Journal of Constructional Steel Research 224 (2025) 109085

Contents lists available at ScienceDirect

Journal of Constructional Steel Research

journal homepage: www.elsevier.com/locate/jcsr

Data-driven design approaches for hollow section columns—Database

analysis and implementation
Hyeyoung Koh a,b , Hannah B. Blum b ,∗
a Department of Civil and Environmental Engineering, Washington State University, WA, United States
b
Department of Civil and Environmental Engineering, University of Wisconsin - Madison, WI, United States

ARTICLE INFO ABSTRACT

Keywords: Structural engineering has a plethora of existing data from previous experiments and computational modeling
Data-driven analysis results, yet the benefits of employing data methods in structural engineering are still largely unexplored. As a
Steel hollow section columns test case to demonstrate the use of data-driven design approaches in structural engineering, this study applies
Interpolation
both conventional interpolation and advanced machine learning techniques, Extreme Gradient Boosting and
High-strength steel
Multi-layer Perceptron (MLP), to estimate capacity strength of SHS and RHS columns using a comprehensive
Cold-formed and hot-rolled steel
Machine learning
database consisting of 695 experimental results and 3,794 finite element (FE) analysis results. The database
covers a wide range of material and geometric properties, including steel grades ranging from normal-strength
to high-strength steel, cross-sectional dimensions, member slenderness, and forming process (cold-formed or
hot-rolled). The impact of data source (experiment or FE models) and ratios of training to testing sets on the
model prediction accuracy are explored. The best model predictions are also compared to predictions from
established design standards including AISC 360 and Eurocode 3. It was found that the MLP model performed
the best among the data driven models and the MLP predictions across the range of member slenderness ratios,
and steel grades, and forming methods performed better than either established design standard, indicating
the potential benefits of using advanced data methods. To demonstrate the future potential of how data-driven
design methods can enhance structural engineering design, the developed models and database are available
in a public repository and a practical example of how to use the database is detailed.

1. Introduction time-efficient alternative to physical experiments [1]. Consequently,

the steel structural industry and steel design standards, such as
Big data is going to transform the civil engineering industry with the [2–5], have embraced finite element (FE) analysis for structural design.
help of its vast variety of tools and techniques. In the field of structural To effectively handle and process big data, the design of appropri-
engineering, big data can be leveraged for descriptive and predictive ate data structures becomes crucial. Data structures provide efficient
analytics. Descriptive analytics enables the extraction of meaningful storage and organization of data, enabling users to access and update
insights from raw data, interpreting the relationships between various information seamlessly. However, the industry has often relied on
features that influence complex structural scenarios and behaviors. On specialists to handle the overwhelming volume of field data for tasks
the other hand, predictive analytics employs data-driven models to such as data review and mapping, which can be time-consuming and
forecast the likelihood of future outcomes, enabling the prediction of
resource-intensive. Additionally, compared to other subdisciplines in
structural responses in conjunction with physics-based models. Data
civil engineering, structural engineering has yet to fully capitalize on
analytics facilitates the extraction and analysis of behavioral data and
the advancements in data analytics [6]. Incorporating programming
trends from large databases, offering valuable insights to inform design
language-based analytical methods and data structures into structural
and structural decision-making.
design practices can propel the field of data analytics within the
Recent advancements in computational capabilities have made com-
prehensive datasets readily accessible, encompassing a multitude of structural engineering community.
observations and features. Structural engineers can now harness this Data-driven design approaches, including interpolation and regres-
wealth of data acquired through physical experiments and robust sion, have broad applications for predictive analysis using available
computational models. Computational models offer a cost-effective and data. However, when dealing with large datasets while maintaining

∗ Corresponding author.
E-mail addresses: [email protected] (H. Koh), [email protected] (H.B. Blum).

https://doi.org/10.1016/j.jcsr.2024.109085
Received 8 January 2024; Received in revised form 8 October 2024; Accepted 13 October 2024
Available online 4 November 2024
0143-974X/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

tractable computational complexity, traditional interpolation and re- • Design rules such as AISC 360 [31] do not consider the forma-
gression techniques face challenges as they are primarily derived from tion methods of hollow sections, and the design of cold-formed
univariate predictions [7]. In recent years, machine learning (ML) tech- hollow sections follow the design rules for hot-rolled sections.
niques have demonstrated significant advantages in estimating complex AISC-360 produced unconservative predictions compared to ac-
relationships among features. This presents an opportunity for struc- tual results for cold-formed steel hollow section columns [40,43,
tural design, optimization, structural health monitoring, and structural 45–50]. Therefore, the current design provisions fail to accurately
reliability, and other related fields to leverage the vast amounts of avail- estimate the capacity of SHS and RHS columns, particularly for
able data [8–10]. ML has been successfully implemented in predicting cold-formed steel members. The discrepancies observed in these
the responses of steel structural components [11–22], showcasing its predictions emphasize the need for an accurate prediction model
ability to enhance regression accuracy in predicting the capacity of that considers the buckling capacity of RHS and SHS columns
steel components. Furthermore, first order reliability methods (FORM)
fabricated using different methods such as hot-rolling and cold-
have been used to calibrate ML models related to structural design,
forming. As a note, Eurocode 3 [3] uses different column capacity
such as for concrete filled steel tubes [23,24], concrete beams [25,26],
models for hot-rolled and cold-formed sections, therefore this
and buckling restrained braces [27]. This shows the potential for using
issue is not considered for Eurocode 3.
ML-enhanced structural design.
• The absence of a unified design equation leads to inconsistent pre-
In accordance with various steel design codes [2–5], steel design
dictions for structural members approaching failure mode bound-
and analysis can use both physical testing and numerical simulations.
Several studies [23,24,26–29] incorporated data from experiments and aries [51]. Hence, there is a need for a unified model that encom-
numerical analysis for ML training, but they did not provide cru- passes various geometric (e.g., member and element slenderness)
cial statistical details for each data source (experimental, numerical, and material properties (e.g., yield strength), as well as forming
and combined). This missing information includes key metrics, such processes, which may affect the governing failure modes.
as mean and coefficient of variation of parameters and correlations Recent research has leveraged ML to predict the capacity of hollow
between parameters, which impact ML training results.
section columns with various cross-sections and materials. Research on
A data structure for structural design and data-driven models uti-
carbon steel has focused on cold-formed elliptical hollow sections [52],
lizing a comprehensive database that incorporates both experimental
cold-formed oval hollow sections [53], and both cold-formed and hot-
and numerical data is created for the buckling capacity of square and
rolled SHS and RHS [54]. Similarly, studies have explored ML for
rectangular hollow section (SHS and RHS) columns. Hollow sections
stainless steel hollow section columns, such as cold-formed circular
are widely employed as structural elements in buildings and other
hollow sections [55,56] and cold-formed SHS and RHS [51]. Notably,
structures due to their favorable strength-to-weight ratio, resulting in
a large dataset from literature. The primary manufacturing processes these studies primarily evaluated their models against Eurocode 3
for SHS and RHS members include hot-rolling and cold-forming. Less design provisions, thus the limitations outlined in AISC 360 for hollow
common manufacturing techniques, such as welding four steel plates section columns were not addressed. Addressing a broader range of
into a box shape or welding two channel sections tip-to-tip [30], are structural design codes would significantly enhance the applicability
beyond the scope of this study. of data-driven design approaches in the literature. Additionally, this
Design provisions for SHS and RHS members under compression study incorporates interpolation techniques alongside ML algorithms.
are codified in structural steel design codes, such as the American This approach investigates how data-driven methods can be effectively
code (AISC 360 [31]) and the European code (Eurocode 3 [3]). These utilized in structural design and provides valuable insights into when
codes incorporate experimental and numerical results for different more complex ML techniques might be necessary.
cross-section types and sizes, accounting for geometric imperfections, This study focuses on the potential of data-driven methods in struc-
to establish the buckling curves specified in the design provisions. For tural design and its role in incorporating new materials and forming
instance, AISC 360 [31] employs a single buckling curve based on processes into structural steel design. The objectives of this study are
numerical results of columns with initial out-of-straightness, as inves- as follows:
tigated by Tide [32], Bjorhovde [33], and Galambos [34]. Eurocode
3 (EC 3) [3] provides five separate buckling curves derived from a • To design a comprehensive data structure that encompasses both
combination of experimental and numerical findings [35–39]. While experimental and computational data. As a test case, the buckling
these codified provisions have widespread adoption in the structural capacity of square and rectangular hollow section (SHS and RHS)
steel industry, there remain several challenges that can be addressed to columns is investigated. The developed database includes diverse
enhance the accuracy and efficiency of the design rules: steel grades, geometric properties, and both cold-forming and hot-
rolling forming processes. Therefore, prediction models based on
• Although high-strength steel grades have become available due this database can address the limitations of current design rules
to advancements in steel manufacturing technology, the design
for hollow section columns and provide a framework for updating
rules for these grades are typically simple extensions of those for
future design specifications.
normal-strength steels, failing to account directly for the influence
• To develop data-driven models using conventional interpolation
of material yield strength. Meng and Gardner [40] proposed an
techniques and advanced machine learning algorithms. Initially, a
SHS and RHS column design approach that incorporates the effect
multivariate linear interpolation model is developed to assess the
of material yield strength by modifying the imperfection factor as
need for employing more sophisticated data-driven techniques. A
a continuous function of yield strength. While EC 3 [41] provides
codifications for steel grades up to 700 MPa for hollow sections, comparison and discussion of required model complexity for large
AISC 360 [31] has limited provisions for hollow sections with datasets is included.
steel grades exceeding 485 MPa. The grades and limitations for • To provide practical applications and demonstrate the imple-
applicable steel materials for hollow structural sections can be mentation of data-driven methods in steel design. The provided
found in Chapter A3.1 of AISC 360 [31]. The predicted capacities guidance aims to encourage structural designers and fabricators
of high-strength steel columns [40,42–44] based on EC 3 [3] to embrace data analytics and leverage its benefits in struc-
or AISC 360 [31] were found to be overly conservative, poten- tural design. The data and models are coded in a user-friendly
tially leading to inefficient and costly designs. Addressing this computational notebook format and made available in a GitHub
issue requires the development of a predictive model capable of repository. These open-source tools enable future modifications
accommodating a wide range of steel grades. and updates as more data becomes available.

