Thin-Walled Structures 184 (2023) 110518

Contents lists available at ScienceDirect

Thin-Walled Structures
journal homepage: www.elsevier.com/locate/tws

Full length article

Predicting the buckling behaviour of thin-walled structural elements using

machine learning methods
Seyed Mohammad Mojtabaei a,c ,∗, Jurgen Becque b , Iman Hajirasouliha a , Rasoul Khandan c
a
Department of Civil and Structural Engineering, The University of Sheffield, Sheffield S1 3JD, UK
b
Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, UK
c
College of Engineering and Physical Sciences, Aston University, Birmingham B4 7ET, UK

ARTICLE INFO ABSTRACT

Keywords: The design process of thin-walled structural members is highly complex due to the possible occurrence of
Thin-walled members multiple instabilities. This research therefore aimed to develop machine learning algorithms to predict the
Machine learning buckling behaviour of thin-walled channel elements subjected to axial compression or bending. Feed-forward
Artificial Neural Network (ANN)
multi-layer Artificial Neural Networks (ANNs) were trained, in which the input variables comprised the cross-
K-fold cross-validation
sectional dimensions and thickness, the presence/location of intermediate stiffeners, and the element length.
Buckling resistance
Modal decomposition
The output data consisted of the elastic critical buckling load or moment, while also providing an immediate
modal decomposition of the buckled shape into the traditionally defined ‘pure’ buckling mode categories (i.e.
local, distortional and global buckling). The sample output for training was prepared using a combination of the
Finite Strip Method (FSM) and the Equivalent Nodal Force Method (ENFM). The ANN models were subjected
to a K-fold cross-validation technique and the hyperparameters were tuned using a grid search technique. The
results indicated that the trained algorithms were capable of predicting the elastic critical buckling loads and
carrying out the modal decomposition of the critical buckled shapes with an average accuracy (𝑅2 -value) of
98%. The influence of the various channel parameters on the output was assessed using the SHapley Additive
exPlanations (SHAP) method.

1. Introduction effective width concept in their fundamental approach, but specify

separate strength curves for distortional buckling. These traditional
Cold-formed steel (CFS) structural elements, manufactured near code-specified design rules are quite prescriptive, however, and tend to
room temperature from thin steel plate, possess tangible advantages, become cumbersome and tedious when applied to geometrically more
such as high strength-to-weight and stiffness-to-weight ratios, ease of advanced cross-sections, which may contain features such as multiple
handling and transportation, a flexible manufacturing process capa- intermediate web/flange stiffeners and complex lip stiffeners. The fact
ble of producing a variety of cross-sectional shapes, and recyclability that these design rules rely on traditional distinctions between, for
without loss of quality, which in turn promotes sustainability [1,2].
instance, ‘flanges’ and ‘webs’ even pre-empts application to some cross-
However, their limited wall thickness results in a heightened suscepti-
sections with non-traditional, novel and innovative geometries. For
bility to instabilities and generates a need to account for these through
these reasons, they are sometimes seen as an impediment to further
a design process of increased complexity.
development, optimization and innovation in the field.
Traditional standard-prescribed design methods are typically based
on the effective width approach, pioneered by von Karman [3], in their To a large extent because of the above issues, the Direct Strength
treatment of cross-sectional instability. The plate elements constituting Method (DSM) has enjoyed a steady rise in popularity as an alternative
the cross-section are thereby treated as hinged along their adjoining design method since it was first proposed by Schafer and Peköz in
lines, and interaction between plate elements is conveniently ignored. 1998 [6]. The DSM relies on the determination of the individual
The Eurocode EN1993-1-3 [4] follows this approach. Additionally, (elastic) local, distortional and global buckling stresses of the member,
distortional buckling is dealt with using a column buckling model of a and combines this information with the yield stress of the material to
relevant cross-sectional subassembly, with the restraint exerted by the define a slenderness value corresponding to each type of instability.
remainder of the cross-section represented by a Winkler foundation. Statistically calibrated strength curves are then used to determine the
The North-American design specifications (AISI-S100) [5] also use the ultimate capacity. The DSM has historically been closely linked to the

∗ Corresponding author.
E-mail address: [email protected] (S.M. Mojtabaei).

