Review

Application of Artificial Intelligence to Support Design and

Analysis of Steel Structures
Sina Sarfarazi , Ida Mascolo * , Mariano Modano and Federico Guarracino *

Department of Structural Engineering, University of Naples “Federico II”, 80125 Naples, Italy;
[email protected] (S.S.); [email protected] (M.M.)
* Correspondence: [email protected] (I.M.); [email protected] (F.G.)

Abstract: In steel structural engineering, artificial intelligence (AI) and machine learning
(ML) are improving accuracy, efficiency, and automation. This review explores AI-driven
approaches, emphasizing how AI models improve predictive capabilities, optimize perfor-
mance, and reduce computational costs compared to traditional methods. Inverse Machine
Learning (IML) is a major focus since it helps engineers to minimize reliance on itera-
tive trial-and-error by allowing them to identify ideal material properties and geometric
configurations depending on predefined performance targets. Unlike conventional ML
models that focus mostly on forward predictions, IML helps data-driven design generation,
enabling more adaptive engineering solutions. Furthermore, underlined is Explainable
Artificial Intelligence (XAI), which enhances model transparency, interpretability, and trust
of AI. The paper categorizes AI applications in steel construction based on their impact
on design automation, structural health monitoring, failure prediction and performance
evaluation throughout research from 1990 to 2025. The review explores challenges such as
data limitations, model generalization, engineering reliability, and the need for physics-
informed learning while examining AI’s role in bridging research and real-world structural
applications. By integrating AI into structural engineering, this work supports the adoption
of ML, IML, and XAI in structural analysis and design, paving the way for more reliable
and interpretable engineering practices.

Academic Editors: Wuyi Ming,

Keywords: machine learning; inverse machine learning; explainable AI; AI-based
Xiaoke Li and Wenbin He performance evaluation; data-driven design automation; steel structures
Received: 27 February 2025
Revised: 27 March 2025
Accepted: 2 April 2025
Published: 4 April 2025 1. Introduction
Citation: Sarfarazi, S.; Mascolo, I.; Structural engineering depends on both exact design specification and extensive anal-
Modano, M.; Guarracino, F. ysis to ensure that infrastructure and buildings satisfy safety and performance criteria.
Application of Artificial Intelligence to Although conventional structural engineering depends on analytical calculations, experi-
Support Design and Analysis of Steel
mental testing, and numerical simulations, modern developments demand a change toward
Structures. Metals 2025, 15, 408.
https://doi.org/10.3390/
data-driven methods and automation. Like they once moved from hand calculations to
met15040408 computational tools like finite element (FE) analysis, engineers have to adopt fresh ap-
proaches to solve problems as technology develops. The increasing application of artificial
Copyright: © 2025 by the authors.
Licensee MDPI, Basel, Switzerland.
intelligence (AI) and machine learning (ML) in structural engineering [1–3] presents a
This article is an open access article similar challenge and demands experts to become familiar with AI-driven tools. These new
distributed under the terms and technologies reinforce decision-making with adaptive, data-supported solutions rather than
conditions of the Creative Commons replacing conventional knowledge; they are instead extensions of engineering intuition [4].
Attribution (CC BY) license
Structural engineering guarantees design and analysis accuracy by means of methodical
(https://creativecommons.org/
processes and uniform guidelines. These approaches have been included into software
licenses/by/4.0/).

Metals 2025, 15, 408 https://doi.org/10.3390/met15040408

Metals 2025, 15, 408 2 of 41

tools over time to automatically complete compliance checks and regular computations.
Large-scale steel buildings, however, often call for high-performance workstations and li-
censed tools since they demand significant computational resources, specialized simulation
software, and expert oversight.
Among the deterministic, physics-based models supporting most conventional struc-
tural engineering approaches are reliability-based optimization frameworks, Computa-
tional Fluid Dynamics (CFD), and the Finite Element Method (FEM). These reliable methods
provide engineers with complete awareness of mechanical behavior, stress distribution,
failure modes, and safety margins. However, the application of structural systems be-
comes more challenging and computationally expensive as they get larger and include
nonlinearities, material heterogeneity, or advanced loading scenarios. Even with strong
simulation tools, each analysis iteration requires manual intervention, time, and expert
interpretation—factors that can slow down the design and optimization process. Another
paradigm comes from AI—in particular from ML. ML models learn from patterns in past
datasets, derived from experiments, simulations, or hybrid sources, instead of solving
governing equations repeatedly. Once trained, these models can save computational time,
let engineers rapidly investigate large design areas, and offer almost instantaneous predic-
tions for new designs. This capacity underlies early-stage design, parametric sensitivity
studies, and multi-objective optimization where speed and adaptability are particularly
critical. From a structural perspective, ML techniques have been used to precisely predict
material strength, buckling resistance, joint behavior, and global system performance. From
the AI side, the use of surrogate modeling, ensemble learning, and neural networks has
made it possible to capture highly nonlinear relationships that are difficult to express
analytically. Acting as a fast, approximative evaluator, trained ML models enable real-time
structural assessment, uncertainty-based optimization, or integration with digital twin
systems. This synergy between physics-based rigor and data-driven intelligence marks a
potential revolution in how future structures are conceived, assessed, and optimized.
Analyzing a structure and designing it, have different uses and needs for different
skill sets in structural engineering. Analysis is mostly concerned with how a structure
responds to external forces; design is the process of developing answers that fit performance
criteria and practical limitations. These two disciplines sometimes have different logical
frameworks. Analysis depends on scientific ideas and deterministic models while design
is adaptive, iterative, and affected by outside events including material availability, cost,
and safety rules. One challenge of engineering education is that students are mostly taught
mathematical and computational methods, which stress strict formulations and simulations.
However, when they enter the industry, they must deal with real-world design problems
that require creative problem-solving, adaptability, and decision-making under uncertainty.
This gap emphasizes the need of bridging analytical precision with design intuition so
that engineers may properly move from knowledge of structural behavior to developing
practical, efficient, creative designs that satisfy real-world needs [5].
Particularly in inverse design, ML is becoming more popular since it lets engineers
directly decide ideal material properties, geometric layouts, and performance criteria
depending on predefined structural goals [4,6]. Inverse Machine Learning (IML) allows a
data-driven approach to quickly find the best design solutions in steel constructions, not
only iteratively improving designs through simulations.
Apart from predictive powers, the growing complexity of ML models has generated
questions about interpretability and transparency in structural engineering uses [7]. Ex-
plainable machine learning (XML) has become increasingly important in order to meet
these difficulties since it guarantees that ML-driven models produce justifiable, clear, and
understandable results [8]. By means of explainability techniques including feature impor-
Metals 2025, 15, 408 3 of 41

tance analysis and SHapley Additive exPlanations (SHAP) [9], engineers can better grasp
the contribution of various design parameters in decision-making, rendering ML-based
structural analysis more reliable and useful. The identification of possible biases, validation
of model reliability, and enhancement of decision-making confidence in safety-critical
applications including load-bearing assessments, failure predictions, and material selection
also quite heavily depend on XML techniques [10].
Typically addressing different engineering challenges, three main categories define
ML techniques: supervised learning, unsupervised learning, and reinforcement learning.
Supervised learning, which relies on labeled datasets, is particularly effective for regression
tasks such as estimating material properties and classification problems like detecting struc-
tural damage [11]. Conversely, unsupervised learning is useful for predictive maintenance
and structural design optimization since it finds hidden trends in engineering data without
depending on predefined labels [12]. Though less often used, reinforcement learning is
very important for decision-making under uncertainty—especially in dynamic structural
analysis [13].
Table 1 provides a comparative overview of these three categories—highlighting their
underlying logic, typical algorithms, structural applications, and key limitations.

Table 1. Overview of main ML categories.

Applications in
Common
ML Category Core Principle Structural Limitations
Algorithms
Engineering
Material property Requires large,
Learns from labeled Linear Regression,
estimation, damage labeled datasets;
Supervised Learning datasets to predict Random Forest,
classification, load limited in
outcomes ANN
prediction extrapolation
Structural health
Finds hidden Results may lack
monitoring, design
Unsupervised patterns or K-means, PCA, clear interpretation;
clustering, pattern
Learning groupings in Autoencoders requires expert
discovery in sensor
unlabeled data analysis
data
Real-time structural
Learns via
control, High computational
trial-and-error
Reinforcement Q-learning, Deep decision-making cost; limited
interactions to
Learning Q-Networks under uncertainty, adoption due to
maximize rewards
adaptive load training complexity
over time
redistribution

In structural engineering, ML first emerged in the late 1980s when researchers started
working on design tasks using artificial neural networks (ANN) [14]. Recent advances in
data-driven design methods, ML-driven predictive modeling, and optimization techniques
have expanded its opportunities. Although more and more studies on AI in civil and
structural engineering are available, there is still a dearth of targeted studies methodically
analyzing ML uses in steel constructions. Emphasizing improved accuracy, efficiency, and
automation, this review shows how AI is transforming steel structure engineering and
synthesizing recent developments from 1994 to 2025. Previous studies have mostly focused
on particular AI techniques in structural engineering, including pattern recognition for
structural health monitoring, genetic algorithms in optimization problems, ML applications
in concrete structures, and steel–concrete composite structures. Although these works offer
insightful analysis, their scope is usually more general and they do not particularly address
the function of ML in steel structure engineering.
Metals 2025, 15, 408 4 of 41

This review paper provides a concentrated study of AI applications in design and

analysis of steel structures. It also underlines the increasing relevance of IML, which allows
engineers to directly choose optimal material properties and design criteria independent
of iterative trial-and-error processes. By learning complicated interactions between input
parameters and structural behavior, ML models can accelerate the design process, increase
prediction accuracy, and maximize structural performance unlike conventional simulation-
based techniques. This review provides a basic resource for researchers and engineers
by aggregating and critically evaluating the most recent ML-driven innovations in steel
structure engineering, thus bridging the gap between conventional computing methods
and modern AI-enhanced approaches. The aim of this work is to enable the acceptance of
ML, IML, and XML in steel structure design and analysis so guaranteeing more effective,
transparent, and data-driven engineering solutions.

Objectives of This Review

• Providing a Focused Overview of ML in Steel Structure Engineering:
This paper attempts to give a thorough study of the part ML plays in the design and
analysis of steel buildings. This work will show how ML-driven approaches are improving
structural performance prediction, material selection, load-bearing capacity estimate, and
failure assessment by aggregating significant developments. The paper will also discuss
how ML is changing conventional design methods, thus enabling structural engineering to
be more efficient, data-driven, and optimized.
• Exploring the Role of IML in Steel Structure Design:
This work tries to show a thorough investigation of the part ML does in the design
and analysis of steel buildings. IML allows engineers, free from trial-and-error methods,
to directly choose ideal design parameters including geometric configurations, material
properties, and performance criteria unlike conventional iterative approaches. This review
will examine how application of IML is improving steel structure optimization.
• Investigating XML for Transparent Structural Engineering
The growing complexity of ML models raises questions about interpretability and
confidence in structural engineering uses. This paper will review XML techniques to show
how engineers might better understand and validate AI-driven predictions. By means
of better model transparency, XML can help to identify significant influencing factors,
guarantee safety compliance, and improve reliability in steel analysis and design.
• Addressing Challenges, Limitations, and Future Directions
Even if ML has great promise for structural engineering, several issues have to be
resolved before general acceptance. Important questions including data shortage, model
interpretability, generalization, and integration of physics-informed learning to improve
engineering accuracy are investigated in this review. Moreover, covered are problems with
feature engineering, uncertainty quantification, and regulatory approval of ML-driven
models in structural safety assessments. Future research directions are also under discus-
sion, underlining the need of hybrid AI-physics models, industry-driven validation, and
responsible AI application in design and analysis processes.

2. Research Data Extraction Process

This research evaluates steel building design and analysis through bibliometric meth-
ods and content analysis to investigate the implementation of AI techniques.
Scopus’s extensive scholarly coverage enabled this study to compile a comprehensive
dataset for examining the impact of AI on the design and analysis of steel structures. The
Metals 2025, 15, 408 5 of 41

selected papers included applications of ML, IML, deep learning, fuzzy logic, genetic algo-
rithms, optimization, surrogate modeling, and XML across various domains of structural
engineering. Special attention was given to steel building components such as cold-formed
steel, stainless steel, high-strength steel, beams, columns, plates, trusses, connections, and
full steel frames. The dataset also included research on design strategies—such as structural
optimization, topology optimization, reliability analysis, uncertainty quantification, and
performance-based design—and structural analysis techniques including FEM, numerical
simulations, structural health monitoring, load capacity evaluation, fatigue assessment,
buckling behavior, seismic performance, wind load effects, and fire resistance.
The following inclusion criteria ensured both scientific quality and thematic relevance:
only peer-reviewed journal articles were selected; the search was limited to publications
between 1994 and 2025; all chosen papers explicitly focused on the application of artificial
intelligence in steel structural engineering, excluding studies centered solely on concrete,
timber, or hybrid systems; and the selected studies were required to demonstrate either
methodological innovation or practical relevance in structural modeling, performance
prediction, or design optimization. A keyword-based search was conducted in Scopus
using combinations of terms such as “steel structures”, “machine learning”, “artificial
intelligence”, “predictive modeling”, and “structural optimization”. Duplicate records,
non-English publications, editorials, and studies without accessible full texts were excluded.

Overview of Research Contributions in AI-Driven Steel Structure Studies

A two-stage screening process was applied. In the first stage, papers were screened
depending on title, abstract, and keywords to ensure relevance. The second stage involved
a thorough review to select studies focusing on ML-driven structural analysis and design
strategies. The bibliometric analysis identified 2291 English-language publications, as
shown in Table 2, highlighting the growing scholarly attention on AI’s role in structural
engineering. This collection shows a wide international collaboration, with contributions
from 1277 researchers affiliated with 1367 institutions across 85 countries. Covering from
1994 to 2025, the dataset captures three decades of increasing interest in AI-driven structural
performance evaluation and optimization. Of the collected works published across 159
scientific journals, 2227 are research articles and 64 are review papers. The citation impact is
also noteworthy, with a total of 50,893 citations and an average of 22.21 citations per paper.
An annual citation rate of 2035.72 further emphasizes the rapid acceptance and industrial
relevance of AI-based technologies in steel construction engineering.

Table 2. Core Data Insights.

Detailed Overview Outcomes

Citations 50,893
Authors 1277
Organization 1367
Countries 85
Journal 159
Documents 2291
Average citations per year 2035.72
Average citations per document 22.21
Time span 1994–2025

The Citation Overview graph (Figure 1) displays the annual publication and citation
distribution in the field of AI-driven steel structure research. From 2000 to about 2015, the
results show a slow but steady increase in research activity despite rather low numbers
of publications and citations. Starting in 2015, however, both measures clearly speed
 Metals 2025, 15, 408 6 of 41

forward, a reflection of structural engineering’s increasing integration of AI and data-

driven approaches. This trend fits world developments in computational modeling, deep
025, 15, x FOR PEER REVIEW learning, and ML, which have made AI-based solutions more pragmatic 6 ofand
41 successful for
engineering use.

Figure 1. Citation and publication

Figure 1. Citationtrends over time. Bar
and publication colorover
trends intensity
time.reflects the number
Bar color intensityofreflects
pub- the number of
lications (darkerpublications
= more publications).
(darker = more publications).

Figure 2 showsPublications
the publication trends
pointing of AI-driven
a shift toward AIstructural
acceptanceengineering research
in the field began to appear around
across several journals from 2000 to 2024. Among the most prominent
2018. This development confirms that the research community is placing sources, only the
increasing empha-
top five journalssis with
on AIthe highest number
applications of publications
in structural are considered:
analysis, optimization, andEngineering
design. Interest continues
Structures, Structures, Automation in Construction, Journal of Building Engineering,
to rise, reaching its highest levels between 2022 and 2024. Citation counts follow a similar
and
Thin-Walled Structures. These journals are important venues for disseminating innovative
trend, showing a marked increase from 2015 onward and peaking in 2024. This trend
research, since their major contributions to the development of AI applications in struc-
underlines both the growing academic weight of AI-related research and its relevance in
tural engineering have played a major contribution in advancing the field. The statistics
show a slow riseaddressing
in research challenging
activity inengineering
the early problems.
years followed by a sudden surge in
publications after 2015,
Figure suggesting
2 shows thean increasing
publicationinterest
trends of in AI-driven
AI applications within
structural this
engineering research
field. Engineering Structures and Structures show the most notable rise among the
across several journals from 2000 to 2024. Among the most prominent sources, only the
sources, especially starting
top five fromwith
journals 2020the forward.
highest This trend of
number implies that in steel
publications structure Engineering
are considered:
engineering, these publications
Structures, now serve
Structures, as maininvenues
Automation for publishing
Construction, Journalresearch
of Buildingon AI.
Engineering, and
Particularly noteworthy
Thin-Walled is the fast rise in
Structures. publications
These journals arewithin Structures
important after
venues for2020, which
disseminating innovative
reflects a recent surge in
research, research
since contributions
their major mosttolikely
contributions driven by the
the development growing
of AI ac- in structural
applications
ceptance of ML, optimization
engineering havestrategies,
played a majorand computational
contribution inmodeling
advancing inthe
structural
field. Thede-
statistics show a
sign. slow rise in research activity in the early years followed by a sudden surge in publications
Though at a rather slower pace, other journals including Automation in Construction,
after 2015, suggesting an increasing interest in AI applications within this field. Engineering
Journal of Building Engineering, and Thin-Walled Structures—also exhibit a consistent
Structures and Structures show the most notable rise among the sources, especially starting
upward trend in AI-related publications. Particularly from 2018 onward, these journals
from 2020as
have gained popularity forward.
AI-based This trend implies
approaches are that in steel structure
increasingly applied engineering,
across multiple these publications
now serve as main venues for publishing research on
domains, including structural design, optimization, and performance assessment. This AI. Particularly noteworthy is the fast
steady rise in publication frequency
rise in publications suggests
within that AI after
Structures methods2020,are expanding
which reflectsbeyond
a recenttra-
surge in research
ditional design principles into broader applications, such as structural resilience evalua-
contributions most likely driven by the growing acceptance of ML, optimization strategies,
tion, real-time monitoring,
and computationaland automated
modelingconstruction.
in structural design.
Metals2025,
Metals 15,408
2025,15, x FOR PEER REVIEW 77 of 41
of 41

Figure2.2.Annual
Figure Annualpublication
publicationtrends
trendsininAI-driven
AI-drivenstructural
structuralengineering
engineering across
across thethe top
top five
five contrib-
contribut-
uting
ing journals
journals by by volume
volume (2000–2024).
(2000–2024).