2
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

• To discuss the existing design rules for estimating the capacity

of SHS and RHS columns. Given that the database covers a wide
range of material and geometric properties, as well as formation
processes, a comparison between the predictions obtained from
the proposed machine learning models and the current design
rules including AISC 360 [31] and Eurocode 3 [3] can offer
valuable insights and recommendations for improving the design
provisions.
To facilitate the data-driven analysis, an extensive database is con-
structed, consisting of 695 experimental results and 3794 FE analy-
sis results of square and rectangular hollow section (SHS and RHS)
columns from the existing literature. The investigated columns cover
a wide range of geometric and material properties, including different
steel grades and forming processes. A detailed procedure is provided
for developing interpolation and machine learning regression models
tailored to the constructed database. Specifically, an 𝑁-dimensional
linear interpolation model, Extreme Gradient Boosting [57], and Multi-
layer Perceptron algorithms are employed. The performance of these
models is evaluated in terms of computation time and prediction ac-
curacy. To facilitate practical implementation, guidance is offered for
utilizing the proposed data-driven method, ensuring that practitioners
can readily apply the computational coding. Finally, the developed
machine learning models for SHS and RHS columns are assessed by
comparing their predictions with the capacity estimates obtained from
the existing design methods in AISC 360 [31] and Eurocode 3 [3]. The
results demonstrate the high accuracy and low variance of the predic-
tions obtained by the developed machine learning models, highlighting
their potential for consistently estimating the strength of hollow section Fig. 1. Graphical illustration of the machine learning algorithms (a) Extreme gradient
columns with various steel grades and forming processes. This paper ex- boosting (b) Multi-layer perceptron.
plores the potential of data-driven methodologies within the structural
design field, and it is crucial to acknowledge that this represents an
initial step in a longer development process for machine learning-based points, ensuring that the triangle edges do not intersect. As a result,
design. Key challenges remain in improving structural engineers’ con- the interpolator forms a patchwork of triangular faces across the grid’s
fidence with machine learning algorithms and ensuring equivalent or extent. It should be noted that triangulation with linear interpola-
better reliability levels as those currently provided by published Spec- tion produces optimal results when data points are evenly distributed
ifications, considering that required target levels of structural safety throughout the grid area [64].
are not consistently met with different load combinations [58]. By
encouraging data sharing through repositories, data-driven structural 2.1.2. Extreme gradient boosting
design with extensive real-world data could be facilitated. Extreme Gradient Boosting (XGBoost) is a decision tree-based
model, as shown in Fig. 1a. Decision tree algorithms [65] construct
2. Background regression models in the form of tree-like structures. Starting from
the root of the tree, which encompasses the entire training set, the
2.1. Data-driven approaches algorithm splits the root node into internal nodes based on conditional
observations. The final prediction is determined by a sequence of
Given that samples obtained from experiments or numerical analy- decision rules that dictate how to divide subsets from the root node
ses are often sparsely and unevenly distributed, this study employed to leaf nodes, which are not further divided. The XGBoost algorithm
linear multivariate interpolation to predict the buckling capacity of operates by iteratively adding and training new trees to fit the residual
errors of the previous iteration. In essence, it builds a model comprising
hollow section columns at unsampled locations. Furthermore, two ma-
multiple decision trees and employs a gradient boost training approach,
chine learning (ML) algorithms, which are Extreme Gradient Boosting
which enhances a weaker model by combining it with other weaker
(XGBoost) [57] and Multi-layer Perceptron (MLP), were implemented
trees, resulting in a robust prediction model. Thus, XGBoost delivers
to investigate whether complex models offer improved performance.
more accurate and efficient predictions compared to tree algorithms
These algorithms have demonstrated success in similar regression tasks
that lack boosting.
involving the prediction of strength in steel and reinforced concrete
structural members and systems [20,51,59–61]. The implementation 2.1.3. Multi-layer perceptron
of these algorithms was carried out using the open-source libraries A Multi-layer Perceptron (MLP) is composed of interconnected neu-
scipy [62] and scikit-learn [63] in Python 3. rons that transmit information to each other, similar to the human
brain and nervous system, enabling the implementation of function
2.1.1. Linear multivariate interpolation approximations. The network can be divided into three primary layers:
Linear interpolation is a straightforward mathematical method used an input layer, hidden layers, and an output layer, as shown in Fig. 1b.
to estimate values within a given range based on existing data points. The input layer serves as the initial layer of the network, incorporating
This method involves constructing linear polynomials to generate new input features that contribute to generating an output. The network ne-
data points. In this study, piecewise-linear interpolation was employed, cessitates at least one hidden layer, which performs computations and
which involves creating meshes through the triangulation of scattered operations on the preceding layer using a mathematical function that
data to determine interpolated values. The interpolation model gen- involves weights and values. These weights, assigned to the connections
erates triangles by connecting lines between irregularly spaced data between the layers, determine the significance of the neurons.

3
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 2. Cross section geometry for SHS and RHS.

Fig. 3. Buckling curves provided in the Eurocode 3 [3] and AISC 360 [31].

2.2. Design rules for RHS and SHS columns

2.2.2. Eurocode 3 Part 1-1
2.2.1. ANSI/AISC 360-22 Chapter 6 of Eurocode 3 (EC 3) [3] outlines column design pro-
Chapter E of the ANSI/AISC 360–22 specification [31] outlines the visions. Buckling resistance (𝑁𝑢,𝐸 𝐶 ) is calculated differently based on
design criteria for compression members. For SHS and RHS sections, cross-section classification into four types: Class 1 has sufficient ro-
the provisions for design capacity are dependent on the classification of tation capacity for plastic hinge formation, Class 2 develops plastic
the cross-section. Nonslender elements are characterized by a width-to- moment resistance with limited rotation, Class 3 has both plastic and
√ elastic distributions, and Class 4 tends to experience local buckling
thickness ratio (𝑏∕𝑡) less than or equal to 1.4 𝐸∕𝐹𝑦 , where 𝑏 represents
before reaching full yield strength. For compact sections (Classes 1–
the width of flat regions and 𝑡 is the thickness (refer to Fig. 2),
√ 3), buckling resistance uses gross area (𝐴𝑔 ) (Eq. (5)), while for slender
while slender elements have a 𝑏∕𝑡 ratio greater than 1.4 𝐸∕𝐹𝑦 . The
sections (Class 4), it uses effective area (𝐴𝑒 ) (Eq. (6)).
design capacity (𝑁𝑢,𝐴𝐼 𝑆 𝐶 ) of the section is determined using Eq. (1)
𝜒𝐸 𝐶 𝐹𝑦 𝐴𝑔
for nonslender cross-sections and Eq. (2) for cross-sections with slender 𝑁𝑢,𝐸 𝐶 = for Classes 1-3 cross-sections (5)
𝛾𝑀1
elements:
𝜒𝐸 𝐶 𝐹𝑦 𝐴𝑒
𝑁𝑢,𝐸 𝐶 = for Class 4 cross-sections (6)
𝑁𝑢,𝐴𝐼 𝑆 𝐶 = 𝜙𝑐 𝜒𝐴𝐼 𝑆 𝐶 𝐹𝑦 𝐴𝑔 for SHS and RHS with nonslender elements 𝛾𝑀1
(1) where 𝜒𝐸 𝐶 is the EC 3 buckling reduction factor. In most cases, the
partial factor 𝛾𝑀1 , which can be interpreted as a safety factor, is
considered as 1.0 according to the recommendations in Eurocode 3 [3].
𝑁𝑢,𝐴𝐼 𝑆 𝐶 = 𝜙𝑐 𝜒𝐴𝐼 𝑆 𝐶 𝐹𝑦 𝐴𝑒 for SHS and RHS with slender elements (2)
Fig. 3 presents the five buckling curves defined by EC 3 where the
where 𝜙𝑐 is the resistance factor for compression, which is defined buckling reduction factor is plotted against the relative slenderness 𝜆𝑐 .
as 0.9 in the AISC 360 specification [31]; 𝜒𝐴𝐼 𝑆 𝐶 is the AISC column When the relative slenderness is equal to or less than 0.2 (𝜆𝑐 ≤ 0.2),
only cross-sectional checks are required, and the effects of buckling
buckling reduction factor; 𝐹𝑦 is the yield strength of the material; 𝐴𝑔 is
can be neglected. These buckling curves have been calibrated based
the gross cross-sectional area; 𝐴𝑒 is the effective cross-sectional area.
on extensive experimental and numerical data for various cross sec-
The column buckling reduction factor, 𝜒𝐴𝐼 𝑆 𝐶 , can be computed
tions, drawing on the Ayrton–Perry formulation [66] and considering
using either Eq. (3) or Eq. (4), depending on the member slenderness
√ previous studies [35–39].
ratio 𝐿𝑐 ∕𝑟 or the relative slenderness ratio 𝜆𝑐 = (𝐿𝑐 ∕𝜋 𝑟) 𝐹𝑦 ∕𝐸: Each buckling curve has its own designated imperfection factor 𝛼
√
𝐹𝑦
𝐿 𝐸 that depends on several factors including the forming process, yield
𝜒𝐴𝐼 𝑆 𝐶 = 0.658 𝐹𝑒 for 𝑐 ≤ 4.71 (or 𝜆𝑐 ≤ 1.5) (3)
𝑟 𝐹𝑦 strength of the material, type of cross-section, cross-sectional height-
√ to-width ratio, and flange thickness. According to EC 3 [3], hot-rolled
𝐹𝑒 𝐿𝑐 𝐸 hollow sections with steel grades ranging from 235 MPa to 460 MPa
𝜒𝐴𝐼 𝑆 𝐶 = 0.877 for > 4.71 (or 𝜆𝑐 > 1.5) (4)
𝐹𝑦 𝑟 𝐹𝑦 adopt buckling curve 𝑎, with an 𝛼 value of 0.21. For high-strength hot-
rolled hollow sections with nominal yield strength between 460 MPa
where 𝐿𝑐 is the effective member length determined by 𝐿𝑐 = 𝐾 𝐿; 𝐾
and 700 MPa, buckling curve 𝑎𝑜 is adopted, with an 𝛼 value of 0.13.
is the effective length factor; 𝐿 is the member length; 𝐸 is Young’s Cold-formed hollow sections adopt buckling curve 𝑐, with an 𝛼 value of
modulus; 𝑟 is the radius of gyration of the cross section; 𝐹𝑒 is the elastic 0.49. It is worth noting that the height-to-width ratio is not considered
buckling stress, which can be calculated by (𝜋 2 𝐸)∕(𝐿𝑐 ∕𝑟)2 . in estimating the 𝛼 value for hollow sections.
Fig. 3 provides the AISC 360 column buckling curve, which is de-
veloped based on an initial member out-of-straightness of 𝐿𝑐 ∕1500 [32– 3. Database
34], and is applicable to all forming processes. However, for the design
of hollow sections, AISC 360 is restricted to steel grades with a yield Data-driven models undergo a training procedure to recognize pat-
strength up to 485 MPa (70 ksi). It should be noted that AISC 360 does terns in data. Therefore, it is crucial to use reliable training data to
allow for the general use of steel with yield strength up to 690 MPa develop accurate data-driven models. In this study, a database of steel
(100 ksi). square hollow sections (SHS) and rectangular hollow sections (RHS)

