Downloaded from https://iranpaper.

ir
https://www.tarjomano.com https://www.tarjomano.com

State-of-the-Art Review

Machine-Learning Applications in Structural

Response Prediction: A Review
Aref Afshar 1; Gholamreza Nouri 2; Shahin Ghazvineh 3; and
Seyed Hossein Hosseini Lavassani 4

Abstract: Structural health monitoring (SHM) is an important and practical procedure for ensuring the structural integrity and serviceability
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

of civil engineering structures such as bridges, buildings, and dams. Model-driven or data-driven strategies for structural response prediction
are now widely combined with advances in engineering for use in SHM applications. Engineers have recently demonstrated increasing
interest in using machine learning (ML) and artificial intelligence (AI) to achieve a variety of benefits and possibilities, notably for predicting
structural reactions. This study serves as a comprehensive overview of the use of ML applications for structural response prediction in the
context of SHM for civil engineering structures, with a particular focus on ML, deep learning (DL), and meta-heuristic algorithms. Accord-
ingly, this study summarizes existing knowledge, presents concepts in a simple way, highlights trends, provides methodological insights, and
provides a valuable resource for researchers, stakeholders, and decision-makers to benefit from. It is observed that the use of ML, DL, and
meta-heuristic algorithms to predict the response of civil engineering structures within an acceptable accuracy range can be employed for
SHM, resulting in improved speed, efficiency, and accuracy compared to conventional approaches. DOI: 10.1061/PPSCFX.SCENG-1292.
© 2024 American Society of Civil Engineers.
Author keywords: Civil engineering structures; Structural response prediction; Machine learning (ML); Deep learning (DL); Meta-heuristic
algorithms; Structural health monitoring (SHM).

Introduction sensing technologies, system identification, data mining, and con-

dition assessment (Brownjohn et al. 2011; Doroudi et al. 2022;
Bridges, buildings, dams and pipelines are instances of civil engi- Gharehbaghi et al. 2021). SHM is a practical procedure for guar-
neering structures. They are considered the backbones of cities and anteeing the integrity and serviceability of civil structures and infra-
signify a country’s level of progress and economic might and even structure (Ghazvineh et al. 2021; Limongelli 2020). The process of
public welfare. An investigation of the design, monitoring and aging, material degradation, and unpredictably high excitation
maintenance of these assets as well as conserving them not only loads all pose risks to the health and safety of such structural sys-
protects human lives but also can advance national development tems that can result in irreversible damage, failure, and collapse.
(Feng and Feng 2018; Thacker et al. 2019). For this reason, design Model-driven or data-driven approaches are commonly used for
codes are generally updated yearly and researchers supply analysis structural response prediction and for monitoring structural health and
methodologies for different types of structures (Bakht and Mufti both have benefits and drawbacks (Entezami et al. 2020). Artificial
2018). Structural behavior prediction demands updating and alter- intelligence (AI) and machine learning (ML) applications have been
ation of structural regulations and technical standards. Over the last developed to give civil engineers a range of benefits and opportunities
decade, as the number of engineering data sets for structural parts (Manzoor et al. 2021; Spencer et al. 2019). AI is concerned with the
and civil infrastructures has grown, a shortcut for improving engi- research, design and implementation of intelligent computers. Tradi-
neering knowledge that had previously been based on earlier input tional methods of modeling and optimizing large structure systems
data sets has arisen. necessitate a large amount of processing power and AI-based solu-
Structural health monitoring (SHM) is a relatively new strategy tions can provide opportunities for quickly addressing civil engineer-
that has been adopted in academic and practical research that uses ing challenges (Lu et al. 2012). AI and ML have transformed civil
1
structure health monitoring, outperforming traditional methods like
Faculty of Engineering, Kharazmi Univ., Tehran 15719-14911, Iran. finite element (FE) analysis and manual inspections. They excel at
ORCID: https://orcid.org/0000-0003-4937-0953. Email: std_ArefAfshar@
data-driven insights, enhancing accuracy in structural health assess-
khu.ac.ir
2
Associate Professor, Faculty of Engineering, Kharazmi Univ., Tehran
ments, and predicting responses to various conditions. Advantages
15719-14911, Iran (corresponding author). ORCID: https://orcid.org/0000 include real-time monitoring, nonintrusive testing, adaptability for
-0002-2936-599X. Email: [email protected] anomaly detection, and reduced downtime. These techniques opti-
3
Faculty of Engineering, Kharazmi Univ., Tehran 15719-14911, Iran. mize efficiency, automate monitoring, and integrate with internet
ORCID: https://orcid.org/0000-0002-2714-1300. Email: std_Shahinqazvine@ of things (IoT) for early warnings (Zinno et al. 2022).
khu.ac.ir Since its inception in the 1940s, ML has been used in real-
4
Associate Professor, Faculty of Engineering, Kharazmi Univ., Tehran world applications such as medical diagnosis, speech recognition,
15719-14911, Iran. Email: [email protected]
weather forecasting, and autonomous vehicles. As a subtype of AI,
Note. This manuscript was published online on March 18, 2024.
Discussion period open until August 18, 2024; separate discussions ML has shown promise for data-driven challenges with the focus
must be submitted for individual papers. This paper is part of the on development of prediction algorithms and to support decision-
Practice Periodical on Structural Design and Construction, © ASCE, making in uncertain environments (Thai 2022). It is now possible to
ISSN 1084-0680. complete larger computations in seconds and obtain improved

© ASCE 03124002-1 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

AI Response Structure
accuracy with big data due to the supercomputing capacities supplied
by rapid breakthroughs in processor units (Adeli 2020). Hence, ML Or Or Or
is becoming more popular in engineering applications day-by-day. Machine Learning And Prediction And Infrastructure
In ML problems, three approaches are based on the training pro-
cess: supervised learning, unsupervised learning and reinforcement Or Or Or

learning (RL). The first is the most frequent type of algorithm in Deep Learning Forecast Building
which training is done using lab data. In engineering applications,
supervised ML is commonly used to solve regression and classifica- Fig. 1. Keywords used for data mining.
tion problems. In unsupervised learning, unlabeled data can be used
for clustering, compression, and feature extraction. RL allows an
agent to learn through interactive trial-and-error which incorporates
feedback from its own actions and experiences (Lee et al. 2018). visualizing it using VOS software (Van Eck and Waltman 2010).
Hung and Adeli presented the first use of ML in structural en- This software maps current knowledge and its advancement in the
given field by using large academic data sets to measure the re-
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

gineering in the 1980s by implementing artificial neural networks

(ANN) to construct steel beams (Hung and Adeli 1994). This was search impact and citation processes. Scopus was chosen as the
followed by Vanluchene and Sun, who were interested in perfor- literature database because it covers a wide range of topics in
mance evaluation and employed backpropagation (BP) neural net- construction-related research and contains a larger number of
works to locate loads on beams, design concrete beams and for journal publications (Martinez et al. 2019).
analysis of simple support beams (Vanluchene and Sun 1990). The Scopus keyword search was configured to return any pub-
ANN is now widely employed for structural applications such as lications that contained the specified keywords in the title, abstract
analysis and design, optimization problems, response prediction, or selected keywords. The keywords chosen are shown in Fig. 1.
crack detection, reliability problems and SHM. A lack of data sets The research focused on current literature and was based on the
for training presents an obstacle to the validation of ML methods; development history of AI applications in the field of structural re-
thus, numerous research and engineering contributions have been sponse prediction. The data acquired were monitored and irrelevant
made to extending the database platform through libraries such as items were removed, such as studies in the field of medicine that
Kaggle, Mendeley, and DataCenterHub (Thai 2022). mentioned the terms structure, deep learning, and machine learn-
Support vector machines (SVM) were introduced in 1992 as a ing. The third criterion was language. This study only looked at
modeling tool that can be used to predict test results. Empirical data English-language publications. Non-peer-reviewed studies such
modeling is useful in structural engineering because it enables re- as book reviews or editorials also were excluded.
searchers and engineers to predict test results (Çevik et al. 2015). Because this study focused on structure-based investigations,
Deep learning (DL) dominated the scientific world with a break- irrelevant articles subsequently were filtered out based on their
through in 2006 for training multilayer neural networks and hereti- content. Those that passed screening were considered for the
cal feature extraction with the introduction of the convolution bibliometric analysis. The initial search yielded over 37,000 docu-
neural network (CNN). Deep neural networks (DNN) utilize GPU ments. After removal of articles that did not comply with the years
computing architectures and, as the amount of data for training has of publication, subject area and language, only 1,769 articles re-
increased, deep networks are able to achieve greater accuracy in mained. However, for the purposes of this review, we chose a
complicated issues than standard ML techniques. Recurrent neural representative selection of 97 papers that best addressed the subject.
networks (RNNs) and long-term short-term memory (LSTM) are These 97 papers were picked for their relevance, quality, and
the most often utilized models for prediction applications. impact on this topic, and they served as references to support the
Boosting techniques enable ML methods to enhance perfor- narrative of the review. Fig. 2 shows the research methodology
mance and speed up algorithms by reinforcing the learners, which workflow.
is especially important with big data sets. The first step in such Fig. 3 shows the countries from which most publications origi-
methods was the boosting algorithm, AdaBoost, in 1996. Other nated. China and the United States of America contributed the
methods include extreme gradient boosting (XGBoost) and cat- most, accounting for over 55% of the articles examined in this
egorical gradient boosting (CatBoost). Most studies on response study focusing on articles published between 2011 and 2023. Fig. 4
prediction use a combination of ML and behavioral analysis shows an increasing trend, soaring from 2019, when the number of
(Damirchilo et al. 2021; Osman et al. 2021; Xu et al. 2022). papers on this subject nearly doubled.
The current work presents an overview of the use of ML for Keyword co-occurrence analysis was used to link the knowl-
structural response prediction with a focus on civil engineering edge domains between structural response prediction and AI, as
structures where the most recent works from 2011 to 2023 were keywords comprise the core of each academic publication. The
reviewed. The rest of the paper is organized as follows. The next VOS was used to create a map based on distance, where distance
section discusses bibliometric data mining on recent papers to find indicates the strength of the link between two fields (Oraee et al.
the main contribution of ML for responsive prediction. The section 2017). Small gaps between knowledge domains indicates stronger
on computer-aided response prediction methods reviews the three links. The size of the object tag is determined as the number of
main areas for response prediction and highlights the pros and cons publications in which the keyword has been identified. For author-
of each method. The conclusion summarizes the main findings and key analysis, the minimum number of co-occurrences was set to 15.
key factors from the literature review to pave the ground for future In addition, using 20, 30 and 50 as the criteria, two index-key and
research in this field. all-key analyses were performed twice.
Fig. 5 shows that the most commonly used keywords are general
keywords for machine learning, and deep learning. Additionally,
Bibliometric Data Mining structural health monitoring, prediction, and forecasting are the
domain keywords mostly used by authors. Figs. 6 and 7 show
The first stage in the review process was to conduct a thorough that the Scopus auto-indexing algorithm has identified the term
search of academic articles relevant to the field of study and forecasting as the main keyword for the literature. Moreover,

© ASCE 03124002-2 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

while the term machine learning (machine learning and machine- all keywords, including author and index keywords, forecasting
learning combined) occurred more frequently than deep learning, was the term with which the majority of other keywords were
learning systems along with learning algorithms were also fre- linked. Furthermore, deep learning had one of the strongest links
quently used. In general, as seen in Figs. 8 and 9, after considering in the literature, indicating its importance for the area of study.
Analysis of the data mining results found that the most common
AI approaches for response prediction could be categorized into
workflow the groups of machine learning, deep learning, and metaheuristic
algorithms. These categories thus were used to classify every paper
Data Collection

in this study.
The initial search from
37,000 documents
the Scopus dataset
Computer-Aided Response Prediction Methods
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Machine Learning
Filter irrelevant 35,231 of 37,000 Because civil structures and infrastructures play such an important
documents documents were filtered role in the economy and in facilitating daily living for the global
population, AI and ML are currently employed to monitor the
Screening

Screened documents 1,769 documents health of civil engineering structures (Das et al. 2016; Flah et al.
2021). ML is commonly used for scalable urban infrastructure sta-
Excluded documents
1,672 of 1,769 documents tus assessment, which uses data sources for prediction and assess-
were excluded
ment (Alipour and Harris 2020; Cosgun 2023; Doroudi et al. 2023).
The capacity to forecast is one of the most useful aspects of ML.
Remaining documents 97 documents
For example, ML can predict cable forces, which are critical for
SHM of long-span cable-supported bridges (Dong et al. 2018;
Li and Ou 2016). Another significant aspect of SHM on bridges,
Included

Documents included in
structures, and infrastructure is the measurement and forecast of
97 documents temperature to help to avert damage. Research has shown that
the review
temperature prediction is possible using ML techniques (Han et al.
2021).
ANN, SVM, SVR, GPR, Tadesse et al. used neural networks (NNs) to predict the mid-
ML MLR, LR, MLP, RF, DT,
span deflections of composite bridges with flexible shear connec-
Infrastructure XGBoost, GTB, GBR,
tions under service loading (Tadesse et al. 2012). The three NNs
Category

CNN, DNN, RNN, LSTM, comprised those for simply supported bridges, two-span continu-
Building DL GA, ConvLSTM, PhyCNN, ous bridges and three-span continuous bridges. Their NNs were
AttLSTM, DynNet
multilayered feed-forward networks with neurons that were totally
Others PSO, WOA, DQN, connected in the feed-forward direction in all layers. Closed-form
Metaheuristic
TLBO, HHO, ABC, solutions also were provided on the basis of the constructed NNs.
These NNs took cracking and shear lag in concrete slabs into con-
sideration. The results showed that the NNs could predict midspan
Fig. 2. Research methodology workflow.
deflections and that the errors were small.

China

United States

India

United Kingdom

South Korea

Canada

Australia

Italy

Germany

Saudi Arabia

0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750
Documents

Fig. 3. Countries from which most publications originated.

© ASCE 03124002-3 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

600

500

400
Documents

300
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

200

100

0
2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023
Year

Fig. 4. Country contributions from 2011 to 2023.

Fig. 5. Author-keyword co-occurrence network with a threshold of 15.

© ASCE 03124002-4 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 6. Index-keyword co-occurrence network with a threshold of 20.

Fig. 7. Index-keyword co-occurrence network with a threshold of 50.

© ASCE 03124002-5 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 8. All-keyword co-occurrence network with a threshold of 30.