Figure 3atshows
Though theslower
a rather worldwide distribution
pace, other journalsofincluding
structuralAutomation
engineeringinresearch moti-
Construction,
vated byofAI
Journal and identifies
Building important
Engineering, andcontributing
Thin-Walled nations. Reflecting its
Structures—also stronga investment
exhibit consistent
in data-driven
upward trend inapproaches
AI-relatedand computational
publications. technologies,
Particularly Chinaonward,
from 2018 (23.6%) these
leads journals
the field,
followed by the United States (10.3%), which continues to be a major center
have gained popularity as AI-based approaches are increasingly applied across multiple for structural
analysis and
domains, design.structural
including Iran (8.8%)design,
and India (5.7%)—both
optimization, and actively advancing
performance structuralThis
assessment. per-
formance evaluation and optimization—are other major contributors.
steady rise in publication frequency suggests that AI methods are expanding beyond tradi- Similarly, the
United Kingdom (5.0%) and South Korea (5.0%) have made significant contributions, with
tional design principles into broader applications, such as structural resilience evaluation,
a focus on enhancing the accuracy and efficiency of the engineering process. Apart from
real-time monitoring, and automated construction.
these leading countries, Australia (3.3%), Vietnam (3.1%), Canada (2.8%), and Turkey
Figure
(2.7%) have3also
shows the worldwide
played instrumental distribution of structural
roles in shaping engineering
AI-driven structuralresearch moti-
engineering
vated by AI and identifies important contributing nations. Reflecting its
research. The “Other Countries” category, accounting for 29.7%, emphasizes the global strong investment
reach
in and growing
data-driven relevance
approaches andofcomputational
artificial intelligence applications
technologies, in structural
China (23.6%) leads engineer-
the field,
ing across many regions.
followed by the United States (10.3%), which continues to be a major center for structural
analysis and design. Iran (8.8%) and India (5.7%)—both actively advancing structural
performance evaluation and optimization—are other major contributors. Similarly, the
United Kingdom (5.0%) and South Korea (5.0%) have made significant contributions, with
a focus on enhancing the accuracy and efficiency of the engineering process. Apart from
these leading countries, Australia (3.3%), Vietnam (3.1%), Canada (2.8%), and Turkey (2.7%)
have also played instrumental roles in shaping AI-driven structural engineering research.
The “Other Countries” category, accounting for 29.7%, emphasizes the global reach and
growing relevance of artificial intelligence applications in structural engineering across
many regions.
Metals 2025, 15, x FOR PEER REVIEW 8 of 41
Metals 2025, 15, 408 8 of 41

Figure 3. Global Distribution of AI-Driven Structural Engineering Publications by Country.

Figure 3. Global Distribution of AI-Driven Structural Engineering Publications by Country.
Visualizing the most often used terms in AI-driven structural engineering research,
Visualizing
the keyword the most often
co-occurrence networkused(Figure
terms 4) in shows
AI-driventheirstructural engineering
interconnections research,
and thematic
the keyword co-occurrence network (Figure 4) shows their interconnections and thematic
clusters. Comprising several color-coded clusters, each with a different research focus,
clusters. Comprising several color-coded clusters, each with a different research focus, the
the network is each keyword’s size reflects its importance in the dataset; conversely, the
network is each keyword’s size reflects its importance in the dataset; conversely, the prox-
proximity and connecting lines show the strength of terms’ relationships. Key words in-
imity and connecting lines show the strength of terms’ relationships. Key words including
cluding
“machine “machine
learning”, learning”, “prediction”,
“prediction”, “model”, “model”, “structure”,
“structure”, “analysis”,
“analysis”, “method”,
“method”, and
and “ge-
netic algorithm” appear most at the center of the network. This implies that research in
“genetic algorithm” appear most at the center of the network. This implies that research in
this domain
this domain heavily
heavily emphasizes
emphasizes predictive
predictive modeling,
modeling, structural
structural analysis,
analysis, and
and algorithmic
algorithmic
optimization. Emphasizing
optimization. Emphasizing its its use
use inin structural
structural performance
performance enhancement,
enhancement, the the term
term “ge-
“ge-
netic algorithm”
netic algorithm” is is tightly
tightly related
relatedto to“optimization”,
“optimization”,“frame”,
“frame”,and and“optimal
“optimaldesign”.
design”. The
The
redcluster
red clusterisiscentered
centeredaround
around“machine
“machinelearning”,
learning”, “prediction”,
“prediction”, andand “strength”,
“strength”, indicat-
indicating
aing a focus
focus on AI-driven
on AI-driven approaches
approaches for capacity
for capacity estimation,
estimation, structural
structural behavior
behavior model-
modeling,
ing, and material property analysis. This cluster includes terms such
and material property analysis. This cluster includes terms such as “fire resistance”, “col- as “fire resistance”,
“column”,
umn”, “shear“shear strength”,
strength”, and and “capacity
“capacity prediction”,
prediction”, showing
showing the the
rolerole
of AIof in
AIevaluating
in evaluat-
ing load-bearing capacities and failure mechanisms in steel structures. The blue cluster
load-bearing capacities and failure mechanisms in steel structures. The blue cluster em-
emphasizes “structural health monitoring”, “detection”, and “classification”, indicating a
phasizes “structural health monitoring”, “detection”, and “classification”, indicating a
strong research focus on damage detection, real-time assessment, and digital twin tech-
strong
nology.research
This fitsfocus on damage
the growing detection,
acceptance of real-time
computerassessment, and digital
vision techniques twin learning
and deep technol-
ogy. This fits the growing acceptance of computer vision techniques
for automated structural inspection and reliability assessment. The green cluster—which and deep learning
centers
for on “genetic
automated algorithm”,
structural inspection“frame”, and “optimization
and reliability assessment. design”—represents
The green cluster—which another
important area. This points to the focus of research on seismic reliability analysis,
centers on “genetic algorithm”, “frame”, and “optimization design”—represents struc-
another
tural optimization
important area. Thismethods
points todriven
the focusby of
AI,research
and improvement of steel frame
on seismic reliability performance.
analysis, structural
Terms like “finite
optimization methods element simulation”,
driven by AI, and“deep learningofmethod”,
improvement steel frameandperformance.
“graph neural net-
Terms
work” point to the way AI is being combined with cutting-edge
like “finite element simulation”, “deep learning method”, and “graph neural network” computational ap-
proaches.
point to theInway
engineering
AI is being applications,
combined structural reliability,
with cutting-edge performance approaches.
computational evaluation, and In
uncertainty quantification all depend on this integration. The keyword
engineering applications, structural reliability, performance evaluation, and uncertainty network (Figure
4) shows generally the increasing importance of AI, deep learning, and optimization
quantification all depend on this integration. The keyword network (Figure 4) shows
methods in structural engineering research. The grouping of related keywords empha-
generally the increasing importance of AI, deep learning, and optimization methods in
sizes how AI is being used in predictive modeling, performance analysis, structural mon-
structural engineering research. The grouping of related keywords emphasizes how AI
itoring, and optimization, determining the course of data-driven engineering solutions.
is being used in predictive modeling, performance analysis, structural monitoring, and
optimization, determining the course of data-driven engineering solutions.
Metals 2025, 15, x FOR PEER REVIEW 9 of 41
Metals 2025, 15, 408 9 of 41

Figure
Figure4.4.Keyword
KeywordCo-Occurrence
Co-OccurrenceNetwork
NetworkininAI-Driven
AI-DrivenStructural
StructuralEngineering
EngineeringResearch.
Research.

The
Theintegration
integrationofofAI
AIininstructural
structuralengineering
engineeringhas hasobviously
obviouslyevolved
evolvedoveroverthe
thelast
last
two decades. Early years mostly focused on rule-based systems, genetic algorithms,
two decades. Early years mostly focused on rule-based systems, genetic algorithms, and and
shallow
shallowneural
neural networks—mostly
networks—mostly used used for
fordesign
designoptimization
optimizationororbasic
basiccapacity
capacity predic-
prediction.
tion. As field
As the the field matured,
matured, researchers
researchers gradually
gradually shifted
shifted toward
toward more
more sophisticated
sophisticated meth-
methods,
ods, including support vector machines, ensemble learning models, and metaheuristic op-
including support vector machines, ensemble learning models, and metaheuristic opti-
timization frameworks. After 2015, with increasing computational resources and access to
mization frameworks. After 2015, with increasing computational resources and access
larger datasets, deep learning techniques—particularly convolutional and recurrent neu-
to networks—began
ral larger datasets, deep learning studies
to dominate techniques—particularly
focused on structural convolutional and recurrent
health monitoring, dam-
neural networks—began to dominate studies focused on structural health
age detection, and time-series forecasting. Reflecting the need for transparency in safety- monitoring,
critical
damage domains, most
detection, andrecently the interest
time-series has grown
forecasting. into Explainable
Reflecting the need for AItransparency
(XAI) and hy-in
brid approaches.
safety-critical domains, most recently the interest has grown into Explainable AI (XAI) and
Three
hybrid main applications—ML, IML, and XML—are investigated in our scientometric
approaches.
analysis.
ThreeSections 3–5 include representative
main applications—ML, IML, and studies
XML—are thatinvestigated
demonstrate in how these ap-
our scientomet-
proaches tackle
ric analysis. complex
Sections 3–5engineering problems to studies
include representative show anthat
extensive view of
demonstrate AI applica-
how these ap-
tions in the field.
proaches tackle complex engineering problems to show an extensive view of AI applications
in the field.
3. An Overview of ML Application in Steel Structures
3. AnTo maintain
Overview focus
of and
MLensure scientificin
Application clarity,
Steelthis review concentrates on super-
Structures
vised learning algorithms, which represent the most widely applied category in structural
To maintain focus and ensure scientific clarity, this review concentrates on super-
engineering tasks. Although other machine learning paradigms such unsupervised and
vised learning algorithms, which represent the most widely applied category in structural
reinforcement learning have great value, a thorough review of all current ML models
engineering
would exceedtasks. Although
the scope other machine
and intended lengthlearning paradigms
of this paper. such unsupervised
Therefore, a curated setandof
algorithms with a demonstrated applicability to structural prediction, optimization,
reinforcement learning have great value, a thorough review of all current ML and
models would
evaluation
exceed theisscope
examined in the sections
and intended that
length of follow.
this paper.To help better
Therefore, navigate
a curated setthe hierarchy
of algorithms
and
withrelationships among
a demonstrated these supervised
applicability learning
to structural techniques,
prediction, Figure 5 offers
optimization, a graphicalis
and evaluation
summary
examinedofintheir classification.
the sections that follow. To help better navigate the hierarchy and relationships
among these supervised learning techniques, Figure 5 offers a graphical summary of their
classification.
Metals 2025, 15, 408 10 of 41

Figure 5. Hierarchical classification of supervised ML algorithms commonly applied in structural

Figure 5. Hierarchical classification of supervised ML algorithms commonly applied in structural
engineering.
engineering.
3.1. Supervised ML Algorithms
3.1.1.
3.1. Regression
Supervised MLAlgorithms
Algorithms
3.1.1. One
Regression Algorithms
of the simplest regression models, linear regression, uses a straight-line equation
to build a direct link between
One of the simplest regression input and output
models, linear variables.
regression,This
usesapproach is categorized
a straight-line equation
as simple linear regression (for a single predictor) or multiple linear
to build a direct link between input and output variables. This approach is categorized regression (for manyas
simple linear regression (for a single predictor) or multiple linear regression (for many
predictors) depending on the count of input variables. Expanding on regression techniques,
predictors)
multivariate depending
regressiononextends
the count of input
multiple linearvariables. Expanding
regression on regression
by concurrently tech-
predicting
niques,
several multivariate
output variables,regression extendstomultiple
thus helping understand linear regression by concurrently
interdependencies between inputs pre-
dicting several output variables, thus helping to understand
and outputs in complex systems. By including higher-degree polynomial terms in theinterdependencies between
inputs andpolynomial
equation, outputs in complex
regression systems.
unlikeBy including
linear models higher-degree polynomial
captures nonlinear terms in
relationships.
the equation, polynomial regression unlike linear models captures
Lasso regression adds an L1 regularization term to handle problems with correlated input nonlinear relation-
ships. Lasso
features, regression
helping addsless
to shrink an relevant
L1 regularization
coefficients term to handle
toward problems
zero and with correlated
so accomplish feature
input features, helping to shrink less relevant coefficients toward zero and so accomplish
selection. Another type of regularized linear regression, ridge regression, maintains all
feature selection. Another type of regularized linear regression, ridge regression, main-
variables in the model by using L2 regularization to lower the impact of less important
tains all variables in the model by using L2 regularization to lower the impact of less im-
features. Although these methods are meant for continuous variable prediction, logistic
portant features. Although these methods are meant for continuous variable prediction,
regression
logistic is extensively
regression applied for
is extensively classification
applied problems problems
for classification by estimating the probability
by estimating the
probability
of an outcome of an outcome
and mapping and mapping
inputs inputscategories.
to discrete to discrete Incategories.
structuralInengineering,
structural engi-each
neering, each of these
of these regression regression
techniques hastechniques
unique uses; haslinear
unique uses; linear
regression regression for
is appropriate is appro-
simple
priate for simple
predictions like predictions
load-bearing like load-bearing
estimates, estimates,
polynomial polynomial
regression regression
is helpful for is helpful
modeling
for modeling
nonlinear nonlinear
material material
behavior, and behavior,
logistic and logistichelps
regression regression helpsclassification
in binary in binary classifi-
tasks
cation tasksevaluating
including includingstructural
evaluatingsafety structural safety compliance
compliance [15,16]. [15,16].

3.1.2.
3.1.2.Decision
DecisionTree
Tree
Considered
Consideredaa generally
generallyused
used MLML method
methodwithwithsimplicity
simplicityand andinterpretability
interpretability[17],
[17],
decision tree (DT) analysis is particularly useful for classification and regression
decision tree (DT) analysis is particularly useful for classification and regression activities.
activi-
This
ties. approach methodically
This approach divides
methodically difficult
divides datasets
difficult into smaller
datasets pieces.pieces.
into smaller A DT A is DT
a non-
is a
parametric model that divides the input space into discrete areas, each matching
nonparametric model that divides the input space into discrete areas, each matching a par- a partic-
ular decision result. The structure comprises a root node, which acts as the starting point
ticular decision result. The structure comprises a root node, which acts as the starting point
for decision-making, branches reflecting various conditions, decision nodes performing
for decision-making, branches reflecting various conditions, decision nodes performing
feature-based tests, and leaf nodes, indicating final predictions. The learning process al-
feature-based
ternately tests,the
divides anddataset
leaf nodes, indicating
depending onfinal predictions.
chosen criteria, The
suchlearning processMean
minimizing alter-
nately divides
Squared the dataset
Error (MSE) depending
in regression on chosen
problems. Withcriteria, such
each leaf minimizing
node serving asMean
a lastSquared
classi-
fication or prediction, the path from root to leaf defines the decision rules within the tree.
Error (MSE) in regression problems. With each leaf node serving as a last classification
The aim is to build
or prediction, a compact
the path tree with
from root few
to leaf decision
defines thenodes such
decision thatwithin
rules predictive accuracy
the tree. The
is maintained. Avoiding overfitting depends on appropriate stopping criteria that limitis
aim is to build a compact tree with few decision nodes such that predictive accuracy
maintained. Avoiding overfitting depends on appropriate stopping criteria that limit tree
depth or stop too frequent divisions. DTs demand minimal data preparation, are quite
flexible, able of managing both numerical and categorical data. Their sensitivity to small
data variations and overfitting sensitivity, however, can cause instability and accuracy
Metals 2025, 15, 408 11 of 41

loss relative to more sophisticated tree-based models like Random Forests and Boosting
Algorithms. Regularizing methods—such as pruning or limiting tree complexity—are
widely used to improve performance, guaranteeing better generalization to unseen data.

3.1.3. Random Forest

Widely used in both classification and regression tasks, Random Forest (RF) [18] is a
potent ensemble learning method. RF improves predictive accuracy by building several
DTs and aggregating their outputs, thus lowering the overfitting risk. The basis of RF, the
idea of ensemble learning, depends on the combined ability of several models to gener-
ate more consistent results. Bootstrapped aggregation (bagging), in which the dataset is
randomly split into several subsets and separate classifiers are trained on each subset, is
one of the most successful ensemble techniques available within RF. This method reduces
variance and enhances model stability by means of some data points appearing in several
subsets. Majority voting (for classification) or averaging (for regression) determines the last
prediction, thus guaranteeing better generalization on fresh data. RF is unique in that it
randomly chooses features during tree building, thus lowering correlation between indi-
vidual trees and improving model performance. For handling high-dimensional datasets
with many input variables—a difficulty that conventional Decision Trees often face—this
randomness makes RF especially efficient. While training many trees is computationally
efficient, producing predictions from a trained RF model can be time-intensive because
the demand to process several trees concurrently. Among RF’s main benefits are its lower
sensitivity to overfitting and greater resilience than of a single Decision Tree. RF is easily
available for users without strong knowledge in ML since it often performs well with
default settings even without great fine-tuning.

3.1.4. Support Vector Machines

Support Vector Machines (SVM) are a widely used ML technique known for their
effectiveness in classification tasks [19]. Originally presented for handling linearly sepa-
rable data, SVM has since evolved to accommodate more complex problems including
nonlinearly separable datasets, regression (Support Vector Regression, SVR), and clustering
(Support Vector Clustering, SVC). Although its main use is still in classification, SVM’s
flexibility lets it be applied in many fields needing strong accuracy and solid decision limits.
Fundamentally, SVM seeks to identify the best separating hyperplane maximizing the
margin separating several data classes.
Its orientation and position are much influenced by the support vectors, the important
data points closest to the hyperplane. SVR uses a similar idea but concentrates on fitting
a function that retains most data points inside a specified margin, thus allowing some
flexibility in error tolerance in regression problems. SVM uses kernel functions and penalty
parameters to improve its capability in situations when data are not linearly separable.
While preserving a wide margin, the penalty parameter adds slack variables that enable
controlled misclassifications, thus balancing accuracy with generalization. Conversely,
kernel functions create a higher-dimensional space from the original input space where
a linear separation is feasible. Commonly used kernels include sigmoid, linear, poisson,
radial basis function (RBF), and sigmoid functions, each with different benefits contingent
on the dataset.

