See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/371331016

Use of AI and ML Algorithms in Developing Closed-Form Formulae for

Structural Engineering Design

Chapter · March 2023

DOI: 10.4018/978-1-6684-5643-9.ch004

CITATIONS READS
0 73

3 authors:

George Markou Nikolaos Bakas

University of Pretoria GRNET
98 PUBLICATIONS 665 CITATIONS 102 PUBLICATIONS 432 CITATIONS

SEE PROFILE SEE PROFILE

Ashley van der Westhuizen

University of Pretoria
5 PUBLICATIONS 6 CITATIONS

SEE PROFILE

All content following this page was uploaded by George Markou on 18 June 2023.

The user has requested enhancement of the downloaded file.

 Use of AI and ML Algorithms in
developing closed-form formulae, for
Structural Engineering Design:
George Markou Nikolaos Bakas
University of Pretoria, South Africa RDC Informatics, Greece
Ashley Megan van der Westhuizen
University of Pretoria, South Africa

ABSTRACT
The design and analysis of structures is performed with the use of national and international design codes
that usually suggest the use of semi-empirical formulae. Often the formulae are oversimplified, in some
cases are not available to engineers or are time-consuming and challenging to implement. The objective of
this chapter is to demonstrate the use of artificial intelligence and machine learning to develop more
accurate formulae for different types of applications related to structural design. The applications that are
discussed in this work include predicting the shear capacity of reinforced concrete slender and deep beams
without stirrups, calculating the fundamental period of reinforced concrete and steel structures, and
predicting the deflection of horizontally curved steel I-beams.

Keywords: Shear Strength, Fundamental Period, Curved Steel I-Beams, Reinforced Concrete, Steel
Structures, Artificial Intelligence, Machine Learning, Design Formulae

INTRODUCTION
The use of artificial intelligence (AI) and machine learning (ML) in structural analysis and design problems
has been increasing due to the methods’ abilities to handle complex nonlinear structural systems under
extreme actions (Markou & Bakas, 2021a, 2021b; Thai, 2022). The purpose of the chapter is to discuss the
use of AI and ML in structural engineering design based on current work performed by the authors and
analyse the future perspectives related to this technology. The applications that are presented herein include
the development of predictive formulae for the calculation of the shear capacity of slender reinforced
concrete (RC) beams without stirrups, predicting the shear capacity of deep beams without stirrups
reinforced with fibre reinforced polymer (FRP) rebars, determining the fundamental period of RC as well
as steel structures, and determining the deflection of curved steel I-beams. In some cases, the formulae
currently available for the determination of the above structural problems are oversimplified or are not
readily available to practising engineers (Ababu et al., 2022), therefore, this is discussed in the relevant
sections of this chapter that follow.
It is the objective of this chapter to show the reader the development of closed mathematical form design
formulae that can predict the mechanical behaviour of structures using an out-of-the-box approach (Bakas
& Markou, 2019), while artificial neural networks (ANN) and other ML methods were adopted. According
to the proposed approach, by using finite element software and nonlinear analysis through the use of
ReConAn FEA (2020) software, datasets are developed that are used for the development of predictive
models. This approach aims to marry state-of-the-art 3D detailed modelling with AI and ML algorithms for
the development of predictive models that exhibit a more accurate calculation of the mechanical response
of structural members and structures. Furthermore, given that AI and ML algorithms are an effective and
efficient tools used to predict analysis outputs of computationally demanding engineering problems, their
use by the civil engineering community has emerged over the last decade and increases exponentially
(Bakas et al., 2021).
A dataset is obtained from the results of the analysis of these models and is then used to train AI and
ML algorithms that can develop predictive models. One of the main obstacles in training and testing models
is the lack of datasets that refer to a specific structural problem. This is attributed to the limitations in
relation to performing a large number of experiments for any structural problem, where a sufficient number
of output data will be made available for the training of our AI and ML models (van der Westhuizen et al.,
2022). For this reason, the proposed approach foresees the substitution of the actual experiment through the
use of 3D detailed finite element analysis, hence, to be able to proceed with this approach an extensive
validation of the adopted finite element algorithm has to be performed. This is also going to be presented
by referencing the research work performed towards validating the research finite element analysis (FEA)
software ReConAn FEA (2020) that was used for the development of the datasets presented in this work.
Additionally, it is important to note that the proposed approach foresees, where feasible, the validation of
the predictive models through the use of additional datasets that derived from physical experiments and
were found in the international literature.
For the case where physical experiments are not available, the accuracy of the proposed predictive
models is validated using out-of-sample data, while for all the structural problems presented herein the
mean absolute percentage error (MAPE) and other error parameters are determined and compared to that
of the design formulae currently available in the international literature. As is going to be shown in this
chapter, the predictive models derived by using AI or ML algorithms proved to outperform design code
formulae currently used to determine the mechanical properties and the strength of structures or their
members. All studies presented in the chapter show strong evidence that a combination of FEA and ML
methods provides reliable predictions that can be used in improving design code formulae currently used
to calculate the shear capacity of slender and deep beams, determining the fundamental period of RC and
steel structures as well as determining the deflection of curved steel I-beams.
The ML algorithms used in the studies discussed include ANN, linear regression (LR), non-linear
higher-order regression (NLR), random forests (RF) and gradient boosting (GB). It is important to note that
the ML algorithms are not a one-size-fits-all solution and different methods are used in each case to find
the most accurate predictive model (Markou & Bakas, 2021b).