The interactions of wind, waves and narrow coastal bridges can such as ground slope and fines content, and the particle size of
be compounded by the transient, nonstationary properties of high liquefiable sediment (Xie et al. 2020). ML techniques for pro-
winds and large waves due to the nonlinear nature of the structural babilistic seismic response prediction and variable relevance inves-
system and fluid–structure interactions (Zhu and Zhang 2016). In tigations of liquefaction-induced lateral spreading in extended
the design of long-span bridges, the prediction of their stochastic pile-shaft-supported bridges were investigated by Wang et al.
structural response under strong winds and waves is crucial, but the (2021b). They considered a number of factors, including soil, struc-
nonstationary characteristics of wind and waves may impair the tural and ground-motion characteristics. Multilinear regression
accuracy (Petersen et al. 2019). (MLR), lasso regression, NN, random forests (RF), and gradient
Fang et al. offered an efficient and reliable methodology for tree boosting (GTB) are five popular ML approaches that were used
computing the stochastic response of a bridge under nonstationary by them for response forecasting of bridges (peak bearing defor-
wind and wave fields using a wind-wave-bridge system (Fang et al. mation and column drift ratio). The results demonstrated that
2020). To save processing expenses, the ML techniques of support NN and GTB, followed by RF, were able to accurately forecast
vector regression (SVR), BP neural network (BPNN), and Gaus- the seismic reactions of the soil bridge systems tested, but Lasso
sian process regression (GPR) were employed to link the stochastic regression and MLR were unable to produce valid estimations.
inputs of the environmental variables to the single output of bridge The regression models generated also were used to perform a
response. Their study employed a cable-stayed bridge to demon- variable importance analysis. The results suggested that intensity
strate the flexibility of the three models as well as the viability of measures (IMs) are more important than soil and structure-related
the suggested framework. All three models were able to represent variables.
the nonlinear response of the bridge to wind and waves with the ML is currently used to forecast bridge reactions. Examples of
lowest coefficients of determination. The SVR model for forecast- this are studies by Kumar et al. on rapid prediction of long-term
ing tower displacement, the BPNN model for predicting internal deflections in steel-concrete composite bridges using a NN model
foundation force, and the GPR model for predicting mid-span dis- and Lim and Chi, who focused on damage estimation using
placement could all be employed to improve computation accuracy. XGBoost (Kumar et al. 2021; Lim and Chi 2019). And Abdu et al.’s
All three of the strongest models had good accuracy. work, along with the utilization of train vertical acceleration, a
Predicting the amount of lateral displacement caused by soil highly sensitive variable, in conjunction with ML models like
liquefaction is a difficult problem. A multitude of parameters influ- the gradient boost regressor (GBR), enables accurate and effective
ence liquefaction-induced lateral spread, including earthquake prediction of bridge pier settlement for high-speed railways (Abdu
magnitude, fault-to-site distance, local soil profile characteristics et al. 2023).

© ASCE 03124002-6 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 9. All-keyword co-occurrence network with a threshold of 50.

Castellon et al. examined wind-induced dynamic reactions and South Korea. The research findings indicate that augmenting avail-
demonstrated that analytical and ML models can be used to reliably able data leads to more accurate predictions. However, it is impor-
estimate buffeting responses (Castellon et al. 2021). Full-scale tant to note that the study acknowledges three limitations: the
observations and simulated data based on a probabilistic model necessity for manual data preprocessing grounded in engineering
of observed wind-f data set from Hardanger Bridge. The normal- expertise, the requirement for minimal user involvement in identi-
ized root mean square error (RMSE) ranged from 1.46 × 10−4 to fying VIV states, and the need to develop separate ML models for
7.21 × 10−3. The ML models trained on the synthetic data sets individual bridges. The paper also provides a roadmap for future
achieved extraordinarily high accuracy, making them suitable for research to surmount these limitations, including automating data
efficient surrogate modeling. The use of ML models on synthetic
data sets [i.e., SVR and multilayer perceptron (MLP) models] were
reasonably accurate. In terms of response prediction, the SVR Hidden
Layer
model beat the MLP model in both simulated and full-scale experi-
ments. MLP is a type of NNs widely used in the field due to its Input
straightforward architecture and powerful learning capabilities. It Layer
has three layers including an input layer, a hidden layer and the
output layer. Fig. 10 demonstrates a typical MLP with multiple in- Input 1
put. Furthermore, the analytical response estimations were more Output
accurate than the multimodal approaches. It was concluded that Layer
models trained with full-scale data sets were less accurate than Input 2
models trained using synthetic data sets.
Kim and Kim conducted a research project centered on the pre-
diction of vortex-induced vibrations (VIVs) in long-span bridges
with the application of GPR (Kim and Kim 2022). Their work
underscores the difficulties in crafting robust ML models for VIV
prediction, primarily due to constraints linked to limited and un- Input n
even data sets. The study introduces a comprehensive framework
encompassing five distinct ML strategies tailored to forecast VIVs,
all while addressing the challenge of working with constrained
data. To validate their approach, the researchers utilized real mon-
Fig. 10. Typical MLP algorithm.
itoring data from the Jindo Bridge, a cable-stayed bridge situated in

© ASCE 03124002-7 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

preprocessing and crafting universal prediction models. An impor- ratio, concrete compressive strength, reflected impulse, reinforcing
tant takeaway from this study is that ML models designed for one steel yield strength, blast scaled distance, slab support, and slab
bridge may not work well for others. type. The data set for this investigation contained 150 data points
Gharehbaghi et al. investigated the efficacy of a wavelet derived from studies found in the open literature. Permutation fea-
weighted least squares SVM (WWLSSVM) and a feed-forward, ture importance, a variable importance measure, was utilized to
backpropagation ANN to predict the inelastic seismic reactions study the effect of each input feature on the output and the effects
of structures (Gharehbaghi et al. 2020). The first three natural peri- of the features were compared to parametric studies available in the
ods of the frame, as well as their combinations, were used as inputs literature. The RF algorithm was used to create the learning model
to the model. Under most circumstances tested, the ANN model and its performance was compared to that of other learning
beat out the other models with a slight advantage in accuracy, es- algorithms (RF; regression plus regression tree; fine plus linear
pecially when a smaller number of samples were employed for SVM, quadratic SVM, cubic SVM, ensemble; bagged trees plus
training. This research looked at prediction of the maximum ensemble; boosted trees plus random forest; regression). A hybrid
cross-sectional rotation at the ends of the beams and columns, classification-regression RF technique was employed to create the
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

the maximum inelastic interstory drift ratio, the maximum inelastic final model. In another study, a hybrid ML model (Hybrid gradient-
displacement at the roof level and the maximum shear force of boosted regression trees algorithm coupled with Henry gas solubil-
stories in a sample reinforced concrete (RC) frame. The findings ity optimization algorithm) was employed to predict the maximum
revealed that both models yielded reasonable and virtually similar displacement of RC beams under blast loading (Almustafa and
seismic response estimates. Nehdi 2022). The evaluation incorporated various performance cri-
Tall-building collapse can be forecast using ML algorithms and teria, demonstrating the model’s effectiveness. Through the analy-
accurate IMs. For example, Bijelić et al. employed regularized lo- sis of feature importance, the researchers identified the influential
gistic regression as a primary technique for identifying variables factors in predicting the behavior of RC beams under blast loading,
that were important for collapse prediction (Bijelić et al. 2020). adding depth to their study. However, it appears that they could
The results of nonlinear response history assessments of a 20-story have simplified the model’s implementation for nonexperts by of-
tall building done using about two million simulated ground move- fering more practical guidance on its practical application.
ments were employed to that purpose. To examine the utility of The primary objective of Demertzis et al.’s study is to propose
different IMs for collapse prediction, the researchers used linear an interpretable ML method for predicting the maximum interstory
regression (LR) and SVMs. Various ground motion IMs were used drift ratio (MIDR) as the damage index for R/C buildings’ seismic
as input qualities, the output variable taking on values of 1 or 0 damage assessment (Demertzis et al. 2023). To achieve this, the
depending on whether or not ground motion-induced collapse has authors conduct an extensive comparative study of various ML al-
occurred. gorithms (LightGBM, GBR, RF, Extra Trees Regressor, k-Nearest
An easy-to-use response estimation method is critical, given the Neighbors Regressor, LR, Bayesian Ridge, Ridge Regression,
need for rapid damage assessment of high-rise buildings following Decision Tree (DT) Regressor, AdaBoost Regressor, Elastic Net,
an earthquake. Barkhordari and Es-haghi studied existing simple Lasso Regression, Orthogonal Matching Pursuit, Huber Regressor,
regression approaches as well as a hybrid technique, ANN and si- Least Angle Regression), employing a sophisticated data set that
mulated annealing (ANN-SA), for assessing structural responses includes 30 R/C buildings. These buildings exhibit varying struc-
(Barkhordari and Es-haghi 2021). OpenSees software was utilized tural parameters, encompassing factors like the number of stories,
to examine four concrete shear wall buildings (15, 20, 25, and structural eccentricity, and base shear ratios. Notably, both seismic
30 stories). A thorough database of inputs (record characteristics) and structural parameters were chosen as inputs for the ML meth-
and outputs was created using 150 seismic records (responses). The ods. This data set includes 3D R/C buildings designed according to
effect of nonlinearity in the available data was evaluated using the EC8 and EC2 standards, subjecting them to nonlinear time history
hybrid ANN-SA algorithm. The model made use of three-layer analyses with 65 real earthquake records. The highlights of the
structures (input, hidden layer and output layer). Predictors in- study include the utilization of historical data for realistic modeling,
cluded maximum acceleration, maximum velocity and earthquake the generalizability of the LightGBM (Light Gradient Boosting
characteristics. Various ML models are used for comparison. Machine) algorithm for multiparametric problems, and the use of
The sensitivity of the seismic demand model to input factors was the GOSS (Gradient One Side Sampling) technique to handle noisy
studied. data effectively. However, there are opportunities for improvement,
Almustafa and Nehdi studied the feasibility of building an ML such as optimizing hyperparameters, expanding the data set to
model to predict the behavior of retrofitted RC slabs subjected to include diverse structural characteristics and seismic records,
blast stress (Almustafa and Nehdi 2021). They looked at the viabil- and exploring data transformation techniques to enhance model
ity of utilizing ML to predict the maximum displacement of fiber- performance.
reinforced polymer (FRP) retrofitted RC slabs under blast loading Guan et al. investigated seismic narrative drift estimates for steel
as a unique way to obtain more accurate and simpler predictions. A moment resisting frames (SMRFs) (Guan et al. 2021). Based on
GPR approach was utilized to create the model and took into ac- previously described approaches, two models for prediction in
count a number of significant elements of the application. Because SMRFs were constructed. The hybrid model (ML-EMKY) is a
relevant data was scarce in the open literature, a novel approach combination of the statistically adjusted elastic multi-degree-of-
based on a tabular generative adversarial network was used to freedom (MDOF) with the known yield-strength (EMKY) model
generate 200 additional synthetic data sets for model training. that integrates mechanics and ML techniques, whereas the purely
Because of the danger of exposing structures to explosive ML data-driven (MLDD) model was developed to provide a direct
detonation, understanding the structural behavior at such times, link between building features and nonlinear story drift demand.
as well as improving blast resistance, is becoming increasingly vital Both are based on 240 earth motion data sets and a set of seismic
(Zhao et al. 2022a). Almustafa and Nehdi used ten (input) features structural responses from 621 current code-based SMRF designs.
to predict the maximum displacement (output) of RC slabs exposed In terms of accuracy, the assessment results showed that the hybrid
to blast loading (Almustafa and Nehdi 2020). These features in- model surpassed the exclusively data-driven or mechanics-based
clude the slab length, width and thickness, steel reinforcement models.

© ASCE 03124002-8 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

While ML predictions will be discussed in civil engineering geometric configurations. The principal seismic drift responses
structures in this article, it must be kept in mind that ML has differ- of the steel frames recorded the highest top and interstory drift val-
ent applications in different sectors of civil engineering and can be ues. Both the ANN and XGBoost models were found to be capable
used in a number of scenarios. ML techniques, for example, have of accurately predicting the seismic drift responses of the steel
been used to detect reaction mechanisms, estimate shear strength, frames. In terms of prediction, the XGBoost model outperformed
predict damage states, classify failure scenarios and estimate shear the ANN model in all scenarios studied. The effect of input param-
strength of beam-column junctions (Aguilar et al. 2016; Anvari eters on the prediction of seismic drift reactions was also investi-
et al. 2023; Mangalathu and Jeon 2018; Morfidis and Kostinakis gated in this work.
2017; Mousavi et al. 2022; Naser et al. 2021). Zhu et al. provided a method for calculating the nonlinear buck-
Junda et al.’s study focuses on the importance of accurately es- ling load of a reticulated shell using ML techniques such as ANN
timating the MIDR and the roof drift ratio in multistory cross- and SVR, which lowered the expense of running numerous non-
laminated timber (CLT) buildings subjected to seismic forces linear buckling evaluations for defective structures (Zhu et al.
(Junda et al. 2023). The study uses data from 69 different CLT 2021). They were able to validate the performance of the prediction
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

structures and identifies six key input features (the spectral accel- approach and a simpler method for building reticulated shells based
eration at the building’s fundamental period, the peak ground on the SVR and proposed the use of the linear kernel technique.
velocity, tuning ratio, behavior factor, wall height, and the wall ANN and SVR having radial basis function (RBF) kernels ap-
subdivision ratio) that significantly influence drift predictions. peared to be quite good at fitting data with nonlinear relationships
The ML models outperform traditional regression approaches, with and the reduction ratios were predicted with high accuracy when
RF showing the best performance. The research also reveals that used as ML techniques.
certain structural parameters have a greater impact on interstory Thaler et al. created a ground-accelerated structure reaction-data
drift than on roof drift. The article offers valuable insights into NN that effectively enhanced Monte Carlo simulation (Thaler et al.
the realm of seismic drift prediction for CLT structures utilizing 2021). They used a small feed-forward NN on the design which
ML methods. Nevertheless, there exist opportunities for enhance- required fewer training samples and, as a result, only a short train-
ment, such as tackling data constraints and applicability issues, ing period to produce accurate response prediction of peak tail drift
improving the clarity of model interpretation, and exploring the ratios. Six input factors were sufficient to provide correct reaction
wider ramifications of CLT building practices. statistics. These were chosen from a collection of realistic earth-
Nguyen et al. employed ML to forecast the seismic responses of quake intensity indicators. The main advantage of this strategy
planar steel moment-resisting frames subjected to ground vibra- was that it avoided the requirement for the NN to extrapolate pre-
tions using ANN and XGBoost, two of the most effective ML ap- dictions in the case of extremely infrequent events which would
proaches (Nguyen et al. 2021). XGBoost is a gradient boosting result in insufficient accuracy at the tail-end of the distribution.
algorithm that builds an ensemble of DT. It combines multiple The proposed technique had an unexpected side effect of decreas-
weak models to produce a stronger prediction. Fig. 11 shows a ing the accuracy of high probability situations. The traditional
typical XGBoost. The four input variables were ground motion method, which used a typical sample range, forecasted this region
intensity, earthquake and soil parameters, 5% critical-damped spec- more accurately. However, when calculating the likelihood of sys-
tral accelerations at the first three natural periods and structural temic breakdown, this trade-off was, at times, missed. In terms of

Input for modified Input for modified Input for modified

tree 1 tree 2 tree k
Iteration K
Iteration 2
Iteration 1

Sum

Predicted value

Fig. 11. Simplified structure of extreme boosting gradient.