4
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Table 1 production, particularly in cold-formed steel tube sections [77,78]. To

Number of samples in the database for each key characteristic of hollow section
estimate the yield strength of cold-formed steel members, this study
columns.
utilizes a weighted average yield strength approach as given in Eq. (7),
Criteria Experiment FE Total
which takes into account the corner strength enhancements:
Hot-rolled 111 1993 2104
Forming
Cold-formed 584 1981 2565 𝐹𝑦 = 𝐶 𝐹𝑦𝑐 + (1 − 𝐶)𝐹𝑦𝑓 (7)
High-strength (𝐹𝑦 ≥ 460 MPa) 351 797 1148
Grade where 𝐹𝑦𝑐 is the yield strength of the corner region, 𝐹𝑦𝑓 is the yield
Normal-strength (𝐹𝑦 < 460 MPa) 344 3177 3521
strength of the flat region, and 𝐶 is the ratio of the total corner cross-
Member Nonslender (𝐿𝑐 ∕𝑟 ≤ 100) 590 3120 3710
slenderness Slender (𝐿𝑐 ∕𝑟 > 100) 105 854 959
sectional area to the total cross-sectional area of the full cross-section
√ for compression members.
Element Nonslender (𝑏∕𝑡 ≤ 1.4 𝐸∕𝐹𝑦 ) 333 2680 3013
√ Fig. 4 provides the distribution of the input features and output,
slenderness Slender (𝑏∕𝑡 > 1.4 𝐸∕𝐹𝑦 ) 362 1294 1656
𝑁𝑢 , grouped by the data category, either experimental or FE. The
histograms include statistical information such as the minimum and
maximum values, mean, and coefficient of variation (CV). Histograms
members under compression was established by gathering data from with a large CV enable data-driven models to fit a wide range of
various sources, including journal papers and technical reports. A total features, and hence a larger CV is desirable. Comparing the distribu-
of 695 experimental results [40,42,43,45–48,50,67–71] and 3974 FE tions for the experimental and FE data, it can be seen that overall
results [40] were assembled. Meng and Gardner [40] developed shell the FE data has a larger density than the experimental data due to
FE models using the Abaqus software [72] and validated their modeling its comparatively larger sample size. For some features, the FE data
method with the buckling test results of hot-rolled and cold-formed supplements sparse regions of the experimental data. For example,
steel SHS and RHS columns. They next used the validated modeling most of the experimental results primarily involve samples with a yield
method to expand the column buckling data pool. Geometrical and ma- strength below 600 MPa. However, the FE data augments the limited
terial nonlinear analyses with imperfections (GMNIA) were performed samples for high-strength steel columns with a yield strength exceeding
to capture the behavior of the columns. Residual stresses were not 900 MPa, thereby expanding the dataset for such materials. For outer
explicitly modeled in their FE models because, in hot-rolled hollow sec- corner radius 𝑟𝑜 , thickness 𝑡, and the member effective length 𝐿𝑐 , the
tions, residual stress magnitudes have been shown to be negligible com- FE data has larger CVs and adds samples to the database where the
pared with the yield strength [73,74]. Moreover, in cold-formed hollow experimental data is limited. Distributions for the input features 𝐻
sections, the effect of the dominant bending residual stresses is already and 𝐵 in the FE data are discrete because the FE simulations [40]
implicitly incorporated into the material stress–strain curves [75,76]. employed only two cross sections including 100 mm × 100 mm and
The FE models were validated against the experimental column test 100 mm × 50 mm. However, it is important to note that the FE
results [40,43] and then utilized in parametric studies, considering a data covers a wider range of thickness 𝑡 and length 𝐿𝑐 , compared
range of steel grades and geometries, to expand the data pool for SHS to the experimental data. This broader range enhances the spectrum
and RHS column buckling. A detailed description of the FE models of both cross-sectional (element) slenderness and member slenderness
can be found in Meng and Gardner [40]. The comprehensive dataset and, consequently, buckling behaviors. Young’s modulus 𝐸 of steel was
is accessible at the Engineers with Data repository (https://github. excluded from the input feature set due to its minimal variation, which
com/EngineerWithData/HollowSectionColumnBucklingTests). This al- is consistent with the literature [79,80]. For additional notes, member
lows future users to continuously refine the analysis by incorporating slenderness, 𝐿𝑐 ∕𝑟, ranges from 3.8 to 194, and element slenderness, 𝐻
new data as it becomes available. The goal of the Engineers with Data (or 𝐵)/𝑡, ranges from 4.8 to 82.5.
repository is to promote data sharing and encourage practitioners to The correlation relationships between the input features and the
contribute their data to facilitate research advancement. output feature (𝑁𝑢 ) were explored. The correlation coefficient 𝜌, as
Table 1 presents the data grouped according to key characteristics of defined in Eq. (8), was utilized to measure the linear correlation:
the hollow section columns, such as the forming process, steel grade, Cov(𝑋1 , 𝑋2 )
element slenderness, and member slenderness. The experimental, FE, 𝜌𝑋1 ,𝑋2 = (8)
𝜎𝑋1 𝜎𝑋2
and combined (total) data show different sample proportions when
grouped based on each criterion. For instance, the experimental dataset where Cov is the covariance, 𝜎𝑋1 and 𝜎𝑋2 are the standard deviations
has a smaller number of samples for hot-rolled columns compared of input features 𝑋1 and 𝑋2 , respectively. The coefficient 𝜌 has a range
to cold-formed columns, while the FE dataset shows similar numbers from −1, indicating a perfect negative correlation, to 1, indicating a
of samples for hot-rolled and cold-formed steel columns. Regarding perfect positive correlation. When a pair of features has a coefficient
member slenderness, the experimental dataset has limited samples for close to 1 or −1, it indicates a strong relationship between the features.
slender columns, whereas the FE dataset complements the data pool by A coefficient of 0 represents that the two features are uncorrelated.
including more slender column samples. Fig. 5 illustrates the correlation matrix of the features. The correla-
Seven input features known to govern the buckling behavior of SHS tion matrix for the experimental data (Fig. 5(a) shows a high correlation
and RHS members under axial compression were considered in this between the response feature 𝑁𝑢 and the cross-section thickness, char-
study. These features include five geometric parameters, one material acterized by a correlation coefficient value of 𝜌 = 0.67. The remaining
property, and the forming process. The forming process was considered input parameters have correlation coefficients below 0.5 in relation to
as an input feature to ensure the development of a consistent prediction the output feature. In the FE data correlation matrix (Fig. 5(b), the
model that is applicable across different manufacturing techniques. buckling capacity 𝑁𝑢 demonstrates a strong relationship with 𝑟𝑜 and 𝑡,
The forming method can significantly affect residual stresses. However, both with a correlation coefficient value of 𝜌 = 0.61. It is worth noting
due to limitations in available experimental and FE data that directly that these two input features are perfectly correlated with each other.
measure or incorporate residual stresses, they were not included as This could be due to the FE data being limited to only two specific cross
input parameters in this study. The geometric parameters consist of sections: 50 mm × 100 mm and 100 mm × 100 mm. The correlation
four cross-section dimensional properties: the overall width (𝐵), overall matrix of the combined data (Fig. 5(c) shows correlation coefficients
height (𝐻), outer corner radius (𝑟𝑜 ), and thickness (𝑡) (Fig. 2), as well that more closely resemble those of the FE data. This likely reflects the
as the effective length (𝐿𝑐 ) of the column. larger size of the FE dataset that comprises 85% of the database. While
In the case of cold-formed steel, the yield strength incorporates this influence might seem intuitive, including the FE data is crucial as it
corner strength enhancements that arise due to cold work during covers a wider range of parameters. Specifically, the strong correlation

5
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 4. Statistical distributions of the input features and output 𝑁𝑢 .

observed between 𝐵 and 𝑁𝑢 in the experimental data is diminished in the FE data comprises only two distinct values for 𝐵, resulting in a small
the combined data due to the lower correlation present in the FE data. CV and high density, whereas the experimental data has a wider range
Furthermore, there is a contrasting correlation relationship between 𝐻 for this parameter. The substantial variation in the experimental data
and 𝐵, where the experimental data shows a negative correlation while contributes to the positive correlation observed between 𝐵 and 𝐻 in
the FE data demonstrates a positive correlation. Fig. 4 illustrates that the combined data.

6
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 5. Correlation matrix of the input features (a) experimental data (b) FE data (c) combined data.