3.1.5. Artificial Neural Networks

Inspired by the organization and capability of the human brain, artificial neural net-
works (ANNs) are a fundamental ML method. Originally designed for pattern recognition
tasks, Rosenblatt’s perceptron model in 1958 helped to form the idea. Computational power
improvements over time have allowed ANNs to develop into more advanced designs able
Metals 2025, 15, 408 12 of 41

to address challenging ML tasks. ANNs have developed into several specialized designs
each intended to solve particular ML problems. The most fundamental form is the feed-
forward neural network (FFNN), in which data moves just from input to output layers.
Incorporating several hidden layers, an advanced form of the Multilayer Perceptron (MLP)
improves processing capability. Conversely, the Radial Basis Function Neural Network
(RBFNN) is especially fit for specialized uses since it uses radial basis functions as activa-
tion mechanisms. Applied in fields including structural engineering, convolutional neural
networks (CNNs) are optimized for image processing and pattern recognition, playing a
major part in crack detection. Designed to manage sequential data, recurrent neural net-
works (RNNs) and their sophisticated variant, Long Short-Term Memory (LSTM) networks,
shine in capturing long-term dependencies. The Adaptive Neuro-Fuzzy Inference System
(ANFIS) is another hybrid method whereby neural networks are combined with fuzzy logic
to improve decision-making capacity in uncertain surroundings [20].
Inspired by the human brain, which has linked layers of neurons processing and
interpreting data, ANN architecture follows Input, hidden, and output layers. The hidden
layers use sophisticated computations to derive significant patterns from unprocessed data.
An ANN’s efficiency mostly relies on elements like activation functions, which control
signal flow across the network. The sigmoid or Rectified Linear Unit (ReLU) controls
neuron activations, influencing the learning process. ANNs need hyperparameter fine-
tuning of the number of layers, learning rate, and training epochs if they are to reach
optimal performance. Correct optimization guarantees against overfitting or underfitting
and guarantees the network can generalize to unseen data effectively.
The capacity of ANNs to replicate the learning mechanisms of the human brain defines
one of their main advantages. Like neurons and synapses, these networks handle data in a
way where experience shapes connections either strengthening or weakening. By means of
deep feature extraction made possible by hidden layers, the network can identify complex
interactions inside data. In difficult problem-solving activities, high predictive accuracy
depends on this layered approach. Modern ML applications depend on ANNs since they
offer a strong framework for adaptive learning by modeling biological neural paths [21].

3.1.6. k-Nearest Neighbor

Mostly used for classification and regression problems, the k-Nearest Neighbor (KNN)
algorithm [22] is among the simplest yet most often used ML methods. KNN preserves the
whole dataset and generates predictions depending on proximity-based decision-making
unlike many other models that build an explicit function or discard training data after
learning. The method assigns a class label depending on the majority vote among the
k-closest data points, sometimes known as “nearest neighbors”, when classifying a new
instance. KNN averages the target values of the closest neighbors to forecast a continuous
value in regression problems. Being a non-parametric method, KNN does not presume any
predefined data distribution. Its efficiency relies on choosing a suitable value for k and
measuring data point similarity using an appropriate distance metric. When no previous
domain knowledge is available, the most often used metric is the Euclidean distance; never-
theless, depending on the dataset properties other metrics, such Manhattan or Minkowski
distance, could also be used. KNN is a common choice, especially for applications where
model transparency is crucial, because of its simplicity and interpretability.

3.1.7. Boosting Algorithms

Combining several weak learners into a more accurate and robust model helps boost-
ing, a ML method, increase predictive performance. Freund and Schapire [23] first proposed
the idea of boosting in the middle of the 1990s, which helped to create AdaBoost, among
Metals 2025, 15, 408 13 of 41

the first adaptive boosting systems. Many sophisticated boosting techniques have been
developed since then to improve computational efficiency as well as accuracy. Among
the notable examples are Friedman’s Gradient Boosting Machine (GBM), which iteratively
improves predictions by minimizing errors via gradient optimization. More recent develop-
ments including Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine
(LightGBM), and Categorical Gradient Boosting (CatBoost) have further optimized boost-
ing techniques by improving speed, lowering memory use, and more effectively handling
categorical data.
First presented in 1996 by Freund and Schapire [23], Adaptive Boosting (AdaBoost)
was among the first boosting methods to show notable success. AdaBoost’s central concept
is to prioritize challenging-to-classify samples, sequentially improving weak learners. The
method starts with training a base model on the whole dataset and evaluates mistakes
afterwards, strengthening the impact of misclassified events. This guarantees that next
models give more of their attention on fixing past errors. Many weak learners are trained
as the process runs, each tackling particular shortcomings of the previous models. The
resultant ensemble model uses a weighted voting system to decide predictions, giving
more weight to more accurate models. Although AdaBoost is quite successful in raising
prediction accuracy, its dependence on reweighting misclassified data makes it vulnerable to
outliers and noisy data, thus compromising general performance. AdaBoost’s efficiency and
ability to improve simpler models help it to still be extensively used in many applications
despite this restriction.
Designed to increase predictive accuracy by building models in a stepwise manner—
where each new model seeks to correct the errors of the previous one—gradient boosting
(GB) [24] is a ML method. GB, unlike AdaBoost, which changes instance weights, trains
each next model on the mistakes left by its predecessor, minimizing the residual errors.
The name “gradient” derives from the technique applied to maximize the loss function—
each iteration moves in the direction most effectively lowering errors. Flexibility is one of
its main advantages since it can be used with different loss functions depending on the
kind of the problem. But because of its incremental character, computational cost may be
high—especially in relation to big datasets.
Designed especially for speed and scalability when handling vast amounts of data,
XGBoost [25] is an advanced application of the gradient boosting framework. It increases
prediction accuracy by consecutively training decision trees, in which each new tree fixes
the mistakes made by the one before it, same as in conventional GB. Still, XGBoost brings
some improvements that boost efficiency. Its regularizing systems are an important aspect,
since they stop overfitting and enhance computational efficiency.
Particularly useful for large-scale ML applications, Light Gradient Boosting Machine
(LightGBM) [26] is a high-performance gradient boosting method intended for speed
and efficiency. Furthermore, unique to LightGBM is its leaf-wise tree growth method,
which deviates from standard level-wise methods. This enables the method to create
increasingly sophisticated, deeper trees, thus improving predictive accuracy. But especially
if the number of leaves is not under control, this rapid tree growth raises the possibility
of overfitting. Despite this difficulty, LightGBM is still among the most effective boosting
methods, since it can manage vast amounts of high computational-efficient data.
CatBoost [27] is a gradient boosting method that handles categorical features using
target-based statistics, instead of one-hot encoding like many other boosting algorithms.
This approach helps reduce the risk of data leakage and overfitting. CatBoost uses an
ordered boosting method, where each model iteration is trained only on the data that
comes before it. This helps produce stable predictions and avoids bias caused by using
future data. CatBoost also uses a symmetric tree structure to maximize the search space
Metals 2025, 15, 408 14 of 41

and speed training while preserving great accuracy. It is a user-friendly option for14many
Metals 2025, 15, x FOR PEER REVIEW of 41
ML projects since it requires less hyperparameter tuning when compared to XGBoost and
LightGBM. CatBoost is especially suited for uses involving structured data with multiple
categorical features since it can effectively manage categorical variables and provide fast,
categorical features since it can effectively manage categorical variables and provide fast,
accurate
accurate results.
results.
Figure
Figure 66 shows
shows thethedistribution
distributionofofML
MLalgorithms
algorithms used
used in in structural
structural engineering
engineering re-
search, based on the reviewed papers published between 1990 and 2025. Neural Networks
research, based on the reviewed papers published between 1990 and 2025. Neural Networks
(NN), with the largest share at 53%, remain the dominant choice across a wide range of
structural applications.
structural applications. Boosting
Boosting Algorithms
Algorithms (BA)
(BA) follow
follow with
with 18%,
18%, clearly
clearly indicating
indicating their
their
growing presence
growing presence in in recent
recent years.
years. Support
Support Vector Machines (SVM)
Vector Machines (SVM) account
account for 8% of
for 8% of the
the
total, while Random Forest (RF) and Decision Trees (DT) each contribute 5%. Regression-
total, while Random Forest (RF) and Decision Trees (DT) each contribute 5%. Regression-
based Approaches
based Approaches(RA) (RA)also represent
also 5%,5%,
represent andand
the remaining 6% are
the remaining 6%grouped under under
are grouped Other
methods, which reflects a combination of various ML techniques. This distribution
Other methods, which reflects a combination of various ML techniques. This distribution under-
scores the strong reliance on NN-based models, while also highlighting the increasing role
underscores the strong reliance on NN-based models, while also highlighting the increasing
of ensemble methods such as BA, RF, and DT in structural engineering research.
role of ensemble methods such as BA, RF, and DT in structural engineering research.

Figure 6. ML techniques applied in the considered research of this review spanning 1990 to 2025.
Figure 6. ML techniques applied in the considered research of this review spanning 1990 to 2025.
3.2. ML for Steel Joints, Connections, and Rotational Stiffness Prediction
3.2. ML for SteelML
Including Joints,
intoConnections, and Rotational
steel connections and jointsStiffness
helps toPrediction
predict structural behavior with
muchIncluding ML into
more accuracy andsteel connections
efficiency, and joints
thus reducing helps to predict
dependence structural
on expensive behavior
experimental
with much
testing and more accuracyempirical
conventional and efficiency, thus reducing
formulations [28–33].dependence
Several researchon expensive
have used exper-
ML
imental testing and conventional empirical formulations [28–33].
techniques to enhance the prediction of moment-rotation behavior, joint stiffness, and Several research have
used MLoftechniques
strength bolted andtowelded
enhance the prediction of moment-rotation behavior, joint stiff-
connections.
ness,Paral
and strength of bolted and welded connections.method for evaluation of semi-rigid
et al. [34] presented a deep learning-based
Paral et al. [34] presented a deep learning-based method for evaluation of semi-rigid
joint condition in steel frames. The model efficiently analyzed global vibration response
joint condition in steel frames. The model efficiently analyzed global vibration response
signals from impulse excitation by combining Convolutional Neural Networks (CNN) with
signals from impulse excitation by combining Convolutional Neural Networks (CNN)
Continuous Wavelet Transform (CWT). Emphasizing moment-rotation (M-φ) correlations,
with Continuous Wavelet Transform (CWT). Emphasizing moment-rotation (M-φ) corre-
Kueh
lations,[35] suggested
Kueh explicit explicit
[35] suggested mathematical formulations
mathematical for steelfor
formulations flush
steelendplate
flush endplatebeam–
beam–column
column connection
connection rotational
rotational stiffnessstiffness prediction.
prediction. Practical
Practical formulations
formulations werewere derivedde-
rived depending on geometric and material characteristics using ANNs and MLR.
depending on geometric and material characteristics using ANNs and MLR. In In an-
another
other study
study on moment-rotation
on moment-rotation behavior,behavior, Tran
Tran [36] [36] examined
examined flush end-plate
flush end-plate connections connec- at
tions at elevated
elevated temperatures
temperatures using FEusing FE simulations
simulations and ANNs andtoANNs
forecastto the
forecast
ultimatethe moment
ultimate
moment
(Mu) and(Mu)
shapeand shape parameters
parameters of an optimalof anM-φoptimal
model.M-φ model. Expanding
Expanding the use ofthe MLuse of ML
in bolted
in bolted connections,
connections, Sarothi etSarothi
al. [37]etcreated
al. [37] created a predictive
a predictive ML framework
ML framework for structural
for structural steel
steel bearing strength estimation of double shear bolted joints. RF
bearing strength estimation of double shear bolted joints. RF showed the best accuracy showed the best accu-
racy (R2 = 0.88)
(R2 = 0.88) aboveabove current
current design
design formulations.
formulations. In In a similar
a similar vein,
vein, Jiangetetal.al.[38]
Jiang [38]investi-
inves-
tigated ML-based failure load and failure mode predictions for bolted connections of high
strength steel. Training eight ML algorithms—including SVM, RF, and XGBoost—the
Metals 2025, 15, 408 15 of 41

gated ML-based failure load and failure mode predictions for bolted connections of high
strength steel. Training eight ML algorithms—including SVM, RF, and XGBoost—the study
obtained a failure mode prediction accuracy of 97.2%, higher than the 67.9–85.3% accuracy
of traditional design codes.
Beyond ANN models, Multi-Gene Genetic Programming (MGGP) has been investi-
gated for moment-rotation prediction. In modeling semi-rigid connection behavior, Rabbani
et al. [39] compared MGGP with ANN-based models and showed that MGGP displayed
better accuracy and generalization capacity.

3.3. ML for Buckling and Stability Analysis

Recent developments in ML and AI have produced more exact, data-driven ap-
proaches for predicting lateral-torsional buckling (LTB), elastic and inelastic buckling, and
stability failure modes in steel buildings. Several studies have also shown how well ANNs
might predict LTB resistance. For LTB resistance in slender steel cellular beams, Ferreira
et al. [40] created an ANN-based predictive model. The findings revealed notable progress
over current analytical models, thus highlighting ML’s structural optimization potential.
The existence of web openings in steel beams presents yet another difficulty for buck-
ling resistance estimate since it can greatly affect post-buckling strength. Developing an
ANN model for web-post buckling resistance and failure modes in steel beams with ellipti-
cally based web openings, Shamass et al. [41] addressed this problem. While web opening
height, width, and radius had negative effects, the study found important geometric pa-
rameters influencing buckling behavior including beam height and web thickness (which
favorably affected resistance). Their results underlined how successfully ANN-based mod-
els capture intricate geometric effects on stability performance. Furthermore, difficult in
LTB prediction are steel beams with perforated web geometries. De Carvalho et al. [42]
investigated the lateral-torsional buckling behavior of I-beams including sinusoidal web
openings. The results imply that engineers using perforated I-sections might find useful
design tools in ANN-enhanced predictive models. Xing et al. [43] presented an ANN-based
model for local buckling prediction in the fire-resistant design of stainless steel beams. The
study trained ANN models optimized using Kruppa’s criteria and k-Fold cross-valuation
by including experimental and FE results into a comprehensive dataset, thus ensuring
robust and reliable buckling predictions for stainless steel I-sections under fire exposure.
Rossi et al. [44] showed how well ANN-based models predicted LTB strength for steel
I-beams. Their FEA parametric study investigated LTB resistance under geometric parame-
ters and loading conditions. The ANN model gave better predictive accuracy than current
design equations. Developing an ML framework for high-strength steel I-section columns,
Cheng et al. [45] addressed the scattered accuracy of conventional design rules. Seven ML
models were trained; Categorical Boosting turned out to be the most accurate method.

3.4. ML for Strength Prediction and Optimization of Cold-Formed Steel Structures

Cold-formed steel (CFS) buildings are rather common because of their great strength-
to-weight ratio, economy, and sustainability. Predicting their strength, stability, and failure
mechanisms still challenges since traditional design approaches sometimes oversimplify
nonlinear interactions, residual stresses, and defects. ML is quickly becoming a great tool
for load-bearing predictions, cross-section optimization, and structural reliability increase.
By addressing constraints in conventional design approaches that mostly concentrate
on global buckling, Xu et al. [46] helped to predict ML-based bearing capacity for cold-
formed stainless steel circular hollow section (CHS) columns. Trained on a database of
280 CHS columns, their ML-based approach showed significant accuracy improvements
over current design codes, enabling improved predictive models in cold-formed steel
Metals 2025, 15, 408 16 of 41

engineering. Nguyen et al. [47] estimated the axial compression capacity of cold-formed
steel oval hollow section columns using ANN and Adaptive Neural Fuzzy Inference System
(ANFIS) models. Comparisons with three current design codes confirmed even more
that ML-based approaches are better in estimating CFS section capacity. Fang et al. [48]
developed a deep learning framework to evaluate the web crippling strength of cold-
formed stainless steel channels addressing web crippling in perforated steel sections.
Their DBN-based predictions exceeded conventional web crippling equations; new design
equations were proposed based on the ML results. Lu et al. [49] developed a regression-
classification ensemble ML model for predicting loading capacity and buckling modes
of cold-formed steel built-up I-section columns. With high accuracy in both capacity
estimation and buckling mode classification, XGBoost stood out among the tested models.
To solve constraints in traditional fire design methods, Shaheen et al. [50] developed an
ML-based predictive model for estimating the mechanical properties of high-strength steel
at elevated temperatures. The work trained Deep Neural Networks (DNNs) on a large
experimental dataset using temperature and chemical composition as input features to
predict ultimate tensile strength, yield strength, 0.2% proof strength, and elastic modulus.
Shahin et al. [51] developed hybrid ML methods for cold-formed steel-lipped channels
to forecast web crippling capacity. Their work honed forecasts by combining ANN with
GA and PSO. The PSO-ANN hybrid model exceeded other methods and offered more
consistent strength estimates than accepted design guidelines. Yılmaz et al. [52] developed
an ML-based predictive model for the load-bearing capacity of lipped channel sections. By
training their model on a dataset of 2240 FE simulations, they found that flange length and
section thickness were the most important parameters, underlining the importance of ML
in geometry-based strength evaluations.

3.5. ML Applications in Steel Frame Design, Optimization, and Damage Detection

For automated visual inspection of steel frame structures, Kim et al. [53] presented
a deep convolutional neural network-based damage locating (DCNN-DL) method. The
DCNN-DL approach found and overlaid damage sites on input images, thus enabling exact,
real-time steel frame inspections. Using ML models, Truong et al. [54] investigated their
applicability for load-carrying capacity prediction of semi-rigid steel structures. Twelve
ML techniques were tested. With regard to error metrics and determination coefficients,
XGBoost showed the best accuracy among the tested models. Aiming to replicate the
decision-making process of an experienced designer, Jahjouh [55] investigated the possibili-
ties of ANNs in the design optimization of steel frames. Optimal structural designs for 2D
steel frames were produced using an adaptive harmony search technique, subsequently
regarded as training data for ANNs. The trained ANN models showed a 99% accuracy in
forecasting appropriate designs during verification. By including real-time model training
and parameter tuning, Shan et al. [56] developed an ML-assisted optimization framework
for steel frame design, thus improving the efficiency of metaheuristic algorithms. The
technique dynamically creates surrogate models to improve convergent accuracy and
convergence.

3.6. ML-Based Structural Health Monitoring in Steel Structures

Pal et al. [57] presented a deep-learning method for structural health monitoring of
a steel frame construction with bolted connections based on vibrations. Developed to
extract discriminative features from time-frequency scalogram images of vibration data, a
convolutional neural network (CNN) classified Bolt loosening conditions into fully loose,
hand-tight, and completely tight (undamaged) categories. Focusing on connection damage
identification using statistical vibration features, Naresh et al. [58] developed an ML-based
Metals 2025, 15, 408 17 of 41

health monitoring system for steel frame buildings. High accuracy in structural damage
detection allowed an SVM model to classify both undamaged and damaged conditions
to show Analyzing natural frequency fluctuations, Vu et al. [59] investigated ML-based
damage identification in steel beams, offering a substitute for conventional structural
health monitoring techniques. FEM generated a dataset of natural frequencies under
several damage scenarios; ML models including ANN, XGB, and RF were trained to predict
damage location, width, and depth.
Although conventional sensor-based monitoring systems—such as strain gauges, fiber
Bragg grating sensors, and accelerometers—have been extensively used for damage de-
tection in steel constructions due to their high precision and long-term stability, they are
often limited to localized measurements and demand direct physical installation and con-
tinuous maintenance. Conversely, for non-contact, wide-area structural inspection, recent
developments in computer vision and deep learning present interesting substitutes. These
AI-powered visual techniques, particularly those using convolutional neural networks
(CNNs), have shown strong performance in tasks such as surface crack segmentation, cor-
rosion detection, and bolt loosening identification. The primary advantage of vision-based
approaches lies in their scalability, cost-effectiveness, and ability to process large amounts
of visual data rapidly. However, they also face practical challenges such as sensitivity to
lighting, occlusion, camera angle, and the need for high-quality annotated datasets. Several
recent studies suggest hybrid approaches that use the strengths of both systems to combine
sensor signals with vision-based outputs, thus enabling stronger condition assessment.
Integrating these two techniques could provide a more robust and interpretable structure
for real-time steel structure monitoring in complex environments as the field develops.
Beyond real-time detection, one of the most exciting frontiers of AI in structural engi-
neering is the prediction of long-term degradation in steel structures including industrial
frameworks, bridges, and high-rise buildings. Recurrent neural networks (RNNs), long
short-term memory (LSTM), and Bayesian networks have been investigated recently in
order to forecast damage accumulation depending on operational load, climate exposure,
and historical sensor data. Models have been trained, for example, to learn patterns from
multi-year monitoring datasets to forecast fatigue crack propagation, coating degradation,
and corrosion depth over time. Though early results are encouraging, such applications
still provide practical difficulties. These comprise limited access to high-quality long-term
field data, environmental condition variability, and the challenge of documenting rare but
critical degradation modes. Recent efforts aiming at overcoming these constraints have
concentrated on hybrid models combining AI with physics-based deterioration models,
providing better interpretability and generalizability. Transfer learning methods let models
developed on one structure be adjusted for others with similar typologies concurrently.
These methods, which are still under development, imply that AI has great possibilities
for lifecycle assessment of steel infrastructure under actual conditions and predictive
maintenance planning.
Especially in the automated damage and failure mode detection in steel construc-
tions, ML techniques have brought significant SHM advances. Deep learning and feature
extraction-based models have remarkably faithfully classified bolt loosening, connection
damage, and other structural failures. Moreover, ML-driven frequency analysis provides a
fast approach to identify damage in steel beams and frames, thus reducing dependency on
traditional sensor-based methods and hand inspections.