https://doi.org/10.1016/j.tws.2022.110518
Received 22 September 2022; Received in revised form 6 December 2022; Accepted 31 December 2022
Available online 7 January 2023
0263-8231/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Finite Strip Method (FSM), where the latter is used as an analysis bending, and subsequently proposed ANN-based predictive equations.
tool to determine the elastic buckling stresses corresponding to the These equations were then used in a follow-up study by Pala and
various modes from the ’signature diagram’, which plots the buckling Caglar [23] to investigate the effects of the geometric parameters (in-
stress against the buckle half-wavelength. This is further explained in cluding the web height, flange width, flange thickness, and inclination
Section 2.1. However, a number of difficulties may arise in practical and length of the lips) on the distortional buckling stress. In another
application [7]. An ‘indistinct minimum’ may be present, meaning that study, an ANN algorithm was trained by Guzelbey et al. [24] using
the local mode minimum in the signature curve is obscured by the an experimental dataset, with the aim of predicting the web crippling
distortional buckling curve, or vice versa. The minima in the signature strength of CFS trapezoidal decks. The results indicated that the ANN
curve also more often than not do not correspond to the ‘pure’ modes, could provide considerably more accurate predictions compared to
but to coupled instabilities. The distortional minimum in particular those obtained from the current design codes.
usually includes a non-negligible contribution of the local mode. These Due to the complexity of calculating distortional buckling stresses
issues have spurred research into the ‘modal decomposition problem’ using classical shell theories, Dias and Silvestre [25] trained an ANN on
with the aim of determining the buckled shapes and buckling stresses an analytically developed dataset, and consequently presented closed-
of the pure local, distortional and global modes, as well as their form expressions to estimate the critical distortional buckling stress and
contributions in a randomly deformed shape. To generate the capability associated half-wavelength of elliptical hollow sections in compression.
of a complete decomposition where all possible deformations within In another relevant study, Tohidi and Sharifi [26] developed an ANN
the deformation space are accounted for, two more types of pure model to predict the ultimate moment capacities of steel I-beams and
modes are typically added: shear modes and (transverse) extension demonstrated that the proposed ANN-based formula is more accurate
modes. Groundbreaking work in this area was conducted by Adany than the existing design codes.
and Schafer [8], who achieved modal decomposition by importing the A soft-computing technique using ANN and Genetic Expression
mechanical descriptions of the buckling modes established in Gener- Programming (GEP) was developed by D’Aniello et al. [27] to predict
alized Beam Theory (GBT) into the FSM. However, GBT is based on the rotational capacity of CFS steel beams with rectangular and square
a number of idealized assumptions (e.g. Vlasov’s assumptions), which hollow sections. Subsequently, predictive models were proposed and
are fundamentally incompatible with the more generally formulated verified against experimental data. Degtyarev [28] trained ANN mod-
mechanical framework of the FSM, leading to an inelegant solution els to predict the elastic shear buckling load and the ultimate shear
which lacks full orthogonality between the modes and needs several strength of CFS channels with slotted webs, and proved that this led
to more accurate predictions compared to the code-prescribed design
‘patch-up’ solutions. More recent solutions to the modal decomposition
equations. In a follow-up study, Degtyarev and Naser [29] compared
problem [9–11] do not suffer from these shortcomings. Among them,
the results of five different machine learning boosting algorithms,
the method of the equivalent nodal forces (ENFM) [11] is the most
including gradient boosting regressor (GBR), extreme gradient boosting
robust and general one. It was therefore employed in this research and
(XGBoost), light gradient boosting machine learning (LightGBM), gradi-
a short description is provided in Section 2.2.
ent boosting with categorical features support (CatBoost), and adaptive
The main goal of this paper is to investigate whether Machine
boosting (AdaBoost), in predicting the elastic shear buckling loads and
Learning can be used to predict the buckling behaviour of CFS members
the shear strength. It was reported that the CatBoost algorithm was
and, additionally, provide a viable and robust solution to the modal
capable of providing the most accurate predictions compared to other
decomposition problem. The field of Artificial Intelligence (AI), and
boosting algorithms.
specifically Machine Learning and Deep Learning, have seen significant
Fang et al. [30,31] trained Deep Belief Network (DBN) algorithms,
development in recent years and are increasingly finding their way
using the results of experimentally validated FE models, to predict the
into structural engineering applications. It has previously been demon-
axial compressive capacity of CFS channel sections with and without
strated that machine learning techniques are capable of providing
holes. Additionally, the predictions obtained from the DBN were used
accurate predictions in highly nonlinear problems with large numbers
to propose enhancement/reduction factors on the axial capacity of such
of parameters [12,13]. Machine learning [14] can be defined as a fam- cross-sections. Zarringol et al. [13] and Xu et al. [32] employed ma-
ily of methods that evaluate the relationship between input and output chine learning algorithms to predict the ultimate strength of concrete-
parameters by detecting latent patterns in data, and consequently use filled CFS tubular columns and stainless steel tubular columns, respec-
the uncovered relationships or patterns to predict future data (i.e. su- tively, under various loading conditions. More recently, Couto [33]
pervised learning). Machine learning can be also employed to carry out trained an ANN model based on the results of FE simulations to
decision-making under uncertainties (e.g. by reinforcement learning to predict the critical buckling moment of tapered beams and compared
find the optimal behaviour in an environment). Compared to the rule- the ANN-based predictions with those given by the available design
based predictive analytics on which conventional structural design is guidelines.
based, machine learning algorithms can be more efficient and powerful Contrary to previous research, this paper focuses on developing
tools by automatically extracting relationships and patterns from large- ANN algorithms to investigate the elastic stability of CFS structural
volume high-dimensional data without relying on data engineering and elements by predicting the elastic critical buckling load and decompos-
domain knowledge. The following paragraphs aim to give an overview ing the associated critical buckled shape into its modal contributions.
of previous research where these approaches have been specifically The training datasets for the elastic critical buckling loads and mo-
employed to CFS. ments were compiled based on FSM output, while the ANN models to
A first group of studies [15–18] have applied AI to the man- achieve modal decomposition were trained on results obtained from the
ufacturing/rolling process of CFS profiles, mostly with the aim of Equivalent Nodal Force Method (ENFM).
predicting the properties of the finished product. Furthermore, a num- The results of this study will potentially prove useful in creating a
ber of research studies have investigated the design and optimization practical design tool for engineers and practitioners and provide them
of CFS elements for different applications, using AI methods. El-Kassas with a largely intuition-based insight into the mechanical behaviour
et al. [19] presented an optimization framework using an Artificial of thin-walled elements, without necessarily requiring any background
Neural Network (ANN) to find the optimum cross-sectional shape for knowledge of programming or complex mechanical concepts. A future
columns. The same researchers also developed neural networks which extension of the presented research could capitalize on the potential
were trained based on the predictions of the BS-5950 Part 5 [20] design to further train the ANN into a method of near universal versatility
standard to predict the failure load of CFS lipped channel sections [21]. by invoking the strengths of other decomposition methods (e.g. the
Pala [22] employed an ANN method to estimate the elastic distortional ‘polarization method’ [10]) in areas where the ENFM lacks applicability
buckling stresses of CFS C-sections under pure compression and pure (e.g. for cross-sections with rounded corners).

2
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 1. Discretization of a thin-walled member into strips.

mid-length. While the degrees of freedom (ui , vi ) and (uj , vj ) determine

the in-plane membrane displacements of a strip located between nodal
lines i and j, the deformations caused by plate bending are determined
by (w, 𝜃i ) and (wj , 𝜃j ). For a strip with pinned boundary conditions,
the shape functions in the longitudinal direction are assumed to be
sinusoidal for v and w, and co-sinusoidal for u. In the transverse
direction, on the other hand, they are linear for u and v, and cubic
for w.
The elastic and geometric stiffness matrices of the member, 𝑲 and
𝑮, are assembled using the local elastic and geometric stiffness matrices
of each strip [34], which leads to the formulation of the following
stability eigenvalue problem:

(𝑲 − 𝜆𝑮) .𝝂 = 0 (1)

where 𝝂 is an eigenvector revealing the buckled shape of the ele-

Fig. 2. Main nodes and sub-nodes in a lipped channel section. ment, and 𝜆 is the eigenvalue indicating the corresponding elastic
buckling stress. The critical buckling stress (𝜎𝑐𝑟 ) for a given buckle
half-wavelength (L) is obtained as the lowest eigenvalue (𝜆𝑐𝑟 ). The plot
2. Theoretical background showing (𝜎𝑐𝑟 ) as a function of (𝐿) is usually called the ‘‘signature curve’’
of the member, and an example is provided in Fig. 4.
This section aims to summarize the basic principles of the FSM and
the ENFM. 2.2. The Equivalent Nodal Force Method (ENFM)

The most straightforward and robust way to achieve modal decom-

2.1. The Finite Strip Method (FSM)
position of buckled shapes is the ‘method of the equivalent forces’
proposed by Becque et al. [11]. In this method, the pure local, distor-
In the FSM [34] thin-walled elements are divided into a number
tional and global buckling modes, with the added shear and transverse
of longitudinal strips, the longitudinal boundaries of which are called extension modes, form a full orthogonal set of basis vectors of the
the ‘nodal lines’ (Fig. 1). In the remainder of this paper, these nodal complete deformation space. This implies that any deformation can
lines will simply be referred to as ‘nodes’. As also illustrated in Fig. 1, a be expressed as a linear combination of these basis vectors. By en-
local (x, y, z) coordinate system associated with each strip was defined, forcing orthogonality of the basis vectors, the decomposition becomes
as well as a global (X, Y, Z) coordinate system. The total number of mathematically unique and its interpretation therefore unambiguous.
nodes and strips are indicated by 𝑁 and 𝑁𝑃 , respectively. The nodes The method hinges on the determination of sets of nodal forces which
of a cross-section can be categorized into three groups, as shown in produce the pure buckled shapes in a first-order linear elastic problem.
Fig. 2: (i) 𝑁𝑚𝑖 internal main nodes, which connect adjacent strips with
non-aligned local y-axes - the number of main nodes depends on the 2.2.1. Local modes
geometry of the cross-section, (ii) 𝑁𝑚𝑒 external main nodes, which are The local buckling modes are characterized by the corners of the
positioned along the free edges of a cross-section (for a lipped channel cross-section (coinciding with the internal main nodes as defined in
section the number of external main nodes 𝑁𝑚𝑒 = 2), and (iii) 𝑁𝑠 sub- Section 2.1) remaining in place during buckling. According to this
nodes which connect adjacent strips with aligned local y-axes. The definition, the local buckling modes can be completely described by
number of sub-nodes can be arbitrarily chosen, with a higher number a subset of degrees of freedom: (i) the rotations 𝜃 of all nodes, and
of subnodes typically resulting in more accurate results [34]. (ii) the out-of-plane displacements w of all nodes except the internal
As shown in Fig. 3, each nodal line has four degrees of freedom, main nodes. Therefore, a set of basis vectors of the local space 𝒗𝐿,𝑖 can
namely the longitudinal x-displacements (u) measured at the ends of be constructed by assembling all possible 4𝑁 × 1 vectors in which all
the nodal line, and the in-plane y-displacement (v), the out-of-plane elements are zero, except for a single element corresponding to one of
z-displacement (w) and the rotation about the x-axis (𝜃), measured at the above-mentioned degrees of freedom, which is set equal to 1. The