© ASCE 03124002-9 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

efficiency, the novel strategy outperformed the traditional Monte determining scour depth was based on the link between cumulative
Carlo method and ensured an accurate prediction at the tail-end frequency and relative error.
of the distribution. This technique proved especially effective at In the field of structural engineering, the decision regarding
predicting events that occur near to structural failure. which prediction algorithms to employ, such as ANN, SVR, and
ML has the potential to predict concrete dam deformation to XGBoost, significantly influences the precision and efficiency of
facilitate SHM. The aging of dams as concrete structures is of sig- forecasts for structural responses. ANNs are renowned for their
nificant environmental concern as many older dams are in the pro- capacity to model intricate and nonlinear data relationships,
cess of deformation (Mata et al. 2021). Radovanovic et al. offered although they require substantial training data and can be seen
the nonlinear autoregressive exogenous (NARX) model, an ANN as complex and opaque models. In contrast, SVR excels at man-
designed to forecast dam displacement and strain and where the aging noisy data and offers controlled model complexity through
prediction variables are displayed as a time series of displacements kernel functions, although the fine-tuning of hyperparameters can
and strains on a daily basis (Radovanovic et al. 2015). It also pro- be intricate. Meanwhile, XGBoost is celebrated for its outstanding
vides a common MLR model for comparison. This research found accuracy, built-in regularization techniques, and the ability to as-
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

that, while modeling deformation in concrete dams, the resource- sess the importance of features, yet it may necessitate substantial
efficient NARX model outperformed the MLR models for real-time computational resources and meticulous hyperparameter tuning.
safety monitoring of gravity and arch dams, primarily because Engineers must thoughtfully consider these pros and cons when
NARX employs historical output and exogenous input variables. choosing an algorithm to ensure the most effective prediction of
Zhang et al. introduced a ML approach for predicting ground structural responses, aligning with the specific demands of their
settlement during shield tunneling that took into account the projects and available data resources.
interplay of variables such as geological conditions, construction
parameters, construction sequences and grouting volume and tim- Deep Learning
ing (Zhang et al. 2020a). In the geological condition inputs, the soil
Deep learning, the second approach, was reviewed for response
profile was represented by the thickness of the soil layer and karst
prediction. As previously noted, ML is the process of teaching a
treatment was represented by a dummy variable. A hybrid NN
computer to recognize hidden patterns in extracted characteristics
model for predicting shield tunneling-induced ground settlement
and targets for categorization or prediction (Lake et al. 2015). As a
was created in their study. The differential evolution approach
result, the success of such algorithms depends heavily on this stage.
was used to identify the best ANN design and hyper parameters.
As the amount of data expands, it becomes more difficult and time-
Adaptive moment estimation (Adam) was used to accelerate the consuming to identify the most important features for a complex
ANN training process. With the use of Adam, a differential evolu- problem. In the realm of AI, DL is a subclass of ML that can learn
tion technique was designed to handle a high number of ANN can- directly from raw data. DL is a representational learning technique
didates using a low number of computer resources. Comparison of that enables a network architecture to independently learn highly
the measured and estimated settlements revealed a high level of abstract features from raw data in order to execute recognition
agreement for both the training and testing sets. or classification tasks (Ye et al. 2019). DL provides excellent ac-
Ahmadianfar et al. used ML to predict the scour depth caused by curacy in packed data sets. DNN (Fig. 13), CNN, RNN and LSTM
waves in a sand bed around vertical piles using three data-driven algorithms are currently employed in various sectors. DL ap-
methods: locally weighted LR (LWLR), support SVR, and MLR proaches, are preferred for prediction, classification, and identifi-
(Ahmadianfar et al. 2021). As shown in Fig. 12, SVR finds the cation applications (Çakiroğlu and Süzen 2020; Ning et al. 2023).
optimum margins while fitting a hyperplane to the set of data. CNNs, shown in Fig. 14, are best known for their capability for
LWLR was employed for the first time to forecast scour depth pattern recognition, primarily used in the field of computer vision
around vertical piles. The Shields parameter, pile Reynolds num- (Ghazvineh et al. 2023). Oh and Kim provided an approach for
ber, grain Reynolds number, Keulegan-Carpenter number and developing an optimum architecture for the use of CNN to calculate
sediment number are commonly used in models to simulate scour
depth. The LWLR model outperformed previous ML models in
terms of accuracy and reliability and the empirical equation for Hidden Layers

Input
Layer

Input 1
Output
Layer
Input 2

Input n

Fig. 12. Functionality of support vector regression. Fig. 13. Typical deep neural network with multiple input.

© ASCE 03124002-10 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Fig. 14. Typical CNN algorithm.

structural responses (strain) for structural safety evaluations (Oh It should be noted that the 1D-CNN was able to forecast layer
and Kim 2021). They conducted a theoretical study on a beamlike drift in subway stations subjected to nonpulselike ground motion,
structure as well as an experimental examination of RC frame con- but not pulselike ground motion. Because of its prediction capabil-
struction which included a shaking table test to evaluate the effi- ity, the 1D-strong CNN can be considered as well suited for use
ciency of their proposed approach. Their findings showed that in stochastic seismic response analysis using probability density
the number of kernels, kernel size, and subsampling size that evolution technique, which is solved using the finite difference ap-
formed a CNN architecture with exceptional strain prediction per- proach. The proposed method for forecasting the nonlinear seismic
formance can be examined automatically using a genetic algorithm reaction of subway stations was computationally cost-effective, but
(GA) solution search. The optimum CNN design was established had drawbacks, as only a time-series of seismic events could be
using dynamic displacement and strain data obtained from the predicted. Other crucial structural reactions, such as strain and
structural study of the target structure and the CNN strain predic- stress distribution and possible local structural failures, were diffi-
tion performance was tested to validate the proposed approach. cult to determine. The distribution of strains (or stresses) could be
With the use of wind turbines, it is difficult to predict how differ- determined by forecasting the time-series of strain (or stress)
ent systems will react in a complicated wind environment. Woo responses for each element of the numerical model, but the com-
et al. proposed the stacked dilated convolutional LSTM (SDCL) putational cost was prohibitive because of the large number of
approach, a data-driven prediction model influenced by physics components.
that uses a sequence of wind fields as input to predict future wind Zhang et al. described a DL strategy for reaction modeling and
turbine responses (predicting energy generation and structural prediction of structural seismic responses (interstory drift) based on
bending moment) (Woo et al. 2019). Their proposed model repre- a LSTM RNN (Zhang et al. 2019). To describe the temporal
sents the spatial and temporal characteristics of wind flow using a dependence and causality of the input and output sequences, the
network architecture that employs a sequence of inductive biases proposed LSTM architecture comprised many LSTM layers and
inspired by turbulent dynamics. An SDCL is made up of a group of FC layers. Two LSTM methods with differing input/output formats,
dilated CNNs linked with a LSTM that capture the spatial and tem- LSTM-f and LSTM-s, were designed. The good performance of
poral development of the turbulence structure in the input wind the proposed LSTM-s was validated by testing it on a 6-story hotel
field. To record the growth of an eddy of a specific size in a tur- structure and a 3-story nonlinear SMRF. In this study, LSTM-s was
bulent wind environment, a CNN with changing dilation ratios and shown to be the most reliable and computationally cost-effective
the corresponding LSTM module a single component SDCL were for modeling and prediction of a nonlinear structural seismic
employed. The SDCL was able to accurately model the emergence
of multiple eddies of various sizes. They demonstrated that a
physics-inspired network design outperformed other typical DL
architectures for processing a complex wind environment and when
forecasting wind turbine responses.
Huang and Chen used 1D CNN and LSTM networks to forecast
the seismic responses of a two-story, three-span subway station us-
ing a data-driven technique to predict layer drift under the structure
(Huang and Chen 2021). LSTMs are a class of NN that can learn
sequence long-term dependencies. They do this through a gating
mechanism (Fig. 15) that allows them to selectively forget irrelative
information while keeping track of useful information. The predic-
tion and extrapolation abilities of the proposed networks were as-
sessed and compared to an MLP model. The seismic reactions of
the subway station was assessed using data about deformation from
the free field where the subway station was located as input. Pre-
diction and extrapolation from the same data set indicated that the
1D-CNN model outperformed both the LSTM and MLP models,
Fig. 15. Internal structure of long-short term memory cell.
with the LSTM performing the worst of the three models.

© ASCE 03124002-11 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

response when compared to LSTM-f and the traditional MLP ap- seismic damage assessment approach that uses ground motion
proach. Overall, the findings showed that the suggested LSTM-s time-frequency distribution graphs as input. Two numerical analy-
methodology was a viable, reliable, robust, and scalable method ses, one for a single structure and one for a regional scenario com-
for nonlinear structural response modeling under high-intensity prising all structures on the Tsinghua University campus were
ground movements. Notably, when the building displayed close- carried out to verify the technique. The results revealed that the
to-linear behavior without plastic deformation, very high forecast proposed method produced accurate prediction results in near
accuracy was achieved. real-time. For a typical one-building study, the CNN-based tech-
The capacity to analyze building performance is made possible nique provided adequate forecast accuracy.
by accurate earthquake response prediction. In order to attain this Data-driven models for damage detection (Ghannadi and
goal, Zhang et al. presented a physics-guided CNN (PhyCNN) for Kourehli 2019) and prediction of dynamic reactions of linear
data-driven structural seismic response modeling (Zhang et al. and nonlinear systems are essential because of their applications
2020c). The goal was to leverage readily available physics (for ex- in probabilistic analysis and inverse problems (Soleimani-
ample, the laws of dynamics) to limit network outputs, alleviate Babakamali and Esteghamati 2022). Sadeghi Eshkevari et al.
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

overfitting issues, reduce the need for large training data sets provided a physics-based RNN model (DynNet) that is able to
and improve the resilience of the trained model for more accurate estimate ground motion reactions such as displacement, velocity,
prediction. The surrogate model was then employed for fragility acceleration and internal force (Eshkevari et al. 2021). Although
analysis based on given limit state criteria. The deep PhyCNN the prediction accuracy is higher over long trajectories, this model
model uses several convolution layers and FC layers as well as has fewer trainable variables than its more advanced siblings.
a graph-based tensor differentiator and physics constraints to Furthermore, differential equation solver approaches influence
analyze the data. The proposed PhyCNN architecture uses ground the construction of the recurrent block and it is believed that this
motion (e.g., ground acceleration) as input and structural responses approach will result in more generic solutions. DynNet was
(e.g., story displacement) as output to learn the feature mapping validated using two nonlinear case studies. The predictability
between the input and output. The performance of this approach of longer trajectories was investigated and it was shown that
was tested using numerical and practical examples based on DynNet maintained the error level stably as the trajectory length
restricted data sets from simulations or field sensing. The Phy- increased in the majority of cases. The results revealed that the
CNN was trained for a 6-story hotel building in San Bernardino, network could accurately capture a range of nonlinear properties
California, utilizing training and validation data sets (containing of dynamic systems without the requirement for prior knowledge
11 and 4 samples, respectively) of ground motion acceleration se- or extremely large data sets. The only inputs necessary for
quences as input and story accelerations as measurement data. The DynNet inference are ground motion acceleration and the initial
results indicated that the proposed deep PhyCNN model for sim- conditions of state.
ulating seismic structural reaction was an effective, dependable and Torky and Ohno presented hybrid DL techniques (hybrid
computationally efficient method. The trained model could also be ConvLSTM-LSTM) for predicting the multi-component nonlinear
used to create a fragility function with which to assess a building’s responses of structures to significant seismic activity (Torky and
serviceability. Overall, the proposed approach is simple in nature Ohno 2021). This is a more comprehensive surrogate model for
and scalable to various types of structures (such as bridges) and for real-time building system identification. An industrial level exam-
other types of hazards. ple of an 8-story structure and the associated field sensing data were
Oh et al. proposed a CNN-based method with which to forecast used to demonstrate the performance of four models, each of which
the seismic reactions of structures (Oh et al. 2020). In their tech- comprised successive combinations of ConvLSTM layers, Conv3D
nique, the CNN was trained using previously measured time- layers, LSTM layers, and fully linked layers. The input character-
histories of structural response data and the time-histories of istics were ground or base sensor measurements and the goal was to
displacement in future earthquakes were predicted using the determine how the superstructure sensors reacted. For model train-
learned CNN. The ASCE benchmark model was utilized to verify ing and testing, 32 genuine earthquake events with highly nonlinear
the correctness of the proposed technique. The CNN was trained structural reactions were used. Large accelerations and Fourier
using the model’s seismic responses, which were collected from a spectrum predictions were used to test the model accuracy. The
variety of artificial earthquakes tailored to a specific geography and models that largely relied on ConvLSTM layers performed poorly
seismic risk. The performance of the trained CNN in predicting in terms of computing speed and accuracy. The models that
seismic reactions was validated by comparing it to the seismic achieved high accuracy used hybrid model upgrades such as flat-
responses of artificial earthquakes that were not used for training. tened LSTM layers and appropriate discrete wavelet transform de-
A 3-story RC frame structure was shaken and the applicability of composition filtering, which is suitable for real-time inference or
the suggested technique was tested using the structural reactions artificial circumstances. The findings showed that the proposed hy-
acquired from the shaking table test. The maximum displacement, brid models were able to indirectly estimate the structural capacity
as well as the time-histories of the displacement responses for curves, providing new opportunities for engineers to assess a build-
earthquakes that were not used for CNN training were accurately ing’s seismic performance. Two of the most significant improve-
predicted. ments from this study were the model selection procedures and
Assessment of seismic damage in a timely and precise manner is models capable of multicomponent prediction.
critical for limiting loss caused by significant events. Traditional Because the development and upgrading of a complex FE model
on-site investigations typically require extended examination to simulate a structure’s response to seismic forces requires a sig-
periods. A number of computational solutions show promise for nificant amount of time and effort, a number of rapid estimate ap-
addressing the flaws of the previous method. Lu et al. suggested proaches to anticipate engineering parameters have been presented.
a quick regional post-event damage assessment approach to pre- Li et al. presented a method for estimating the MIDR which com-
dicting the damage state based on CNNs (Lu et al. 2021). In their bines the interstory drift spectrum with a DL method (Li et al.
method, a scenario bank was created by combining an inventory of 2021a). Two estimates have been included, the first being to esti-
structures, expected ground motion data sets and related damage mate the MIDR as an approximation using the interstory drift spec-
values for a region. The authors proposed a CNN-based post-event trum. Because discrepancies existed between the interstory drift