3.7. ML-Based Prediction of Structural Performance in Steel Components

ML uses in shear strength, bending capacity, moment resistance, and structural stability
prediction in steel components is investigated in this part.
Metals 2025, 15, 408 18 of 41

Li et al. [60] addressed errors in conventional design approaches by developing a

data-driven ML model for shear strength of high-strength friction-grip bolts. Backprop-
agation Neural Networks (BPNN) obtained the lowest coefficient of variation and the
best accuracy (93% goodness of fit) among the several ML techniques the study examined.
Using 108 stainless steel lipped channels and 238 carbon steel Lite Steel sections as test
data, Dissanayake et al. [61] further used ML models to forecast the shear capacity of steel
channel sections. The work evaluated several ML methods and found that SVR gave the
most consistent shear capacity predictions. Aiming to overcome conventional design codes,
Dai et al. [62] concentrated on moment capacity prediction of cold-formed steel channel
beams with edge-stiffened and un-stiffened web holes. A FEM was validated against 12
experimental datasets, and an XGBoost model was trained using a database of 1620 sam-
ples, obtaining a R2 score of ~99%, much above traditional design approaches. Reliability
analysis helped to validate the suggested ML-driven equations, guaranteeing adherence
to AISI safety standards. Liu et al. [63] expanded this study by developing ML-based
models for predicting the ultimate bending moment resistance of high-strength steel (HSS)
welded I-section beams. The results confirmed that ML models greatly outperformed both
European and American design codes.
To estimate the failure loads of normal and high-strength steel welded I-section beam–
columns, Su et al. [64] proposed an ML-based unified design method. Traditional methods
were found to understate or scatter failure load predictions when compared to Eurocode
and American design rules; the ML approach produced more consistent and dependable
results. Utilizing structural mechanics and ANNs to forecast point forces and distributed
loads, Tusnin et al. [65] applied ML for load identification in steel structural systems. The
suggested approach allows final load estimations, structural decomposition, ML-based
projections, and structural integrity monitoring better accuracy in engineering inspections
and structural health monitoring.

3.8. ML-Based Seismic Performance Assessment and Optimization

Aiming at seismic collapse behavior using ML, Sediek et al. [66] developed a database
of over 900 experimental and numerical results for deep wide flange (W-shaped) columns.
The work estimated cumulative inelastic rotation until failure under axial and lateral
loading and categorized failure modes (local, global, and coupled). By raising classification
accuracy by 30%, the ML models exceeded highly ductile limits and demonstrated the
possibilities of data-driven methods in creating future seismic design rules. Using ML
models, Imam et al. [67] investigated how best to forecast the seismic performance of
steel moment-resisting frames (SMRFs). The study assembled a sizable 29,200 data point
collection from 292 structural models, training RF, XGBoost, and ANN. Compared to
conventional nonlinear dynamic analysis techniques, the ML-based models showed great
predictive accuracy.
Addressing constraints in traditional multivariable regression models, Shin and
Kim [68] developed an ML-based framework for estimating the hysteretic behavior of
two-sided clamped steel shear walls (TCSSWs). Under cyclic load, the ML-based model
shown efficiency in automating and enhancing the hysteretic behavior modeling of steel
shear walls. Cho and Han [69] developed a numerical material model to simulate the cyclic
behavior of high-strength steel (HSS), thus extending ML applications in seismic engineer-
ing. Eliminating the need for expensive LCF tests, empirical equations were proposed
to estimate model parameters using just monotonic tensile test data, improving practical
applicability.
Hu et al. [70] presented an ML-aided peak and residual displacement-based design
(PRDBD) method to use self-centering braces to increase the seismic resilience of SMRFs.
Metals 2025, 15, 408 19 of 41

Conventional SMRFs dissipate energy through steel yielding, producing significant residual
displacements that sometimes call for expensive post-earthquake demolition. Samadian
et al. [71] developed surrogate models for seismic and pushover responses of SMRFs,
furthering ML applications in seismic engineering. Providing a computationally efficient
substitute for traditional nonlinear time history and pushover analyses, CatBoost models
emerged as the most effective in the study. Salama [72] investigated ML techniques to
maximize the seismic resilience of vertically irregular steel building. Using XGBoost
with Owl Search Algorithm (OSA) for hyperparameter tuning produced accurate seismic
behavior predictions. The results underlined the great reliance of base shear capacity and
general seismic performance on structural irregularities.

3.9. AI Applications in Real-World Steel Structures

In bridge engineering, AI-based inspection and SHM systems are advancing safety,
automation, and predictive maintenance. Huang et al. [73] developed a CNN-based deep
learning method for structural damage identification (SDI) on steel truss bridges, achieving
high detection accuracy using field data. Iacussi et al. [74] explored a drive-by monitoring
approach, where smart sensor nodes on moving vehicles predicted bridge deflection
shapes with strong agreement to installed sensors. Wang et al. [75] introduced SBDNet,
a segmentation network trained on real crack images, enabling pixel-level fatigue crack
detection on steel bridge surfaces. Svendsen et al. [76] proposed a hybrid SHM framework
combining numerical simulations with experimental data to detect loose bolts and member
failure on full-scale steel bridges under environmental fluctuations. In high-rise buildings,
Zhou et al. [77] monitored a 300 m twin-tower steel frame where smartphone videos were
processed with computer vision to extract vibration patterns. Ghaffari et al. [78] developed
an RNN model trained on multi-hazard FE data to predict the dynamic response of tall
steel structures, achieving less than 1% error compared to actual displacements.
For offshore and tower structures, AI tools have demonstrated strong practical po-
tential. Wang et al. [79] introduced an ensemble deep learning model (CNN + BiLSTM +
SENet) to detect damage in jacket platforms with 95% accuracy, validated by experimental
results. Martzikos et al. [80] used field data from scaled floating steel platforms to train
ANN models predicting mooring line forces under wave loading. Kouchaki et al. [81] used
ML classifiers to detect damage in transmission towers based on vibration data, achieving
up to 96% accuracy even under noise-contaminated conditions. Kiyoki et al. [82] predicted
bending moments in wind turbine towers using only nacelle sensor data, offering a low-cost
alternative to strain-based measurements.
Beyond monitoring, AI is being embedded into design optimization and material
reuse. Vlasenko et al. [83] developed an ensemble ML model to non-destructively estimate
yield strength of reclaimed steel using magnetic sensor readings, enabling sustainable
material reuse with real-time decisions.
Connection and joint evaluation has also benefited from AI integration. Shang
et al. [84] created a machine vision system that automatically quantifies bolt loosening by
detecting nut rotation angles through image processing.
In the area of maintenance and corrosion monitoring, Huang et al. [85] used an
Inception-v3 deep CNN trained on laboratory corrosion images to classify rust severity
levels on steel surfaces. The model achieved performance comparable to expert inspectors
and enables consistent, rapid corrosion assessment across large steel structures such as
bridges, towers, and tanks.
These studies collectively show that AI technologies are now being deployed in real-
world settings for steel structure monitoring, inspection, design, reuse, and maintenance.
As AI continues to mature, its integration into practical structural workflows will likely
Metals 2025, 15, 408 20 of 41

expand, making structural engineering more data-informed, efficient, and adaptive. This
expanding body of validated applications confirms that AI is no longer limited to simulation
environments but is becoming an essential companion to structural engineers in the field.

4. Inverse Machine Learning (IML) for Design Optimization and

Performance Enhancement in Steel Structures
IML has become a powerful tool in structural engineering for rapidly investigat-
ing complex geometrical designs [86]. Particularly for complicated structural systems,
conventional design techniques sometimes involve several iterative changes that can be
computationally expensive and time-consuming. Direct mapping design objectives to opti-
mal parameters helps IML simplify this process and enable a greater spectrum of design
possibilities by lowering the need for major trial-and-error adjustments [87]. This capacity is
especially important for handling complex structural designs where conventional methods
might not be able to provide sufficient effective answers. In inverse modeling, one of the
main challenges is the ill-posed character of inverse problems, whereby solutions could lack
uniqueness or stability [88]. To solve these challenges, several computational techniques
have been developed to ensure strong and consistent predictions, thus improving the
applicability of IML in useful engineering environments.
In structural engineering, inverse problems mostly concern determining design
parameters—such as geometry and material properties—needed to meet specific per-
formance criteria [89]. Conversely, in forward problems structural performance is expected
to be dependent on predefined input criteria. Unlike forward analyses, inverse problems
are often ill-posed, meaning they can show instability or have several possible answers.
Particularly for complex structural designs [90], conventional techniques—such as trial-
and-error changes—rely on iterative adjustments and hence are both time-consuming and
computationally demanding. IML provides a quick replacement that eliminates the need
for constant hand-offering by directly predicting ideal design parameters straight from
historical data and ML models. This greatly simplifies the design and optimization process,
hence faster and more efficient solutions follow [91]. IML is particularly important for
metamaterials and advanced structural systems where the interactions between geometry,
material composition, and mechanical performance are highly nonlinear and complicated.
Its application to designing buildings ideal for impact absorption, improved buckling
resistance, and shape recovery [92] shows its flexibility over many engineering problems.
By converting structural design into a data-driven process, IML speeds innovation in
demanding structural systems, reduces computing costs, and increases accuracy.

IML Application in Steel Structures

Several studies demonstrating the possibilities of IML in steel construction analysis,
fabrication, and performance prediction offer more efficient and statistically driven solu-
tions. IML has been used to maximize steel buildings by improving geometry, mechanical
aspects, and material composition.
Teimouri et al. [92] devised an IML framework for designing plate-lattice structures
(PLSs) idealized for compression strength and specific recovery force (SFR) in large-scale
steel constructions. Their method consisted of three validation sub-frameworks before PLS
optimization, 3D-printed using a new shape memory polymer (SMP).
By considering manufacture defect and imperfections, IML has also been used to
increase the mechanical robustness of steel-based systems. Glaesener et al. [93] investi-
gated how mechanical response of periodic trusses changed with geometric flaws. Using
IML-based defect prediction models, their study examined common defect types over
several steel truss topologies to evaluate their impact on effective stiffness. This method
Metals 2025, 15, 408 21 of 41

actively found and reduced geometric inconsistencies, thus improving the dependability of
steel trusses.
Particularly in high-strength and ultra-high-strength steel alloys, IML methods have
also been heavily applied to maximize steel materials. Using deep neural networks to
improve hohlraum designs for radiation temperature control, McClarren et al. [94] showed
an IML-driven optimization framework in a high-performance design environment. Similar
uses in steel alloy optimization have been inspired by the approach they used whereby a
forward ML model forecasts results and an inverse model finds optimal parameters. Build-
ing on this, Lee et al. [95] suggested an inverse design framework for high-strength steels,
thus attaining an ultimate tensile strength (UTS) above 2 GPa with improved ductility.
Experimentally validated designs with improved mechanical performance resulted from a
GA optimized using the Shannon diversity index investigating a high-dimensional search
space of chemical compositions and austenitizing conditions. Analogous development of
two ML-based materials genome integration systems for inverse microstructure analysis of
steels was accomplished by Wang and Adachi [96]. These models allowed inverse design
of steel microstructures and direct prediction of material properties to reach desired stress–
strain behavior, tensile strength, and elongation. This method helped to create ideal steel
compositions with better mechanical properties. Using microstructure image recognition
to maximize the composition of martensitic and ferritic steels, Pei et al. [97] introduced an
IML-based framework for inverse alloy design, thus improving steel alloy performance.
Their method greatly reduced trial-and-error efforts in steel development by using neural
networks to identify intricate microstructural patterns and predict alloy properties. Dealing
with microstructure-property optimization, Lertkiatpeeti et al. [98] investigated the impact
of martensite phase alignment in dual-phase steels by means of a Markov Chain Monte
Carlo-based inverse analysis coupled with ML models. Using representative volume ele-
ment (RVE) simulations incorporated new microstructure descriptors—such as Moran’s
index, martensite band index, and martensite orientation—and SVR and ANN models
trained on them quantify spatial effects in martensitic phase distributions. Their method
effectively tuned steel microstructures for intended mechanical characteristics. Adachi
et al. [99] examined IML applications in steel material design more broadly and discussed
ML techniques integrating process-property-microstructure relations. The study under-
lined the growing relevance of IML in improving steel material development by means
of automation. Important design factors are the lifetime and fatigue resistance of steel
buildings; IML has been applied to maximize these features. By means of both direct and
inverse analyses, He et al. [100] devised an ML-based framework to estimate fatigue life and
fatigue limits of steels. A Bayesian optimization-based inverse analysis was also conducted
to ascertain fatigue limits. Optimizing double-layer protective coatings used in spacecraft
friction units, Kolesnikov et al. [101] developed an IML-assisted framework. While the
Extra Trees method was used for feature importance analysis, their work combined FEM
with adaptive sampling techniques to build a training set. By allowing coatings based on
target hardness values to be designed, the inverse ML model helped to achieve predictive
accuracies above R2 = 0.96, offering an effective method of designing high-performance
protective layers.
Relevant for hybrid steel-composite structures, IML has helped to maximize the
Automated Fiber Placement (AFP) process in thermoplastic composite manufacture in
manufacturing processes. To increase process control, Islam et al. [102] presented a hybrid
ML framework combining ANNs, virtual sample generation, physics-based numerical
simulations, and experimental data. Using an inverse property estimation framework,
Masurkar et al. [88] estimated elastic constants of orthotropic laminates used in steel-
composite hybrid structures. Their work structured the problem as an inverse task, whereby
Metals 2025, 15, 408 22 of 41

a deep neural network model was developed on time-series simulation data to effectively
infer material properties.
In recent years, IML has been applied to achieve tangible performance gains in real-
world structural components. For example, Challapalli et al. [89] developed an IML-GAN
framework that optimized the internal geometry of lightweight metamaterials, leading to
a 40–120% improvement in load-bearing capacity compared to traditional lattice designs.
Similarly, Shen et al. [103] proposed an inverse design architecture for high-performance
gradient honeycomb steel structures, where IML integration resulted in designs that of-
fered improved stress distribution and reduced material usage under high-impact loads.
Another study by Challapalli et al. [104] demonstrated how IML can be used to design
thin-walled cellular structures made from shape memory polymers, achieving record-high
specific recovery stress and a 200% improvement in normalized mechanical performance
compared to conventional unit cells. While focused on SMPs, the underlying framework is
directly applicable to lightweight steel-composite systems where strength and flexibility
must be balanced. IML has also been used in dual-phase and ultra-high-strength steels
to refine microstructure design, leading to higher tensile strength and optimized material
utilization [105]. These examples show that IML is no longer limited to conceptual or ex-
perimental research; it is actively shaping how structural efficiency, geometry optimization,
and weight minimization are achieved in modern steel engineering. Using IML on specific
structural engineering problems helps to clearly notice its usefulness. In load-bearing
capacity estimate, IML enables engineers to define the target resistance and quickly access
the design parameters—such as cross-section dimensions or material strength—that satisfy
that need. In material selection, IML considers cost, weight, or environmental impact in
addition to guiding the identification of perfect material types or grades meeting perfor-
mance criteria. Without involving many forward simulations, IML guides suitable forms,
profiles, or arrangement patterns, thus enabling geometric optimization. These projects
demonstrate how much IML can reduce demand for repeated analyses and hand-made
iterations by offering a more intelligent and simplified design workflow.
These applications collectively offer not only theoretical frameworks but also clear
numerical evidence for IML’s effectiveness in structural engineering. For example, IML-
driven metamaterial design achieved up to 120% increases in load capacity [89], inverse
alloy optimization exceeded 2 GPa tensile strength with validated experiments [95], and
fatigue life estimation frameworks reached predictive accuracy above R2 = 0.96 [101]. These
results confirm that IML is not merely an abstract modeling concept but a practically
validated tool for performance-driven structural optimization.

5. Enhancing Explainability in ML: Concepts, Methods, and Applications

in Structural Engineering
Explainability in ML is the ability of models to provide open insights into their
decision-making process so ensuring human comprehension of their predictions. Unlike
conventional black-box models in which the internal computations are difficult to under-
stand [106], explainable ML techniques aim to reveal the link between input features and
output preferences. Transparency of model behavior supports explainability to foster re-
sponsibility, trust, and usability in AI applications [11]. One major advantage of explainable
models is that they help to ensure that relevant and major events drive forecasts, reducing
the risk of erroneous or biased influences [8].
Moreover, explainability facilitates the identification of prejudices in training data,
thus advancing fairness and enhancing the quality of decisions made. Open models also
contribute to build resilience and reliability by identifying discrepancies or anomalies that
might compromise predictive accuracy [107,108]. Beyond only technical improvements,
Metals 2025, 15, 408 23 of 41

explainability supports ethical considerations since it helps users to critically evaluate

model decisions and change them to fit policy, fairness, and operational goals [109–111].
While interpretability stresses how easily a human can grasp the reasoning behind a
given prediction, explainability stresses on understanding the internal mechanisms and
interactions inside a model. Although the terms explainability and interpretability are
sometimes used synonymously. Both components are quite crucial in bridging the gap
between advanced ML algorithms and human knowledge so that AI-driven solutions
remain transparent, trustworthy, and practically relevant in many different fields [112–115].
XML aims to generate transparent, interpretable, trustworthy ML models guaranteeing their
ethical and useful application [116]. Especially in high-stakes decision-making processes,
one of its main objectives is to build user confidence by demonstrating that models behave
consistently and regularly. While explainability is only one aspect influencing trust in AI,
giving users tools to assess the reliability of a model will help to raise confidence in its
predictions [117,118]. Apart from this, XAI aims to provide explanations that allow the
identification of possible causal relationships inside data, thus transcending observations
based on correlation [119]. While conventional ML models focus on pattern recognition,
explainable methods help to identify underlying cause-and-effect dynamics, which can then
be validated using specialized inference techniques. Simplifying challenging models into
understandable insights is also fundamental for XAI so that users may grasp how different
features interact and support predictions. By allowing AI-generated decisions to be more
readily available, these justifications help to enable informed human supervision and
improved integration into useful applications [120]. XAI also emphasizes the transferability
of knowledge, that is, the capacity of knowledge to be applied to other problems once
the decision-making process of a model is well known. On the other hand, transferring
models across tasks depends on a clear awareness of their limitations and presumptions so
that predictions remain relevant in new contexts [121,122]. Yet another crucial emphasis of
explainability is keeping confidence and consistency in predictions. Users have to be able
to assess their certainty, especially in situations when model predictions directly influence
outputs produced by AI [123,124]. XAI transparency open communication of constraints
and uncertainty, thus improving the dependability of automated decision-making systems.
The main goal of XAI is to close the gap between technical complexity and human-centered
usability so ensuring that AI solutions remain responsible, interpretable, and compliant
with ethical and operational guidelines.