3
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 3. Degrees of freedom in the FSM.

number of the basis vectors of the local space (equal to the number of
local modes) is equal to 2𝑁 −𝑁𝑚𝑖 . The 𝒗𝐿,𝑖 vectors can then be arranged
as columns in a matrix 𝑯 𝐿 with dimensions 4𝑁 × (2𝑁 − 𝑁𝑚𝑖 ), and any
local mode shape (𝒅 𝐿 ) can be expressed as a linear combination of the
basis vectors (𝒗𝐿 ) [8]:

𝒅𝐿 = 𝑯 𝐿 𝒂 (2)

In the above equation, 𝒂 is a vector of unknown coefficients, which

can be determined for a given loading by solving the following eigen-
value problem:
( 𝑇 )
𝑯 𝐿 𝑲𝑯 𝐿 − 𝜆𝐿 𝑯 𝑇𝐿 𝑮𝑯 𝐿 𝒂 = 𝟎 (3)

The eigenvalues (𝜆𝐿 ) resulting from Eq. (3) are the local buckling
stresses corresponding to the buckled shapes determined from (2) once
the (𝒂) vectors are known.

2.2.2. Distortional modes

To uniquely define the distortional modes, four criteria are im-
posed on the cross-section [11]: (DI ) the nodal forces (𝒇 ) maintain Fig. 4. Typical signature curve.
cross-sectional equilibrium, (DII ) the longitudinal forces acting on the
member are zero, (DIII ) the transverse membrane stresses are zero, and In the FSM, criterion (DIII ) can only be imposed in an averaged
(DIV ) the distortional modes are orthogonal to the local modes. form over the strip width (𝑏), which leads to the following set of
The following set of equations is used to ensure cross-sectional equations [25]:
equilibrium: ( ) ( ) ( )
𝑉1,𝑖 − 𝑉2,𝑖 cos 𝛼𝑖 − 𝑊1,𝑖 − 𝑊2,𝑖 sin 𝛼𝑖 𝑈1,𝑖 + 𝑈2,𝑖 ( 𝜈𝜋 )
⎧ + =0
∑𝑁
𝑏𝑖 2 𝐿
⎪ 𝑓𝑌𝑖 = 0
⎪ 𝑖=1 (𝑖 = 1, … , 𝑁𝑃 ) (6)
⎪
⎪ ∑𝑁
where 𝜈 is Poisson’s ratio and the subscripts ‘1’ and ‘2’ refer to both
⎨ 𝑓𝑍𝑖 = 0 (4)
⎪ 𝑖=1
nodal lines at the boundaries of the strip. Furthermore, 𝛼𝑖 is the angle
⎪𝑁 measured from the local y-axis of strip i to the global Y-axis, measured
⎪∑ 𝑖 𝑖 positive in the counterclockwise direction. The coefficients in Eq. (6)
⎪ (𝑓𝑌 𝑍𝑖 − 𝑓𝑍 𝑌𝑖 ) = 0
⎩ 𝑖=1 can be assembled into a matrix 𝑪 𝜎 , while the degrees of freedom are
contained in a vector 𝒅, allowing Eq. (6) to be translated into:
where 𝑓𝑌𝑖 and 𝑓𝑍𝑖 are the nodal force components in the global Y
and Z directions at node 𝑖, respectively, and 𝑌𝑖 and 𝑍𝑖 are the global 𝑪 𝜎 𝒅 = 𝑪 𝜎 𝑲 −1 𝒇 = 𝑪 2 𝒇 = 𝟎 (7)
coordinates of node 𝑖 relative to an arbitrary origin. It is worth noting
Finally, the orthogonality criterion (DIV ) is expressed mathemati-
that the final equation in Eq. (4) contains a typo in the original
cally as:
formulation [25]. The matrix form of Eq. (4) is written as 𝑪 1 𝒇 = 𝟎,
where 𝑪 1 is a 3 × 4𝑁 matrix. 𝑯 𝑇𝐿 𝑲𝒅 = 𝑯 𝑇𝐿 𝒇 = 𝟎 (8)
Criterion (DII ) can also be expressed in matrix form with the help The forces generating the distortional modes (𝑭 𝐷 ) can thus be
of a reduced 3𝑁 × 1 vector (𝒇̂ ) which contains all nodal forces except determined as the null space of the matrix 𝑪 𝐷 :
the longitudinal end forces:
[ ] ⎛⎡ 𝑪 1 ⎤⎞
𝑰 ( ) ⎜⎢ ⎥⎟
𝒇= 𝒇̂ = 𝑻 𝒇̂ (5) 𝑭 𝐷 = 𝑛𝑢𝑙𝑙 𝑪 𝐷 = 𝑛𝑢𝑙𝑙 ⎜⎢ 𝑪 2 ⎥⎟
̂ (9)
𝟎 ⎜⎢ ⎥⎟
⎜⎢ 𝑇 ⎥⎟
⎝⎣𝑯 𝐿 ⎦⎠
In Eq. (5) 𝑰 is the 3𝑁 × 3𝑁 identity matrix and 𝟎 is an 𝑁 × 3𝑁
matrix containing zeros. Consequently, 𝑻 is a matrix of size 4𝑁 × 3𝑁. where: 𝑭 𝐷 = 𝑻 𝑭̂𝐷