© ASCE 03124002-12 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

spectrum and true answers, the interstory drift spectrum should be In earthquake engineering, seismic evaluations of structural sys-
altered. The second approximation refined the first estimate using tems sometimes include the creation of a fragility curve, which is an
deep CNN (DCNN) to forecast the MIDR of a new seismic event. important tool for estimating and analyzing the vulnerability of
The DCNN was trained using 30 RC building time-history assess- structures to seismic activity. Dang-Vu et al. used a frequency-
ments and 38 interstory drift spectra. To highlight its benefits, the based data-driven prediction model to propose a hybrid technique
suggested technique was compared to four ANN models and one for forecasting the seismic reactions of a typical Piloti-type build-
SVM model. The results showed that DCNN was able to accurately ing. A frequency spectrum of earthquakes was used as the major
forecast MIDR by learning the link between the interstory drift input data for the frequency-based data-driven model. The concur-
spectra and the time-history analysis results. The proposed method rent seismic reactions of many structural components can be pre-
also was used to compute the MIDR of 30 complicated FE models sented as output using the proposed method.
of steel moment-resisting frames. These findings indicated that, in The data set for training the DNN-based model was compiled
the event of new seismic activity, the technique was able to produce from numerical simulation results. The input data were the earth-
a credible estimate of structural MIDR. The suggested technique quake PGA, 200 consecutive frequency points as well as the
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

showed good prediction accuracy and provided a reasonable esti- maximum frequency, magnitude and earthquake duration. The
mate of the MIDR of structures in the event of new seismic activity. output data were the structural reactions, particularly the maxi-
In the study by Wen et al. the authors address the crucial prob- mum interstory drift of all column components and roof drifts.
lem of estimating seismic responses for existing RC frame struc- Structural fragility evaluations obtained using a standard method-
tures, particularly when crucial design information is lacking or ology and the proposed DL-based approach were compared to
missing (Wen et al. 2022). The main goal of the work is to design validate the prediction capabilities of the proposed method.
a CNN-based deep learning model called StruNet that can predict Despite the fact that the conclusions of Dang-Vu et al. were based
seismic response quickly and accurately from limited building data. on a typical Piloti-type building, they demonstrated the efficacy of
StruNet takes as input five easily obtainable building parameters the recommended surrogate model for projecting structural seis-
and ground motion records, predicting as output interstory drift mic reactions of varying degrees of complexity (Dang-Vu et al.
and peak floor acceleration for each floor in the given building. 2022). The proposed model could be useful for developing fragil-
The comprehensive approach of the study, which takes into account ity curves for a portfolio of such buildings with complex structural
162 typical RC frame buildings, and the use of a CNN to train the models where traditional methods of nonlinear dynamic analysis
model using 200 ground movements are its strong points. However, and derivation of fragility curves require significant amounts of
there are some limitations to consider. The study requires prepro- computational effort.
cessing of ground motions to unify record durations and sampling In SHM, online measurement and evaluation of dynamic struc-
frequencies, which restricts its applicability to regular RC frame tural reactions is critical for safety management in order to provide
buildings with similar layouts and dynamic characteristics. Extend- early warning and guarantee safe operation of civil infrastructures.
ing the model to accommodate irregular RC frame structures or Li et al. proposed a deep RNN model for the prediction of structural
RC shear wall structures would necessitate new prediction models. reactions under seismic excitation which was used to build a
Additionally, the study acknowledges the need for further research time-series attention-based RNN encoder-decoder (TSA-RNN-ED)
involving a broader range of building types, structural parameters, (Li et al. 2021b). They introduced a novel data-driven predictive
and ground motions to improve model performance. Lastly, while NN architecture for predicting online structural reactions under
the StruNet model is shown to be effective, future work could ex- seismic excitation. The suggested data-driven model predicted se-
plore alternative NN architectures and advanced DL methods for quential responses using sequence-to-sequence learning from ear-
further enhancement. Overall, this paper offers a promising solution lier multivariate time-series signals. Their proposed architecture
for seismic response prediction in the face of limited building infor- was able to accurately regress excitation-response interactions to
mation, with room for continued refinement and expansion. predict dynamic reactions to future earthquakes while addressing
In civil engineering applications, it is critical to analyze and the demand for real-time forecasting for on-the-ground practical
forecast structural dynamic nonlinear behavior (Liu et al. 2023). implementation. The proposed model was tested using two real-
The piecewise linear least squares (PLLS), fully connected neural world structural problems. The experimental results demonstrated
network (FCNN) and LSTM neural network (LSTMNN) methods that the proposed methodology was accurate and efficient for pre-
were studied by Peng et al. for structural dynamic response appli- dicting the seismic reactions of the structures under consideration.
cations under periodic, impact and seismic loads (Peng et al. 2012). Experiments from real-world case studies of the Shanghai Tower
These approaches are based on autoregression models and time and a wood-frame classroom were conducted using a shaking table.
series estimation models that continue to work even when the struc- The suggested NN model was compared with seven state-of-the-art
ture is stimulated by immeasurable inputs. To test the performance methods for sequence learning and prediction in the trials and their
of each approach objectively, a six-story steel frame was tested in results demonstrated substantial modeling and learning skills in
the same atmosphere and under the same four load conditions. The terms of forecasting multivariate time series data.
time series estimation model outperformed the auto-regression As the ability to accurately forecast the future reaction of civil
model for structural dynamics response prediction and the FCNN structures is crucial to health monitoring and safety evaluation,
outperformed the PLLS approach. The LSTMNN method performed Zhang et al. described a spatiotemporal DL technique based on
the best. The number of training samples had no effect on the per- the convolutional LSTM (ConvLSTM) recurrent network for
formance of the PLLS approach, which is based on classic estimate predicting structural responses based on spatial correlations and
methods; however, several adequate training data sets were required long-term temporal dependencies (Zhang et al. 2021). The pro-
for the FCNN and LSTMNN methods to produce satisfactory per- posed spatiotemporal learning architecture contained a ConvLSTM
formance. Furthermore, the PLLS method was susceptible to noise, layer, a flattened and reshaped layer and two fully connected (FC)
but the FCNN and LSTMNN approaches, which are based on DL, layers to represent the spatial and temporal links of the input and
were highly resilient and were not susceptible to noise. Peng et al. output sequences. The fundamental idea was to learn latent spatio-
reported that the LSTMNN was the most promising of the three temporal features from data using a ConvLSTM network and then
methods when appropriate training data was available. utilize that information to generate a surrogate model for structural

© ASCE 03124002-13 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

response predictions. The performance of the proposed technique Start

was demonstrated on a concrete bridge using strain time-history
data spanning three years. According to the findings, the ConvLSTM
methodology is a promising, reliable and computationally efficient
Population initialize
tool for accurately forecasting the dynamic response of civil struc-
tures in a data-driven manner. The results demonstrated that the pro-
posed ConvLSTM model was a reliable and effective strategy for
predicting long-term and short-term reactions. Evaluate the fitness of
objective function
The Streicker Bridge in Princeton, New Jersey, features a
SHM system. Researchers have conducted various studies em-
ploying different algorithms to predict structural responses under
different conditions in this bridge (Oh et al. 2023; Pereira and
Glisic 2022). Park et al. used data collected by a health monitor- Meeting stopping Yes
End
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

ing system to predict long-term strain at the Streicker Bridge criteria

(Park et al. 2022). They used environmental factors such as tem-
perature, relative humidity, and wind speed as inputs for their
No
proposed CNN model. The study systematically evaluates six
forecasting models, provides a comprehensive analysis of the ef-
fectiveness of different combinations of weather data, and in- Generate new population
creases the robustness of the proposed method, which can have
practical applications in SHM. The study identifies the combina-
tion of relative humidity and temperature as effective for long- Fig. 16. Typical diagram of metaheuristic algorithms.
term strain prediction but does not delve into further analysis
or discussion of why this combination works best, missing an
opportunity for deeper insights.
Leveraging DL algorithms and enhancing existing models to Fig. 16. Feature selection is a critical and visible activity in ML.
predict seismic acceleration responses and detect anomalies in The primary goal of a feature selection problem is to reduce the
bridges offers numerous advantages in effectively monitoring and size of the feature set while maintaining performance accuracy.
assessing the overall health and condition of these critical structures Many ways have been developed to categorize the data sets; how-
(Chen et al. 2023; Zhang et al. 2023a). Liao et al. studied an in- ever, metaheuristic algorithms have sparked much interest for a
novative attention-based LSTM neural network (AttLSTM) for range of optimization problems, which are divided into two main
modeling the seismic response of bridges, aiming to improve pre- groups: single-agent and multiagent-based (Agrawal et al. 2021;
dictions under dynamic loads like earthquakes (Liao et al. 2023). Doroudi and Lavassani 2021).
The AttLSTM utilizes an attention mechanism to select informative Metaheuristics has resulted in significant advancements for
components from sequential data, enhancing learning from limited solving of intractable optimization problems (Dokeroglu et al.
training data. The model’s performance is validated using both 2019). To increase fitness, these algorithms draw inspiration from
numerical simulations and real-world data from girder and cable- a variety of sources, although the majority of them are nature-
stayed bridges. It is compared favorably against the traditional inspired. Because of their numerous uses in engineering and AI,
LSTM model, showing improved accuracy and reliability. The they have been intensively explored (Wong and Ming 2019).
study does not consider site effects on seismic responses, which can Metaheuristics can develop acceptable answers to a difficult prob-
significantly impact the behavior of bridges in earthquake events. lem through trial-and-error in a reasonable amount of time. Because
It is better that future works address this aspect for a more com- the intricacy of the topic of interest makes it impractical to search
prehensive analysis. for every possible solution or combination, the goal is to find a
In the field of structural response prediction, a spectrum of al- reasonable, feasible solution within a reasonable timeframe. There
gorithms, including CNN, LSTM networks, and traditional ML currently is no guarantee that the best answers will be discovered or
techniques, are harnessed to varying degrees of success. CNNs ex- to know whether or not an algorithm will work or why it will work
cel at processing grid-like structural data, leveraging their capacity if it does. The goal is to create an efficient and practical algorithm
to automatically learn hierarchical features, although they may that works most of the time and produces high-quality results
overlook temporal dependencies. In contrast, LSTM networks spe- (Gandomi et al. 2013).
cialize in capturing sequential data and long-term relationships, Wang et al. presented a computational solution to the deflection
making them ideal for time-series structural data but demanding prediction problem for a prestressed concrete (PC) rigid frame
more extensive data sets and computational resources. Traditional bridge with corrugated steel webs (CSWs) constructed using the
algorithms like SVR or RF offer interpretability and adaptability to balanced cantilever method (Wang et al. 2021c). There were three
smaller data sets but may struggle with complex, high-dimensional layers of analysis in the framework: (1) obtain data sets for estab-
structural response modeling. The choice of algorithm hinges on lishing the prediction model in which the input samples and output
the unique characteristics of the structural system and the resources response are generated by Latin hypercube sampling and high-
available, with engineers aiming to achieve a precise balance be- fidelity 3D FE modeling, respectively, (2) establish an optimized
tween predictive accuracy and practicality. NN model that integrates the mind evolutionary computation
(MEC) and BP algorithms to evaluate the model performance, and
(3) obtain the data sets for establishing the prediction model in
Metaheuristic Algorithms which the input samples and output for an actual CSWs-PC bridge
Metaheuristic algorithms also were reviewed to determine their is used to demonstrate the applicability and efficiency of the frame-
performance for prediction aims in civil engineering structures. work. The numerical findings produced by the confirmed FE model
The general flowchart of metaheuristic algorithms is depicted in were compared to the MEC-BP and BP model predictions. In terms

© ASCE 03124002-14 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

of prediction accuracy and efficiency, the MEC-BP model sur- tunneling-induced ground reactions in real-time and overcome the
passed the BP model. Segment deflection could be correctly pre- limits of the empirical, analytical, and numerical models of the do-
dicted using the MEC-BP model with optimal structures. main expert. Five shield operational parameters (thrust, torque,
Predicting surface settlement and tunnel squeezing behavior by grout filling, penetration rate, chamber pressure) and five geologi-
combining ML and meta-heuristic algorithms enhances accuracy cal parameters (modified blow counts of standard and dynamic
and efficiency in urban tunneling, helping construction planning penetration tests of soil layers, modified uniaxial compressive
and risk management. It also offers insights and the potential for strength of weathered rock, groundwater table, soil type at the
ongoing improvements in real-time predictions (Kim et al. 2022; cut) were used to forecast tunneling induced settlement as the
Zhang et al. 2023b; Zhou et al. 2022). Zhang et al. offered a output.
new RL-based optimizer which combined the deep-Q network Analysis of monitoring data produced from sensor measurement
(DQN) and particle swarm optimization (PSO) (Fig. 17) to improve is required for the prediction of concrete arch dam behavior. Math-
the extreme learning machine (ELM)-based tunneling induced set- ematical models based on AI algorithms can be used to simulate
tlement prediction model (Zhang et al. 2020b). In this scenario, the dam displacement as a function of water level, irreversible time
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