5.1. SHapley Additive exPlanations (SHAP)

Different approaches have developed from enhancing the interpretability of ML mod-
els; these can be usually categorized as either model-specific or model-agnostic [125].
Designed to investigate internal dynamics of particular algorithms, model-specific meth-
ods offer insights dependent on the structure of the model. By contrast, model-agnostic
approaches provide a more flexible approach appropriate for any ML model independent
of its underlying design.
One of the most often used model-agnostic interpretability methods [9] is SHAP, which
measures the contribution of particular features to model predictions using ideas from
game theory. This method views the predictive model as a cooperative game based on
Shapley values, in which case features play as players and their influence on the final
prediction is assessed methodically either including or excluding them from the model. The
Shapley value is the average contribution of a feature across all possible combinations of
input variables, guaranteeing a mathematically sound evaluation of feature importance [9].
SHAP is globally, as well as locally, interpretable. Globally, it identifies which traits
have the most influence over a whole dataset, providing a general picture of the model’s
Metals 2025, 15, 408 24 of 41

decision-making pattern [9,126]. SHAP shows how specific input values produced a given
result, clarifying local level individual predictions. Using binary representations, SHAP
produces linear approximations clarifying their influence on predictions by indicating the
presence or absence of features.
Beyond only increasing model transparency, SHAP is quite crucial for the ethical
development of AI since it supports responsibility and fairness. By precisely defining the
function of every feature in decision-making, SHAP helps to detect possible biases and
guarantees that models remain interpretable for end users. Even as ML advances, SHAP
remains a simple tool for enhancing trust and openness in complex predictive systems [127].
SHAP’s surrogate model links complex ML models with human interpretability [11]
using both local and global explanations. SHAP quantifies its individual influence on a
given prediction since it generates Shapley values for each local input feature. This clarifies
for consumers which truly influence an outcome and which have either little or no impact
on it. Globally, SHAP aggregates these local explanations to provide a more complete view
of feature significance over the entire dataset [9,107]. By means of methodically evaluating
the contribution of every feature, SHAP guarantees more transparency in model behavior,
separating between variables influencing decision-making from those less important.

5.2. XML Application in Steel Structures

To address the complexity of several instabilities in channel sections under axial com-
pression or bending, Mojtabaei et al. [128] created an XML-based model to forecast the
buckling behavior of thin-walled structural elements. High accuracy was obtained by
the trained ANN models; SHAP analysis revealed feature contributions. With regard to
material properties and geometry, Hu et al. [129] then developed XML models for the
probabilistic prediction of buckling stress in steel shear panel dampers. While core plate
thickness (t) had little effect, SHAP and feature importance analysis found yield strength
(fy), height-to-width ratio (α), width-to-thickness ratio (β), and initial imperfection (δ) as
critical factors affecting buckling stress. Building on these developments, Hou et al. [130]
developed an interpretable ML model to predict the probabilistic axial buckling strength
of steel circular hollow section members, addressing the impact of geometric, material,
and initial imperfections not explicitly accounted for in current design codes. Parameter
impact was assessed using global sensitivity analysis and SHAP. While greatly lowering
computational time, the ML-based probabilistic model showed that it closely matches nu-
merical and design code predictions and provides a more detailed and effective method for
axial buckling strength evaluation. Extending XML applications to more general structural
systems, Samadian et al. [131] developed meta databases and a surrogate modeling frame-
work for steel moment-resisting frames, providing a computationally efficient alternative
to nonlinear time history analysis for disaster risk management. Feature importance was
evaluated using random forest and SHAP analysis pointing up important elements in
SMRF design. XML studies have also concentrated much on fire resistance. Liu et al. [132]
created an XML-based framework to maximize cold-formed steel wall design for the econ-
omy and forecast their fire resistance time. Although SHAP analysis found important
input parameters, the XGBoost model—optimized with Bayesian tuning—showcased great
predictive accuracy. Further advancing ML applications in fire-resistant design, Tang
et al. [133] developed an XML framework to predict the axial capacity of cold-formed steel
channel sections at elevated temperatures, offering an effective alternative to conventional
techniques. Features were interpreted using SHAP analysis. The results show ML’s poten-
tial to provide faster and more consistent axial capacity predictions as well as to enhance
fire-resistant structural design. Apart from fire resistance, XML has helped to enhance the
design and reliability of perforated and cellular steel sections. Addressing constraints in
Metals 2025, 15, 408 25 of 41

traditional design provisions, Degtyarev et al. [134] presented a probabilistic XML model
based on Natural Gradient Boosting for predicting the resistance of laterally restrained
cellular steel beams. Reliability studies identified the resistance reduction factors needed
for compliance with European and US design frameworks; SHAP analysis was then used
to improve interpretability. Beyond static loading situations, Widanage et al. [135] investi-
gated the use of XML models for predicting blast loads on rigid structures, thus offering
a time-efficient alternative to experimental and numerical methods. Model transparency
was guaranteed using XML methods, thus confirming that predictions follow fundamental
blast physics rules. Anand et al. [136] similarly used XML models to forecast engineering
demand parameters in buckling-restrained braced frames under seismic load, thus reduc-
ing reliance on computationally costly simulations. XML methods emphasized important
seismic parameters affecting BRBF behavior, thus enhancing model transparency. Fan
et al. [137] suggested another creative use for an XML-based framework to forecast the axial
compressive capacity of Σ-shaped cold-formed steel with web opening, thus addressing the
inefficiencies of conventional finite element and experimental techniques. SHAP analysis
revealed important design parameters affecting axial capacity and their interactions, thus
offering interpretability. Extending on connections and fastener-related issues, Sarfarazi
et al. [138] used XML techniques to forecast the shear strength of stainless-steel column web
panels, addressing constraints in current design standards that do not entirely reflect the
strain hardening behavior of stainless steel. Twelve machine learning models were tested;
Extra Trees Regression had the best predictive accuracy. SHAP analysis identified bolt
diameter and the column’s second moment of inertia as the most influential factors in shear
strength. Predicting axial capacity and failure modes, Aloko et al. [139] investigated further
the use of XML models in cold-formed steel built-up columns. Their work showed that,
surpassing conventional strength prediction approaches, ML techniques could efficiently
record buckling interactions and load-bearing behavior. The interpretability of the models
supports even more the relevance of ML in design validation and optimization. Still, it
is difficult to guarantee generalizability over different section geometries and loading
conditions. Furthermore, helping seismic performance optimization are XML techniques.
Gharagoz et al. [140] presented an XML framework including XGBoost for structural
engineering optimizing seismic retrofitting techniques. Considered for seismic energy dis-
sipation and self-centering mechanisms, the framework incorporates the Spring-rotational
Friction Damper system. Transparency of models was improved using interpretable mod-
els, guaranteeing interpretability for use in decision-making. For the quick evaluation
of seismic resistance in steel frames, Su et al. [141] presented an XML-based method in-
cluding active learning techniques. The dataset comprised 1056 seismic response records
from 250 steel frames with varying geometric characteristics and steel grades. Evaluated
were seven ML models (DT, RF, SVM, KNN, ADA, XG, CB); Extreme Gradient Boosting
(XG) obtained the best accuracy (97.79%). SHAP analysis provided interpretability for
seismic response predictions. Applying XML models, Gatheeshgar et al. [142] projected
web-crippling strength in cold-formed steel beams with staggered sloped perforations.
Four ML models—KNN, RF, SVR, ANN—were assessed. Slope length and bearing plate
length were found by SHAP analysis to be main determinants of web-crippling strength.
Sarfarazi et al. [143] presented a hybrid XML framework for examining the mechanical
response of stainless-steel beam-to-column connections.
Beyond their academic importance, these interpretability techniques now serve practi-
cal roles in engineering decision-making. For example, when SHAP values identify specific
parameters (such as flange width, thickness ratio, or bolt diameter) as dominant influencers
in capacity prediction, engineers can confidently prioritize these variables during section
design or material selection. Similarly, in XML-based fire resistance studies, the clear iden-
Metals 2025, 15, 408 26 of 41

tification of wall thickness or insulation parameters helps design safer cold-formed steel
walls without extensive physical testing. In seismic applications, SHAP analysis pinpoints
key damping and stiffness properties, allowing for informed retrofitting or configuration
changes. These tools enable engineers to move from ‘black-box’ models to traceable and
code-aligned decisions, strengthening both model trust and practical design outcomes.

6. Addressing Challenges, Limitations, and Future Directions

Although ML models show very good accuracy in estimating structural responses,
several elements limit their usability in professional engineering practice including data
availability, model reliability, interpretability, integration with conventional design methods,
and regulatory acceptance. This section methodically addresses the shortcomings of current
approaches, investigates the main barriers preventing the full-scale implementation of ML
in structural engineering, and proposes future directions of research to close the distance
between ML-driven discoveries and useful applications.

6.1. Rethinking Model Training: Advancing from Accuracy to Engineering Reliability

6.1.1. Generalization and Overfitting in Structural ML Models
The main issue with current ML applications in steel construction is the excessive
emphasis of model accuracy without considering engineering reliability. Although many
studies reveal low error rates and strong R2 values, these measures do not always translate
into safe and reliable structural designs. Overfitting is a main problem whereby a model
performs remarkably on training data but fails in real-world or unseen conditions. Unlike
conventional engineering models including physical laws (such as equilibrium equations
and constitutive models), many ML techniques depend just on statistical learning. This can
generate physically contradicting predictions whereby results seem statistically valid but
violate fundamental engineering criteria.

6.1.2. Addressing the Gap in Physics-Based Learning

Unlike in computational mechanics and fluid dynamics, where Physics-Informed
Neural Networks (PINNs) and hybrid solvers have gained considerable ground—most
ML applications in structural engineering remain entirely data-driven [144]. Although on
well-curated datasets these black-box models can show remarkable accuracy, they usually
lack embedded physical constraints including equilibrium, compatibility, or constitutive
relations. As a result, such models risk producing outputs that, although numerically plau-
sible, may contradict fundamental engineering principles, especially under extrapolated or
safety-critical conditions [145].
The structural engineering community is starting to investigate Physics-Informed
Machine Learning (PIML) as a more reliable substitute to mitigate this shortcoming. These
models directly include domain-specific physical laws including elasticity, buckling equa-
tions, or dynamic balance into the ML training process. For example, Raissi et al. [145]
introduced the foundational framework of PINNs, in which governing partial differential
equations (PDEs) are embedded in the neural network’s loss function. Haghighat et al. [146]
later extended this method to solid mechanics problems, enabling the network to produce
stress or displacement fields that respect the governing equations of structural behavior.
In parallel, hybrid ML–FEM approaches have emerged that blend neural network-
based surrogates with traditional finite element formulations. These models seek to pre-
serve mechanical integrity of the results while lowering computational effort—especially in
repeated simulations. Recent models, for instance, let ML models act as fast approximators
for particular FEM sub-processes including stress field estimate or crack growth under
variable boundary conditions [145,147].
Metals 2025, 15, 408 27 of 41

A comprehensive review by Habib et al. [144] cataloged over 90 PIML applications tai-
lored to structural PDEs, spanning use cases like stiffness identification, crack propagation,
and damage localization. When training data are limited or difficult to obtain—such as
post-buckling behavior, fatigue evolution, or high-strain-rate responses—these techniques
are especially helpful. Physics-guided constraints in such environments not only stabilize
learning but also strengthen engineering confidence in model outputs.
Physics-based learning must clearly be central in bridging the gap between academic
experimentation and practical adoption as structural ML develops. Including mechanical
rules into data-driven models increases generalizability, transparency, and offers a necessary
basis for future possible regulatory approval.

6.1.3. Future Directions for Reliability-Oriented Model Training

• Creating PIML Models
• Hybrid Engineering-ML Models
• Uncertainty Quantification (UQ)

6.2. The Overlooked Role of Feature Engineering and Domain Expertise

6.2.1. The Difficulties Using Raw Data Without Engineering Context
In structural engineering, feature engineering—the choice and modification of input
variables—is among the most crucial but underappreciated elements of ML modeling.
Many studies direct raw numerical data—such as cross-sectional dimensions, material
strength, and loading conditions—into ML models without regard for the deeper physical
interactions guiding structural behavior. For buckling failure, for instance, just entering the
height, width, and thickness of a steel column might not be sufficient. Rather, engineering
equation-derived features—such as slenderness ratio, moment of inertia, and effective
length factor—may offer more important predictors.

6.2.2. Why Structural Analysis Driven by ML Requires Domain Knowledge

While fields like image recognition [148] or natural language processing [149] where
raw data directly reflects the problem space, structural engineering calls for a complete
awareness of mechanics and materials. Ignorance of domain knowledge might lead to
models that memorize patterns instead of learning fundamental structural behavior. How-
ever, an often overlooked yet critical issue in ML-based structural analysis is the limited
transferability of models across different steel structural typologies. A model trained on
I-section beams under flexure may not perform well when applied to box-section columns
or perforated trusses under axial or torsional loading. This lack of generalizability stems
from fundamental differences in governing mechanics, geometric characteristics, failure
modes, and load interactions. To address this, several approaches have been proposed.
One promising direction is transfer learning, where a pre-trained model is fine-tuned using
smaller datasets from a new typology, preserving learned features while adapting to new
contexts [150]. Another emerging method is meta-learning, which aims to train models
capable of quickly adapting to new structural scenarios with limited data [151]. Domain
adaptation techniques are also under exploration to help models account for the shift in
data distributions when moving from one structural typology to another [152]. While these
techniques offer potential, their use in structural engineering remains limited and largely
exploratory. Therefore, future research must systematically investigate cross-typology
learning to develop truly generalizable and robust AI models for structural applications.

6.2.3. Future Directions for Feature Engineering and Domain-Driven Learning

• Creating feature selection techniques grounded in engineering
• Incorporating Expert Knowledge into Feature Engineering
Metals 2025, 15, 408 28 of 41

• Automated Feature Extraction Techniques

6.3. From Deterministic to Probabilistic Modeling

6.3.1. The Importance of Uncertainty Quantification in Structural Predictions
ML models treat structural properties as fixed parameters rather than stochastic
variables, thus providing single-value predictions. Structural materials do, however, exhibit
natural variation, and operational and environmental uncertainty results in loads never
exactly known. Traditional engineering methods allow for these uncertainties using safety
factors; but, ML models typically ignore uncertainty. For example, a neural network
forecasting the shear strength of a steel joint could produce a value of 500 kN but does not
show a confidence range (e.g., 480–520 kN). The lack of probabilistic insight reduces the
reliability of ML-based predictions in applications related to safety.

6.3.2. The Role of Bayesian ML and Monte Carlo Simulations

Two fundamental components of safety-critical structural design, uncertainty quantifi-
cation and probabilistic risk assessment, can be supported by ML beyond point predictions.
Conventional engineering frameworks like Load and Resistance Factor Design (LRFD) em-
bed probabilistic safety margins, yet most ML models remain deterministic. This disconnect
limits their practical reliability in engineering settings.
To bridge this gap, Probabilistic ML techniques such Bayesian Neural Networks
(BNNs) [153] and Monte Carlo [154] dropout have been adopted to estimate predictive dis-
tributions rather than single values. These techniques help quantify epistemic uncertainty—
arising from limited training data—and allow engineers to evaluate how much trust can be
placed in a model’s prediction [155]. For example, a BNN can estimate that a predicted shear
capacity of 480 kN has a 95% confidence range of 450–510 kN, helping decision-makers
incorporate statistical safety considerations directly into their workflow.
In parallel, Gaussian Process Regression (GPR) is increasingly used to develop proba-
bilistic surrogate models that approximate finite element responses under uncertain loads
or material properties [156]. These surrogates dramatically reduce computational cost
when thousands of simulations are needed, for instance, in fragility curve generation or
reliability-based design optimization (RBDO). When integrated with Monte Carlo simula-
tions or Latin Hypercube Sampling, such models enable full probabilistic risk evaluations
without sacrificing computational efficiency [157].
Embedding explainable AI (XAI) tools—such as SHAP or feature sensitivity analysis—
within these frameworks further enhances their credibility. Engineers can not only assess
the probability of structural failure but also understand which input uncertainties most
influence those risks, improving transparency and accountability [9].
These approaches are now beginning to reshape performance-based structural design,
particularly under seismic and fatigue loading scenarios. While regulatory approval is
still pending, this convergence of AI and probabilistic engineering offers a promising path
toward data-informed, risk-conscious structural decision-making.

6.3.3. Future Directions for Probabilistic and Uncertainty-Aware Modeling

• Adopting Bayesian ML techniques
• Integrating ML with probabilistic design codes
• Developing uncertainty-aware ML models
Metals 2025, 15, 408 29 of 41

6.4. Bridging the Gap Between Research and Practice

6.4.1. Why ML Still Is Not Often Applied in Industry?
Although many studies indicate that ML is better than conventional methods, very
few ML models have been included into commercial engineering tools. Some primary
factors include:
• ML models lack a universal benchmarking system unlike FEA, which follows accepted
verification processes.
• Some deep learning architectures demand significant computational resources, thus
they are useless for real-time analysis.
• Regulatory and Code Compliance Problems; a highly regulated discipline, struc-
tural engineering lacks formal approval of engineering codes using ML-based design
methods.
Another key factor limiting the industrial uptake of AI-driven structural optimization
is the associated computational cost. The training phase—particularly for deep learning
models, evolutionary algorithms, or high-dimensional inverse design frameworks—often
requires access to high-performance computing (HPC) environments, including GPUs or
multi-core CPU clusters [158]. This is especially true when models are coupled with finite
element simulations or need to explore large design spaces through surrogate modeling or
generative methods.
However, it is important to note that the computational burden is mostly concentrated
during the training or model development phase. Once trained, most ML and IML models
are highly efficient at inference and can generate design predictions or optimize solutions
in real time on standard computing hardware [159]. Additionally, the emergence of cloud-
based AI platforms and lightweight optimization strategies (e.g., pruning, quantization,
and knowledge distillation) now offers engineers access to powerful tools without owning
dedicated HPC resources [160]. Nonetheless, to train advanced models or explore multi-
objective optimizations at scale, HPC remains an enabler rather than a barrier, especially in
research or enterprise settings.