4
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

The matrix 𝑪 𝐷 has dimensions (2𝑁 − 𝑁𝑚𝑖 + 𝑁𝑃 + 3) × 3𝑁. A 2.2.6. Modal contributions
constrained eigenvalue problem can then be formulated using the ma- Once the pure modes (𝒅 𝑖 ) (i.e. the local (𝒅 𝐿 ), distortional (𝒅 𝐷 ),
trix 𝑯 𝐷 = 𝑲 −1 𝑭 𝐷 , comprising a set of basis vectors of the distortional global (𝒅 𝐺 ), transverse extension (𝒅 𝑇 𝐸 ) and shear modes (𝒅 𝑆 )) have
space: been determined, their participations in a random deformed shape
( 𝑇 ) 𝒗 (expressed through the local (𝑐𝐿 ), distortional (𝑐𝐷 ), global (𝑐𝐺 ),
𝑯 𝐷 𝑲𝑯 𝐷 − 𝜆𝐷 𝑯 𝑇𝐷 𝑮𝑯 𝐷 𝒂 = 𝟎 (10)
transverse extension (𝑐𝑇 𝐸 ) and shear mode (𝑐𝑆 ) participation factors)
This determines the buckling stresses of the distortional modes (𝜆𝐷 ) can be calculated as:
and the associated modal shapes 𝒅 𝐷 = 𝑯 𝐷 𝒂 under a given loading. It is ∑
noted that for an open unbranched cross-section (e.g. a lipped channel) 𝑐𝑖 = (0.5𝒗𝑇 𝑲𝒅 𝑖 ) (15)
the number of distortional modes is always equal to 𝑁𝑚𝑖 − 2.
where the sum is carried out over all the basis vectors of the considered
2.2.3. Global modes subspace.
To define the pure global modes, similar criteria to those maintained A particularly relevant application is the case where 𝒗 is a vector
for the distortional modes are applied, except that cross-sectional equi- belonging to the FSM output obtained from Eq. (1).
librium of the nodal forces is no longer required: (GI ) no longitudinal
forces are necessary to generate the global modes, (GII ) the transverse 3. Dataset and parameter space
membrane stresses are zero, and (GIII ) the global modes are orthogonal
to both local and distortional modes. Based on this, the forces generat- A dataset was compiled pertaining to 4608 CFS elements with
ing the global modes (𝑭 𝐺 ) can be determined as the null space of the lipped channel sections. The data cover various lengths, cross-sectional
matrix 𝑪 𝐺 : dimensions and thicknesses, and account for the possible presence of
⎛⎡ 𝑪 2 ⎤⎞ intermediate stiffeners in the web and/or flanges. More specifically,
( ) ⎜⎢ ⎥⎟ four cross-sectional shapes were considered, namely an unstiffened
𝑭 𝐺 = 𝑛𝑢𝑙𝑙 𝑪 𝐺 = 𝑛𝑢𝑙𝑙 ⎜⎢ 𝑯 𝑇𝐿 ⎥⎟
̂ (11) lipped channel, a lipped channel with intermediately stiffened flanges,
⎜⎢ ⎥⎟
⎜⎢ 𝑇 ⎥⎟ a lipped channel with a stiffened web, and a lipped channel with
⎝⎣𝑯 𝐷 ⎦⎠
stiffened flanges and web, as shown in Table 1. For each cross-sectional
where: 𝑭 𝐺 = 𝑻 𝑭̂𝐺 shape an identical number of 1170 data points were generated. The
Subsequently, the matrix 𝑯 𝐺 = 𝑲 −1 𝑭 𝐺 , containing a set of basis input data consisted of seven independent parameters, including six
vectors of the global space, is employed to formulate a constrained cross-sectional parameters (see Table 1): the web height (ℎ), the flange
eigenvalue problem: width (𝑏), the lip length (𝑐), the plate thickness (𝑡), and the locations
( 𝑇 ) of the intermediate stiffeners in the flanges (𝑟1 ) and the web (𝑟2 ). The
𝑯 𝐺 𝑲𝑯 𝐺 − 𝜆𝐺 𝑯 𝑇𝐺 𝑮𝑯 𝐺 𝒂 = 𝟎 (12)
remaining parameter was the element length (𝐿). It is noted that the
This yields the global mode shapes 𝒅 𝐺 = 𝑯 𝐺 𝒂 under a given loading intermediate stiffeners always consisted of two 10 mm legs with an
and the corresponding global buckling stresses (𝜆𝐺 ). It is noted that the intersecting angle of 60◦ . The ranges of the selected parameters are
number of global modes is always three. given in Table 1. The output data was generated by performing FSM
and ENFM analyses on the selected CFS elements, and comprised: (i)
2.2.4. Transverse extension modes
the elastic critical buckling load (𝑃𝑐𝑟 ), (ii) the modal contributions to
The transverse extension modes are determined by imposing that:
the critical buckling shape associated with (𝑃𝑐𝑟 ), expressed through the
(TEI ) no longitudinal nodal forces are necessary to generate the trans-
verse extension modes, and (TEII ) the transverse extension modes are local (𝑐𝐿,𝐶 ), distortional (𝑐𝐷,𝐶 ) and global (𝑐𝐺,𝐶 ) participation factors,
orthogonal to the local, distortional and global modes. These criteria (iii) the elastic critical moment (𝑀𝑐𝑟 ), and (iv) the modal contributions
can be mathematically expressed as: to the critical buckling shapes associated with (𝑀𝑐𝑟 ), expressed through
the local (𝑐𝐿,𝐹 ), distortional (𝑐𝐷,𝐹 ) and global (𝑐𝐺,𝐹 ) participation fac-
⎛⎡ 𝑯 𝑇𝐿 ⎤⎞
⎜⎢ ⎥⎟ tors. It should be noted that the modal decomposition results across the
( )
̂
𝑭 ⎜⎢ 𝑇 ⎥⎟
𝑇 𝐸 = 𝑛𝑢𝑙𝑙 𝑪 𝑇 𝐸 = 𝑛𝑢𝑙𝑙 𝑯 𝐷 (13) dataset showed negligible contributions from the transverse extension
⎜⎢ ⎥⎟ and shear modes (less than 1%). Therefore, only the local, distortional
⎜⎢ 𝑇 ⎥⎟
⎝⎣ 𝐺 ⎦⎠
𝑯 and global instabilities were taken into account in this study.
̂
where: 𝑭 𝑇 𝐸 = 𝑻 𝑭 Fig. 5 shows the histograms of the input and output parameters of
𝑇𝐸
The matrix 𝑯 𝑇 𝐸 = 𝑲 −1 𝑭 𝑇 𝐸 , comprising a set of basis vectors of the the dataset. For illustrative purposes, the distributions of the compres-
transverse extension space, can then be used to formulate an eigenvalue sive (𝜆𝑐𝑟,𝐶 ) and flexural (𝜆𝑐𝑟,𝐹 ) cross-sectional slenderness parameters
problem in a similar fashion to Eqs. (3), (10) and (12) to obtain the are also shown in Fig. 5. These slenderness parameters are defined as:
transversely extended modal shapes (𝒅 𝑇 𝐸 = 𝑯 𝑇 𝐸 𝒂) under a given √
𝑃𝑦
loading. 𝜆𝑐𝑟,𝐶 = (16)
𝑃𝑐𝑟
√
2.2.5. Shear modes 𝑀𝑦
As a final category of buckling modes, the shear modes are de- 𝜆𝑐𝑟,𝐹 = (17)
𝑀𝑐𝑟
termined by their orthogonality with all previously obtained modes:
where 𝑃𝑦 and 𝑀𝑦 are the compressive yield load and the yield moment,
⎛⎡ 𝑯 𝑇𝐿 ⎤⎞ respectively. The yield stress of material was assumed to be 350 MPa.
⎜⎢ ⎥⎟ It is noted that the ranges of the input parameters were selected to
( ) ⎜⎢ 𝑯 𝑇 ⎥⎟
⎜ ⎢ 𝐷
⎥⎟ be representative of commercially available channel sections. Addi-
𝑭𝑆 = 𝑛𝑢𝑙𝑙 𝑪 𝑆 = 𝑛𝑢𝑙𝑙 (14)
⎜⎢ 𝑯 𝑇 ⎥⎟ tional advice in this respect was sought from the industrial project
⎜⎢ 𝐺 ⎥⎟ partner [35]. For instance, the member lengths in the dataset were kept
⎜⎢ 𝑇 ⎥⎟
⎝⎣𝑯 𝑇 𝐸 ⎦⎠ within the practically encountered span lengths for CFS (500 mm ≤
The shear modes (𝒅 𝑆 = 𝑯 𝑆 𝒂) under a given loading can then be 𝐿 ≤ 3000 mm), which in turn led to lower modal contributions
obtained by constraining the solutions of Eq. (1) to a linear combination from the global instabilities. This is illustrated by the typical modal
of the basis vectors of the shear space, contained in 𝑯 𝑆 = 𝑲 −1 𝑭 𝑆 . The decomposition for a compressed lipped channel element shown in
number of shear modes is equal to the total number of nodes N. Fig. 6.

5
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Table 1
Cross-sectional shapes and dimensions considered in the dataset (in mm).