DQN-PSO optimizer was used to optimize ELM’s weights and impact and temperature. Liu et al. effectively applied a mathemati-
biases. Unlike previous metaheuristic optimization algorithms cal model based on the kernel-ELM (KELM) method to displace-
which guided particle movement in a crude manner, the DQN- ment prediction modeling of concrete gravity dams using long-term
based optimizer reward rule focused on analysis of the payoff of daily air temperature monitoring data sets (Liu et al. 2021a). The
an agent’s actions, ensuring that the particles, like a smart person, purpose of their work was to validate the applicability of the
always picked the best action in each step as the optimizer DQN- KELM-based model in terms of thermal effects on the prediction
PSO knew what to do and when to execute it, ensuring that the of the behavior of concrete arch dams. The KELM intelligence al-
global optima were reached. A similar hybrid model was used in gorithm recorded an air-temperature data series used for thermal
a real-world tunnel project. The findings revealed that the DQN- effect modeling. The Jaya optimization method was used during
PSO optimizer clearly outperformed classic metaheuristic optimi- the modeling phase to determine the critical parameters of the
zation methods in terms of accuracy and computational cost while KELM model. The KELM and Jaya algorithms worked together
looking for the best global weights and biases for ELM. This ap- to provide an autonomous optimization strategy for improving
proach was able to identify correlations between the influencing modeling efficiency. The results showed that the suggested method
variables and ground reactions through self-practice. was a viable and powerful methodology for predicting the behavior
Their final model could be explicitly stated and used to predict of concrete arch dams.
tunneling induced ground reactions in real-time, making it useful in Peak shear strength estimates of RC shear walls were performed
engineering practice. The hybrid deep RL model, which combines by Parsa and Naderpour using a combination of SVR and meta-
ELM and the DQN-PSO optimizer was able to correctly forecast heuristic optimization techniques such as teaching-learning-based
optimization (TLBO), PSO and Harris-Hawks optimization (HHO)
(Parsa and Naderpour 2021). They compiled a large database of
228 experimental data sets for RC shear walls with nine parameters
Start that included geometric properties, materials and RC shear-wall
forces. The development of models for forecasting the shear
Initialize group of particles strength of three types of RC shear wall (squat, cylindrical, and
thin) was one of the most useful features of their study. The SVR-
TLBO, SVR-PSO, and SVR-HHO models were suggested. The
Evaluate pBest for each particle proposed models outperformed existing equations in terms of
performance evaluations and the ratio of predicted values to exper-
imental values. The results suggested that the proposed models
were accurate. As a result, researchers may be able to use these
Current position is
True Update pBest models to estimate the shear strength of RC shear walls, improve
better than pBest? forecasting accuracy and minimize construction costs.
The artificial bee colony (ABC) method is a newly found swarm
intelligence optimization technique that has been used to train ANN
False models. Asteris and Nikoo used the ABC technique to train a feed-
Assign pBest to gBest
forward NN (FF-ABC-NN) to predict the fundamental period of
vibration in infilled-frame RC structures, which is crucial for earth-
quake design (Asteris and Nikoo 2019). The number of stories,
Compute velocity spans, span length, stiffness of the infill-wall panel and percentage
of apertures inside the infill panel were selected as input parameters
and the value of the vibration period was selected as an output
Update particle position
parameter. The findings suggest that the ABC technique can be
utilized to train FFNNs successfully. The ANN outperformed the
statistical models in terms of ability, flexibility and accuracy when
False Target reached? True End the weights were modified using the ABC technique.
Standard approaches for predicting structural collapse of RC
structures requires time and effort. Traditional learning algorithms
are prone to premature convergence and becoming stuck in local
Fig. 17. Typical flowchart of the PSO algorithm, pBest: the best
optima. To overcome this issue, Chatterjee et al. proposed a
position that the particle has reached thus far. gBest: the best position
PSO-based technique for training the NN (NN-PSO) (Chatterjee
of the particles in the entire population.
et al. 2017). The PSO was used to find the NN weight vector with

© ASCE 03124002-15 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

the smallest RMSE. The proposed NN-PSO classifier was capable utilities rather than time-consuming and expensive experimental
of tackling the problem of forecasting multistory RC structural fail- investigations.
ure by detecting the possibility of its future failure. The experimen- Wang et al. developed an effective and robust soft computing
tal findings were based on a database of 150 RC structures from technique in which ANN were hybridized with GA and PSO to
multistory buildings. To determine the level of creativity of the anticipate the bond strength in concrete-encased steel (CES) con-
NN-PSO model, it was compared to the NN and MLP-FFN clas- struction (Wang et al. 2021a). The results showed that the novel
sifiers. The results showed that suggested NN-PSO classifier out- GA-ANN and PSO-ANN models outperformed both the standard
performed both classifiers and performed well in terms of accuracy. ANN model and existing empirical equations. The PSO-ANN
The proposed system then utilized the PSO-trained ANN (NN-PSO) outperformed the GA-ANN in terms of convergence speed and
classifier to avoid the NN becoming stuck in a local optimum when prediction error due to its unique information-sharing technique.
forecasting the collapse of RC structures. The results of the experi- The findings could be utilized to intelligently analyze the bond
ments showed that the proposed model was superior for forecasting strength in CES without the need for expensive and time-consuming
the structural condition of a multistory RC building. studies.
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Although floating offshore wind turbines (FOWTs) are promis- PSO-ANN was used by Nguyen et al. to forecast horizontal dis-
ing solutions for offshore renewable energy harvesting, they con- placement of columns in a short building following considerable
tinue to face design and maintenance challenges in terms of cost seismic loading (Nguyen et al. 2020). Their research sought to im-
reduction, monitoring, safety and sustainability. Chen et al. em- prove the ANN using PSO for ground response approximation in a
ployed full-scale data (including floater movements, wind, wave short structure. The inputs included the Chi-Chi earthquake dy-
and current data) from a single FOWT in Scotland to undertake namic time, friction angle, dilation angle, unit weight, soil elastic
numerical optimization and dynamic response prediction of thrust modulus, Poisson’s ratio, structure axial stiffness, and bending stiff-
force, blade-tip deformation and tower-top deformation (Chen et al. ness. The output was the horizontal deflection of the columns at
2021). SADA (software-in-the-loop coupled AI approach for dy- their greatest levels. The results showed that, although the PSO-
namic response analysis) employs RL methodologies to enable ANN model was reliable for calculating ground reaction and hori-
numerical software to self-learn. The results revealed that the zontal deflection of structural columns in short structures following
SADA technique allowed the AI-trained numerical model to more earthquake loading, the hybrid PSO-ANN model was a superior
accurately predict the movements of the supporting floater. Further- and more trustworthy ANN model.
more, physical qualities that could not be obtained directly through In the domains of civil engineering and structural response pre-
full-scale measurement but are important to the business could be diction, metaheuristic algorithms have proven to be invaluable
obtained from more realistic perspectives. The capability of SADA tools. These algorithms, such as PSO, are employed in tandem with
to forecast FOWT movement illustrated its ability to predict other ANN to enhance the accuracy of predictions across a spectrum of
significant physical features. structural behaviors, including seismic response, deflection, bond
In seismic load applications, ANN does not provide an accurate strength, and more. The advantages of PSO lie in its ability to ef-
forecast of ground surface reactions. Liu et al. evaluated the appli- ficiently explore complex optimization spaces, often converging
cability of a whale optimization algorithm (WOA) to compute the faster than alternative methods, thus reducing training time.
horizontal deflection of columns in a short structure subjected to However, its sensitivity to parameter settings and the potential
severe seismic loading (Liu et al. 2021b). They provided a quan- for premature convergence to local optima are noteworthy disad-
tum-theory-based version of the WOA that they had optimized for vantages. Nonetheless, when carefully harnessed and tailored to
ground response approximation in short buildings. The Chi-Chi specific structural prediction tasks, the combination of PSO with
earthquake in Taiwan (1999) served as the work input database. ANN demonstrates its effectiveness in optimizing models and im-
The training and testing databases for the ANN algorithm and proving the understanding of structural behavior, ultimately con-
WOA were created using 10,403 FE simulation results as input tributing to safer and more efficient civil engineering practices.
databases. All inputs included the Chi-Chi earthquake dynamic
time, soil elastic modulus, dilation angle, Poisson’s ratio, unit
weight, friction angle, bending stiffness and axial stiffness. The Conclusion
horizontal deflection of the columns at their highest level also
was included in the results. In contrast to the PSO-ANN and In the field of civil engineering, ensuring the stability and health of
ANN techniques, the WOA-ANN method provided reliable structures is of utmost importance, and structural health monitoring
approximation results when calculating column horizontal defor- (SHM) serves as a crucial procedure for maintaining their integrity
mation over time. and serviceability over time. The emergence of advanced prediction
Mehrabi et al. used an adaptive neuro-fuzzy inference system methods, particularly machine learning (ML), has revolutionized
(ANFIS) to forecast the seismic response of FRC rectangular col- the prediction of structural responses, providing a data-driven ap-
umns and determine the promising factor of strength subjected to proach to enhance the field. The current study used biometric data
seismic loads using a combination of PSO and GA (Mehrabi et al. mining to uncover essential methods for predicting structural re-
2021). The results were validated using ELM on 317 data sets, sponses in civil engineering applications. The results showed a sig-
all of which were derived from real-world test results. To predict nificant expansion of ML applications for predicting structural
and optimize the concrete compressive strength and maximum lat- responses over the past thirteen years. Quantitative investigation
eral force, each model used the width, height, fiber fraction ratio, indicated that machine learning, deep learning, and metaheuristic
maximum lateral force, concrete compressive strength, fiber yield algorithms were the principal methodologies used. Table 1 provides
strength, and shear span ratio as inputs. The results indicated that a comprehensive overview of investigations centered around struc-
ANFIS–PSO accurately predicted the lateral load. The results of the tural response prediction in civil engineering. Within this table, you
mechanical performance prediction showed that ELM accurately will find concise descriptions of each investigation, highlighting the
predicted the compressive strength. Both the ANFIS–GA and structural type, application, and methodology employed. This ex-
ANFIS–PSO techniques demonstrated reliable prediction perfor- tensive compilation truly reflects an extraordinary exploration of
mance which would encourage researchers to employ prediction the ever-evolving domain of structural response prediction.

© ASCE 03124002-16 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

© ASCE

Table 1. Summary of review

Category Structure type Application Methodology Reference
Infrastructure Composite bridges Predicting the midspan deflections NN Tadesse et al. (2012)
Dam Forecasting displacement and Strain NARX Radovanovic et al.
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

(2015)
Wind turbine Predicting energy generation and structural bending SDCL Woo et al. (2019)
moment
Bridges Damage estimation XGBoost Lim and Chi (2019)
Cable-stayed bridge Predicting tower displacement, internal foundation force, SVR, BPNN, GPR Fang et al. (2020)
mid-span displacement
Tunneling Settlement prediction DQN-PSO Zhang et al. (2020b)
Tunnel Predicting ground settlement during shield tunneling ANN Zhang et al. (2020a)
Turbine Prediction thrust force, blade tip deformation and tower top SADA Chen et al. (2021)
deformation
Steel-concrete composite bridges Rapid prediction of long-term deflections ANN Kumar et al. (2021)
Bridge Estimating buffeting responses SVR, MLP Castellon et al. (2021)
Bridge Predicting peak bearing deformation and column drift ratio MLR, LR,NN, RF, GTB Wang et al. (2021b)
Bridge Predicting deflection MEC-BP Wang et al. (2021c)
Concrete bridge Strain prediction ConvLSTM Zhang et al. (2021)
Subway station Layer drift prediction 1-D CNN, LSTM Huang and Chen (2021)
Concrete arch dam Predicting displacement KELM Liu et al. (2021a)
Tunnel Predicting the severity of tunnel squeezing WOA-SVM, ANN, SVM, GA Zhou et al. (2022)
Tunneling Prediction of surface settlement SVR, RF, GBM, XGBoost, LightGBM Kim et al. (2022)
Streicker bridge (pedestrian bridge on Prediction of strain Probabilistic neural network Pereira and Glisic (2022)
03124002-17

Princeton University campus)

Streicker bridge (pedestrian bridge on Prediction of strain CNN Park et al. (2022)
Princeton University campus)
Aqueduct Predicting the displacement and crack openings ELM, evolutionary ELM (E-ELM), BP, LSTM, MLR, Jiang et al. (2022)
PSO-ELM
Bridge Estimation of shear force, bending moment and curvature ANN, LSTM Li et al. (2022)
ductility
Jindo Bridge (a twin-cable stayed bridge Predicting the root mean square acceleration of the bridge GPR Kim and Kim (2022)
in South Korea) deck
Bridge Estimating the pier settlement GPR, extra tree regressor, RF Abdu et al. (2023)
Streicker bridge (pedestrian bridge on Prediction of strain CNN Oh et al. (2023)
Princeton University campus)
A cable-stayed bridge, two-span Prediction of displacement LSTM, AttLSTM Liao et al. (2023)
continuous girder bridge
Bridge Prediction of Seismic acceleration response Conv1D-LSTM, XGBoost, RF Zhang et al. (2023a)
Xijiang Bridge (in Zhaoqing City, Strain forecasting Bidirectional long- and short-term time-series network with Chen et al. (2023)
Pract. Period. Struct. Des. Constr.