6.4.2. Future Directions for Bridging the Gap Between Research and Practice
• Creating Industry-Standard Validation Benchmarks
• Developing ML-Integrated Engineering Software
• Regulatory Frameworks for AI in Structural Engineering

6.5. Expanding the Role of IML in Structural Design

6.5.1. IML and Generative AI as the Next Frontier in Structural Design Automation
While most structural ML applications remain focused on forward prediction, IML
and generative AI are reshaping the early stages of structural design by allowing engineers
to define performance targets and automatically generate optimized design configurations.
This shift holds major implications for steel construction, where structural complexity,
load-bearing demands, and geometric constraints often lead to highly iterative and labor-
intensive design cycles.
In generative design, AI models such as Generative Adversarial Networks (GANs) and
Variational Autoencoders (VAEs) are trained to synthesize novel structural topologies based
on desired performance criteria. These models do not merely search for optimum values in
a fixed design space; instead, they learn the underlying structural logic and generate entirely
new configurations that satisfy both architectural intent and engineering performance. For
example, Challapalli et al. [104] demonstrated an inverse design framework using GANs to
generate cellular structures with record-breaking recovery stress, revealing how generative
models can produce designs that outperform conventional configurations.
Metals 2025, 15, 408 30 of 41

In the context of steel frameworks, ML-guided topology optimization—which couples

finite element analysis with reinforcement learning or surrogate modeling—has emerged as
a powerful technique for generating lightweight, efficient structural forms. AI algorithms
explore design variants that would be computationally prohibitive to evaluate through
brute-force methods. This has enabled engineers to optimize member layout, cross-sectional
geometry, and bracing patterns in ways that reduce weight and material usage without
sacrificing stiffness or stability.
Despite these advances, practical challenges remain. Generative models often produce
non-code-compliant geometries, requiring post-processing to ensure structural safety and
constructability. Additionally, real-world datasets for training such models are scarce,
and integration with BIM and FEM workflows is still underdeveloped. Nonetheless, the
growing synergy between IML, generative design, and performance-based engineering
is paving the way for a new paradigm in steel structure design, where AI supports both
creativity and compliance.

6.5.2. Future Directions for Inverse and Generative AI in Structural Design

• Integration of GANs and VAEs into Structural Topology Optimization
• Coupling Generative Design with FEM and BIM Systems
• Data-Driven Evolutionary Design Platforms for Custom Steel Systems

6.6. Regulatory Acceptance and Validation of AI Models in Structural Engineering

6.6.1. Current Regulatory Limitations and Engineering Concerns
Although AI and ML are rapidly advancing structural engineering, they are not
formally approved in safety-critical design by any major code authority including the
Eurocodes, AISC, or ASCE yet. Current building codes still rely on traditional analytical or
empirical formulas that have been tested and validated over decades. These codes have not
yet adapted to the use of complex, data-driven models like AI, especially in their “black-
box” form, where internal logic may not be transparent or physically interpretable [161].
For now, any AI-based model applied in engineering has to be thoroughly tested and
demonstrated to satisfy the same safety criteria established in the codes. This typically
means comparing AI predictions with experimental data or ensuring that they align with
the safety factors prescribed by standards. Particularly in predicting structural resistance
when trained on large, high-quality datasets, several studies have shown that AI can achieve
either similar or even better accuracy than conventional codes [162]. Notwithstanding this
technical promise, there is still no official path for AI models to be certified or approved as
direct substitutes for code-based design equations. To address this issue, researchers have
proposed the idea of “engineering equivalence”. This method develops an ML model to
replicate the output of a known code equation, thus guaranteeing consistent and logical
behavior with accepted engineering principles [163]. If this equivalence is achieved, the
AI model might be accepted as a substitute for the original equation. Although this idea
is still at an early stage, it shows promise for future updates to codes. At the same time,
international standards organizations such as ISO/IEC JTC1 SC42 are working on high-level
AI quality and trust guidelines that could later be adapted to civil and structural engineering
needs [164,165]. Even so, there are currently no specific clauses in design codes that allow
general approval of AI models. In practice, if AI is used, it must typically be justified under
“alternative design methods” and backed by strong validation. Explainability is one of
the key causes of this uncertainty. Engineering authorities are naturally wary of adopting
models that perform well but lack any evident justification for their projections. Given the
generalizability of AI models [166], this becomes even more critical.
Metals 2025, 15, 408 31 of 41

The lack of comprehensive training data is another issue. Especially catastrophic

structural failures are rare and costly to test. Training datasets thus are often small, biased,
or lacking important failure scenarios [167]. There are also concerns about accountability:
if an AI-guided decision results in a structural failure, it is unclear who bears the legal or
ethical responsibility—the software developer, the design engineer, or the authority that
accepted the design [161].
These unresolved problems keep most AI applications in structural engineering in a
supporting role. Although they help with design exploration, optimization, and monitoring,
they are not yet accepted as replacements for code-approved methods in final decision-
making. To move forward, researchers are concentrating on creating explainable and hybrid
models, combining AI with physics-based rules, and suggesting more precise validation
and certification guidelines [159,164]. AI will remain a complimentary but not central
player in the structural engineering code environment until those efforts develop.

6.6.2. Role of Explainable AI (XAI) in Meeting Regulatory Demands

Explainable AI (XAI) models—those that integrate techniques like SHAP, and global
sensitivity analysis—are especially promising for regulatory consideration. Their trans-
parency and interpretability let engineers to trace model predictions to specific design
inputs, satisfying key principles of engineering traceability. This transparency can support
formal validation by offering clear reasoning, which is essential for approval in safety-
critical applications. While no design code formally certifies XAI models for structural
design, their fit with rising trustworthiness criteria makes them attractive candidates for
future code inclusion. XAI tools remain useful as interpretive aids until codified paths de-
velop, bridging the gap between high-performance AI models and the regulatory demand
for transparent, verifiable decision-making.
Moving forward, the structural engineering community must work toward devel-
oping formal AI regulatory frameworks tailored for design approval, certification, and
standardization. These frameworks should define quality control protocols, minimum inter-
pretability thresholds, validation benchmarks against code-based methods, and pathways
for certifying hybrid models that combine AI with physical principles. Establishing such
standards will require coordinated input from engineers, AI researchers, code authorities
(e.g., Eurocodes, AISC), and software developers. Without these regulatory foundations,
the full industrial adoption of AI in structural design will remain restricted to non-critical
applications or limited to use under alternative design justifications.

6.6.3. Future Directions for Regulatory Approval and Model Certification

• Establishing formal AI regulatory frameworks tailored for civil/structural engineering
• Defining interpretability thresholds and validation protocols
• Creating certification pathways for hybrid models combining AI and physical princi-
ples

6.7. Practical Barriers to Industry Adoption of AI in Structural Engineering

6.7.1. Data Scarcity, Trust, and the Need for Validation
Although many scholarly studies show the strength of AI models in structural en-
gineering, their actual application in practice is still rather restricted. Data availability
presents one of the key difficulties. Many engineering companies find that high-quality
structural datasets are either difficult to gather or nonexistent. Many datasets are kept
privately by businesses or organizations, and experimental results are often few. Most
AI models stay limited to particular cases or academic use [162,164] without access to
large and well-structured data. Lack of interpretability is another main problem. Most
engineers are taught to operate using transparent models based on exact physical principles.
Metals 2025, 15, 408 32 of 41

Engineers often find AI models unreliable when they make predictions without displaying
how or why, particularly in safety-critical contexts. This lack of trust becomes a serious
concern when the model output could affect structural performance, risk evaluation, or
code compliance [168]. The need for engineering-level validation forms a third practical
obstacle. In academic research, an AI model may be sufficient to exhibit great test data accu-
racy. In industry, however, models are expected to satisfy code-based safety margins, pass
experimental comparisons, and offer repeatable results under several design environments.
These expectations are not only technical but also linked to professional liability and
public safety. As long as validation procedures for AI remain unclear or unsupported by
formal standards, adoption will remain slow [161,165,167].

6.7.2. Future Directions for Industry Adoption of AI in Structural Engineering

• Enhancing access to shared structural datasets
• Promoting model transparency in safety-critical use cases
• Aligning industrial validation with regulatory expectations

6.8. Advancing Seismic Modeling Through Multi-Physics AI

6.8.1. Limitations of Current Seismic AI Models
Most structural AI applications today treat seismic performance as a simplified dy-
namic problem—often ignoring the coupling of multi-physical factors such as temperature
variation, soil-structure interaction, and real-time degradation during earthquakes. This
reductionist approach limits prediction accuracy in real-world scenarios. As structural
systems become more complex, and the demand for resilient infrastructure increases,
integrating AI with multi-physics modeling becomes important for seismic design and
real-time monitoring.

6.8.2. Emerging Research on Multi-Physics AI Integration

Recent studies have explored the fusion of AI with multi-physics finite element meth-
ods, including soil-fluid–structure interaction, thermal-mechanical behavior under seismic
excitation, and even coupled damage evolution. These models allow AI to learn from
both physical laws and real-time data, creating better generalization under rare or extreme
events [168,169].
Recent advancements in seismic performance prediction for steel structures highlight
the effectiveness of physics-informed machine learning (PiML). By embedding fundamen-
tal physical laws into neural networks, particularly using techniques like LSTM, these
models reliably predict nonlinear seismic behaviors such as interstory drift, plastic hinge
formation, and collapse mechanisms. This integration enhances prediction accuracy and
interpretability compared to traditional methods. However, computational demand and
data calibration challenges still limit their widespread adoption in design offices.
Despite promising progress, multi-physics AI models require improved data infras-
tructure, standardized validation protocols, and transparent uncertainty handling to gain
broader trust and usability in seismic safety-critical decisions.

6.8.3. Feature Direction for Advancing Seismic Modeling Through Multi-Physics AI

• Coupling PINNs with seismic time-history simulations
• Creating benchmark datasets for earthquake-driven ML training
• Developing real-time digital twins for seismic assessment
• Standardizing validation protocols in Eurocode 8, ASCE 41, etc.
Together, ML, IML, and XAI provide a complementary framework for reshaping
structural engineering processes. Their integration has the potential to modernize design
Metals 2025, 15, 408 33 of 41

strategies while upholding safety, transparency, and regulatory integrity. To summarize

their respective strengths and challenges, Table 3 provides a comparative overview of these
three approaches in terms of their structural applications, benefits, and limitations. The
future trajectory of AI in structural engineering depends on bridging the gap between
academic advances and industry practices through explainable methods, hybrid physics-
guided models, and reliable validation procedures. Furthermore, AI will increasingly
contribute to sustainable, climate-adaptive design and life-cycle optimization. Inverse
ML will redefine how engineers generate solutions from desired outcomes, opening new
frontiers in structural optimization. Achieving this vision requires cross-disciplinary collab-
oration, industrial validation, and stronger AI literacy in engineering education.

Table 3. Summary of ML, IML, and XAI.

Key Structural
Technique Strengths Limitations
Applications
- Rapid prediction of
structural responses once
- Often acts as a “black box” - Performance prediction
trained.
with low interpretability. under complex loads.
- Capable of handling large,
- Requires large and - Data-driven structural
nonlinear,
Machine Learning (ML) high-quality labeled health monitoring.
high-dimensional datasets.
datasets. - Load-carrying capacity
- Useful for surrogate
- Limited generalization and failure mode
modeling, failure
outside trained domains. prediction.
classification, and load
capacity estimation.
- Inverse problems can be
- Directly maps
ill-posed and unstable. - Automated design of
performance goals to
- Often requires cross-sections and steel
optimal design parameters.
regularization or surrogate profiles.
Inverse Machine Learning - Reduces manual iteration
models to ensure - Topology optimization.
(IML) in parametric design.
convergence. - Material and
- Efficient for optimization
- Experimental validation microstructure tuning in
in multi-variable,
still limited in structural steel alloy design.
constrained problems.
contexts.
- Code validation and
- Still emerging in
- Automated design of transparency for AI-driven
regulatory practice.
cross-sections and steel designs.
- Trade-off between
profiles. - SHAP/LIME
complexity and
Explainable AI (XAI) - Topology optimization. interpretation of failure
explainability.
- Material and risk.
- Interpretations can be
microstructure tuning in - Engineering decision
misused if not
steel alloy design. support in safety-critical
domain-verified.
systems.

7. Conclusions
Including AI into structural engineering transforms a paradigm change in the design,
analysis, and optimization of steel constructions. Over the past three decades, fast devel-
opment of ML techniques has given engineers great tools to increase predictive accuracy,
automate challenging tasks, and derive data-driven insights. But rather than only replac-
ing traditional engineering knowledge, the evolution of AI in this field is about raising
its capacity to negotiate the complexity of contemporary building challenges. Using AI,
engineers can adopt adaptive, real-time solutions that account for uncertainty, variability,
and evolving design constraints, moving beyond traditional simulation-based approaches.
Metals 2025, 15, 408 34 of 41

Particularly in fields where more conventional methods have limits, one of the most
transforming consequences of AI is its ability to enable more exact decision-making. From
load-bearing studies and failure predictions to material selection and seismic performance
evaluations, AI models have shown their capacity to provide engineers more complete
knowledge of structural behavior. Particularly with IML, engineers can let AI select the
best design solutions and directly specify target performance criteria, thus adding a new
dimension in optimization. This shift represents a fundamental move away from trial-and-
error methods, streamlining the design process and reducing dependence on repetitive
simulations
Notwithstanding these advances, the path toward widespread acceptance of AI in
structural engineering remains a long one. Practical implementation of AI-driven models
still requires greater transparency, standardized practices, and alignment with regulatory
codes to ensure their reliability and safety in real-world applications. Moreover, the
demand for high-quality, diverse datasets is still a main challenge, since AI models are only
as successful as the data they are trained on. Overlooking issues such as data limitations,
overfitting, and model interpretability will confine the role of AI in structural engineering
to academic research, rather than enabling its practical application
Looking ahead, AI will influence developing fields including digital twin technology,
sustainability, and automated building methods beyond its traditional applications, impact-
ing steel structure engineering. Real-time monitoring systems allow AI-powered design
frameworks to interact and apply proactive structural assessments and predictive mainte-
nance programs. Moreover, as engineering education evolves, AI literacy will become an
essential skill for future professionals, ensuring that engineers not only know how to use
AI tools but also critically assess their outputs within the context of engineering principles.
In structural engineering, AI is fundamentally reshaping the approach to design and
analysis, going beyond mere technological advancement. Though challenges still exist, the
continuous development of AI-driven technologies offers interesting paths for acquiring
more intelligent, resilient, and efficient steel buildings. By supporting cooperation between
academia, industry, and regulatory authorities, the engineering community can maximize
AI’s possibilities and so guarantee that its application stays rooted in safety, reliability, and
engineering intuition.

Author Contributions: S.S.: Writing—original draft, Writing—review and editing, Data curation,
Software. I.M.: Supervision, Visualization. M.M.: Supervision. F.G.: Supervision. All authors have
read and agreed to the published version of the manuscript.

Funding: This research was funded by Italian Ministry of Education, University and Research (MIUR),
Departments of Excellence (grant number L.232/2016), Italian Ministry of Education, University and
Research (MIUR), Next-Generation EU-Prin (PNRR) (grant number 2022P7PF8J) and Italian Ministry
of Education, University and Research (MIUR), Next-Generation EU-Prin (PNRR) (grant number
P2022Y9ZJ2).

Data Availability Statement: The data presented in this study are available on request from the
corresponding author.

Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Thai, H.-T. Machine learning for structural engineering: A state-of-the-art review. Structures 2022, 38, 448–491. [CrossRef]
2. Sun, H.; Burton, H.V.; Huang, H. Machine learning applications for building structural design and performance assessment:
State-of-the-art review. J. Build. Eng. 2021, 33, 101816. [CrossRef]
3. Salehi, H.; Burgueño, R. Emerging artificial intelligence methods in structural engineering. Eng. Struct. 2018, 171, 170–189.
[CrossRef]
Metals 2025, 15, 408 35 of 41

