mutual information for response prediction of the 2nd World Conference on Earthquake Engineering, Tokyo,
of large buildings. Engineering Structures, 178, 603–615. Japan (pp. 781–798).
Radford, A., Metz, L., & Chintala, S. (2016). Unsupervised representa- Tang, Z., Chen, Z., Bao, Y., & Li, H. (2019). Convolutional neu-
tion learning with deep convolutional generative adversarial net- ral network-based data anomalydetection method using multiple
works. Proceedings of the 4thInternational Conference on Learning information for structural health monitoring. StructuralControl
Representations, ICLR 2016 – ConferenceTrack Proceedings. and Health Monitoring, 26(1), e2296.
Rafiei, M. H., & Adeli, H. (2016). A novel machine learning model Wang, F., Song, G., & Mo, Y. L. (2021). Shear loading detection
for estimation of sale prices of real estate units. Construction Engi- of through-bolts in bridge structures using a percussion-based
neering and Management, 142(2), 04015066. one-dimensional memory-augmented convolutional neural net-
Rafiei, M. H., & Adeli, H. (2017). NEEWS: A novel earthquake early work. Computer-Aided Civil and Infrastructure Engineering, 36(3),
warning system using neural dynamic classification and neu- 289–301.
ral dynamic optimization model. Soil Dynamics and Earthquake Wang, T., Zhang, Z., & Li, Y. (2019). EarthquakeGen: Earthquake
Engineering, 100, 417–427. generator using generative adversarial networks. In D. Bevc, & O.
Rafiei, M. H., & Adeli, H. (2018a). Novel machine learning model for Nedorub (Eds.) SEG technical program expanded abstracts 2019,
construction cost estimation taking into account economic vari- (pp. 2674–2678). Society of Exploration Geophysicists.
ables and indices. Journal of Construction Engineering and Man- Xu, B., Wang, N., Chen, T., & Li, M. (2015). Empirical evaluation
agement, 144(12), 04018106. of rectified activations in convolutional network. arXiv preprint
Rafiei, M. H., & Adeli, H. (2018b). A novel unsupervised deep learn- arXiv:1505.00853.
ing model for global and local health condition assessment of Yang, D., & Zhou, J. (2015). A stochastic model and synthesis for near-
structures. Engineering Structures, 156, 598–607. fault impulsive ground motions. Earthquake Engineering & Struc-
Rafiei, M. H., Khushefati, W. H., Demirboga, R., & Adeli, H. (2017). tural Dynamics, 44, 243–264.
Supervised deep restricted Boltzmann machine for estimation Yu, Y., Cai, C. S., & Liu, Y. (2021). Probabilistic vehicle weight esti-
of concrete compressive strength. ACI Materials Journal, 114(2), mation using physics-constrained generative adversarial network.
237–244. Computer-Aided Civil and Infrastructure Engineering, 36(6), 781–
Rezaeian, S., & Der Kiureghian, A. (2010). Simulation of synthetic 799.
ground motions for specified earthquake and site characteris- Zhang, K., He, Z., Zheng, L., Zhao, L., & Wu, L. (2021). A generative
tics. Earthquake Engineering and Structural Dynamics, 39, 1155– adversarial network for travel times imputation using trajectory
1180. data. Computer-Aided Civil and Infrastructure Engineering, 36(2),
Rofooei, F. R., Mobarake, A., & Ahmadi, G. (2001). Generation of 197–212.
artificial earthquake records with a nonstationary Kanai-Tajimi Zhou, Z., & Adeli, H. (2003). Wavelet energy spectrum for time-
model. Engineering Structures, 23(7), 827–837. frequency localization of earthquake energy. International Journal
Sajedi, S. O., & Liang, X. (2021). Uncertainty-assisted deep vision of Imaging Systems and Technology, 13(2), 133–140.
structural health monitoring. Computer-Aided Civil and Infras-
tructure Engineering, 36(2), 126–142.
Shahjouei, A., & Ghodrati Amiri, G. (2011). Appraisal of the
evolutionary-based methodologies in generation of artificial earth- How to cite this article: Matinfar, M., Khaji, N.,
quake time histories. Computational Methods in Civil Engineering, & Ahmadi, G. (2023). Deep convolutional
2, 145–152. generative adversarial networks for the generation
Shen, J., Xiong, X., Li, Y., He, W., Li, P., & Zheng, X. (2021). Detect-
of numerous artificial spectrum-compatible
ing safety helmet wearing on construction sites with bounding-box
earthquake accelerograms using a limited number
regression and deep transfer learning. Computer-Aided Civil and
Infrastructure Engineering, 36(2), 180–196. of ground motion records. Computer-Aided Civil
Sirca, G. F. Jr., & Adeli, H. (2004). A neural network-wavelet and Infrastructure Engineering, 38, 225–240.
model for generating artificial accelerograms. International Jour- https://doi.org/10.1111/mice.12852
nal of Wavelets, Multiresolution and Information Processing, 2,
217–235.