Guangdong Province, China), Huangpu attention mechanism (BiLSTNet-A)

Bridge (in Guangzhou, China)
Tunneling Predicting surface settlement Relevance vector machine (RVM), principal component Zhang et al. (2023b)
analysis-RVM (PCA-RVM), PSO-RVM, PCA-PSO-RVM
Building RC structures Predicting the structural collapse NN-PSO Chatterjee et al. (2017)
RC frames Prediction of the damage state using prediction of ANN Morfidis and Kostinakis
maximum interstory drift ratio (2017)
RC frames Predicting the fundamental period FF-ABC-NN Asteris and Nikoo (2019)
6-story hotel building, 3-story steel frame Predicting interstory drifts LSTM-s, LSTM-f Zhang et al. (2019)

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

© ASCE

Table 1. (Continued.)
Category Structure type Application Methodology Reference
18-story RC frame Prediction of the maximum cross-sectional rotation, the WWLSSVM, feed-forward, backpropagation ANN Gharehbaghi et al.
maximum inelastic interstory drift ratio, the maximum (2020)
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

inelastic displacement at the roof level, and the maximum

shear force of stories
Short building Estimating the horizontal displacement PSO-ANN Nguyen et al. (2020)
6-story hotel building predicting story displacements PhyCNN Zhang et al. (2020c)
3-story RC frame structure Forecasting maximum displacement CNN Oh et al. (2020)
20-story RC building Collapse prediction regularized logistic regression Bijelić et al. (2020)
Concrete shear wall buildings (15, 20, 25, Predicting maximum acceleration, maximum velocity, and hybrid ANN Simulated Annealing (ANN-SA) Barkhordari and
and 30-story) earthquake characteristics Es-haghi (2021)
6-story steel frame Predicting displacement, strain, stress PLLS, FCNN, LSTMNN Peng et al. (2012)
8-story structure Predicting acceleration Responses Hybrid ConvLSTM-LSTM Torky and Ohno (2021)
Short structure Computing the horizontal displacement of columns WOA-ANN Liu et al. (2021b)
Three-story-two-bay frame structure Prediction of peak story drift ratios feed-forward NN Thaler et al. (2021)
Steel frames Estimating seismic story drift demand MLDD, ML-EMKY Guan et al. (2021)
Steel frame Predicting interstory drifts ANN, XGBoost Nguyen et al. (2021)
Piloti-type building Forecasting maximum interstory drift and roof drifts DNN Dang-Vu et al. (2022)
RC frame, beam-like structure Estimating strain responses CNN Oh and Kim (2021)
Tsinghua University campus Predicting the Damage state CNN Lu et al. (2021)
4DOF shear building system Predicting displacement, force, acceleration DynNet Eshkevari et al. (2021)
Multistory concentrated-mass shear Estimating the maximum interstory drift ratio DCNN Li et al. (2021a)
model, 30 steel frame
03124002-18

Concrete-encased steel structures Predicting the bond strength GA-ANN, PSO-ANN Wang et al. (2021b)
Shanghai Tower, one-story wood-frame Forecasting displacement responses, short-term time-series TSA-RNN-ED Li et al. (2021b)
classroom predictions
Buildings Damage predictions LR, SVR, GBR, RF Ghimire et al. (2022)
Concrete Buildings Estimating probabilistic seismic demand models deep learning model using an encoder–decoder model Soleimani-Babakamali
(Dense, CNN and LSTM based networks) and Esteghamati (2022)
RC frame structures Predicting interstory drift (IDR) and peak floor acceleration StruNet (a CNN-based prediction model) Wen et al. (2022)
(PFA)
30 R/C buildings Estimating the maximum interstory drift ratio in order to LightGBM, GBR, RF, Extra Trees Regressor, k-Nearest Demertzis et al. (2023)
assess the seismic damage of the buildings Neighbors Regressor, LR, Bayesian Ridge, Ridge
Regression, DT, AdaBoost, Elastic Net, Lasso Regression,
Orthogonal Matching Pursuit, Huber Regressor, Least
Angle Regression
Reinforced concrete (RC) buildings Predicting the performance of the existing buildings against ANN, RF, SVM, DT Cosgun (2023)
earthquakes
2 to 9 story steel moment-resisting frames Predicting the maximum interstory drift ratio XGBoost, RF, Bagging Regressor, Extra-Trees Regressor, Kazemi et al. (2023)
Pract. Period. Struct. Des. Constr.

GBM, Histogram-based GBR (HistGBR), AdaBoost,

Regression-based k-Nearest Neighbors, Voting Regressor,
Nu-SVR (NuSVR), Extremely Randomized Tree
Regressor, ANNs
A SDOF system, a steel building frame, Nonlinear time-history predictions of seismic responses LSTM, 2D CNN, WaveNet Ning et al. (2023)
a multi-component bridge structure
Six-story building, four cases of the SDOF Predicting displacement, velocity, acceleration, and LSTM, physics-informed long short-term memory Liu et al. (2023)
system restoring force (PI-LSTM), physics-guided CNN (PhyCNN),
physics-guided LSTM (PhyLSTM)
Cross-laminated timber buildings Estimation of the maximum interstory drift ratio and the MLR, Regression Trees, RF, K-nearest Neighbor, SVR Junda et al. (2023)
roof drift ratio

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

The reliability and accuracy of the ML-based prediction models

Mousavi et al. (2022)

Mehrabi et al. (2021)
Almustafa and Nehdi

Almustafa and Nehdi

Almustafa and Nehdi

Bolandi et al. (2022)

Parsa and Naderpour
Aguilar et al. (2016)
are heavily dependent on the performance of the learning algo-

Anvari et al. (2023)

Zhao et al. (2022b)

Zhao et al. (2022b)

Ahmadianfar et al.
Reference

Zhu et al. (2021)

rithms and the properties of the training data sets employed. Each

Wu et al. (2023)
He et al. (2023)
of these algorithms has its own set of strengths and weaknesses.
ANNs excel at modeling complex data but require substantial train-
(2020) ing data. SVR handles noisy data with controlled complexity but

(2021)

(2021)

(2021)

(2022)
needs hyperparameter tuning. XGBoost offers high accuracy but
demands computational resources. In structural response predic-
tion, CNNs process gridlike data well but may overlook temporal

LR, k-Nearest Neighbor, DT, RF, SVM, MLP, LightBoost,

dependencies, while LSTMs capture sequences but require more

Hybrid GBR trees coupled with Henry gas solubility

data and resources. Traditional methods like SVR and RF offer in-

MLP, SVM, DT, RF, extremely randomized trees

terpretability but may struggle with complexity. Engineers must
choose algorithms based on project needs and data availability.
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

In civil engineering, metaheuristic algorithms like PSO enhance

prediction accuracy but require careful tuning.
RF, hybrid classification-regression RF

Generative expression programming

SVR-TLBO, SVR-PSO, SVR-HHO

Table 1 shows that ANN and SVR for machine learning, CNN
Methodology

and LSTM for deep learning, and various combinations of PSO in

Generative adversarial network

meta-heuristic algorithms for structural response prediction were
the algorithms used most frequently. It was observed that, with
ANFIS–GA, ANFIS–PSO

the use of these three types of algorithms, the response of the civil
optimization algorithm
SVM, GPR, RF, BP

structures could be predicted to be within an acceptable accuracy

LWLR, SVR, MLR

SVR, ANN, GPR

range and could be used for different areas of SHM. Because of the
increased usage of ML in various fields, we expect it to be em-
ANN, SVR

ployed in more and more sophisticated fields of civil engineering

AdaBoost

in the future, although generative models have demonstrated re-

ANN

CNN
GPR

markable efficacy in various domains, their adoption within the

SHM community remains limited. As a forward-looking direction
for the field, we propose a shift of emphasis toward emerging
Forecasting concrete compressive strength and maximum

generative models, including transformers, diffusion models, and

generative adversarial networks (GANs). These models are more
Predicting the failure pattern (damage prediction)

scalable and thus more suitable for transfer learning, enhancing

their adaptability to diverse SHM scenarios.
the peak displacement response
Estimating the nonlinear buckling load

Predicting the maximum displacement

Predicting the maximum displacement

the maximum deformation

Predict the maximum displacement

Data Availability Statement

Application

Predicting maximum displacement

Predicting the total shear strength

possible damage states
Peak shear strength estimation

stress distributions
Estimating the shear strength

All data, models, and code generated or used during the study
Estimating the scour depth

appear in the published article.

References
lateral force

Abdu, D. M., G. Wei, and W. Yang. 2023. “Assessment of railway bridge

Predicting
Predicting
Predicting

Predicting

pier settlement based on train acceleration response using machine

learning algorithms.” Structures 52 (Jun): 598–608. https://doi.org/10
.1016/j.istruc.2023.03.167.
Adeli, H. 2020. “Four decades of computing in civil engineering.” In
Proc., CIGOS 2019, Innovation for Sustainable Infrastructure, 3–11.
New York: Springer.
Concrete-filled steel tubes columns
Nonstructural exterior infill walls

Agrawal, P., H. F. Abutarboush, T. Ganesh, and A. W. Mohamed. 2021.

“Metaheuristic algorithms on feature selection: A survey of one decade
RC beams under blast loading
RC slabs under blast loading

RC beams with FRP sheets

Structure type

Steel-plate composite walls

of research (2009–2019).” IEEE Access 9 (Mar): 26766–26791. https://

Reinforced masonry walls

FRC rectangular columns

FRP retrofitted RC slabs

doi.org/10.1109/ACCESS.2021.3056407.
Aguilar, V., C. Sandoval, J. M. Adam, J. Garzón-Roca, and G. Valdebenito.
RC bridge columns

2016. “Prediction of the shear strength of reinforced masonry walls

Reticulated shell

RC shear walls

using a large experimental database and artificial neural networks.”

Struct. Infrastruct. Eng. 12 (12): 1661–1674. https://doi.org/10.1080
Steel plates

/15732479.2016.1157824.
RC slabs

Ahmadianfar, I., M. Jamei, and X. Chu. 2021. “Prediction of local scour

Table 1. (Continued.)

Pile

around circular piles under waves using a novel artificial intelligence

approach.” Mar. Georesour. Geotechnol. 39 (1): 44–55. https://doi
.org/10.1080/1064119X.2019.1676335.
Alipour, M., and D. K. Harris. 2020. “A big data analytics strategy for scal-
Category

able urban infrastructure condition assessment using semi-supervised

Others

multi-transform self-training.” J. Civ. Struct. Health Monit. 10 (2):

313–332. https://doi.org/10.1007/s13349-020-00386-4.

© ASCE 03124002-19 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Almustafa, M. K., and M. L. Nehdi. 2020. “Machine learning model for Structures 50 (Apr): 1994–2003. https://doi.org/10.1016/j.istruc.2023
predicting structural response of RC slabs exposed to blast loading.” .02.127.
Eng. Struct. 221 (Oct): 111109. https://doi.org/10.1016/j.engstruct Damirchilo, F., A. Hosseini, M. Mellat Parast, and E. H. Fini. 2021.
.2020.111109. “Machine learning approach to predict international roughness index
Almustafa, M. K., and M. L. Nehdi. 2021. “Machine learning prediction using long-term pavement performance data.” J. Transp. Eng. Part B.
of structural response for FRP retrofitted RC slabs subjected to blast Pavements 147 (4): 04021058. https://doi.org/10.1061/JPEODX
loading.” Eng. Struct. 244 (Oct): 112752. https://doi.org/10.1016/j .0000312.
.engstruct.2021.112752. Dang-Vu, H., Q. D. Nguyen, T. Chung, J. Shin, and K. Lee. 2022.
Almustafa, M. K., and M. L. Nehdi. 2022. “Novel hybrid machine learning “Frequency-based data-driven surrogate model for efficient prediction
approach for predicting structural response of RC beams under blast of irregular structure’s seismic responses.” J. Earthquake Eng. 26 (14):
loading.” Structures 39 (May): 1092–1106. https://doi.org/10.1016/j 7319–7336. https://doi.org/10.1080/13632469.2021.1961940.
.istruc.2022.04.007. Das, S., P. Saha, and S. K. Patro. 2016. “Vibration-based damage detection
Anvari, A. T., S. Babanajad, and A. H. Gandomi. 2023. “Data-driven pre- techniques used for health monitoring of structures: A review.” J. Civ.
diction models for total shear strength of reinforced concrete beams Struct. Health Monit. 6 (3): 477–507. https://doi.org/10.1007/s13349
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

with fiber reinforced polymers using an evolutionary machine learning -016-0168-5.