Fig. 7 present the correlations between the input parameters and The ANN procedure can mathematically be expressed as follows:
(i) the elastic critical buckling loads (𝑃𝑐𝑟 ), and (ii) the corresponding ( )
∑ 𝑗 𝑗−1
modal contributions (𝑐𝐿,𝐶 , 𝑐𝐷,𝐶 , 𝑐𝐺,𝐶 ). Fig. 8 explores this correlation 𝑎𝑗𝑖 = 𝑓 𝑗 𝑤𝑖𝑘 𝑎𝑘 + 𝑏𝑗𝑖 (18)
for the elastic critical buckling moments (𝑀𝑐𝑟 ) and their corresponding 𝑘
modal contributions (𝑐𝐿,𝐹 , 𝑐𝐷,𝐹 , 𝑐𝐺,𝐹 ). Based on the coefficients of In the above equation, the 𝑎 values are called the activations, and
determination (𝑅2 ) generally poor correlations are observed between 𝑓 is the activation function which decides whether a neuron should
the input and output parameters, especially when it comes to the be activated or not within the network (see Section 4.4). The 𝑤 and
participation factors. An exception can be found for the plate thickness 𝑏 parameters represent the weights and biases of the ANN model,
(𝑡) where considerably higher correlations exist with 𝑃𝑐𝑟 (𝑅2 = 0.853) respectively. The subscripts 𝑗 and 𝑖 denote the 𝑗 𝑡ℎ layer and 𝑖𝑡ℎ node,
and 𝑀𝑐𝑟 (𝑅2 = 0.749). This poor correlation points to machine learning while 𝑘 represents the number of nodes in the (𝑗 − 1)𝑡ℎ layer which
as a potentially more viable avenue to obtain reliable predictions. are connected to the 𝑖𝑡ℎ node in the 𝑗 𝑡ℎ layer. In the input layer, each
𝑎1𝑖 value is essentially equal to the 𝑖𝑡ℎ input parameter (𝑥𝑖 ). In the
4. Machine learning methods output layer, where 𝑗 = 𝑛 (𝑛 is the total number of layers in the
network), the calculated activation values (𝑎𝑗𝑖 ) constitute the predicted
output values (𝑦′ ). The weights and biases are initially assumed at the
4.1. Overview of Artificial Neural Networks (ANN)
beginning of the training process and then learned by the network using
an algorithm based on the ‘‘backward propagation of errors’’ [36].
A particular kind of machine learning, the feedforward multilayer
Following the calculation of the activations (𝑎𝑗𝑖 ) for all layers of the
ANN, was employed in this study. The main advantage of the ANN
network, a cost function 𝐽 (𝑦, 𝑦′ ) is determined for the model based on
approach is that the training process is carried out on a collection
the original output data (𝑦) and the predicted output values (𝑦′ ) of all
of representative examples without requiring a well-defined process training samples. This cost function can take on various forms, such as
to algorithmically convert the input to the output data. The ANN the Mean Absolute Error (MAE), the Mean Absolute Percentage Error
is inspired by the biological structure of the human brain and has (MAPE) and the Mean Squared Error (MSE). The training process of an
a parallel-distributed architecture with a number of interconnected ANN is aimed at searching for those values of the weights and biases
nodes, commonly referred to as neurons. The neurons are arranged in which minimize the cost function. In this study the Gradient Descent
input, hidden and output layers, and each neuron is connected to all of method [37] was used for this purpose, where the following gradients
the neurons in the next layer via weighted connections. In the ANN are numerically calculated:
computational process, the data is first fed into the neural network
𝜕𝐽 (𝑤, 𝑏)
through the input layer which communicates with the hidden layers. 𝐺𝑤 = (19)
𝜕𝑤
Nodes in the hidden layer combine data from various neurons in the 𝜕𝐽 (𝑤, 𝑏)
input layer with appropriate weights. In the next stage, these weighted 𝐺𝑏 = (20)
𝜕𝑏
inputs are summed up for each neuron of the hidden layer and then
The weights and biases are then updated in each iteration (t ) using
passed to the next layer through an activation function, along with a
the following equations:
bias. This process is continued until the last hidden layer is reached,
which is linked to the output layer where the outputs are retrieved. It 𝑤𝑡 = 𝑤𝑡−1 − 𝛼𝐺𝑤,𝑡−1 (21)
should be noted that the connections between the nodes do not form
𝑏𝑡 = 𝑏𝑡−1 − 𝛼𝐺𝑏,𝑡−1 (22)
closed loops and information flows unidirectionally within the network.
The calculation process of an ANN is schematically represented in where 𝛼 is the chosen learning rate, most often in the range between
Fig. 9. 0.0 and 1.0. The influence of 𝛼 on the performance of the model

6
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 5. Histograms of: (a–g) input parameters, (h–i) cross-sectional slenderness and (j–q) output parameters.

was investigated in Section 5.1. It is noted that this ANN process has In traditional ANN methods the available dataset is divided into
previously been successfully implemented in several AI-related studies training, validation and test sets to avoid instances of overfitting. How-
with deep-learning frameworks [30,38,39]. ever, this can noticeably reduce the number of data points available
for training and make the model greatly dependent on the selection
4.2. K-fold cross-validation of the samples into training, validation and test sets. In the K-fold
cross-validation approach, on the other hand, the dataset is randomly
K-fold cross-validation is mainly employed in applied machine shuffled and then divided into K folds. K−1 folds are used to train
learning to estimate the accuracy of the model for unseen data. While the model and the remaining fold (the test set) is employed for the
implementing K-fold cross-validation into the network can be com- evaluation. In a permutative approach, each of the K folds is then
putationally expensive, it can provide significantly more information used, in turn, as the test set, as illustrated in Fig. 10. This implies
about the performance of the model and consequently help to increase that each sample is given an opportunity to be used in the test set
its accuracy. In addition, the K-fold cross-validation technique can be once and is used to train the model K−1 times. In each iteration, an
especially beneficial when the network is required to predict multiple evaluation score is calculated and retained. The performance of the
outputs (as is here the case when predicting the modal contributions in model is then taken as the average of the evaluation scores. In this
the buckled shapes). study, the network was subjected to 5-fold cross-validation, as shown

7
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

of (i) a single value representing the elastic critical buckling load

(𝑃𝑐𝑟 ) or bending moment (𝑀𝑐𝑟 ), and (ii) a vector containing the three
participation factors of the critical compressive (𝑐𝐿,𝐶 , 𝑐𝐷,𝐶 , 𝑐𝐺,𝐶 ) or
flexural (𝑐𝐿,𝐹 , 𝑐𝐷,𝐹 , 𝑐𝐺,𝐹 ) buckled shapes. To improve the performance
of the machine learning algorithms, the dataset was preprocessed. In
particular, the input parameters of the dataset were standardized using
the following equation:
𝑥−𝜇
𝑥̂ = (23)
𝜎
where 𝑥̂ and 𝑥 are the standardized and original values of the input
parameter, and 𝜇 and 𝜎 denote the mean and the standard deviation
of 𝑥. In addition, it was observed that the histograms of the output
data exhibited skewed distributions. This was especially evident for the
modal contributions. Therefore, the logarithmic values of the output
data were used to provide a more uniform distribution:

𝑦̂ = log(1 + 𝑦) (24)

where 𝑦̂ and 𝑦 are the logarithmic and original values of the output
Fig. 6. Typical modal decomposition of FSM output for a compressed lipped channel data, respectively.
element.

4.3.2. Performance metrics

Machine learning algorithms can be evaluated using various perfor-
in Fig. 10. In each iteration the dataset was divided into training and
mance metrics. In this study, the ANNs were evaluated based on the
test sets in an 80%–20% proportion.
values of the MSE (mean squared error), the MAPE (mean absolute
4.3. Data preparation percentage error) and 𝑅2 (coefficient of determination), defined as:

1∑
𝑛

4.3.1. Feature standardization and output transformation 𝑀𝑆𝐸 = (𝑦 − 𝑦′ )2 (25)

𝑛 𝑖=1
As previously discussed in Section 3, the input parameters (𝑥)
100 ∑ || 𝑦 − 𝑦′ ||
𝑛
in this study consisted of the seven geometric features of the CFS
𝑀𝐴𝑃 𝐸 = (%) (26)
element (ℎ, 𝑏, 𝑐, 𝑡, 𝑟1 , 𝑟2 , 𝐿), while the output parameters (𝑦) consisted 𝑛 𝑖=1 || 𝑦 ||

Fig. 7. Correlations between the input parameters and (i) the elastic critical buckling load (𝑃𝑐𝑟 ), and (ii) the modal contributions for compressive elements (L: Local, D: Distortional
and G: Global).