4. Model development first, second, and third layers respectively, was chosen. For weight
optimization, the Limited-memory Broyden–Fletcher–Goldfarb–Shanno
4.1. Model parameters (LBFGS) algorithm was selected due to its memory efficiency when
dealing with large datasets, compared to standard BFGS, stochastic
A multivariate interpolation model was developed in the form of gradient descent (sgd), and Adam optimizers. The regularization pa-
𝑁𝑢 = 𝑓 (𝐻 , 𝐵 , 𝑟𝑜 , 𝑡, 𝐿𝑐 , 𝐹𝑦 , 𝐹 𝑜𝑟𝑚𝑖𝑛𝑔), where the seven input features rameter (alpha) was set to 0.00005 after evaluating values between
were used to approximate the buckling capacity 𝑁𝑢 . The interpolant 0.0005 and 0.00005. A constant learning rate was chosen over adaptive
was constructed by triangulating the input data with Qhull [81], a or inverse scaling learning rate strategies. The maximum number of
command-line tool used for computing the convex hull. The convex iterations was set to 1000.
hull represents the smallest convex shape that contains a set of points in
space. To create meshes for interpolation, Delaunay triangulation [82], 4.2. Computation time and accuracy
a method of dividing space into triangles based on a set of points where
no point is inside the circumcircle of any triangle, was employed. In order to optimize the train–test split, the three data-driven mod-
The parameters for XGBoost and MLP were determined through els were trained on nine data set sizes, ranging from 10% to 90% of the
hyperparameter tuning using a grid search algorithm with 10-fold total data, and tested on the remaining data. The data was randomly
cross-validation. The grid search algorithm explores different param- split into training and test sets, and it was ensured that the same sets
eter combinations from a given set and identifies the optimal combi- were used for both interpolation and machine learning algorithms to
nation based on a scoring metric. The coefficient of determination was maintain consistency.
considered the metric in this study. This technique helps to find the After developing the three data-driven models, the performance of
best parameter values and reduces errors caused by random sampling each model was evaluated using the test set. The predictive perfor-
of the training set. In cross-validation, the training set is divided into mance was assessed using evaluation metrics, including the coefficient
𝑘 subsets (folds), with 𝑘 − 1 subsets used for model fitting and the of determination (𝑅2 ) and root mean square error (RMSE), which can
remaining subset for validation. The process is repeated, and the model be calculated using Eqs (9) and (11), respectively:
performance is averaged over the 𝑘 folds. In this study, the training set Fig. 6a presents a comparison of the computation times required
was divided into ten folds (𝑘 = 10). for developing the data-driven models. In addition to the model de-
To achieve optimal performance from the XGBoost model, various velopment time, the interpolation model also includes the time for
hyperparameter configurations were investigated. The sub-sample ratio the tessellation of triangles in the data. When examining the model
was explored with values of 0.4, 0.5, and 0.6, and a subsample ratio development time for the three models, it is evident that the XGBoost
of 0.5 was chosen. The maximum depth of a tree was evaluated from
model outperforms the other models as the size of the training set
depths of 2 to 8. A maximum depth of 8 was selected. The minimum
increases. While interpolation has a solid theoretical foundation in one
sum of instance weight was investigated with values of 1, 3, 5, and 7,
dimension [83], implementing interpolation becomes challenging when
and 5 was chosen. The learning rate was explored with values of 0.15,
dealing with large data sizes while maintaining tractable computational
0.2, and 0.3, and 0.3 was selected. Finally, the regularization parameter
complexity [7]. Thus, the time required for the triangulation process
(gamma) was evaluated with values of 0.0, 0.05, and 0.1. A gamma
significantly increases as the size of the training set grows. Although
value of 0.1 was chosen. To mitigate overfitting, early stopping was
the MLP model demands a longer computational time than the inter-
employed during the XGBoost model training. Early stopping monitors
polation model during model development, the interpolation model
the model’s performance on a test set and stops the training if no
ultimately needs a higher total computational cost. This is due to the
improvement is observed after a fixed number of iterations.
need for triangulation, which arranges the data for model training,
The input features for the development of the MLP model were
prior to the development of the interpolation model. For instance, when
standardized using the Z-score method. This standardization involved
90% of the data was used for model development, the interpolation
subtracting the mean and dividing by the standard deviation for each
model required 16 times the computation time of the MLP model and
value of each feature. Standardization was performed because neural
6430 times more than that of the XGBoost model.
network models are sensitive to the magnitude of feature values, as they
Fig. 6b presents the comparison of training and testing accuracy,
are developed based on the distances between features. The parameters
assessed using the 𝑅2 score, which can be calculated using Eq. (9).
for the MLP algorithm, determined through hyperparameter optimiza- ∑𝑚 2
tion. The sizes of the hidden layers were explored using a combination 1 (𝑦𝑖 − 𝑦𝑝𝑖 )
𝑅2 = 1 − ( )2 (9)
of permutations of [50,100,150] neurons, resulting in configurations ∑𝑚 1 ∑𝑚
1 𝑦𝑖 − 𝑚 1 𝑦𝑖
with 2 and 3 hidden layers. Additionally, the full configuration of [50,
100, 150] neurons was evaluated. Ultimately, a network architecture where 𝑚 is the total number of test data, 𝑦𝑖 is the test value, and
with three hidden layers, containing 100, 150, and 50 neurons in the 𝑦𝑝𝑖 is the predicted capacity by the prediction models. The 𝑅2 value

7
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 6. Performance comparison of the data-driven models developed based on the experimental data for (a) computation time (b) computation accuracy.

efficiency and accuracy underscores the necessity of implementing

enhanced data-driven approaches, such as machine learning methods.

4.3. Overfitting

Prior to finalizing the ML models, they underwent further inspection

for overfitting through a 10-fold cross-validation on the training set.
Fig. 7 illustrates box plots displaying the cross-validation accuracy
of the two ML models trained with the combined data. Each box
represents the interquartile range (IQR), with the median value denoted
by an orange horizontal line at the center of the box. Horizontal bars
outside the box indicate the maximum and minimum scores, defined
as values greater than 1.5 × IQR above the third quartile and values
less than 1.5 × IQR below the first quartile, respectively. Any outliers
that exceed the range of the maximum and minimum accuracy are
illustrated as circles. The MLP model showed higher cross-validation
Fig. 7. Cross-validation accuracy. accuracy than the XGBoost model, with a smaller variance of accuracy
across the 10 folds. The median accuracies of the XGBoost and MLP
models were 0.985 and 0.995, respectively, which were either equal
represents the predictive capability of the proposed function on the test to or slightly higher than the test set accuracy of the corresponding
set data. models. This confirms that the models do not suffer from overfitting
The performances of the training set and test set were evaluated issues and are sufficiently robust, demonstrating their generalization
and compared to assess the models’ generalization capacity. This ap- capability in predicting the buckling capacity of SHS and RHS columns.
proach is widely used to determine the prediction accuracy of machine 5. Data-driven model results
learning models. It is observed that the data-driven models show higher
accuracy curves for the training set in comparison to the test set, as the 5.1. Predictive performance
latter comprises unseen data. When examining the test performance
of the models, the MLP model demonstrates superior performance, After developing the three data-driven models, the performance of
achieving the highest accuracy of 0.995 when 80% of the data is each model was evaluated using the test set. Fig. 8 presents a visual
used for training. The XGBoost model shows higher accuracy when comparison of the predicted buckling capacity by the developed models
at least 40% of the data is used for training. Considering that the and the actual capacity obtained from experiments or FE analyses in
XGBoost model demonstrated the fastest training time in Fig. 6a, it the test set. For each model, the train–test splits that produced the
proves to be more efficient than the interpolation model in terms of best performance in Fig. 6b were adopted: 6:4, 7:3, and 8:2 for the
both computational cost and accuracy. The interpolation model shows interpolation, XGBoost, and MLP models, respectively. The coefficients
the highest test accuracy only when the training set comprises 10% of of determination (𝑅2 ) are provided separately to evaluate the perfor-
the data. However, as 10% of the available data is a small sample size, mance of the experimental data and FE data. A perfect prediction is
its applicability is limited to a narrow range of parameters, and using represented by an 𝑅2 value of 1.0. Due to the limited sample size of the
10% data for training is considered inappropriate as a final model since experimental data compared to the FE data, it demonstrated relatively
multiple sample sets in the test set could fall outside its range. lower performance. Comparing the performance of the developed data-
Based on the comparisons of time and accuracy, it can be in- driven models, the MLP model showed the best performance, achieving
ferred that the interpolation model may provide inefficient predictive higher 𝑅2 scores of 0.998 and 0.981 for the FE and experimental
performance due to its higher computation time and lower accuracy data, respectively. The high performance in the test data highlights
compared to the machine learning models. Given that structural de- the models’ generalization ability, referring to their capability to ac-
sign often involves large datasets, this comparison of computational curately predict unseen data. The maximum performance difference

8
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 8. Comparison of the predicted and actual 𝑁𝑢 of hollow section columns.

in 𝑅2 between the two ML models is minimal, with a value of 0.011

for the FE data and 0.014 for the experimental data. The interpolation
model displayed the lowest accuracy, with scores of 0.959 and 0.892
for the FE data and experimental data, respectively. Overall, the models
showed higher accuracy for the FE data compared to the experimental
data, and the performance of the combined data closely resembled the
𝑅2 values observed for the FE data. This indicates the importance of
including combined data in the training dataset to cover the entire
feature range, which might be missing from experimental data alone. It
should be noted that experiments are crucial for validating FE models,
although they are typically more limited in scope compared to FE
models.