approach.” Eng. Struct. 276 (Feb): 115292. https://doi.org/10.1016/j Demertzis, K., K. Kostinakis, K. Morfidis, and L. Iliadis. 2023. “An inter-
.engstruct.2022.115292. pretable machine learning method for the prediction of R/C buildings’
Asteris, P. G., and M. Nikoo. 2019. “Artificial bee colony-based neural net- seismic response.” J. Build. Eng. 63 (Mar): 105493. https://doi.org/10
work for the prediction of the fundamental period of infilled frame .1016/j.jobe.2022.105493.
structures.” Neural Comput. Appl. 31 (9): 4837–4847. https://doi.org/10 Dokeroglu, T., E. Sevinc, T. Kucukyilmaz, and A. Cosar. 2019. “A survey
.1007/s00521-018-03965-1. on new generation metaheuristic algorithms.” Comput. Ind. Eng.
Bakht, B., and A. Mufti. 2018. “Evaluation of one hundred and one instru- 137 (Nov): 106040. https://doi.org/10.1016/j.cie.2019.106040.
mented bridges suggests a new level of inspection should be established Dong, J., X. Yan, and S. Li. 2018. “Cable force monitoring and prediction
in the bridge design codes.” J. Civ. Struct. Health Monit. 8 (1): 3. https:// for cable group of long-span cable-supported bridges.” J. Civ. Struct.
doi.org/10.1007/s13349-017-0256-1. Health Monit. 8 (4): 597–605. https://doi.org/10.1007/s13349-018
Barkhordari, M., and M. Es-haghi. 2021. “Straightforward prediction -0301-8.
for responses of the concrete shear wall buildings subject to ground Doroudi, R., and S. H. H. Lavassani. 2021. “Connection of coupled build-
motions using machine learning algorithms.” Int. J. Eng. 34 (7): ings: A state-of-the-art review.” Structures 33 (Oct): 1299–1326. https://
1586–1601. https://doi.org/10.5829/ije.2021.34.07a.04. doi.org/10.1016/j.istruc.2021.05.017.
Bijelić, N., T. Lin, and G. G. Deierlein. 2020. “Efficient intensity mea- Doroudi, R., S. H. H. Lavassani, and M. Shahrouzi. 2023. “Predicting
sures and machine learning algorithms for collapse prediction of tall acceleration response of super-tall buildings by support vector regres-
buildings informed by SCEC CyberShake ground motion simulations.” sion.” Struct. Des. Tall Special Build. 32 (16): e2049. https://doi.org/10
Earthquake Spectra 36 (3): 1188–1207. https://doi.org/10.1177/8755 .1002/tal.2049.
293020919414. Doroudi, R., S. H. H. Lavassani, M. Shahrouzi, and M. Dadgostar. 2022.
Bolandi, H., X. Li, T. Salem, V. N. Boddeti, and N. Lajnef. 2022. “Bridging “Identifying the dynamic characteristics of super tall buildings by multi-
finite element and deep learning: High-resolution stress distribution pre- variate empirical mode decomposition.” Struct. Control Health Monit.
diction in structural components.” Front. Struct. Civ. Eng. 16 (11): 29 (11): e3075. https://doi.org/10.1002/stc.3075.
1365–1377. https://doi.org/10.1007/s11709-022-0882-5. Entezami, A., H. Sarmadi, and B. Saeedi Razavi. 2020. “An innovative
Brownjohn, J. M. W., A. De Stefano, Y.-L. Xu, H. Wenzel, and A. E. Aktan. hybrid strategy for structural health monitoring by modal flexibility
2011. “Vibration-based monitoring of civil infrastructure: Challenges and clustering methods.” J. Civ. Struct. Health Monit. 10 (5): 845–859.
and successes.” J. Civ. Struct. Health Monit. 1 (3): 79–95. https://doi https://doi.org/10.1007/s13349-020-00421-4.
.org/10.1007/s13349-011-0009-5. Eshkevari, S. S., M. Takáč, S. N. Pakzad, and M. Jahani. 2021. “DynNet:
Çakiroğlu, M. A., and A. A. Süzen. 2020. “Assessment and application of Physics-based neural architecture design for nonlinear structural response
deep learning algorithms in civil engineering.” El-Cezeri J. Sci. Eng. modeling and prediction.” Eng. Struct. 229 (Mar): 111582. https://doi.org
7 (2): 906–922. https://doi.org/10.31202/ecjse.679113. /10.1016/j.engstruct.2020.111582.
Castellon, D. F., A. Fenerci, and O. Øiseth. 2021. “A comparative study of Fang, C., H. Tang, Y. Li, and J. Zhang. 2020. “Stochastic response of a
wind-induced dynamic response models of long-span bridges using ar- cable-stayed bridge under non-stationary winds and waves using differ-
tificial neural networks, support vector regression and buffeting theory.” ent surrogate models.” Ocean Eng. 199 (Mar): 106967. https://doi.org
J. Wind Eng. Ind. Aerodyn. 209 (Feb): 104484. https://doi.org/10.1016/j /10.1016/j.oceaneng.2020.106967.
.jweia.2020.104484. Feng, D., and M. Q. Feng. 2018. “Computer vision for SHM of civil infra-
Çevik, A., A. E. Kurtoğlu, M. Bilgehan, M. E. Gülşan, and H. M. structure: From dynamic response measurement to damage detection–A
Albegmprli. 2015. “Support vector machines in structural engineering: review.” Eng. Struct. 156 (Mar): 105–117. https://doi.org/10.1016/j
A review.” J. Civ. Eng. Manage. 21 (3): 261–281. https://doi.org/10 .engstruct.2017.11.018.
.3846/13923730.2015.1005021. Flah, M., I. Nunez, W. Ben Chaabene, and M. L. Nehdi. 2021. “Machine
Chatterjee, S., S. Sarkar, S. Hore, N. Dey, A. S. Ashour, and V. E. Balas. learning algorithms in civil structural health monitoring: A systematic
2017. “Particle swarm optimization trained neural network for structural review.” Arch. Comput. Methods Eng. 28 (4): 2621–2643. https://doi
failure prediction of multistoried RC buildings.” Neural Comput. Appl. .org/10.1007/s11831-020-09471-9.
28 (8): 2005–2016. https://doi.org/10.1007/s00521-016-2190-2. Gandomi, A. H., X.-S. Yang, S. Talatahari, and A. H. Alavi. 2013. Meta-
Chen, C., L. Tang, Q. Xiao, L. Zhou, H. Wang, Z. Liu, C. Xing, Y. Liu, heuristic applications in structures and infrastructures. Oxford, UK:
J. Chen, and Z. Jiang. 2023. “Unsupervised anomaly detection for Newnes.
long-span bridges combining response forecasting by deep learning Ghannadi, P., and S. S. Kourehli. 2019. “Data-driven method of damage
with Td-MPCA.” Structures 54 (Aug): 1815–1830. https://doi.org/10 detection using sparse sensors installation by SEREPa.” J. Civ. Struct.
.1016/j.istruc.2023.06.033. Health Monit. 9 (4): 459–475. https://doi.org/10.1007/s13349-019
Chen, P., C. Jia, C. Ng, and Z. Hu. 2021. “Application of SADA method -00345-8.
on full-scale measurement data for dynamic responses prediction of Gharehbaghi, S., H. Yazdani, and M. Khatibinia. 2020. “Estimating inelas-
Hywind floating wind turbines.” Ocean Eng. 239 (Mar): 109814. tic seismic response of reinforced concrete frame structures using
https://doi.org/10.1016/j.oceaneng.2021.109814. a wavelet support vector machine and an artificial neural network.”
Cosgun, C. 2023. “Machine learning for the prediction of evaluation of Neural Comput. Appl. 32 (8): 2975–2988. https://doi.org/10.1007
existing reinforced concrete structures performance against earthquakes.” /s00521-019-04075-2.

© ASCE 03124002-20 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Gharehbaghi, V. R., E. Noroozinejad Farsangi, M. Noori, T. Yang, S. Li, Li, C., H. Li, and X. Chen. 2022. “Fast seismic response estimation of tall
A. Nguyen, C. Málaga-Chuquitaype, P. Gardoni, and S. Mirjalili. 2021. pier bridges based on deep learning techniques.” Eng. Struct. 266 (Sep):
“A critical review on structural health monitoring: Definitions, methods, 114566. https://doi.org/10.1016/j.engstruct.2022.114566.
and perspectives.” Arch. Comput. Methods Eng. 29 (Oct): 1–27. https:// Li, H., and J. Ou. 2016. “The state of the art in structural health monitoring
doi.org/10.1007/s11831-021-09665-9. of cable-stayed bridges.” J. Civ. Struct. Health Monit. 6 (1): 43–67.
Ghazvineh, S., G. Nouri, V. Gharehbaghi, S. H. H. Lavasani, E. N. https://doi.org/10.1007/s13349-015-0115-x.
Farsangi, and M. Noori. 2023. “7 Vibration-based damage detection Li, J., Z. He, and X. Zhao. 2021a. “A data-driven building’s seismic re-
using a novel hybrid CNN-SVM approach.” In Data-centric structural sponse estimation method using a deep convolutional neural network.”
health monitoring: Mechanical, aerospace and complex infrastructure IEEE Access 9 (Mar): 50061–50077. https://doi.org/10.1109/ACCESS
systems, 137. Berlin: Walter de Gruyter GmbH & Co KG. https://doi .2021.3065837.
.org/10.1515/9783110791426-007. Li, T., Y. Pan, K. Tong, C. E. Ventura, and C. W. de Silva. 2021b. “Attention-
Ghazvineh, S., G. Nouri, S. H. H. Lavassani, V. Gharehbaghi, and A. based sequence-to-sequence learning for online structural response fore-
Nguyen. 2021. “Application of 2-D convolutional neural networks casting under seismic excitation.” IEEE Trans. Syst. Man Cybern.: Syst.
for damage detection in steel frame structures.” Preprint, submitted 52 (4): 2184–2200. https://doi.org/10.1109/TSMC.2020.3048696.
Liao, Y., R. Lin, R. Zhang, and G. Wu. 2023. “Attention-based LSTM
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

October 29, 2021. http://arxiv.org/abs/2110.15895.

Ghimire, S., P. Guéguen, S. Giffard-Roisin, and D. Schorlemmer. 2022. (AttLSTM) neural network for seismic response modeling of bridges.”
“Testing machine learning models for seismic damage prediction at Comput. Struct. 275 (Jan): 106915. https://doi.org/10.1016/j.compstruc
a regional scale using building-damage dataset compiled after the 2015 .2022.106915.
Gorkha Nepal earthquake.” Earthquake Spectra 38 (4): 2970–2993. Lim, S., and S. Chi. 2019. “Xgboost application on bridge management
https://doi.org/10.1177/87552930221106495. systems for proactive damage estimation.” Adv. Eng. Inf. 41 (Aug):
Guan, X., H. Burton, M. Shokrabadi, and Z. Yi. 2021. “Seismic drift de- 100922. https://doi.org/10.1016/j.aei.2019.100922.
mand estimation for steel moment frame buildings: From mechanics- Limongelli, M. P. 2020. “SHM for informed management of civil structures
based to data-driven models.” J. Struct. Eng. 147 (6): 04021058. and infrastructure.” J. Civ. Struct. Health Monit. 10 (5): 739–741.
https://doi.org/10.1061/(ASCE)ST.1943-541X.0003004. https://doi.org/10.1007/s13349-020-00439-8.
Han, Q., Q. Ma, J. Xu, and M. Liu. 2021. “Structural health monitoring Liu, F., J. Li, and L. Wang. 2023. “PI-LSTM: Physics-informed long short-
research under varying temperature condition: A review.” J. Civ. Struct. term memory network for structural response modeling.” Eng. Struct.
Health Monit. 11 (1): 149–173. https://doi.org/10.1007/s13349-020 292 (Oct): 116500. https://doi.org/10.1016/j.engstruct.2023.116500.
-00444-x. Liu, X., F. Kang, C. Ma, and H. Li. 2021a. “Concrete arch dam behavior
He, J., L. Jiang, L. Jiang, T. Wen, Y. Hu, W. Guo, and J. Sun. 2023. prediction using kernel-extreme learning machines considering thermal
“Estimation of blast-induced peak response of concrete-filled double- effect.” J. Civ. Struct. Health Monit. 11 (2): 283–299. https://doi.org/10
skin tube columns by intelligence-based technique.” Thin-Walled .1007/s13349-020-00452-x.
Struct. 186 (Mar): 110670. https://doi.org/10.1016/j.tws.2023.110670. Liu, Z., L. Zhang, J. Li, and M. Mamluki. 2021b. “Predicting the seismic
response of the short structures by considering the whale optimization
Huang, P., and Z. Chen. 2021. “Deep learning for nonlinear seismic re-
algorithm.” Energy Rep. 7 (Nov): 4071–4084. https://doi.org/10.1016/j
sponses prediction of subway station.” Eng. Struct. 244 (Oct): 112735.
.egyr.2021.06.095.
https://doi.org/10.1016/j.engstruct.2021.112735.
Lu, P., S. Chen, and Y. Zheng. 2012. “Artificial intelligence in civil engi-
Hung, S.-L., and H. Adeli. 1994. “Object-oriented backpropagation and its
neering.” Math. Probl. Eng. 2012 (Dec): 145974. https://doi.org/10
application to structural design.” Neurocomputing 6 (1): 45–55. https://
.1155/2012/145974.
doi.org/10.1016/0925-2312(94)90033-7.
Lu, X., Y. Xu, Y. Tian, B. Cetiner, and E. Taciroglu. 2021. “A deep learning
Jiang, S., L. Zhao, and C. Du. 2022. “Structural deformation prediction
approach to rapid regional post-event seismic damage assessment using
model based on extreme learning machine algorithm and particle swarm
time-frequency distributions of ground motions.” Earthquake Eng.
optimization.” Struct. Health Monit. 21 (6): 2786–2803. https://doi.org
Struct. Dyn. 50 (6): 1612–1627. https://doi.org/10.1002/eqe.3415.
/10.1177/14759217211072237.
Mangalathu, S., and J.-S. Jeon. 2018. “Classification of failure mode and
Junda, E., C. Málaga-Chuquitaype, and K. Chawgien. 2023. “Interpretable prediction of shear strength for reinforced concrete beam-column joints
machine learning models for the estimation of seismic drifts in CLT using machine learning techniques.” Eng. Struct. 160 (Apr): 85–94.
buildings.” J. Build. Eng. 70 (Jun): 106365. https://doi.org/10.1016/j https://doi.org/10.1016/j.engstruct.2018.01.008.
.jobe.2023.106365. Manzoor, B., I. Othman, S. Durdyev, S. Ismail, and M. H. Wahab. 2021.
Kazemi, F., N. Asgarkhani, and R. Jankowski. 2023. “Predicting seismic “Influence of artificial intelligence in civil engineering toward sustain-
response of SMRFs founded on different soil types using machine able development–A systematic literature review.” Appl. Syst. Innova-
learning techniques.” Eng. Struct. 274 (Jan): 114953. https://doi.org/10 tion 4 (3): 52. https://doi.org/10.3390/asi4030052.
.1016/j.engstruct.2022.114953. Martinez, P., M. Al-Hussein, and R. Ahmad. 2019. “A scientometric analy-
Kim, D., K. Kwon, K. Pham, J.-Y. Oh, and H. Choi. 2022. “Surface sis and critical review of computer vision applications for construction.”
settlement prediction for urban tunneling using machine learning Autom. Constr. 107 (Mar): 102947. https://doi.org/10.1016/j.autcon.2019
algorithms with Bayesian optimization.” Autom. Constr. 140 (Aug): .102947.
104331. https://doi.org/10.1016/j.autcon.2022.104331. Mata, J., F. Salazar, J. Barateiro, and A. Antunes. 2021. “Validation of ma-
Kim, S., and T. Kim. 2022. “Machine-learning-based prediction of vortex- chine learning models for structural dam behaviour interpretation and
induced vibration in long-span bridges using limited information.” prediction.” Water 13 (19): 2717. https://doi.org/10.3390/w13192717.
Eng. Struct. 266 (Sep): 114551. https://doi.org/10.1016/j.engstruct Mehrabi, P., S. Honarbari, S. Rafiei, S. Jahandari, and M. Alizadeh Bidgoli.
.2022.114551. 2021. “Seismic response prediction of FRC rectangular columns using
Kumar, S., K. Patel, S. Chaudhary, and A. Nagpal. 2021. “Rapid prediction intelligent fuzzy-based hybrid metaheuristic techniques.” J. Ambient In-
of long-term deflections in steel-concrete composite bridges through a tell. Hum. Comput. 12 (11): 10105–10123. https://doi.org/10.1007
neural network model.” Int. J. Steel Struct. 21 (2): 590–603. https://doi /s12652-020-02776-4.
.org/10.1007/s13296-021-00458-1. Morfidis, K., and K. Kostinakis. 2017. “Seismic parameters’ combinations
Lake, B. M., R. Salakhutdinov, and J. B. Tenenbaum. 2015. “Human-level for the optimum prediction of the damage state of R/C buildings using
concept learning through probabilistic program induction.” Science neural networks.” Adv. Eng. Software 106 (Apr): 1–16. https://doi.org
350 (6266): 1332–1338. https://doi.org/10.1126/science.aab3050. /10.1016/j.advengsoft.2017.01.001.
Lee, J. H., J. Shin, and M. J. Realff. 2018. “Machine learning: Overview of Mousavi, M., A. TohidiFar, and A. Alvanchi. 2022. “BIM and machine
the recent progresses and implications for the process systems engineer- learning in seismic damage prediction for non-structural exterior infill
ing field.” Comput. Chem. Eng. 114 (Jun): 111–121. https://doi.org/10 walls.” Autom. Constr. 139 (Jul): 104288. https://doi.org/10.1016/j
.1016/j.compchemeng.2017.10.008. .autcon.2022.104288.