8
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 8. Correlations between the input parameters and (i) the elastic critical buckling moment (𝑀𝑐𝑟 ) and (ii) the modal contributions in flexure (L: Local, D: Distortional and G:
Global).

Fig. 9. Calculation process of ANN.

9
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 10. 5-fold cross-validation.

Fig. 11. Architecture of the ANNs with one, two and three hidden layers for the predictions of (a) elastic buckling resistances and (b) modal contributions.

∑𝑛 set before training, and include the learning rate, the number of neu-
(𝑦 − 𝑦′ )2
𝑅2 = 1 − ∑𝑖=1
𝑛 (27) rons, the number of layers, the activation functions and the optimizer.
2
𝑖=1 (𝑦 − 𝑦) In this study, the optimal hyperparameters were determined using a
where 𝑛 is the number of samples and 𝑦 is the mean value of the well-known hyperparameter tuning method called ‘grid search’ [40],
𝑦-values. which is based on a trial-and-error process. The following ranges of
hyperparameters were studied:

4.4. Tuning of hyperparameters • One, two and three hidden layers were examined, as shown in
Fig. 11.
The performance and accuracy of the ANN model highly depends • The number of neurons was varied from 10 to 100 in intervals of
on the network parameters (the so-called ‘hyperparameters’) which are 11.

10
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 12. Activation functions.

• The learning rate was set to 0.1, 0.2 and 0.3. respectively, was considered, and the number of neurons and the
• A wide variety of activation functions are available for imple- learning rates were varied in order to study their influence. As shown
mentation into the ANN. In this study, four different activation in Fig. 13, the cost function converged at approximately 40 neurons
functions were examined, including linear (Lin), rectified linear when the elastic critical buckling load (𝑃𝑐𝑟 ) or bending moment (𝑀𝑐𝑟 )
(Rel), log-sigmoid (Sig) and hyperbolic tangent-sigmoid (Tan). was predicted, while convergence was achieved at about 50 neurons
These functions are illustrated in Fig. 12. While all four activation when modal decomposition was the target of the ANN model. A further
functions were examined for the hidden layers, only the Sig and increase in the number of neurons resulted in negligible changes in
the Tan activation functions were studied in the nodes of the the cost functions. It can also be seen from Fig. 13 that the value of
output layer. Using the Sig function can be an appropriate option the learning rate generally had a negligible effect on the performance
in terms of computational costs and accuracy where output values of the network. Using either MAPE or MSE as the performance metric
should always be positive (Fig. 12). The Tan function, on the for the models also did not result in any substantial difference in the
other hand, has a steeper slope compared to the other activation convergence rate or model performance.
functions and therefore results in higher updates in the weights Fig. 14 illustrates the influence of the activation functions and the
and a potential for faster learning. number of hidden layers on the overall performance of ANN models
• As discussed in Section 4.1, a cost function can be used to repre- with 40 neurons and a learning rate of 0.3. The labels in Fig. 14
sent the performance of the machine learning algorithm. In this (e.g. ‘Lin-Sig’) indicate the type of activation function used in the
study, the MSE and the MAPE cost functions were employed to hidden layers, followed by the type of activation function in the output
assess the accuracy of the ANN models. layer. It can be seen that the types of activation function significantly
affected the performance of the models, as measured by the MSE and
Following a sensitivity analysis, the number of epochs was set to
MAPE. Using the Sig activation function in the output layer always
1000 in all ANN models. This represents the number of passes of the
resulted in ANN models with significant remaining errors. It was also
entire training dataset that the machine learning algorithm completes
concluded that the ANN models with the Lin activation function within
in order to update the biases and weights. In addition, the default
the input layers were incapable of providing accurate solutions. On the
weight and bias initializers were implemented in all ANN models using
other hand, the best performance was achieved when the Tan activation
the ‘‘Glorot’’ and ‘‘Zeros’’ functions, respectively [41].
function was utilized in both the hidden and the output layers. The
5. Results and discussions results also demonstrate that training the ANN models with one hidden
layer may lead to relatively high error levels compared to those with
In this study, the ANN models were developed and tuned in the two and three hidden layers. However, using a larger number of hidden
MATLAB software package [42]. In total, 1440 different ANN models layers considerably increases the computational cost of the training
were evaluated for each prediction target in order to tune the hy- process.
perparameters. Based on these results the most efficient models with
the highest performance were identified. The relative importance of 5.2. Selected networks
the various input parameters on the predicted output values was also
assessed using the Shapley (SHAP) method [43]. Following the tuning of the hyperparameters, the ANN model with
the best performance was selected for each prediction target. A sum-
5.1. Effects of hyperparameters mary of the selected ANN models is provided in Table 2, including
the features of the hyperparameters and the model performance in
In a first step, an ANN model with one hidden layer, using the terms of MSE and MAPE. While the number of hidden layers was set
Sig and Tan activation functions for the hidden and output layers, to two for the predictions of the elastic critical buckling loads (𝑃𝑐𝑟 )

11
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 13. Influence of number of neurons and learning rate on cost functions.

and bending moments (𝑀𝑐𝑟 ), the selected ANN models for the modal respectively). Figs. 15 to 18 explore the relationship between the
decomposition of the compressive (𝑐𝐿,𝐶 , 𝑐𝐷,𝐶 , 𝑐𝐺,𝐶 ) and flexural (𝑐𝐿,𝐹 , network predictions and the actual responses obtained from the FSM
𝑐𝐷,𝐹 , 𝑐𝐺,𝐹 ) buckled shapes contained three hidden layers. The optimum and ENFM by means of a linear regression analysis. The statistical
number of neurons was found to be 40 and 50 for the ANN models indicators for the ratios of the ANN predictions to the actual responses
predicting the elastic buckling resistance and the modal decomposition, are also presented in these figures, including the coefficient of deter-
respectively. For all selected ANN models the learning rate was 0.3, and mination (𝑅2 ), the coefficient of variation (COV) and the mean. For
the Tan activation function was implemented for both the hidden and a perfect fit, all data should fall along a 45◦ line, as the network
the output layers. The MAPE was used as the cost function during the outputs would be equal to the actual responses. It should be noted
training process. that for each prediction problem the model was independently re-
In general, the elastic critical buckling loads and bending moments trained about 5 times to obtain the most accurate results. This was
were more accurately predicted by the ANN models (with MAPEs of done because each training starts from different initial weights and
2.75% and 2.98%) than the modal decompositions of the compressive biases and this randomized aspect can produce networks of slightly
and flexural buckled shapes (with MAPEs of 19.93% and 28.05%, different performance. The best performing model was retained. An

12
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 14. Influence of activation functions and number of hidden layers on cost functions.