5.2. Feature importance

To enhance model interpretability, feature importance analysis was

employed. This analysis identifies the most important factors and re-
dundant factors for SHS and RHS column capacity estimation. Fea- Fig. 9. Performance curves for different ML models.
ture importance analysis increases both explainability, which discovers
meaning between input data and model outputs, and interpretability,
which discovers causal relationships between these inputs and the out- depending on the number of features in a feature set, with the labels
puts. Furthermore, the feature importance results can be benchmarked of data points representing the feature ranking results. For example,
against mechanics-based principles [9]. Moreover, this process helps the performance of the third point from the left represents the 𝑅2
to improve computational efficiency without compromising prediction score of the model that considers only the top three ranked features
accuracy by excluding the least important features from model fitting. in the dataset. According to the curve, the XGBoost and MLP models
In this study, the permutation method is used to measure feature derived different feature orders but both identified effective length 𝐿𝑐
importance, which has been found to be accurate and fast in identifying and thickness 𝑡 as significant parameters. The forming process is the
important features affecting steel frame behaviors [84]. least significant in predicting the buckling capacity of SHS and RHS
The permutation approach estimates feature importance through columns. This is because only two values for the forming process (hot-
repeated shuffling of data values. The process begins by training a rolling and cold-forming) were considered in the data, thus this variable
model to establish a baseline performance. Subsequently, the values provided limited information for model fitting.
within a single feature group are randomly shuffled, and then the
model’s performance is reevaluated. The importance 𝐽 is the difference 6. Comparison to the current design rules
between the model accuracy obtained from the original database (base-
line performance) and the dataset with permuted feature values of 𝑋𝑖 . This section evaluates the performance of the proposed ML model
The importance score can be calculated by Eq. (10): by comparing its results with two widely used column buckling design
codes: the US code, AISC 360 [31], and the European code, Eurocode
𝐽 (𝑋𝑖 ) = accur acy f or dat aset wit hout per mut at ion
(10) 3 [3]. These design codes employ specific equations based on factors
− accur acy f or per mut ed dat aset of 𝑋𝑖 such as slenderness, steel grade, and manufacturing process. AISC
360 [31] restricts the application of its design rules to hollow sections
If an important feature needed for prediction is absent when those with nominal yield strength up to 485 MPa, whereas EC 3 [3] permits
values of the feature are permuted, there will be a large decrease steel grades up to 700 MPa. The proposed ML method considers a
in model accuracy. Likewise, when an unimportant feature is absent wide range of yield strengths, geometric properties, and manufacturing
during permutation, there will be minimal or no decrease in model methods (cold-forming or hot-rolling), offering a consistent approach
accuracy. This study repeated the shuffling ten times for each feature for estimating the buckling capacity of various hollow section columns.
𝑋𝑖 and used the average of 𝐽 (𝑋𝑖 ) as the importance of each feature.
Feature rankings were obtained based on the importance score 6.1. Capacity prediction
estimated by the permutation method. The ML models were trained and
tested on a feature set that includes the top 𝑖 features, where 𝑖 is the Fig. 10 compares the prediction-to-test capacity ratios (𝑁𝑢,𝑝𝑟𝑒𝑑 ∕
number of important features. Fig. 9 provides the 𝑅2 performance curve 𝑁𝑢,𝑡𝑒𝑠𝑡 ) obtained from the proposed MLP model, developed using an 8:2

9
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 10. Comparisons of predictions by AISC 360 [31], Eurocode 3 [3], and the proposed machine learning model (a) normal-strength steel (b) high-strength steel.

train–test split, and the current design provisions across the range of where the capacity ratio is greater than 1.0. This indicates that the
√
the relative member slenderness ratio 𝜆𝑐 (where 𝜆𝑐 = (𝐿𝑐 ∕𝜋 𝑟) 𝐹𝑦 ∕𝐸). formation methods have a significant effect on the capacity. It should be
The entire data set was utilized even when they are beyond the scope noted that AISC 360 [31] does not account for the different buckling
of current SHS and RHS design codes. Moreover, the root mean square capacities between formation techniques. On the other hand, the EC
error (RMSE) was calculated for the combined data within the 𝜆𝑐 range data points show an increased level of scatter for cold-formed steel
with increments of 0.3 using Eq. (11). RMSE reflects the spread of compared to hot-rolled steel but still remain on the conservative side,
prediction errors, indicating the concentration of data around the line where the capacity ratio is less than 1.0. Eurocode 3 [3] applies a
of best fit. different imperfection factor 𝛼 to estimate the column strength depend-
√
√ ing on the formation process, as discussed in Section 2.2. A larger
√1 ∑ 𝑚
RMSE = √ (𝑦 − 𝑦𝑝𝑖 )2 (11) imperfection factor is applied to cold-formed steel compared to hot-
𝑚 𝑚=1 𝑖
rolled steel, resulting in more conservative predictions for cold-formed
steel in the EC design provisions. The provided comparisons between
The predicted capacity of the existing design rules was estimated us-
the proposed ML model and the current design methods in this section
ing the equations previously introduced in Section 2.2. The MLP model
demonstrate the consistency and good performance of the proposed ML
was selected for the ML prediction due to its superior performance com-
model in predicting both hot-rolled and cold-formed column capacities
pared to the XGBoost model as demonstrated in Section 5.1. Fig. 10(a)
over the range of the relative slenderness ratio 𝜆𝑐 .
and (b) present the comparisons of the data divided into two categories:
normal-strength steel (𝐹𝑦 < 460 MPa) and high-strength steel (𝐹𝑦 ≥
6.2. Reliability-based calibration
460 MPa). Both comparisons indicate that the existing design rules
result in a relatively large scatter band across the entire range of 𝜆𝑐 ,
while the ML results provide a better prediction with a narrower scatter The first order reliability method [34,85,86] was used to evaluate
band. The AISC and EC design provisions show significant errors for the reliability of the proposed ML method. The resistance factor 𝜙
members in the relative slenderness range between 0.6 and 1.5, which was estimated for four groups of the data: Normal-strength steel, high-
represents the inelastic buckling range where the effects of residual strength steel, hot-rolled steel, and cold-formed steel. Chapter K of AISI
stresses and geometric imperfections are prominent. The RMSE plot S100 [2], the North American Specification for the Design of Cold-
for high-strength steel predictions (Fig. 10(b)) further demonstrates Formed Steel Structural Members, provides a simplified, yet robust
that the ML model achieves the highest accuracy throughout the range method to determine the resistance factor for a series of tests or rational
of 𝜆𝑐 . The proposed ML model demonstrates high accuracy for both engineering analyses based on the first order reliability method, where
normal- and high-strength steel column predictions. 𝜙 is determined by Eq. (12):
√
2 2 2 2
Fig. 11 compares the predictions generated by the current design 𝜙 = 𝐶𝜙 (𝑀𝑚 𝐹𝑚 𝑃𝑚 )𝑒−𝛽𝑜 𝑉𝑀 +𝑉𝐹 +𝐶𝑃 𝑉𝑃 +𝑉𝑄 (12)
provisions and the proposed ML model by dividing the samples into two
groups based on the formation techniques: hot-rolled and cold-formed.
In Fig. 11(a), which presents the predictions for hot-rolled steel, the As provided by the AISI S100 Chapter K, this paper adopts a calibra-
ML model shows the best performance with a small variance and a low tion coefficient 𝐶𝜙 of 1.52 and a coefficient of variation of load effect
RMSE score, followed by the EC and AISC models. The AISC and EC 𝑉𝑄 of 0.21 for LRFD. Fabrication factors 𝐹𝑚 and 𝑉𝐹 of 1.0 and 0.05,
predictions for hot-rolled steel are predominantly on the conservative respectively, for compression members are adopted. Material factors
side, where 𝑁𝑢,𝑝𝑟𝑒𝑑 ∕𝑁𝑢,𝑡𝑒𝑠𝑡 ≤ 1.0, and the AISC predictions show a 𝑀𝑚 and 𝑉𝑀 of 1.1 and 0.1, respectively, are adopted [2,87]. These
relatively wide scatter band and a larger error (RMSE) compared to the factors reflect the statistical distributions of dimensions and material
EC predictions. However, for the cold-formed steel results in Fig. 11(b), properties. The target reliability index 𝛽𝑜 for structural members is set
a large number of AISC predictions fall in the unconservative zone, at 2.5 for LRFD.

10
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Fig. 11. Comparisons of predictions by AISC 360 [31], Eurocode 3 [3], and the proposed machine learning model (a) hot-rolled steel (b) cold-formed steel.

Table 2
Resistance factors.
Normal-strength steel High-strength steel Hot-rolled steel Cold-formed steel Entire data
𝜙 0.86 0.91 0.96 0.88 0.90

∑
The mean value of professional factor 𝑃𝑚 is calculated as (1∕𝑛) 𝑛𝑖=1 the subgroups. Future studies should investigate if it may be beneficial
(𝑅𝑡,𝑖 ∕𝑅𝑛,𝑖 ), where 𝑅𝑡,𝑖 represents the tested strength predicted by the to use different resistance factors for various groupings, especially after
proposed ML model in this paper, and 𝑅𝑛,𝑖 is the nominal strength augmenting the dataset with additional samples.
estimated by AISC 360 [31]. The coefficient of variation of test results
𝑉𝑃 is 𝑠𝑐 ∕𝑃𝑚 , where 𝑠𝑐 is the standard deviation of 𝑅𝑡,𝑖 ∕𝑅𝑛,𝑖 , where 𝑉𝑃 7. Practical application
should not be less than 0.065. The correction factor 𝐶𝑃 accounts for
(1+1∕𝑛)(𝑛−1) This section aims to encourage stakeholders such as fabricators
the number of tests: for 𝑛 ≥ 4 and 5.7 for 𝑛 = 3 (minimum
𝑛−3 and practitioners to utilize data-driven methods for steel design and
number of tests permitted). In this analysis, 𝑛 is the number of sample analysis. In order to facilitate this, the analysis models and database
data in each of the four analysis groups. developed in this study have been made publicly available on GitHub,
Table 2 summarizes the resistance factor values determined using a code-hosting platform that enables version control and collabora-
the reliability method, with grouping criteria as outlined in Section 6.1. tion. The GitHub repository for this project can be accessed at https:
The resistance factors vary by group and range from 0.86 to 0.96. The //github.com/EngineerWithData/HollowSectionColumnBucklingTests.
resistance factor for all data groups combined is 0.90, which is the same The analysis methods have been implemented using Google Colab
value specified in AISC 360 [31]. (short for Collaboratory, https://colab.google/), which provides a user-
friendly environment for writing and executing Python code directly
6.3. Future work through a web browser. This platform is particularly well-suited for ma-
chine learning, data analysis, and educational purposes. By leveraging
As specific geometric and material conditions of hollow sections Colab, users can train data-driven models directly in the cloud, facil-
were explored, it is recommended to use the proposed data-driven ap- itating seamless collaboration without compromising the integrity of
proaches within the ranges of the data utilized in this study. To improve the original project. The provided guidance on implementing the data-
the effectiveness of data-driven methods for design and analysis of driven models will assist stakeholders in understanding the application
hollow section columns in the future, it is advisable to collect additional of data analytics in structural engineering and facilitate the adoption
of this modern technology. Additionally, as all project data has been
data covering a wider range of properties and incorporate the new
stored in the accessible repository, future researchers and users can
data into the analysis. Particularly, expanding the FE data beyond the
easily update and modify the codes as necessary by accessing the saved
current focus on 50 mm × 100 mm and 100 mm × 100 mm cross
versions on GitHub.
sections to include larger hollow sections can enhance the robustness
of hollow section column analysis. Moreover, this study considered the 7.1. Data structure
formation method (hot-rolled and cold-formed) as a binary variable,
even though residual stresses could significantly differ between the two The data structure, represented by the file dataset.csv, is orga-
methods. Including residual stress patterns or magnitudes as an input nized in a searchable manner and grouped based on the input features.
parameter may improve the data-driven model. The columns of the data encompass the input features, along with
The reliability analysis indicated that while the resistance factor the output feature, which corresponds to the buckling capacity. For
for all data matched the current resistance factor specified in AISC example, the following codes import the data file from the repository
360 [31] for compression members, the resistance factors varied among and assort the data based on the input features:

11
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

Lc_s = (Lc-np.mean(X[‘$L_{c}$’]))/np.std(X[‘$L_{
data_url = ‘’https :// raw. githubusercontent .com/ c}$’])
EngineerWithData / Fy_s = (Fy-np.mean(X[‘$F_y$ ’]))/np.std(X[‘$F_y$ ’
HollowSectionColumn - ])
BucklingTests /main/
dataset .csv ’ input = pd. DataFrame ([[Forming_s , H_s , B_s , ro_s
data = pd. read_csv (data_url , header =0) , t_s , Lc_s , Fy_s]],
data_x = data[[‘Forming ’, ‘H’, ‘B’, ‘r_o ’, ‘t’, columns = data_x . columns )
‘L_c ’, ‘F_y ’]] y_pred = mlp. predict ( input ) #get prediction
using the scaled
features
The data samples were labeled as Cold-formed and y_pred_de = y_pred *np.std(y).item ()+np.mean(y).
Hot-rolled based on their respective forming processes. To fa- item () #De - normalize
cilitate machine-readability, label encoding was applied to convert y_pred
these forming process labels into a numerical representation. Specifi- print(‘ Buckling capacity = {:6.1f} kN ’. format (
cally, Cold-formed and Hot-rolled were replaced with 0 and 1, y_pred_de [0]))
respectively, by implementing the following code:

8. Conclusion
data[‘Forming ’]. replace ([‘Cold - formed ’, ‘Hot-
rolled ’],[0, 1], inplace This paper presents a test case for using data-driven approaches in
=True) structural engineering by predicting the buckling capacity of square and
rectangular hollow section (SHS and RHS) members under compres-
sion. The data-driven models were developed using linear interpolation
7.2. Code for analysis and two machine learning techniques: Extreme Gradient Boosting (XG-
Boost) and Multi-layer Perceptron (MLP). The development of these
This section aims to provide practical guidelines for practitioners to models was based on a comprehensive database consisting of 695
utilize the proposed models in obtaining the buckling capacity of hol- experimental results and 3794 finite element (FE) analysis results of
low section columns. The corresponding code files for the XGBoost and SHS and RHS columns. The database covered a wide range of material
MLP models, namely XGB.ipynb and MLP.ipynb respectively, can and geometric properties, including steel grades ranging from normal-
be accessed and opened using Google Colab. The repository contains strength to high-strength steel, cross-sectional dimensions, member
comprehensive instructions on how to train data-driven models and slenderness, and forming process (cold-formed or hot-rolled). The com-
generate predictions. Each notebook for the XGBoost and MLP models putational time and accuracy of the three developed data-driven models
consists of five code blocks, with the fifth block dedicated to deriving were evaluated. The best-performing model was validated by compar-
the buckling capacity, which is particularly valuable for practitioners. ing its predictions with the existing design rules for columns provided
In the case of the MLP model, the fifth block is outlined as follows. in the US code (AISC 360) and the European code (Eurocode 3).
To obtain the buckling capacity, users simply need to input the values Additionally, practical applications of these models were demonstrated
of the desired hollow section column’s input features. The example to assist practitioners in understanding the utilization of this modern
code below presents the input features for a cold-formed square hollow technology.
section that is 1700 mm long with a width and height of 150 mm, The major findings regarding using data-driven models are as fol-
an outer corner radius of 9 mm, a thickness of 5 mm, and a yield lows:
strength of 600 MPa. To ensure compatibility with the trained model,
input features should fall within the ranges specified in Section 3. Once 1. The computation time and accuracy of the three data-driven
the input features are provided, the data is standardized since the MLP models showcased the advantages of machine learning tech-
model was developed using the scaled data as presented in Section 4. A niques. The conventional interpolation model required consid-
data structure, input, is created to carry the input features, which is erable time to create meshes from the discrete data, resulting
then utilized to generate the prediction value, y_pred_de. To obtain in a significant execution time for its development. Moreover,
the final buckling capacity value (y_pred_de), denormalization is when evaluating the accuracy using the coefficient of determina-
applied to y_pred. Finally, the nominal buckling capacity value is tion (𝑅2 ), the interpolation model exhibited lower performance
printed. compared to the machine learning models.
2. The MLP model obtained the highest accuracy among the three
# enter values of the input features examined data-driven models. The 𝑅2 score for the test set that
Forming = 0 # use 0 for cold - formed steel and 1 includes both experimental and FE samples was 0.995.
for hot - rolled steel
3. It was shown that the machine learning models created herein do
H = 150 # overall height , mm
B = 150 # overall width , mm not suffer from overfitting issues and possess good generalization
ro = 9 # outer corner radius , mm capabilities.
t = 5 # thickness , mm The major findings regarding the data output comparisons are as
Lc = 1700 # the effective length , mm
follows:
Fy = 600 # yield strength , MPa
1. The developed data-driven models were evaluated separately
# scale the input features based on three data groups: (1) experimental data only, (2) FE
Forming_s = ( Forming -np.mean(X[‘Forming ’]))/np.
data only, and (3) combined data. Overall, the performance on
std(X[‘ Forming ’])
H_s = (H-np.mean(X[‘$H$ ’]))/np.std(X[‘$H$ ’]) the FE data was the highest among the three data groups, while
B_s = (B-np.mean(X[‘$B$ ’]))/np.std(X[‘$B$ ’]) the experimental data resulted in the poorest performance. This
ro_s = (ro-np.mean(X[‘$r_o$ ’]))/np.std(X[‘ $r_o$ ’ can be attributed to the relatively limited sample size of the
]) physical test results in the training set and the greater inherent
t_s = (t-np.mean(X[‘$t$ ’]))/np.std(X[‘$t$ ’]) uncertainty in the input features compared to the FE models.