© ASCE 03124002-21 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Naser, M. Z., S. Thai, and H.-T. Thai. 2021. “Evaluating structural response Thacker, S., D. Adshead, M. Fay, S. Hallegatte, M. Harvey, H. Meller, N.
of concrete-filled steel tubular columns through machine learning.” O’Regan, J. Rozenberg, G. Watkins, and J. W. Hall. 2019. “Infrastructure
J. Build. Eng. 34 (Feb): 101888. https://doi.org/10.1016/j.jobe.2020 for sustainable development.” Nat. Sustainability 2 (4): 324–331. https://
.101888. doi.org/10.1038/s41893-019-0256-8.
Nguyen, H., H. Moayedi, L. K. Foong, H. A. H. Al Najjar, W. A. W. Jusoh, Thai, H.-T. 2022. “Machine learning for structural engineering: A state-of-
A. S. A. Rashid, and J. Jamali. 2020. “Optimizing ANN models with the-art review.” Structures 38 (Apr): 448–491. https://doi.org/10.1016/j
PSO for predicting short building seismic response.” Eng. Comput. .istruc.2022.02.003.
36 (3): 823–837. https://doi.org/10.1007/s00366-019-00733-0. Thaler, D., M. Stoffel, B. Markert, and F. Bamer. 2021. “Machine-learning-
Nguyen, H. D., N. D. Dao, and M. Shin. 2021. “Prediction of seismic drift enhanced tail end prediction of structural response statistics in earth-
responses of planar steel moment frames using artificial neural network quake engineering.” Earthquake Eng. Struct. Dyn. 50 (8): 2098–2114.
and extreme gradient boosting.” Eng. Struct. 242 (Mar): 112518. https:// https://doi.org/10.1002/eqe.3432.
doi.org/10.1016/j.engstruct.2021.112518. Torky, A. A., and S. Ohno. 2021. “Deep learning techniques for predicting
Ning, C., Y. Xie, and L. Sun. 2023. “LSTM, WaveNet, and 2D CNN for nonlinear multi-component seismic responses of structural buildings.”
nonlinear time history prediction of seismic responses.” Eng. Struct. Comput. Struct. 252 (Aug): 106570. https://doi.org/10.1016/j.compstruc
286 (Jul): 116083. https://doi.org/10.1016/j.engstruct.2023.116083.
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

.2021.106570.
Oh, B. K., and J. Kim. 2021. “Optimal architecture of a convolutional neu- Van Eck, N., and L. Waltman. 2010. “Software survey: VOSviewer, a
ral network to estimate structural responses for safety evaluation of the computer program for bibliometric mapping.” Scientometrics 84 (2):
structures.” Measurement 177 (Jun): 109313. https://doi.org/10.1016/j 523–538. https://doi.org/10.1007/s11192-009-0146-3.
.measurement.2021.109313. Vanluchene, R., and R. Sun. 1990. “Neural networks in structural engineer-
Oh, B. K., H. S. Park, and B. Glisic. 2023. “Time-dependent structural ing.” Comput.-Aided Civ. Infrastruct. Eng. 5 (3): 207–215. https://doi
response estimation method for concrete structures using time informa- .org/10.1111/j.1467-8667.1990.tb00377.x.
tion and convolutional neural networks.” Eng. Struct. 275 (Jan): Wang, X., Z. Li, and A. Shafieezadeh. 2021b. “Seismic response prediction
115193. https://doi.org/10.1016/j.engstruct.2022.115193. and variable importance analysis of extended pile-shaft-supported
Oh, B. K., Y. Park, and H. S. Park. 2020. “Seismic response prediction bridges against lateral spreading: Exploring optimized machine learning
method for building structures using convolutional neural network.” models.” Eng. Struct. 236 (Jun): 112142. https://doi.org/10.1016/j
Struct. Control Health Monit. 27 (5): e2519. https://doi.org/10.1002/stc .engstruct.2021.112142.
.2519.
Wang, X., Y. Liu, and H. Xin. 2021a. “Bond strength prediction of concrete-
Oraee, M., M. R. Hosseini, E. Papadonikolaki, R. Palliyaguru, and M.
encased steel structures using hybrid machine learning method.” Struc-
Arashpour. 2017. “Collaboration in BIM-based construction networks:
tures 32 (Aug): 2279–2292. https://doi.org/10.1016/j.istruc.2021.04.018.
A bibliometric-qualitative literature review.” Int. J. Project Manage.
Wang, X., C. Miao, and X. Wang. 2021c. “Prediction analysis of deflection
35 (7): 1288–1301. https://doi.org/10.1016/j.ijproman.2017.07.001.
in the construction of composite box-girder bridge with corrugated
Osman, A. I. A., A. N. Ahmed, M. F. Chow, Y. F. Huang, and A. El-Shafie.
steel webs based on MEC-BP neural networks.” Structures 32 (Aug):
2021. “Extreme gradient boosting (Xgboost) model to predict the
691–700. https://doi.org/10.1016/j.istruc.2021.03.011.
groundwater levels in Selangor Malaysia.” Ain Shams Eng. J. 12 (2):
Wen, W., C. Zhang, and C. Zhai. 2022. “Rapid seismic response prediction
1545–1556. https://doi.org/10.1016/j.asej.2020.11.011.
of RC frames based on deep learning and limited building information.”
Park, H. S., T. Hong, D.-E. Lee, B. K. Oh, and B. Glisic. 2022. “Long-term
Eng. Struct. 267 (Sep): 114638. https://doi.org/10.1016/j.engstruct
structural response prediction models for concrete structures using
.2022.114638.
weather data, fiber-optic sensing, and convolutional neural network.”
Expert Syst. Appl. 201 (Sep): 117152. https://doi.org/10.1016/j.eswa Wong, W., and C. I. Ming. 2019. “A review on metaheuristic algorithms:
.2022.117152. Recent trends, benchmarking and applications.” In Proc., 2019 7th Int.
Parsa, P., and H. Naderpour. 2021. “Shear strength estimation of reinforced Conf. on Smart Computing & Communications (ICSCC), 1–5. New York:
concrete walls using support vector regression improved by Teaching– IEEE.
learning-based optimization, Particle Swarm optimization, and Harris Woo, S., J. Park, J. Park, and L. Manuel. 2019. “Wind field-based short-
Hawks Optimization algorithms.” J. Build. Eng. 44 (Dec): 102593. term turbine response forecasting by stacked dilated convolutional
https://doi.org/10.1016/j.jobe.2021.102593. LSTMs.” IEEE Trans. Sustainable Energy 11 (4): 2294–2304. https://
Peng, H., J. Yan, Y. Yu, and Y. Luo. 2012. “Time series estimation based on doi.org/10.1109/TSTE.2019.2954107.
deep Learning for structural dynamic nonlinear prediction.” Structures Wu, T.-Y., R.-T. Wu, P.-H. Wang, T.-K. Lin, and K.-C. Chang. 2023.
29 (Feb): 1016–1031. https://doi.org/10.1016/j.istruc.2020.11.049. “Development of a high-fidelity failure prediction system for reinforced
Pereira, M., and B. Glisic. 2022. “A hybrid approach for prediction of long- concrete bridge columns using generative adversarial networks.” Eng.
term behavior of concrete structures.” J. Civ. Struct. Health Monit. Struct. 286 (Jul): 116130. https://doi.org/10.1016/j.engstruct.2023.116130.
12 (4): 891–911. https://doi.org/10.1007/s13349-022-00582-4. Xie, Y., M. Ebad Sichani, J. E. Padgett, and R. DesRoches. 2020. “The
Petersen, Ø. W., O. Øiseth, and E.-M. Lourens. 2019. “The use of inverse promise of implementing machine learning in earthquake engineering:
methods for response estimation of long-span suspension bridges with A state-of-the-art review.” Earthquake Spectra 36 (4): 1769–1801.
uncertain wind loading conditions.” J. Civ. Struct. Health Monit. 9 (1): https://doi.org/10.1177/8755293020919419.
21–36. https://doi.org/10.1007/s13349-018-0319-y. Xu, J., D. Dang, Q. Ma, X. Liu, and Q. Han. 2022. “A novel and robust data
Radovanovic, S., V. Milivojevic, D. Divac, and N. Milivojevic. 2015. anomaly detection framework using LAL-AdaBoost for structural
“Prediction of concrete dam deformation using artificial neural net- health monitoring.” J. Civ. Struct. Health Monit. 12 (2): 305–321.
work.” In Proc., 4th Int. Conf. on Soft Computing Technology in Civil, https://doi.org/10.1007/s13349-021-00544-2.
Structural and Environmental Engineering, 1–4. Prague, Czech Repub- Ye, X., T. Jin, and C. Yun. 2019. “A review on deep learning-based struc-
lic: Civil-Comp Press. tural health monitoring of civil infrastructures.” Smart Struct. Syst
Soleimani-Babakamali, M. H., and M. Z. Esteghamati. 2022. “Estimating 24 (5): 567–585. https://doi.org/10.12989/sss.2019.24.5.567.
seismic demand models of a building inventory from nonlinear static Zhang, D., Y. Chen, C. Zhang, G. Xue, J. Zhang, M. Zhang, L. Wang, and
analysis using deep learning methods.” Eng. Struct. 266 (Sep): 114576. N. Li. 2023a. “Prediction of seismic acceleration response of precast
https://doi.org/10.1016/j.engstruct.2022.114576. segmental self-centering concrete filled steel tube single-span bridges
Spencer, B. F., Jr., V. Hoskere, and Y. Narazaki. 2019. “Advances in based on machine learning method.” Eng. Struct. 279 (Mar): 115574.
computer vision-based civil infrastructure inspection and monitoring.” https://doi.org/10.1016/j.engstruct.2022.115574.
Engineering 5 (2): 199–222. https://doi.org/10.1016/j.eng.2018.11.030. Zhang, K., H.-M. Lyu, S.-L. Shen, A. Zhou, and Z.-Y. Yin. 2020a. “Evolu-
Tadesse, Z., K. Patel, S. Chaudhary, and A. Nagpal. 2012. “Neural net- tionary hybrid neural network approach to predict shield tunneling-
works for prediction of deflection in composite bridges.” J. Constr. induced ground settlements.” Tunnelling Underground Space Technol.
Steel Res. 68 (1): 138–149. https://doi.org/10.1016/j.jcsr.2011.08.003. 106 (Dec): 103594. https://doi.org/10.1016/j.tust.2020.103594.

© ASCE 03124002-22 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Zhang, P., H. Li, Q. P. Ha, Z.-Y. Yin, and R.-P. Chen. 2020b. loads.” Eng. Struct. 273 (Dec): 115104. https://doi.org/10.1016/j.engstruct
“Reinforcement learning based optimizer for improvement of predicting .2022.115104.
tunneling-induced ground responses.” Adv. Eng. Inf. 45 (Mar): 101097. Zhao, W., P. Chen, X. Liu, and L. Wang. 2022b. “Impact response predic-
https://doi.org/10.1016/j.aei.2020.101097. tion and optimization of SC walls using machine learning algorithms.”
Zhang, R., Z. Chen, S. Chen, J. Zheng, O. Büyüköztürk, and H. Sun. 2019. Structures 45 (Nov): 390–399. https://doi.org/10.1016/j.istruc.2022
“Deep long short-term memory networks for nonlinear structural seis- .09.036.
mic response prediction.” Comput. Struct. 220 (Aug): 55–68. https://doi Zhou, J., S. Zhu, Y. Qiu, D. J. Armaghani, A. Zhou, and W. Yong. 2022.
.org/10.1016/j.compstruc.2019.05.006. “Predicting tunnel squeezing using support vector machine optimized
Zhang, R., Y. Liu, and H. Sun. 2020c. “Physics-guided convolutional neu- by whale optimization algorithm.” Acta Geotech. 17 (4): 1343–1366.
ral network (PhyCNN) for data-driven seismic response modeling.” https://doi.org/10.1007/s11440-022-01450-7.
Eng. Struct. 215 (Jul): 110704. https://doi.org/10.1016/j.engstruct.2020 Zhu, J., and W. Zhang. 2016. “Numerical simulation of wind and wave
.110704. fields for coastal slender bridges.” J. Bridge Eng. 22 (3): 04016125.
Zhang, R., L. Meng, Z. Mao, and H. Sun. 2021. “Spatiotemporal deep https://doi.org/10.1061/(ASCE)BE.1943-5592.0001002.
learning for bridge response forecasting.” J. Struct. Eng. 147 (6): Zhu, S., M. Ohsaki, and X. Guo. 2021. “Prediction of non-linear buckling
04021070. https://doi.org/10.1061/(ASCE)ST.1943-541X.0003022. load of imperfect reticulated shell using modified consistent imperfec-
Downloaded from ascelibrary.org by UCLA Digital Coll Svcs on 03/18/24. Copyright ASCE. For personal use only; all rights reserved.

Zhang, Y., Z. Wang, H. Kuang, F. Fu, and A. Yu. 2023b. “Prediction tion and machine learning.” Eng. Struct. 226 (Jan): 111374. https://doi
of surface settlement in shield-tunneling construction process using .org/10.1016/j.engstruct.2020.111374.
PCA-PSO-RVM machine learning.” J. Perform. Constr. Facil. 37 (3): Zinno, R., S. S. Haghshenas, G. Guido, K. Rashvand, A. Vitale, and A.
04023012. https://doi.org/10.1061/JPCFEV.CFENG-4363. Sarhadi. 2022. “The state of the art of artificial intelligence approaches
Zhao, C., Y. Zhu, and Z. Zhou. 2022a. “Machine learning-based ap- and new technologies in structural health monitoring of bridges.” Appl.
proaches for predicting the dynamic response of RC slabs under blast Sci. 13 (1): 97. https://doi.org/10.3390/app13010097.

© ASCE 03124002-23 Pract. Period. Struct. Des. Constr.

Pract. Period. Struct. Des. Constr., 2024, 29(3): 03124002