excellent fit was obtained for the elastic critical buckling loads (𝑃𝑐𝑟 ) slightly inaccurate predictions of the global participation factors in
and bending moments (𝑀𝑐𝑟 ), with 𝑅2 > 0.99. On the other hand, the bending (Fig. 18). Finally, it should be noted that ANN models are
predictions for the modal decomposition of the compressive (𝑐𝐿,𝐶 , 𝑐𝐷,𝐶 , generally expected to be less accurate for data outside the ranges of
𝑐𝐺,𝐶 ) and flexural (𝑐𝐿,𝐹 , 𝑐𝐷,𝐹 , 𝑐𝐺,𝐹 ) buckled shapes showed a slightly the selected training dataset.
lower level of accuracy, with 𝑅2 > 0.95. This is attributed to the fact
that predicting the modal decomposition is a multi-output regression 5.3. Impact of input parameters on critical buckling resistance and modal
problem, while only a single-output ANN model is required for the decomposition
elastic buckling resistances. In addition, as shown in the histograms of
the output data (Fig. 5), the distributions of the modal contribution The Shapley value is based on a solution concept in game theory,
results were sometimes skewed, which in turn resulted in less accurate which is used to determine the contribution of each player in a coalition
predictions. This was especially evident for the contributions of the or a cooperative game [44]. The Shapley value represents the relative
global (i.e. lateral–torsional) buckling mode in the critical buckled importance or influence of a specific feature (i.e. an input parameter)
shapes of the flexural elements, which were less than 10% across on the model predictions (𝐹 ). To compute the Shapley value associated
the whole training dataset (Fig. 5q). This was consciously accepted, with the 𝑖𝑡ℎ feature, this feature is first excluded from the feature set
since CFS beam elements are typically used in applications where they (𝑥), and predictions are obtained for all possible subsets S of features
are laterally supported by floor/roof diaphragms and consequently that can be formed: 𝑆 ⊆ 𝑥∖ {𝑖}. If we indicate the prediction of a model
restrained against lateral–torsional buckling. However, it did result in with a subset S of features by 𝐹𝑆 , and the prediction of another model

13
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Table 2
Summary of the selected ANN models for buckling behaviour of CFS elements.
Element type Target Hyperparameter features Performance metrics
No. hidden layers No. neurons Activation function Learning rate Cost function MSE MAPE (%)
Compressive 𝑃𝑐𝑟 2 40 Tan 0.3 MAPE 1.19 (kN)2 2.75
𝐶𝐿,𝑐 , 𝐶𝐷,𝑐 , 𝐶𝐺,𝑐 3 50 Tan 0.3 MAPE 23.77 19.93
Flexural 𝑀𝑐𝑟 2 40 Tan 0.3 MAPE 10.15 (kN m)2 2.98
𝐶𝐿,𝑓 , 𝐶𝐷,𝑓 , 𝐶𝐺,𝑓 3 50 Tan 0.3 MAPE 42.54 28.05

Fig. 15. Performance of the selected ANN model for the prediction of the elastic critical Fig. 17. Performance of the selected ANN model for the prediction of the elastic critical
buckling load (𝑃𝑐𝑟 ). buckling moment (𝑀𝑐𝑟 ).

Fig. 16. Performance of the selected ANN model for the prediction of the modal Fig. 18. Performance of the selected ANN model for the prediction of the modal
contributions in the critical buckled shape of thin-walled compressive members (𝑐𝐿,𝐶 , contributions in the critical buckled shape of thin-walled flexural members (𝑐𝐿,𝐹 , 𝑐𝐷,𝐹 ,
𝑐𝐷,𝐶 , 𝑐𝐺,𝐶 ). 𝑐𝐺,𝐹 ).

with the 𝑖𝑡ℎ feature added to this subset as 𝐹𝑆 ⋃{𝑖} , then the marginal 6. Summary and conclusions
( )
contribution of the 𝑖𝑡ℎ feature can be quantified as: [𝐹𝑆 ⋃{𝑖} 𝑥𝑆 ⋃{𝑖} −
( ) Machine learning algorithms were developed using Artificial Neu-
𝐹𝑆 𝑥𝑆 ] (where 𝑥𝑆 represents the input features in the subset 𝑆, where
ral Networks (ANN) to predict the elastic critical buckling loads and
the 𝑖𝑡ℎ feature is absent). The Shapley value is then computed using a
bending moments of thin-walled structural elements, as well as the con-
weighted average over all possibilities: tributions of the various ‘pure’ buckling modes in the critical buckled
∑ |𝑆|! (|𝑥| − |𝑆| − 1)! ( ) ( ) shape. The selected dataset consisted of 4608 samples of thin-walled
𝜙𝑖 = [𝐹𝑆 ⋃{𝑖} 𝑥𝑆 ⋃{𝑖} − 𝐹𝑆 𝑥𝑆 ] (28)
𝑆⊆𝑥∖{𝑖}
|𝑥|! channels, considering various cross-sectional geometries, locations of
intermediate stiffeners and element lengths as the input data. The
In the above equation, |𝑆| and |𝑥| represent the number of elements output data for training, i.e. the elastic critical buckling loads and
in the subset 𝑆 and the total number of features, respectively. The the modal contributions in the buckled shapes, were prepared based
procedure is supported in MATLAB by the ‘shapley’ function. on the results of the Finite Strip Method (FSM) and the Equivalent
Fig. 19 shows the mean absolute Shapley values calculated over the Nodal Force Method (ENFM). The study showed that the tuning of
whole data set. It can be seen that the element thickness (𝑡) always has the hyperparameters significantly affects the performance of the ANN
the highest influence on the buckling behaviour of the CFS members models. However, properly tuned ANN models for the prediction of the
compared to other input parameters. Other influential input parameters elastic buckling loads and the modal decomposition demonstrated a
were the cross-sectional dimensions (ℎ, 𝑏, 𝑐), which had an important high level of accuracy with coefficients of determination of over 0.99
effect on the elastic critical buckling load/moment and the local and and 0.95, respectively, revealing machine learning as a viable solu-
distortional modal contributions, while the length of the element (𝐿) tion technique for this highly non-linear problem. The slightly lower
logically had the second-highest impact on the contributions of the accuracy of the ANN models in predicting modal decompositions was
global modes. partially attributed to the skewed distribution of the global buckling

14
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

Fig. 19. Importance of the input parameters in the prediction of output values.

mode participations within the dataset, which itself was a consequence References
of only considering practical lengths. In addition, it was shown that
the prediction of modal contributions, which is a multi-regression [1] D.T. Phan, S.M. Mojtabaei, I. Hajirasouliha, J. Ye, J.B.P. Lim, Coupled element
problem, is more sensitive to the size and quality of the training dataset and structural level optimisation framework for cold-formed steel frames, J.
Construct. Steel Res. (2019) 105867.
compared to single-regression problems. The above stated accuracies
[2] S.M. Mojtabaei, J. Becque, I. Hajirasouliha, Local buckling in cold-formed steel
can be expected within the practical ranges of geometric parameters moment-resisting bolted connections: Behavior, capacity, and design, J. Struct.
listed in Table 1. Eng. 146 (2020) 04020167.
[3] T. von Karman, E.E. Sechler, L.H. Donnell, The strength of thin plates in
CRediT authorship contribution statement compression, ASME Appl. Mech. Trans. 54 (1932) 53–70.
[4] CEN, Eurocode 3: Design of Steel Structures, Part 1.3: General Rules—
Seyed Mohammad Mojtabaei: Writing – original draft, Validation, Supplementary Rules for Cold Formed Members and Sheeting, European Comittee
Software, Methodology, Investigation, Formal analysis, Data curation, for Standardization, Brussels, 2005.
[5] AISI S100- 16, North American Specification for the Design of Cold-Formed Steel
Conceptualization. Jurgen Becque: Writing – review & editing, Val-
Structural Members, American Iron and Steel Institute (AISI), Washington, DC,
idation, Methodology, Conceptualization. Iman Hajirasouliha: Vali- USA, 2016.
dation, Supervision, Resources, Project administration, Conceptualiza- [6] B.W. Schafer, T. Peköz, Computational modeling of cold-formed steel: character-
tion. Rasoul Khandan: Supervision, Resources, Project administration, izing geometric imperfections and residual stresses, J. Construct. Steel Res. 47
Conceptualization. (1998) 193–210.
[7] B.W. Schafer, Review: The Direct Strength Method of cold-formed steel member
Declaration of competing interest design, J. Construct. Steel Res. 64 (2008) 766–778.
[8] S. Ádány, B.W. Schafer, Buckling mode decomposition of single-branched open
cross-section members via finite strip method: Derivation, Thin-Walled Struct. 44
The authors declare that they have no known competing finan- (2006) 563–584.
cial interests or personal relationships that could have appeared to [9] V. Karakonstantis, J. Becque, Modal classification and decomposition of instabil-
influence the work reported in this paper. ities in thin-walled structural members, in: J.K. Paik, S.J. Kim, C.H. Song (Eds.),
Proceedings of the 7th International Conference on Thin-Walled Structures,
Data availability The Korea Ship and Offshore Research Institute, Pusan National University,
South-Korea, 2014.
[10] J. Becque, X. Li, A new approach to modal decomposition of buckled shapes,
Data will be made available on request.
Structures 4 (2015) 2–12.
[11] J. Becque, X. Li, B. Davison, Modal decomposition of coupled instabilities: The
Acknowledgments method of the equivalent nodal forces, Thin-Walled Struct. 143 (2019) 106229.
[12] Y. Xu, M. Zhang, B. Zheng, Design of cold-formed stainless steel circular
The authors would like to thank the EPSRC for their financial hollow section columns using machine learning methods, Structures 33 (2021)
support for this work via the EPSRC Fellowship Award. 2755–2770.