12
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

2. The performance of the developed MLP model was further as- References
sessed by comparing its predictions to those obtained from the
existing design rules for SHS and RHS columns. The ratios of [1] H. Koh, H.B. Blum, A review of current practice for testing by analysis of cold-
predicted-to-tested capacity were computed for the US code formed steel structures, Structures 37 (2022) 871 – 880, http://dx.doi.org/10.
1016/j.istruc.2022.01.017.
(AISC 360), the European code (Eurocode 3), and the proposed
[2] AISI S100-16, North American Specification for the Design of Cold-Formed Steel
MLP model. These comparisons were conducted separately for Structural Members, AISI, 2016.
(1) normal- and high-strength steel columns and (2) hot-rolled [3] European Committee for Standardization, Eurocode 3—Design of Steel
and cold-formed columns. The MLP model consistently per- Structures—Part 1-1: General Rules and Rules for Buildings, EN 1993-1-1, British
formed well, providing capacity ratios close to 1.0 with minimal Standard, 2022.
[4] European Committee for Standardization, Eurocode 3—Design of Steel
variance across the entire range of member slenderness ratios
Structures—Part 1-3: General Rules - Supplementary Rules for Cold-Formed
and for various steel grades and forming methods. In contrast, Members and Sheeting, EN 1993-1-3, British Standard, 2024.
the existing design methods (ASIC 360 and EC 3) exhibited in- [5] AS/NZS 4600, Cold-Formed Steel Structures, Standards Australia, 2018.
consistent performance with relatively large root mean squared [6] W. El-Dakhakhni, Data analytics in structural engineering, J. Struct. Eng. (2021)
errors and variance. http://dx.doi.org/10.1061/(ASCE)ST.1943-541X.0003112.
3. Reliability analyses were conducted using the first order reliabil- [7] T.C. Lux, L.T. Watson, T.H. Chang, Y. Hong, K. Cameron, Interpolation of
sparse high-dimensional data, Numer. Algorithms 88 (1) (2021) 281–313, http:
ity method provided in AISI S100 to estimate the resistance fac- //dx.doi.org/10.1007/s11075-020-01040-2.
tors of four groups of data: normal-strength steel, high-strength [8] H. Salehi, R. Burgueño, Emerging artificial intelligence methods in structural
steel, hot-rolled steel, and cold-formed steel. The resistance fac- engineering, Eng. Struct. 171 (2018) 170–189, http://dx.doi.org/10.1016/j.
tors for these groups range from 0.86 to 0.96, and the resistance engstruct.2018.05.084.
factor for all data is 0.90 which aligns with AISC 360’s resistance [9] H. Sun, H.V. Burton, H. Huang, Machine learning applications for building
structural design and performance assessment: State-of-the-art review, J. Build.
factor for compression members. Eng. 33 (2021) 101816, http://dx.doi.org/10.1016/j.jobe.2020.101816.
To encourage structural engineers and stakeholders to embrace [10] H.-T. Thai, Machine learning for structural engineering: A state-of-the-art review,
in: Struct, Vol. 38, Elsevier, 2022, pp. 448–491, http://dx.doi.org/10.1016/j.
data-driven steel design and analysis, the database and models de-
istruc.2022.02.003.
scribed herein are publicly available in a GitHub repository where they [11] M.R. Sheidaii, R. Bahraminejad, Evaluation of compression member buckling and
can be executed directly through a web browser in a computational post-buckling behavior using artificial neural network, J. Constr. Steel Res. 70
notebook format. Step-by-step instructions were provided so users can (2012) 71–77, http://dx.doi.org/10.1016/j.jcsr.2011.10.020.
gain hands-on experience in data-driven analysis. This paper demon- [12] Z. Fang, K. Roy, B. Chen, C.-W. Sham, I. Hajirasouliha, J.B. Lim, Deep learning-
based procedure for structural design of cold-formed steel channel sections
strates how data-driven design can be a viable approach to enhance
with edge-stiffened and un-stiffened holes under axial compression, Thin-Walled
existing design provisions. Struct. 166 (2021) 108076, http://dx.doi.org/10.1016/j.tws.2021.108076.
This paper explores the potential of data-driven methods in struc- [13] Y. Pu, E. Mesbahi, Application of artificial neural networks to evaluation of
tural design, acknowledging that this represents an early step in a larger ultimate strength of steel panels, Eng. Struct. 28 (8) (2006) 1190–1196, http:
journey towards machine learning-aided design. Fostering data sharing //dx.doi.org/10.1016/j.engstruct.2005.12.009.
[14] M. Kumar, N. Yadav, Buckling analysis of a beam–column using multilayer per-
through repositories could accelerate data-driven structural design by
ceptron neural network technique, J. Franklin Inst. 350 (10) (2013) 3188–3204,
enabling the incorporation of extensive real-world data. http://dx.doi.org/10.1016/j.jfranklin.2013.07.016.
[15] M. Abambres, K. Rajana, K.D. Tsavdaridis, T.P. Ribeiro, Neural network-based
CRediT authorship contribution statement formula for the buckling load prediction of I-section cellular steel beams, Comput
8 (1) (2018) 2, http://dx.doi.org/10.3390/computers8010002.
[16] V.V. Degtyarev, K.D. Tsavdaridis, Buckling and ultimate load prediction models
Hyeyoung Koh: Writing – original draft, Visualization, Method-
for perforated steel beams using machine learning algorithms, J Build Eng 51
ology, Investigation, Formal analysis, Conceptualization. Hannah B. (2022) 104316, http://dx.doi.org/10.1016/j.jobe.2022.104316.
Blum: Writing – review & editing, Supervision, Project administration, [17] Y. Dai, K. Roy, Z. Fang, B. Chen, G.M. Raftery, J.B. Lim, A novel machine
Methodology, Funding acquisition, Conceptualization. learning model to predict the moment capacity of cold-formed steel channel
beams with edge-stiffened and un-stiffened web holes, J Build Eng 53 (2022)
104592, http://dx.doi.org/10.1016/j.jobe.2022.104592.
Funding statement
[18] M. Pala, A new formulation for distortional buckling stress in cold-formed steel
members, J. Constr. Steel Res. 62 (7) (2006) 716–722, http://dx.doi.org/10.
Support for this research was provided by the University of Wiscon- 1016/j.jcsr.2005.09.011.
sin - Madison Office of the Vice Chancellor for Research, United States [19] M. D’Aniello, E.M. Güneyisi, R. Landolfo, K. Mermerdaş, Analytical prediction of
and Graduate Education with funding from the Wisconsin Alumni available rotation capacity of cold-formed rectangular and square hollow section
beams, Thin-Walled Struct. 77 (2014) 141–152, http://dx.doi.org/10.1016/j.tws.
Research Foundation, United States.
2013.09.015.
[20] S.H. Kim, X. Song, C. Cho, C.H. Lee, Strength prediction of steel CHS X-joints
Declaration of competing interest via leveraging finite element method and machine learning solutions, J. Constr.
Steel Res. 176 (2021) 106394, http://dx.doi.org/10.1016/j.jcsr.2020.106394.
The authors declare that they have no known competing finan- [21] K. Jiang, O. Zhao, Unified machine-learning-assisted design of stainless steel
bolted connections, J. Constr. Steel Res. 211 (2023) 108155, http://dx.doi.org/
cial interests or personal relationships that could have appeared to
10.1016/j.jcsr.2023.108155.
influence the work reported in this paper. [22] J. Rahman, A.M. Billah, P. Arafin, K. Islam, M.L. Nehdi, Design-focused inter-
pretable machine learning models for compressive capacity prediction of gusset
Acknowledgments plate connections, Eng. Struct. 298 (2024) 117038, http://dx.doi.org/10.1016/j.
engstruct.2023.117038.
[23] M. Zarringol, H.-T. Thai, M. Naser, Application of machine learning models
The authors would like to thank Prof. Leroy Gardner from Imperial
for designing CFCFST columns, J. Constr. Steel Res. 185 (2021) 106856, http:
College London for sharing his research group’s experimental and finite //dx.doi.org/10.1016/j.jcsr.2021.106856.
element database on steel hollow sections. [24] V.V. Degtyarev, H.-T. Thai, Design of concrete-filled steel tubular columns using
data-driven methods, J. Constr. Steel Res. 200 (2023) 107653, http://dx.doi.org/
Data availability 10.1016/j.jcsr.2022.107653.
[25] T.G. Wakjira, M. Ibrahim, U. Ebead, M.S. Alam, Explainable machine learning
model and reliability analysis for flexural capacity prediction of RC beams
The link to the data is in the paper. strengthened in flexure with FRCM, Eng. Struct. 255 (2022) 113903, http:
//dx.doi.org/10.1016/j.engstruct.2022.113903.