4. Málaga-Chuquitaype, C. Machine learning in structural design: An opinionated review. Front. Built Environ. 2022, 8, 815717.
[CrossRef]
5. Haber, E.; Ascher, U.M.; Oldenburg, D. On optimization techniques for solving nonlinear inverse problems. Inverse Probl. 2000,
16, 1263. [CrossRef]
6. Gallet, A.; Rigby, S.; Tallman, T.N.; Kong, X.; Hajirasouliha, I.; Liew, A.; Liu, D.; Chen, L.; Hauptmann, A.; Smyl, D. Structural
engineering from an inverse problems perspective. Proc. R. Soc. A Math. Phys. Eng. Sci. 2022, 478, 20210526. [CrossRef]
7. Barredo Arrieta, A.; Díaz-Rodríguez, N.; Del Ser, J.; Bennetot, A.; Tabik, S.; Barbado, A.; Garcia, S.; Gil-Lopez, S.; Molina,
D.; Benjamins, R.; et al. Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities, and challenges toward
responsible AI. Inf. Fusion 2020, 58, 82–115. [CrossRef]
8. Angelov, P.P.; Soares, E.A.; Jiang, R.; Arnold, N.I.; Atkinson, P.M. Explainable artificial intelligence: An analytical review. Wiley
Interdiscip. Rev. Data Min. Knowl. Discov. 2021, 11, e1424. [CrossRef]
9. Lundberg, S.M.; Lee, S.I. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst. 2017, 30, 2493.
10. Minh, D.; Wang, H.X.; Li, Y.F.; Nguyen, T.N. Explainable artificial intelligence: A comprehensive review. Artif. Intell. Rev. 2022, 55,
3503–3568. [CrossRef]
11. Jiang, T.; Gradus, J.L.; Rosellini, A.J. Supervised machine learning: A brief primer. Behav. Ther. 2020, 51, 675–687. [CrossRef]
[PubMed]
12. Alloghani, M.; Al-Jumeily, D.; Mustafina, J.; Hussain, A.; Aljaaf, A.J. A systematic review on supervised and unsupervised machine
learning algorithms for data science. In Supervised and Unsupervised Learning for Data Science; Springer: Cham, Switzerland, 2020;
pp. 3–21.
13. Nian, R.; Liu, J.; Huang, B. A review on reinforcement learning: Introduction and applications in industrial process control.
Comput. Chem. Eng. 2020, 139, 106886. [CrossRef]
14. Adeli, H. Neural networks in civil engineering: 1989–2000. Comput.-Aided Civ. Infrastruct. Eng. 2001, 16, 126–142. [CrossRef]
15. Song, B.; Zhou, R.; Ahmed, F. Multi-modal machine learning in engineering design: A review and future directions. J. Comput.
Inf. Sci. Eng. 2024, 24, 010801. [CrossRef]
16. Daneshfar, R.; Esmaeili, M.; Mohammadi-Khanaposhtani, M.; Baghban, A.; Habibzadeh, S.; Eslamian, S. Advanced machine
learning techniques: Multivariate regression. In Handbook of Hydroinformatics; Elsevier: Amsterdam, The Netherlands, 2023; pp.
1–38. [CrossRef]
17. Mienye, I.D.; Jere, N. A survey of decision trees: Concepts, algorithms, and applications. IEEE Access 2024, 12, 86716–86727.
[CrossRef]
18. Schonlau, M.; Zou, R.Y. The random forest algorithm for statistical learning. Stata J. 2020, 20, 3–29. [CrossRef]
19. Tanveer, M.; Rajani, T.; Rastogi, R.; Shao, Y.H.; Ganaie, M.A. Comprehensive review on twin support vector machines. Ann. Oper.
Res. 2024, 339, 1223–1268. [CrossRef]
20. Abd-elaziem, A.H.; Soliman, T.H. A multi-layer perceptron (MLP) neural network for stellar classification: A review of methods
and results. Int. J. Adv. Appl. Comput. Intell. 2023, 3, 54216.
21. Abiodun, O.I.; Jantan, A.; Omolara, A.E.; Dada, K.V.; Mohamed, N.A.; Arshad, H. State-of-the-art in artificial neural network
applications: A survey. Heliyon 2018, 4, e00938. [CrossRef]
22. Halder, R.K.; Uddin, M.N.; Uddin, M.A.; Aryal, S.; Khraisat, A. Enhancing K-nearest neighbor algorithm: A comprehensive
review and performance analysis of modifications. J. Big Data 2024, 11, 113. [CrossRef]
23. Freund, Y.; Schapire, R.E. Experiments with a new boosting algorithm. ICML 1996, 96, 148–156.
24. Friedman, J.H. Greedy function approximation: A gradient boosting machine. Ann. Stat. 2001, 29, 1189–1232. [CrossRef]
25. Chen, T.; Guestrin, C. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794. [CrossRef]
26. Ke, G.; Meng, Q.; Finley, T.; Wang, T.; Chen, W.; Ma, W.; Liu, T.Y. LightGBM: A highly efficient gradient boosting decision tree.
Adv. Neural Inf. Process. Syst. 2018, 30, 1998–2007. [CrossRef]
27. Prokhorenkova, L.; Gusev, G.; Vorobev, A.; Dorogush, A.V.; Gulin, A. CatBoost: Unbiased boosting with categorical features.
In Advances in Neural Information Processing Systems; Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N.,
Garnett, R., Eds.; Curran Associates, Inc.: Red Hook, NY, USA, 2018; Volume 31, pp. 6638–6648. Available online: https:
//proceedings.neurips.cc/paper/2018/file/14491b756b3a51daac41c24863285549-Paper.pdf (accessed on 1 April 2025).
28. Mascolo, I.; Guarracino, F.; Sarfarazi, S.; Della Corte, G. A proposal for a simple characterization of stainless steel connections
through an equivalent yield strength. Structures 2024, 68, 107043. [CrossRef]
29. Sarfarazi, S.; Shamass, R.; Mascolo, I.; Della Corte, G.; Guarracino, F. Some considerations on the behaviour of bolted stainless-steel
beam-to-column connections: A simplified analytical approach. Metals 2023, 13, 753. [CrossRef]
30. Sarfarazi, S.; Shamass, R.; Della Corte, G.; Guarracino, F. Assessment of design approaches for stainless-steel joints through an
equivalent FE modelling technique. ce/papers 2022, 5, 271–281. [CrossRef]
Metals 2025, 15, 408 36 of 41

31. Sarfarazi, S.; Saffari, H.; Fakhraddini, A. Shear behavior of panel zone considering axial force for flanged cruciform columns. Civ.
Eng. Infrastruct. J. 2020, 53, 359–377. [CrossRef]
32. Sarfarazi, S.; Fakhraddini, A.; Modaresahmadi, K. Evaluation of panel zone shear strength in cruciform columns, box-columns
and double web-columns. Int. J. Struct. Civ. Eng. 2016, 5, 52–56.
33. Saffari, H.; Sarfarazi, S.; Fakhraddini, A. A mathematical steel panel zone model for flanged cruciform columns. Steel Compos.
Struct. 2016, 20, 851–867. [CrossRef]
34. Paral, A.; Roy, D.K.S.; Samanta, A.K. A deep learning-based approach for condition assessment of semi-rigid joint of steel frame.
J. Build. Eng. 2021, 34, 101946. [CrossRef]
35. Kueh, A.B.H. Artificial neural network and regressed beam-column connection explicit mathematical moment-rotation expres-
sions. J. Build. Eng. 2021, 43, 103195. [CrossRef]
36. Tran, V.L. Investigating the behavior of steel flush endplate connections at elevated temperatures using FEM and ANN. Int. J.
Steel Struct. 2022, 22, 1433–1451. [CrossRef]
37. Sarothi, S.Z.; Ahmed, K.S.; Khan, N.I.; Ahmed, A.; Nehdi, M.L. Predicting bearing capacity of double shear bolted connections
using machine learning. Eng. Struct. 2022, 251, 113497. [CrossRef]
38. Jiang, K.; Liang, Y.; Zhao, O. Machine-learning-based design of high-strength steel bolted connections. Thin-Walled Struct. 2022,
179, 109575. [CrossRef]
39. Rabbani, A.; Ghiami Azad, A.R.; Rahami, H. Enhancing prediction of the moment-rotation behavior in flush end plate connections
using Multi-Gene Genetic Programming (MGGP). Struct. Eng. Mech. 2024, 91, 643–656. [CrossRef]
40. Ferreira, F.P.V.; Shamass, R.; Limbachiya, V.; Tsavdaridis, K.D.; Martins, C.H. Lateral–torsional buckling resistance prediction
model for steel cellular beams generated by artificial neural networks (ANN). Thin-Walled Struct. 2022, 170, 108592. [CrossRef]
41. Shamass, R.; Ferreira, F.P.V.; Limbachiya, V.; Santos, L.F.P.; Tsavdaridis, K.D. Web-post buckling prediction resistance of steel
beams with elliptically-based web openings using artificial neural networks (ANN). Thin-Walled Struct. 2022, 180, 109959.
[CrossRef]
42. Silva de Carvalho, A.; Hosseinpour, M.; Rossi, A.; Martins, C.H.; Sharifi, Y. New formulas for predicting the lateral–torsional
buckling strength of steel I-beams with sinusoidal web openings. Thin-Walled Struct. 2022, 181, 110067. [CrossRef]
43. Xing, Z.; Wu, K.; Su, A.; Wang, Y.; Zhou, G. Intelligent local buckling design of stainless steel I-sections in fire via Artificial Neural
Network. Structures 2023, 58, 105356. [CrossRef]
44. Rossi, A.; Hosseinpour, M.; Silva de Carvalho, A.; Martins, C.H.; Sharifi, Y. Lateral torsional capacity of steel beams in different
loading conditions by neural network. Proc. Inst. Civ. Eng. Struct. Build. 2023, 177, 892–910. [CrossRef]
45. Cheng, J.; Li, X.; Jiang, K.; Li, S.; Su, A.; Zhao, O. Machine-learning-assisted design of high-strength steel I-section columns. Eng.
Struct. 2024, 308, 118018. [CrossRef]
46. Xu, Y.; Zhang, M.; Zheng, B. Design of cold-formed stainless steel circular hollow section columns using machine learning
methods. Structures 2021, 33, 2755–2770. [CrossRef]
47. Nguyen, T.H.; Tran, N.L.; Nguyen, D.D. Prediction of axial compression capacity of cold-formed steel oval hollow section columns
using ANN and ANFIS models. Int. J. Steel Struct. 2022, 22, 1–26. [CrossRef]
48. Fang, Z.; Roy, K.; Padiyara, S.; Chen, B.; Raftery, G.M.; Lim, J.B.P. Web crippling design of cold-formed stainless steel channels
under interior-two-flange loading condition using deep belief network. Structures 2023, 47, 1967–1990. [CrossRef]
49. Lu, Y.; Wu, B.; Li, W.; Zhou, T.; Li, Y. Regression-classification ensemble machine learning model for loading capacity and buckling
mode prediction of cold-formed steel built-up I-section columns. Thin-Walled Struct. 2024, 205, 112427. [CrossRef]
50. Shaheen, M.A.; Presswood, R.; Afshan, S. Application of machine learning to predict the mechanical properties of high-strength
steel at elevated temperatures based on the chemical composition. Structures 2023, 52, 17–29. [CrossRef]
51. Shahin, R.I.; Ahmed, M.; Liang, Q.Q.; Yehia, S.A. Predicting the web crippling capacity of cold-formed steel lipped channels
using hybrid machine learning techniques. Eng. Struct. 2024, 309, 118061. [CrossRef]
52. Yılmaz, Y.; Demir, S.; Öztürk, F. Predicting the load-bearing capacity of lipped channel section cold-formed steel profiles under
combined effects using machine learning. Structures 2024, 66, 106898. [CrossRef]
53. Kim, B.; Yuvaraj, N.; Park, H.W.; Sri Preethaa, K.R.; Pandian, R.A.; Lee, D.-E. Investigation of steel frame damage based on
computer vision and deep learning. Autom. Constr. 2021, 132, 103941. [CrossRef]
54. Truong, V.-H.; Pham, H.-A.; Van, T.H.; Tangaramvong, S. Evaluation of machine learning models for load-carrying capacity
assessment of semi-rigid steel structures. Eng. Struct. 2022, 273, 115001. [CrossRef]
55. Jahjouh, M. An experience-based artificial neural network in the design optimization of steel frames. Eng. Res. Express 2022, 4,
045031. [CrossRef]
56. Shan, W.; Liu, J.; Zhou, J. Integrated method for intelligent structural design of steel frames based on optimization and machine
learning algorithm. Eng. Struct. 2023, 284, 115980. [CrossRef]
57. Pal, J.; Sikdar, S.; Banerjee, S. A deep-learning approach for health monitoring of a steel frame structure with bolted connections.
Struct. Control Health Monit. 2022, 29, e2873. [CrossRef]
Metals 2025, 15, 408 37 of 41

58. Naresh, M.; Kumar, V.; Pal, J. A machine learning approach for health monitoring of a steel frame structure using statistical
features of vibration data. Asian J. Civ. Eng. 2024, 25, 39–49. [CrossRef]
59. Vu, V.T.; Thom, D.V.; Tran, T.D. Identification of damage in steel beam by natural frequency using machine learning algorithms.
Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 2024, 238, 9644–9659. [CrossRef]
60. Li, H.; Yin, X.; Sha, L.; Yang, D.; Hu, T. Data-driven prediction model for high-strength bolts in composite beams. Buildings 2023,
13, 2769. [CrossRef]
61. Dissanayake, M.; Nguyen, H.; Poologanathan, K.; Perampalam, G.; Upasiri, I.; Rajanayagam, H.; Suntharalingam, T. Prediction of
shear capacity of steel channel sections using machine learning algorithms. Thin-Walled Struct. 2022, 175, 109152. [CrossRef]
62. Dai, Y.; Roy, K.; Fang, Z.; Chen, B.; Raftery, G.M.; Lim, J.B.P. A novel machine learning model to predict the moment capacity of
cold-formed steel channel beams with edge-stiffened and un-stiffened web holes. J. Build. Eng. 2022, 53, 104592. [CrossRef]
63. Liu, J.-Z.; Li, S.; Guo, J.; Xue, S.; Chen, S.; Wang, L.; Zhou, Y.; Luo, T.X. Machine learning (ML)-based models for predicting the
ultimate bending moment resistance of high-strength steel welded I-section beam under bending. Thin-Walled Struct. 2023, 191,
111051. [CrossRef]
64. Su, A.; Cheng, J.; Li, X.; Zhong, Y.; Li, S.; Zhao, O.; Jiang, K. Unified machine-learning-based design method for normal and
high-strength steel I-section beam–columns. Thin-Walled Struct. 2024, 199, 111835. [CrossRef]
65. Tusnin, A.R.; Alekseytsev, A.V.; Tusnina, O.A. Load identification in steel structural systems using machine learning elements:
Uniform length loads and point forces. Buildings 2024, 14, 1711. [CrossRef]
66. Sediek, O.A.; Wu, T.-Y.; McCormick, J.; El-Tawil, S. Prediction of seismic collapse behavior of deep steel columns using machine
learning. Structures 2022, 40, 163–175. [CrossRef]
67. Imam, M.H.; Mohiuddin, M.; Shuman, N.M.; Oyshi, T.I.; Debnath, B.; Liham, M.I.M.H. Prediction of seismic performance of steel
frame structures: A machine learning approach. Structures 2024, 69, 107547. [CrossRef]
68. Shin, D.-H.; Kim, H.-J. Machine learning-based prediction of hysteretic behaviors of two-side clamped steel shear walls. Structures
2024, 60, 105875. [CrossRef]
69. Cho, E.; Han, S.W. A numerical model simulating cyclic behavior of high-strength steel. Adv. Struct. Eng. 2024, 27, 1490–1508.
[CrossRef]
70. Hu, S.; Zhu, S.; Alam, M.S.; Wang, W. Machine learning-aided peak and residual displacement-based design method for
enhancing seismic performance of steel moment-resisting frames by installing self-centering braces. Eng. Struct. 2022, 271, 114935.
[CrossRef]
71. Samadian, D.; Muhit, I.B.; Occhipinti, A.; Dawood, N. Surrogate models for seismic and pushover response prediction of steel
special moment resisting frames. Eng. Struct. 2024, 314, 118307. [CrossRef]
72. Salama, A.H.E.S. Optimization seismic resilience: A machine learning approach for vertical irregular buildings. Asian J. Civ. Eng.
2024, 25, 6233–6248. [CrossRef]
73. Huang, M.; Zhang, J.; Li, J.; Deng, Z.; Luo, J. Damage identification of steel bridge based on data augmentation and adaptive
optimization neural network. Struct. Health Monit. 2024; in press. [CrossRef]
74. Iacussi, L.; Chiariotti, P.; Cigada, A. AI-enhanced IoT system for assessing bridge deflection in drive-by conditions. Sensors 2025,
25, 158. [CrossRef]
75. Wang, X.; Yue, Q.; Liu, X. SBDNet: A deep learning-based method for the segmentation and quantification of fatigue cracks in
steel bridges. Adv. Eng. Inform. 2025, 65, 103186. [CrossRef]
76. Svendsen, B.T.; Øiseth, O.; Frøseth, G.T.; Rønnquist, A. A hybrid structural health monitoring approach for damage detection in
steel bridges under simulated environmental conditions using numerical and experimental data. Struct. Health Monit. 2022, 22,
540–561. [CrossRef]
77. Zhou, K.; Duan, M.-G.; Wu, Z.-L.; Zhi, L.-H.; Hu, F. Dynamic behavior monitoring of twin supertall buildings during Super
Typhoon Soksuri using social sensing data. J. Build. Eng. 2024, 95, 110119. [CrossRef]
78. Ghaffari, A.; Shahbazi, Y.; Mokhtari Kashavar, M.; Fotouhi, M.; Pedrammehr, S. Advanced predictive structural health monitoring
in high-rise buildings using recurrent neural networks. Buildings 2024, 14, 3261. [CrossRef]
79. Wang, M.; Incecik, A.; Tian, Z.; Zhang, M.; Kujala, P.; Gupta, M.; Krolczyk, G.; Li, Z. Structural health monitoring on offshore
jacket platforms using a novel ensemble deep learning model. Ocean Eng. 2024, 301, 117510. [CrossRef]
80. Martzikos, N.; Ruzzo, C.; Malara, G.; Fiamma, V.; Arena, F. Applying neural networks to predict offshore platform dynamics. J.
Mar. Sci. Eng. 2024, 12, 2001. [CrossRef]
81. Kouchaki, M.; Salkhordeh, M.; Mashayekhi, M.; Mirtaheri, M.; Amanollah, H. Damage detection in power transmission towers
using machine learning algorithms. Structures 2023, 56, 104980. [CrossRef]
82. Kiyoki, S.; Yoshida, S.; Rushdi, M.A. Machine learning-based prediction of 2 MW wind turbine tower loads during power
production based on nacelle behavior. Energies 2025, 18, 216. [CrossRef]
83. Vlasenko, T.; Hutsol, T.; Vlasovets, V.; Glowacki, S.; Nurek, T.; Horetska, I.; Kukharets, S.; Firman, Y.; Bilovod, O. Ensemble
learning based sustainable approach to rebuilding metal structures prediction. Sci. Rep. 2025, 15, 1210. [CrossRef]
Metals 2025, 15, 408 38 of 41