15
S.M. Mojtabaei, J. Becque, I. Hajirasouliha et al. Thin-Walled Structures 184 (2023) 110518

[13] M. Zarringol, H.-T. Thai, M.Z. Naser, Application of machine learning models [29] V.V. Degtyarev, M.Z. Naser, Boosting machines for predicting shear strength of
for designing CFCFST columns, J. Construct. Steel Res. 185 (2021) 106856. CFS channels with staggered web perforations, Structures 34 (2021) 3391–3403.
[14] T.M. Mitchell, Machine Learning, McGraw-Hill, New York, 2010. [30] Z. Fang, K. Roy, J. Mares, C.-W. Sham, B. Chen, J.B.P. Lim, Deep learning-based
[15] J. Ghaisari, H. Jannesari, M. Vatani, Artificial neural network predictors for axial capacity prediction for cold-formed steel channel sections using Deep Belief
mechanical properties of cold rolling products, Adv. Eng. Softw. 45 (2012) Network, Structures 33 (2021) 2792–2802.
91–99. [31] Z. Fang, K. Roy, B. Chen, C.-W. Sham, I. Hajirasouliha, J.B.P. Lim, Deep learning-
[16] A. Brahme, M. Winning, D. Raabe, Prediction of cold rolling texture of steels based procedure for structural design of cold-formed steel channel sections
using an Artificial Neural Network, Comput. Mater. Sci. 46 (2009) 800–804. with edge-stiffened and un-stiffened holes under axial compression, Thin-Walled
[17] M.R. Jamli, N.M. Farid, The sustainability of neural network applications within Struct. 166 (2021) 108076.
finite element analysis in sheet metal forming: A review, Measurement 138 [32] Y. Xu, B. Zheng, M. Zhang, Capacity prediction of cold-formed stainless steel
(2019) 446–460. tubular columns using machine learning methods, J. Construct. Steel Res. 182
[18] Y. Wang, C. Li, L. Peng, R. An, X. Jin, Application of convolutional neural
(2021) 106682.
networks for prediction of strip flatness in tandem cold rolling process, J. Manuf.
[33] C. Couto, Neural network models for the critical bending moment of uniform
Process. 68 (2021) 512–522.
and tapered beams, Structures 41 (2022) 1746–1762.
[19] E.M.A. El-Kassas, R.I. Mackie, A.I. El-Sheikh, Using neural networks in
[34] Y.K. Cheung, L.G. Tham, Finite Strip Method, CRC Press, Boca Raton, USA, 1998.
cold-formed steel design, Comput. Struct. 79 (2001) 1687–1696.
[35] BW Industries, Carnaby Industrial Estate, Lancaster Rd, Carnaby, Bridlington, UK,
[20] BS5950, British Standards Institution, Structural Use of Steelwork in Building,
YO15 3QY. http://www.bw-industries.co.uk.
Part 5: Code of Practice for Design of Cold-Formed Sections, BSI, London, UK,
[36] S.E. Dreyfus, Artificial neural networks, back propagation, and the Kelley-Bryson
1987.
[21] E.M.A. El-Kassas, R.I. Mackie, A.I. El-Sheikh, Using neural networks to predict gradient procedure, J. Guid. Control Dyn. 13 (1990) 926–928.
the design load of cold-formed steel compression members, Adv. Eng. Softw. 33 [37] S. Ruder, An overview of gradient descent optimization algorithms, 2017, arXiv:
(2002) 713–719. 1609.04747.
[22] M. Pala, A new formulation for distortional buckling stress in cold-formed steel [38] A. Parvat, J. Chavan, S. Kadam, S. Dev, V. Pathak, A survey of deep-learning
members, J. Construct. Steel Res. 62 (2006) 716–722. frame-works, in: International Conference on Inventive Systems and Control,
[23] M. Pala, N. Caglar, A parametric study for distortional buckling stress on ICISC, 2017, pp. 1–7, arXiv:1609.04747.
cold-formed steel using a neural network, J. Construct. Steel Res. 63 (2007) [39] Z. Fang, K. Roy, Q. Ma, A. Uzzaman, J.B.P. Lim, Application of deep learning
686–691. method in web crippling strength prediction of cold-formed stainless steel
[24] I.H. Guzelbey, A. Cevik, A. Erklig, Prediction of web crippling strength of cold- channel sections under end-two-flange loading, Structures 33 (2021) 2903–2942.
formed steel sheetings using neural networks, J. Construct. Steel Res. 62 (2006) [40] P. Liashchynskyi, P. Liashchynskyi, Grid search, random search, genetic
962–973. algorithm: a big comparison for NAS, 2019, arXiv:1912.06059.
[25] J.L.R. Dias, N. Silvestre, A neural network based closed-form solution for the [41] X. Glorot, Y. Bengio, Understanding the difficulty of training deep feedforward
distortional buckling of elliptical tubes, Eng. Struct. 33 (2011) 2015–2024. neural networks, in: Proceedings of the 13th International Conference on
[26] S. Tohidi, Y. Sharifi, Neural networks for inelastic distortional buckling capacity Artificial Intelligence and Statistics (AISTATS), Vol. 9, Sardinia, Italy, 2010,
assessment of steel I-beams, Thin-Walled Struct. 94 (2015) 359–371. JMLR: W & CP 9.
[27] M. D’Aniello, E.M. Güneyisi, R. Landolfo, K. Mermerdaş, Analytical prediction of [42] Mathworks, Matlab R2021b, Mathworks Inc, 2021.
available rotation capacity of cold-formed rectangular and square hollow section [43] S.M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions,
beams, Thin-Walled Struct. 77 (2014) 141–152. Adv. Neural Inf. Process. Syst. (2017) 4765–4774.
[28] V.V. Degtyarev, Neural networks for predicting shear strength of CFS channels [44] E. Štrumbelj, I. Kononenko, Explaining prediction models and individual
with slotted webs, J. Construct. Steel Res. 177 (2021) 106443. predictions with feature contributions, Knowl. Inf. Syst. 41 (2014) 647–665.

16