13
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

[26] V.-L. Tran, D.-K. Thai, J.-K. Kim, Machine learning-based model for moment [53] T.-H. Nguyen, N.-L. Tran, D.-D. Nguyen, Prediction of axial compression capacity
capacity prediction and reliability analysis of PSC beams, in: Struct, Vol. 62, of cold-formed steel oval hollow section columns using ANN and ANFIS models,
Elsevier, 2024, 106181, http://dx.doi.org/10.1016/j.istruc.2024.106181. Int. J. Steel Struct. 22 (1) (2022) 1–26, http://dx.doi.org/10.1007/s13296-021-
[27] M.F. Tamimi, A.A. Alshannaq, I. Mu’ath, Sensitivity and reliability assessment of 00557-z.
buckling restrained braces using machine learning assisted-simulation, J. Constr. [54] A. Toffolon, M.A. Kraus, A. Taras, Deep learning based method for the prediction
Steel Res. 211 (2023) 108187, http://dx.doi.org/10.1016/j.jcsr.2023.108187. of the buckling resistance of SHS and RHS, CE Pap. 4 (2–4) (2021) 1076–1084,
[28] Z. Lyu, J. Zhang, N. Zhao, Q. Xiang, Y. Song, J. Li, A comparative study on the http://dx.doi.org/10.1002/cepa.1398.
performance of FEM, RA and ANN methods in strength prediction of pallet-rack [55] Y. Xu, M. Zhang, B. Zheng, Design of cold-formed stainless steel circular hollow
stub columns, Int. J. Steel Struct. 20 (2020) 1509–1526, http://dx.doi.org/10. section columns using machine learning methods, in: Struct, Vol. 33, Elsevier,
1007/s13296-020-00386-6. 2021, pp. 2755–2770, http://dx.doi.org/10.1016/j.istruc.2021.06.030.
[29] M. Dissanayake, H. Nguyen, K. Poologanathan, G. Perampalam, I. Upasiri, H. [56] I. Abarkan, M. Rabi, F.P.V. Ferreira, R. Shamass, V. Limbachiya, Y.S. Jweihan,
Rajanayagam, T. Suntharalingam, Prediction of shear capacity of steel channel L.F.P. Santos, Machine learning for optimal design of circular hollow section
sections using machine learning algorithms, Thin-Walled Struct. 175 (2022) stainless steel stub columns: A comparative analysis with eurocode 3 predic-
109152, http://dx.doi.org/10.1016/j.tws.2022.109152. tions, Eng. Appl. Artif. Intell. 132 (2024) 107952, http://dx.doi.org/10.1016/j.
[30] L. Gardner, N. Saari, F. Wang, Comparative experimental study of hot-rolled engappai.2024.107952.
and cold-formed rectangular hollow sections, Thin-Walled Struct. 48 (7) (2010) [57] T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, in: Proc. of the
495–507, http://dx.doi.org/10.1016/j.tws.2010.02.003. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
[31] American Institute of Steel Construction, Specification for Structural Steel Mining, 2016, pp. 785–794.
Buildings, ANSI/AISC 360-22, Chicago, IL, 2022. [58] D. Akchurin, R. Sabelli, R.D. Ziemian, B.W. Schafer, ASD and LRFD: Reliability
[32] R. Tide, Reasonable column design equations, in: Proceedings 1985 Annual comparison for designs subjected to wind loads, J. Constr. Steel Res. 213 (2024)
Technical Session, Structural Stability Research Council, Bethlehem, PA, 1985. 108327, http://dx.doi.org/10.1016/j.jcsr.2023.108327.
[33] R. Bjorhovde, Columns: from theory to practice, ENG J-AISC 25 (1) (1988) [59] J. Kiani, C. Camp, S. Pezeshk, On the application of machine learning techniques
21–34, http://dx.doi.org/10.62913/engj.v25i1.1269. to derive seismic fragility curves, Comput. Struct. 218 (2019) 108–122, http:
[34] T.V. Galambos, Guide to Stability Design Criteria for Metal Structures, fifth ed., //dx.doi.org/10.1016/j.compstruc.2019.03.004.
John Wiley & Sons, 1998. [60] J. Rahman, K.S. Ahmed, N.I. Khan, K. Islam, S. Mangalathu, Data-driven shear
[35] H. Beer, G. Schulz, Bases théoriques des courbes européennes de flambement, strength prediction of steel fiber reinforced concrete beams using machine
Construct. Métal. 3 (1970) 37–57. learning approach, Eng. Struct. 233 (2021) 111743, http://dx.doi.org/10.1016/
[36] J. Strating, H. Vos, Computer simulation of the ECCS buckling curve using j.engstruct.2020.111743.
a Monte–Carlo method, in: Proc. of the International Colloquium on Column [61] V.V. Degtyarev, Neural networks for predicting shear strength of CFS channels
Strength, 1972. with slotted webs, J. Constr. Steel Res. 177 (2021) 106443, http://dx.doi.org/
[37] D. Sfintesco, A. Carpena, Experimental bases of the ECCS column curves, in: 10.1016/j.jcsr.2020.106443.
International Colloquium on Stability Introductory Report, (Liege, Washington),
[62] P. Virtanen, R. Gommers, T.E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau,
1977, pp. 68–75.
E. Burovski, P. Peterson, W. Weckesser, J. Bright, S.J. van der Walt, M. Brett,
[38] T.M. Chan, L. Gardner, K.H. Law, Structural design of elliptical hollow sections:
J. Wilson, K.J. Millman, N. Mayorov, A.R.J. Nelson, E. Jones, R. Kern, E.
A review, Proc. Inst. Civ. Eng. Struct. Build. 163 (6) (2010) 391–402, http:
Larson, C.J. Carey, İ. Polat, Y. Feng, E.W. Moore, J. VanderPlas, D. Laxalde, J.
//dx.doi.org/10.1680/stbu.2010.163.6.391.
Perktold, R. Cimrman, I. Henriksen, E.A. Quintero, C.R. Harris, A.M. Archibald,
[39] A. Taras, R. Greiner, New design curves for lateral–torsional buckling—Proposal
A.H. Ribeiro, F. Pedregosa, P. van Mulbregt, SciPy 1.0 Contributors, SciPy 1.0:
based on a consistent derivation, J Const Steel Res 66 (5) (2010) 648–663,
Fundamental algorithms for scientific computing in python, Nature Methods 17
http://dx.doi.org/10.1016/j.jcsr.2010.01.011.
(2020) 261–272, http://dx.doi.org/10.1038/s41592-019-0686-2.
[40] X. Meng, L. Gardner, Behavior and design of normal-and high-strength steel SHS
[63] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.
and RHS columns, J. Struct. Eng. 146 (11) (2020) 04020227, http://dx.doi.org/
Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Scikit-learn: Machine
10.1061/(ASCE)ST.1943-541X.0002728.
learning in python, J. Mach. Learn. Res. 12 (2011) 2825–2830.
[41] European Committee for Standardization, Eurocode 3—Design of Steel
[64] C.-S. Yang, S.-P. Kao, F.-B. Lee, P.-S. Hung, Twelve different interpolation
Structures—Part 1-12: General - High Strength Steels, EN 1993-1-12, Brussels,
methods: A case study of surfer 8.0, in: Proceedings of the XXth ISPRS Congress,
Belgium, 2007.
Vol. 35, 2004, pp. 778–785.
[42] G. Sedlacek, B. Kuhn, J. Rondal, P. Boeraeve, Buckling Behaviour of Hot-Formed
[65] J.R. Quinlan, Induction of decision trees, Mach. Learn. 1 (1) (1986) 81–106,
SHS in High Strength Steel Grade E-460, Tech. Rep., Cidect Report 2T-2/99,
http://dx.doi.org/10.1007/BF00116251.
Comite International Pour le Developpement et l’Etude de la Construction
[66] W. Ayrton, J. Perry, On struts, Engineer 62 (464–465) (1886) 513–515.
Tubulaire, Altendorf, Switzerland, 1999.
[67] J. Rondal, Contribution à l’Étude de la Stabilité des Profils Creux à Parois Minces
[43] J.-L. Ma, T.-M. Chan, B. Young, Experimental investigation on stub-column
(Ph.D. thesis), Université de Liège, Faculté des Sciences Appliquées, 1984.
behavior of cold-formed high-strength steel tubular sections, J. Struct. Eng. 142
(5) (2016) 04015174, http://dx.doi.org/10.1061/(ASCE)ST.1943-541X.0001456. [68] J. Barber, P. Birkemoe, An Experimental Investigation of the Column Behaviour
[44] J. Wang, L. Gardner, Flexural buckling of hot-finished high-strength steel SHS of Cold-Formed Stress-Relieved Hollow Structural Steel Sections, Department of
and RHS columns, J. Struct. Eng. 143 (6) (2017) 04017028, http://dx.doi.org/ Civil Engineering, University of Toronto, 1978.
10.1061/(ASCE)ST.1943-541X.0001763. [69] R.M. Sully, G.J. Hancock, Behavior of cold-formed SHS beam-columns, J.
[45] P. Guiaux, Essais De flambement sur Profils Creux Formes a Froid, Carres et Struct. Eng. 122 (3) (1996) 326–336, http://dx.doi.org/10.1061/(ASCE)0733-
Circulaires, Tech. Rep., CIDECT 72/28/F, Altendorf, Switzerland, 1972. 9445(1996)122:3(326).
[46] R. Bjorhovde, Strength and Behavior of Cold-Formed HSS Columns, Tech. Rep., [70] L. Pavlovčič, B. Froschmeier, U. Kuhlmann, D. Beg, Finite element simulation of
Rep. No. 65, Univ. of Alberta, Edmonton, AB, Canada, 1977. slender thin-walled box columns by implementing real initial conditions, Adv.
[47] M. Braham, J.P. Grimault, C.C. Massonnet, J. Mouty, J. Rondal, Flambement Eng. Softw. 44 (1) (2012) 63–74, http://dx.doi.org/10.1016/j.advengsoft.2011.
des Profils Creux a Parois Minces. Cas des Profils Rectangulaires Charges 05.036.
Axialment, Tech. Rep., CIDECT 2H 79/19. Altendorf, Switzerland: Committee for [71] SSAB, Axial Resistance of Double Grade (S355, S420) Hollow Sections Manu-
International Development and Education on Construction of Tubular Structures, factured by SSAB, Statistical Evaluation Based on Tests, Tech. Rep., Helsinki,
1980. Finland, 2014.
[48] P.W. Key, S.W. Hasan, G.J. Hancock, Column behavior of cold-formed hollow [72] Dassault Systems, Abaqus/CAE, V6.16, Dassault Systems, Johnston, RI, 2015.
sections, J. Struct. Eng. 114 (2) (1988) 390–407, http://dx.doi.org/10.1061/ [73] K. Law, L. Gardner, Lateral instability of elliptical hollow section beams, Eng.
(ASCE)0733-9445(1988)114:2(390). Struct. 37 (2012) 152–166, http://dx.doi.org/10.1016/j.engstruct.2011.12.008.
[49] G. Sedlacek, J. Rondal, P. Boeraeve, N. Stranghöner, R. Schneider, D. Grotmann, [74] J. Wang, S. Afshan, M. Gkantou, M. Theofanous, C. Baniotopoulos, L. Gardner,
Buckling Behaviour of a New Generation of Cold Formed Hollow Sections, Tech. Flexural behaviour of hot-finished high strength steel square and rectangular
Rep., Final CIDECT Report 2R-2/96, Aachen University, Aachen, Germany, 1996. hollow sections, J. Constr. Steel Res. 121 (2016) 97–109, http://dx.doi.org/10.
[50] B. Somodi, B. Kövesdi, Flexural buckling resistance of cold-formed HSS hollow 1016/j.jcsr.2016.01.017.
section members, J. Constr. Steel Res. 128 (2017) 179–192, http://dx.doi.org/ [75] K. Rasmussen, G. Hancock, Design of cold-formed stainless steel tubular mem-
10.1016/j.jcsr.2016.08.014. bers. II: Beams, J. Struct. Eng. 119 (8) (1993) 2368–2386, http://dx.doi.org/10.
[51] Y. Xu, B. Zheng, M. Zhang, Capacity prediction of cold-formed stainless steel 1061/(ASCE)0733-9445(1993)119:8(2368).
tubular columns using machine learning methods, J. Constr. Steel Res. 182 [76] M. Jandera, L. Gardner, J. Machacek, Residual stresses in cold-rolled stainless
(2021) 106682, http://dx.doi.org/10.1016/j.jcsr.2021.106682. steel hollow sections, J. Constr. Steel Res. 64 (11) (2008) 1255–1263, http:
[52] T.-H. Nguyen, D.-X. Nguyen, T.-T.T. Nguyen, V.-L. Phan, D.-D. Nguyen, Machine //dx.doi.org/10.1016/j.jcsr.2008.07.022.
learning models for predicting the axial compression capacity of cold-formed [77] R.B. Cruise, L. Gardner, Strength enhancements induced during cold forming
steel elliptical hollow section columns, Asian J. Civ. Eng. 25 (2) (2024) of stainless steel sections, J. Constr. Steel Res. 64 (11) (2008) 1310–1316,
1935–1947, http://dx.doi.org/10.1007/s42107-023-00886-w. http://dx.doi.org/10.1016/j.jcsr.2008.04.014.

14
H. Koh and H.B. Blum Journal of Constructional Steel Research 224 (2025) 109085

[78] S. Afshan, B. Rossi, L. Gardner, Strength enhancements in cold-formed structural [83] E.W. Cheney, W.A. Light, A Course in Approximation Theory, vol. 101, Amer
sections—Part I: Material testing, J. Constr. Steel Res. 83 (2013) 177–188, Math Soc, 2009.
http://dx.doi.org/10.1016/j.jcsr.2012.12.008. [84] H. Koh, H.B. Blum, Machine learning-based sensitivity of steel frames with
[79] F.M. Bartlett, R.J. Dexter, M.D. Graeser, J.J. Jelinek, B.J. Schmidt, T.V. Galam- highly imbalanced and high-dimensional data, Eng. Struct. 259 (2022) 114126,
bos, Updating standard shape material properties database for design and http://dx.doi.org/10.1016/j.engstruct.2022.114126.
reliability, Eng. J. AISC 40 (1) (2003) 2–14, http://dx.doi.org/10.62913/engj. [85] L.-E. Hsiao, W.-W. Yu, T.V. Galambos, AISI LRFD method for cold-formed steel
v40i1.800. structural members, J. Struct. Eng. 116 (2) (1990) 500–517, http://dx.doi.org/
[80] B. Ellingwood, J.G. MacGregor, T.V. Galambos, C.A. Cornell, Probability based 10.1061/(ASCE)0733-9445(1990)116:2(500.
load criteria: Load factors and load combinations, J. Struct. Div. 108 (5) (1982) [86] W.B. Hall, T. Peköz, Probabilistic evaluation of test results, in: Proceedings of
978–997, http://dx.doi.org/10.1061/JSDEAG.0005959. the Ninth International Specialty Conference on Cold-Formed Steel Structures,
[81] C. Barber, D. Dobkin, Qhull: Software for computing convex hulls and related 1988.
structures, 2020, http://www.qhull.org/, Version 8.0.0 [Software]. [87] T.V. Galambos, M. Ravindra, Proposed criteria for load and resistance factor
[82] B. Delaunay, Sur la sphère vide, Bull. Acad. Sci. URSS 6 (1934) 793–800. design, Struct. Div. 15 (1) (1978) 8–17, http://dx.doi.org/10.62913/engj.v15i1.
304.

15