84. Shang, Z.; Qin, X.; Zhang, Z.; Jiang, H. Bolt loosening and preload loss detection technology based on machine vision. Buildings
2024, 14, 3897. [CrossRef]
85. Huang, X.; Duan, Z.; Hao, S.; Hou, J.; Chen, W.; Cai, L. A deep learning framework for corrosion assessment of steel structures
using Inception v3 model. Buildings 2025, 15, 512. [CrossRef]
86. Badini, S.; Regondi, S.; Pugliese, R. Enhancing mechanical and bioinspired materials through generative AI approaches. Next
Mater. 2025, 6, 100275. [CrossRef]
87. Arridge, S.; Maass, P.; Öktem, O.; Schönlieb, C.-B. Solving inverse problems using data-driven models. Acta Numer. 2019, 28,
1–174. [CrossRef]
88. Masurkar, F.; Aggarwal, S.; Tham, Z.W.; Zhang, L.; Yang, F.; Cui, F. Estimating the elastic constants of orthotropic composites
using guided waves and an inverse problem of property estimation. Appl. Acoust. 2024, 216, 109750. [CrossRef]
89. Challapalli, A.; Patel, D.; Li, G. Inverse machine learning framework for optimizing lightweight metamaterials. Mater. Des. 2021,
208, 109937. [CrossRef]
90. Liao, W.; Lu, X.; Huang, Y.; Zheng, Z.; Lin, Y. Automated structural design of shear wall residential buildings using generative
adversarial networks. Autom. Constr. 2021, 132, 103931. [CrossRef]
91. Danhaive, R.; Mueller, C.T. Design subspace learning: Structural design space exploration using performance-conditioned
generative modeling. Autom. Constr. 2021, 127, 103664. [CrossRef]
92. Teimouri, A.; Challapalli, A.; Konlan, J.; Li, G. Machine learning assisted design and optimization of plate-lattice structures with
superior specific recovery force. Giant 2024, 18, 100282. [CrossRef]
93. Glaesener, R.N.; Kumar, S.; Lestringant, C.; Butruille, T.; Portela, C.M.; Kochmann, D.M. Predicting the influence of geometric
imperfections on the mechanical response of 2D and 3D periodic trusses. Acta Mater. 2023, 254, 118918. [CrossRef]
94. McClarren, R.G.; Tregillis, I.L.; Urbatsch, T.J.; Dodd, E.S. High-energy density hohlraum design using forward and inverse deep
neural networks. Phys. Lett. A 2021, 396, 127243. [CrossRef]
95. Lee, J.-Y.; Kim, S.-H.; Jeong, H.-B.; Lee, K.; Cho, K.; Lee, Y.-K. Inverse design of high-strength medium-Mn steel using a machine
learning-aided genetic algorithm approach. J. Mater. Res. Technol. 2024, 33, 2672–2682. [CrossRef]
96. Wang, Z.-L.; Adachi, Y. Property prediction and properties-to-microstructure inverse analysis of steels by a machine-learning
approach. Mater. Sci. Eng. A 2019, 744, 661–670. [CrossRef]
97. Pei, Z.; Rozman, K.A.; Doğan, Ö.N.; Wen, Y.; Gao, N.; Holm, E.A.; Hawk, J.A.; Alman, D.E.; Gao, M.C. Machine-learning
microstructure for inverse material design. Adv. Sci. 2021, 8, 2101207. [CrossRef]
98. Lertkiatpeeti, K.; Janya-Anurak, C.; Uthaisangsuk, V. Effects of spatial microstructure characteristics on mechanical properties of
dual-phase steel by inverse analysis and machine learning approach. Comput. Mater. Sci. 2024, 245, 113311. [CrossRef]
99. Adachi, Y.; Chen, T.-T.; Sun, F. A review on inverse analysis models in steel material design. MGE Adv. 2024, 2, e71. [CrossRef]
100. He, L.; Wang, Z.; Akebono, H.; Sugeta, A. Machine learning-based predictions of fatigue life and fatigue limit for steels. J. Mater.
Sci. Technol. 2021, 90, 9–19. [CrossRef]
101. Kolesnikov, I.; Pashkov, D.M.; Belyak, O.A.; Guda, A.A.; Danilchenko, S.A.; Manturov, D.S.; Novikov, E.S.; Kudryakov, O.V.;
Guda, S.A.; Soldatov, A.V.; et al. Design of double-layer protective coatings: Finite element modeling and machine learning
approximations. Acta Astronaut. 2023, 204, 869–877. [CrossRef]
102. Islam, F.; Wanigasekara, C.; Rajan, G.; Swain, A.; Prusty, B.G. An approach for process optimization of the Automated Fibre
Placement (AFP) based thermoplastic composites manufacturing using machine learning, photonic sensing and thermo-mechanics
modelling. Manuf. Lett. 2022, 32, 10–14. [CrossRef]
103. Shen, X.; Yan, K.; Zhu, D.; Hu, Q.; Wu, H.; Qi, S.; Yuan, M.; Qian, X. Inverse machine learning framework for optimizing gradient
honeycomb structure under impact loading. Eng. Struct. 2024, 309, 118079. [CrossRef]
104. Challapalli, A.; Konlan, J.; Li, G. Inverse machine learning discovered metamaterials with record high recovery stress. Int. J. Mech.
Sci. 2023, 244, 108029. [CrossRef]
105. Kusampudi, N.; Diehl, M. Inverse design of dual-phase steel microstructures using generative machine learning model and
Bayesian optimization. Int. J. Plast. 2023, 171, 103776. [CrossRef]
106. Confalonieri, R.; Coba, L.; Wagner, B.; Besold, T.R. A historical perspective of explainable artificial intelligence. WIREs Data Min.
Knowl. Discov. 2021, 11, e1391. [CrossRef]
107. Taffese, W.Z.; Zhu, Y.; Chen, G. Explainable AI based slip prediction of steel-UHPC interface connected by shear studs. Expert
Syst. Appl. 2025, 259, 125293. [CrossRef]
108. Wang, Z.; Liu, T.; Long, Z.; Wang, J.; Zhang, J. Predicting the drift capacity of precast concrete columns using an explainable
machine learning approach. Eng. Struct. 2023, 282, 115771. [CrossRef]
109. Lai, D.; Demartino, C.; Xiao, Y. Interpretable machine-learning models for maximum displacements of RC beams under impact
loading predictions. Eng. Struct. 2023, 281, 115723. [CrossRef]
110. Karathanasopoulos, N.; Singh, A.; Hadjidoukas, P. Machine learning-based modelling, feature importance and Shapley additive
explanations analysis of variable-stiffness composite beam structures. Structures 2024, 62, 106206. [CrossRef]
Metals 2025, 15, 408 39 of 41

111. Wang, S.; Liu, J.; Wang, Q.; Dai, R.; Chen, K. Prediction of non-uniform shrinkage of steel-concrete composite slabs based on
explainable ensemble machine learning model. J. Build. Eng. 2024, 88, 109002. [CrossRef]
112. Le, H.-A.; Le, D.-A.; Le, T.-T.; Le, H.-P.; Le, T.-H.; Hoang, H.-G.T.; Nguyen, T.-A. An extreme gradient boosting approach to
estimate the shear strength of FRP reinforced concrete beams. Structures 2022, 45, 1307–1321. [CrossRef]
113. Zhang, S.; Lei, H.; Zhou, Z.; Wang, G.; Qiu, B. Fatigue life analysis of high-strength bolts based on machine learning method and
SHapley Additive exPlanations (SHAP) approach. Structures 2023, 51, 275–287. [CrossRef]
114. Junda, E.; Málaga-Chuquitaype, C.; Chawgien, K. Interpretable machine learning models for the estimation of seismic drifts in
CLT buildings. J. Build. Eng. 2023, 70, 106365. [CrossRef]
115. Parvizi, M.; Nasserasadi, K.; Tafakori, E. Development of fragility functions of low-rise steel moment frame by artificial neural
networks and identifying effective parameters using SHAP theory. Structures 2023, 58, 105315. [CrossRef]
116. Shahnazaryan, D.; O’Reilly, G.J. Next-generation non-linear and collapse prediction models for short- to long-period systems via
machine learning methods. Eng. Struct. 2024, 306, 117801. [CrossRef]
117. Liu, T.; Cakiroglu, C.; Islam, K.; Wang, Z.; Nehdi, M.L. Explainable machine learning model for predicting punching shear
strength of FRC flat slabs. Eng. Struct. 2024, 301, 117276. [CrossRef]
118. Ribeiro, M.T.; Singh, S.; Guestrin, C. Why should I trust you? Explaining the predictions of any classifier. arXiv 2016,
arXiv:1602.04938.
119. Kim, B.; Glassman, E.; Johnson, B.; Shah, J. iBCM: Interactive Bayesian Case Model Empowering Humans via Intuitive Interaction.
MIT-CSAIL Tech. Rep. 2015, 30, 03z.
120. Rani, P.; Liu, C.; Sarkar, N.; Vanman, E. An empirical study of machine learning techniques for affect recognition in human–robot
interaction. Pattern Anal. Appl. 2006, 9, 58–69. [CrossRef]
121. Huysmans, J.; Dejaeger, K.; Mues, C.; Vanthienen, J.; Baesens, B. An empirical evaluation of the comprehensibility of decision
table, tree and rule-based predictive models. Decis. Support Syst. 2011, 51, 141–154. [CrossRef]
122. Szegedy, C.; Zaremba, W.; Sutskever, I.; Bruna, J.; Erhan, D.; Goodfellow, I.; Fergus, R. Intriguing properties of neural networks.
arXiv 2014, arXiv:1312.6199.
123. James, G.; Witten, D.; Hastie, T.; Tibshirani, R. An Introduction to Statistical Learning; Springer: New York, NY, USA, 2013.
[CrossRef]
124. Yu, B.; Chen, X.; Gupta, A.; Ribeiro, C. Stability. Bernoulli 2013, 19, 1484–1500. [CrossRef]
125. Samek, W.; Müller, K.R. Towards explainable artificial intelligence. In Explainable AI: Interpreting, Explaining and Visualizing Deep
Learning; Samek, W., Montavon, G., Vedaldi, A., Hansen, L., Müller, K.R., Eds.; Springer: Cham, Switzerland, 2019; Volume 11700,
pp. 5–22. [CrossRef]
126. Došilović, F.K.; Brčić, M.; Hlupić, N. Explainable artificial intelligence: A survey. In Proceedings of the 41st International
Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), Opatija, Croatia, 21–25
May 2018; pp. 210–215. [CrossRef]
127. Ahmed, G.; Jeon, G.; Piccialli, F. From artificial intelligence to explainable artificial intelligence in Industry 4.0: A survey on what,
how, and where. IEEE Trans. Ind. Inform. 2022, 18, 5031–5042. [CrossRef]
128. Mojtabaei, S.M.; Becque, J.; Hajirasouliha, I.; Khandan, R. Predicting the buckling behaviour of thin-walled structural elements
using machine learning methods. Thin-Walled Struct. 2023, 184, 110518. [CrossRef]
129. Hu, S.; Wang, W.; Lu, Y. Explainable machine learning models for probabilistic buckling stress prediction of steel shear panel
dampers. Eng. Struct. 2023, 288, 116235. [CrossRef]
130. Hou, Z.; Hu, S.; Wang, W. Interpretable machine learning models for predicting probabilistic axial buckling strength of steel
circular hollow section members considering discreteness of geometries and material. Adv. Struct. Eng. 2024, 28, 828–844.
[CrossRef]
131. Samadian, D.; Muhit, I.B.; Occhipinti, A.; Dawood, N. Meta databases of steel frame buildings for surrogate modelling and
machine learning-based feature importance analysis. Resilient Cities Struct. 2024, 3, 20–43. [CrossRef]
132. Liu, K.; Yu, M.; Liu, Y.; Chen, W.; Fang, Z.; Lim, J.B.P. Fire resistance time prediction and optimization of cold-formed steel walls
based on machine learning. Thin-Walled Struct. 2024, 203, 112207. [CrossRef]
133. Tang, P.; Dai, Y.; Lu, C.; Hu, S. A machine learning framework for predicting the axial capacity of cold-formed steel face-to-face
built-up channel sections at elevated temperatures. Structures 2024, 68, 107144. [CrossRef]
134. Degtyarev, V.V.; Hicks, S.J.; Ferreira, F.P.V.; Tsavdaridis, K.D. Probabilistic resistance predictions of laterally restrained cellular
steel beams by natural gradient boosting. Thin-Walled Struct. 2024, 205, 112367. [CrossRef]
135. Widanage, C.; Mohotti, D.; Lee, C.K.; Wijesooriya, K.; Meddage, D.P.P. Use of explainable machine learning models in blast load
prediction. Eng. Struct. 2024, 312, 118271. [CrossRef]
136. Anand, T.P.; Pandikkadavath, M.S.; Mangalathu, S.; Sahoo, D.R. Machine learning models for seismic analysis of buckling-
restrained braced frames. J. Build. Eng. 2024, 98, 111398. [CrossRef]
Metals 2025, 15, 408 40 of 41

137. Fan, X.; Yang, L.; Zhao, X.; Yan, G.; Yang, Y.; Zhang, H.; Chen, S. Prediction of axial compressive capacity and interpretability
analysis of web perforated Σ-shaped cold-formed steel. Structures 2024, 70, 107880. [CrossRef]
138. Sarfarazi, S.; Shamass, R.; Guarracino, F.; Mascolo, I.; Modano, M. Advanced predictive modeling of shear strength in stainless-
steel column web panels using explainable AI insights. Results Eng. 2024, 24, 103454. [CrossRef]
139. Aloko, M.N.; De Risi, R.; De Luca, F. Capacity prediction and failure mode classification of cold-formed steel built-up columns
using machine learning methods. Thin-Walled Struct. 2025, 210, 112873. [CrossRef]
140. Gharagoz, M.M.; Noureldin, M.; Kim, J. Explainable machine learning (XML) framework for seismic assessment of structures
using Extreme Gradient Boosting (XGBoost). Eng. Struct. 2025, 327, 119621. [CrossRef]
141. Su, A.; Cheng, J.; Wang, Y.; Pan, Y. Machine learning-based processes with active learning strategies for the automatic rapid
assessment of seismic resistance of steel frames. Structures 2025, 72, 108227. [CrossRef]
142. Gatheeshgar, P.; Ranasinghe, R.S.S.; Simwanda, L.; Meddage, D.P.P.; Mohotti, D. Machine learning prediction of web-crippling
strength in cold-formed steel beams with staggered slotted perforations. Structures 2025, 71, 108079. [CrossRef]
143. Sarfarazi, S.; Shamass, R.; Guarracino, F.; Modano, M. Exploring the stainless-steel beam-to-column connections response: A
hybrid explainable machine learning framework for characterization. Front. Struct. Civ. Eng. 2025, 19, 34–59. [CrossRef]
144. Habib, A.; Houri, A.A.; Junaid, M.T.; Barakat, S. A systematic and bibliometric review on physics-based neural networks
applications as a solution for structural engineering partial differential equations. Structures 2024, 69, 107361. [CrossRef]
145. Raissi, M.; Perdikaris, P.; Karniadakis, G.E. Physics-informed neural networks: A deep learning framework for solving forward
and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 2019, 378, 686–707. [CrossRef]
146. Haghighat, E.; Raissi, M.; Moure, A.; Gomez, H.; Juanes, R. A physics-informed deep learning framework for inversion and
surrogate modeling in solid mechanics. Comput. Methods Appl. Mech. Eng. 2021, 379, 113741. [CrossRef]
147. Lu, L.; Meng, X.; Mao, Z.; Karniadakis, G.E. DeepXDE: A deep learning library for solving differential equations. SIAM Rev. 2021,
63, 208–228. [CrossRef]
148. Pak, M.; Kim, S. A review of deep learning in image recognition. In Proceedings of the 4th International Conference on Computer
Applications and Information Processing Technology (CAIPT), Kuta Bali, Indonesia, 8–10 August 2017; pp. 1–3. [CrossRef]
149. Olsson, F. A Literature Survey of Active Machine Learning in the Context of Natural Language Processing. Swed. Inst. Comput.
Sci. Tech. Rep. 2009, 1, 59.
150. Weiss, K.; Khoshgoftaar, T.M.; Wang, D. A survey of transfer learning. J. Big Data 2016, 3, 9. [CrossRef]
151. Hospedales, T.; Antoniou, A.; Micaelli, P.; Storkey, A. Meta-learning in neural networks: A survey. IEEE Trans. Pattern Anal. Mach.
Intell. 2021, 44, 5149–5169.
152. Farahani, A.; Voghoei, S.; Rasheed, K.; Arabnia, H.R. A brief review of domain adaptation. In Advances in Data Science and
Information Engineering: Proceedings from ICDATA 2020 and IKE 2020; Springer: Cham, Switzerland, 2021; pp. 877–894.
153. Khodabakhshian, A.; Puolitaival, T.; Kestle, L. Deterministic and probabilistic risk management approaches in construction
projects: A systematic literature review and comparative analysis. Buildings 2023, 13, 1312. [CrossRef]
154. Harrison, R.L. Introduction to Monte Carlo simulation. AIP Conf. Proc. 2010, 1204, 17. [CrossRef]
155. Gal, Y.; Ghahramani, Z. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In Proceedings
of the 33rd International Conference on Machine Learning (ICML), New York, NY, USA, 20–22 June 2016; Volume 48, pp.
1050–1059. Available online: https://proceedings.mlr.press/v48/gal16.html (accessed on 23 March 2025).
156. Marrel, A.; Iooss, B. Probabilistic surrogate modeling by Gaussian process: A review on recent insights in estimation and
validation. Reliab. Eng. Syst. Saf. 2024, 247, 110094.
157. Sudret, B.; Marelli, S.; Wiart, J. Surrogate models for uncertainty quantification: An overview. In Proceedings of the 11th European
Conference on Antennas and Propagation (EUCAP), Paris, France, 19–24 March 2017; IEEE: Piscataway, NJ, USA, 2017; pp.
793–797.
158. García-Risueño, P.; Ibáñez, P.E. A review of high performance computing foundations for scientists. Int. J. Mod. Phys. C 2012, 23,
1230001.
159. Parekh, R.; Mitchell, O. Progress and obstacles in the use of artificial intelligence in civil engineering: An in-depth review. Int. J.
Sci. Res. Arch. 2024, 13, 1059–1080.
160. Lundberg Patel, D.; Raut, G.; Cheetirala, S.N.; Nadkarni, G.N.; Freeman, R.; Glicksberg, B.S.; Timsina, P.; Klang, E. Cloud
platforms for developing generative AI solutions: A scoping review of tools and services. arXiv 2024, arXiv:2412.06044.
161. American Society of Civil Engineers (ASCE). Policy Statement 573—Artificial Intelligence and Engineering Responsibility; American
Society of Civil Engineers: Reston, VA, USA, 2024. Available online: https://www.asce.org/advocacy/policy-statements/ps573--
-artificial-intelligence-and-engineering-responsibility (accessed on 25 March 2025).
162. American Society of Civil Engineers (ASCE). AI and Civil Engineering; American Society of Civil Engineers: Reston, VA, USA,
2024. Available online: https://www.asce.org/topics/ai-and-civil-engineering (accessed on 25 March 2025).
163. Regona, M.; Yigitcanlar, T.; Xia, B.; Li, R.Y.M. Opportunities and adoption challenges of AI in the construction industry: A
PRISMA review. J. Open Innov. Technol. Mark. Complex. 2022, 8, 45. [CrossRef]
Metals 2025, 15, 408 41 of 41

164. International Organization for Standardization/International Electrotechnical Commission (ISO/IEC) Joint Technical Committee
1, Subcommittee 42 (JTC 1/SC 42). Artificial Intelligence—Overview of Trustworthiness; ISO: Geneva, Switzerland, 2023. Available
online: https://www.iso.org/committee/6794475.html (accessed on 25 March 2025).
165. WorkOrb. Barriers to Adopting AI in AEC Firms. WorkOrb Blog. 2024. Available online: https://www.workorb.com/blog/
barriers-to-adopting-ai-in-aec-firms (accessed on 25 March 2025).
166. RICS. To AI or Not to AI: Five Trends in the Adoption of AI in Construction; Royal Institution of Chartered Surveyors: London,
UK, 2024. Available online: https://www.rics.org/news-insights/wbef/to-ai-or-not-to-ai-five-trends-in-the-adoption-of-ai-in-
construction (accessed on 25 March 2025).
167. American Society of Civil Engineers (ASCE). What Do Civil Engineers Need to Know About Artificial Intelligence? Civ. Eng. Mag.
2024, 94, pp. 46–53. Available online: https://www.asce.org/publications-and-news/civil-engineering-source/civil-engineering-
magazine/issues/magazine-issue/article/2024/11/what-do-civil-engineers-need-to-know-about-artificial-intelligence (ac-
cessed on 25 March 2025).
168. Habib, A.; Yildirim, U. Developing a physics-informed and physics-penalized neural network model for preliminary design of
multi-stage friction pendulum bearings. Eng. Appl. Artif. Intell. 2022, 113, 104953. [CrossRef]
169. Yao, H.; Gao, Y.; Liu, Y. FEA-Net: A physics-guided data-driven model for efficient mechanical response prediction. Comput.
Methods Appl. Mech. Eng. 2020, 363, 112892. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.