BACKGROUND

Machine learning
The higher-order NLR ML algorithm used extensively in the research work presented herein can be seen
below.
Algorithm: Higher Order Regression (Gravett et al., 2021)
Input: XX (matrix of Independent Variables), YY (Vector of Dependent Variable),
nlf (number of nonlinear features to be kept in the model)
Output: Prediction Formulae
1. Create all nonlinear features* (anlf)
2. For i from 1 to nlf do
3. For j from 1 to anlf do
4. Add jth feature to the model
5. Calculate Prediction Error, MAPEj**
6. End
7. Keep in the model the jth feature which yields the minimum prediction error
8. End
Return: Prediction Formula
*with all inter-items combinations up to the 3rd degree, **Mean Absolute Percentage Error (MAPE).
 The LR and NLR models are useful because they generate a closed-form formula whereas the other
algorithms develop predictive models that are numerical. According to the numerous parametric
investigations performed by the authors, it was always found that the LR model is unable to capture the
non-linear behaviour of variables so more complex ML algorithms are required in achieving this task. For
the needs of the numerical investigations related to the accuracy of the proposed predictive models, the ML
algorithms produce the MAPE, which is used as an objective metric for measuring the abilities of the
proposed models to accurately describe the different at-hand engineering applications. It should be noted
that in each case 85% of the database was used for training purposes and 15% for testing. The formulae
obtained are then validated using out-of-sample results that were not used in the training nor in the testing
procedure, where in some cases the validation phase included solely datasets that were derived from
experimental results.

Shear capacity of reinforced concrete beams without stirrups

Shear failure in structural beam members occurs in a brittle manner and is usually attributed to
insufficient shear resistance in the web. It is important to design members, like beams, for shear resistance
to avoid these failures that provide no warning to the occupants (Spijkerman et al., 2021). Various design
codes divide beams into two categories, namely, slender beams and deep beams. This difference is
performed using the span-to-depth ratio (a/d) which is found to be directly proportional to the shear strength
of a concrete beam. In some codes, the a/d classifies a beam as slender when it is larger than 2 (Tureyen &
Frosch, 2002; El-Slayed & Benmokrane, 2005; El-Slayed et al., 2006; AlHamaydeh et al., 2022),
whereas according to Eurocode 2 the ratio a/d must be larger than 3. Unlike slender beams, deep beams
develop the so-called arch action which is a mechanism that foresees the transfer of loads to the supports
through diagonal compression struts (Yang et al., 2003). It should also be noted that the arch action
experienced by these beams as well as the use of FRP rebars within the concrete domain is not accounted
for in most current design formulae. The use of FRP rebars instead of conventional steel reinforcement is
proposed as an alternative due to their high immunity to corrosion, increasing the service life of concrete
(Abed et al., 2012; Dhahrir, 2017, 2018). It is important to investigate how the behaviour differs when
compared to steel-reinforced deep beams and these differences should be considered when developing
revised formulae (Razaqpur & Isgor, 2006; AlHamaydeh et al., 2022).
Currently, the main approach used in determining the shear capacity of RC beams without stirrups is
using formulae derived from experimental data. The use of this method, however, is limited by the
availability of experimental data that covers the vast combination of beam geometries, load cases and
boundary conditions (Markou & Bakas, 2021; Spijkerman et al., 2021). An in-depth literature review for
determining the shear capacity of RC beams can be found in AlHamaydeh et al. (2022). A summary of
the formulae that were used in the comparative study that will be discussed in this chapter can be seen
in Equations 1, 2 and 3.
ACI 318-19 𝑁𝑢 1
𝑉𝑐 = (8 ∙ 𝜆𝑠 ∙ 𝜆 ∙ (𝜌𝑤 )1/3 ∙ √𝑓′𝑐 + )𝑏 ∙ 𝑑
6𝐴𝑔 𝑤
ACI 440 5 2
𝑉𝑐 = ( ∙ 𝑘) 0.17√𝑓′𝑐 ∙ 𝑏𝑤 ∙ 𝑑
2
Eurocode 2 1 3
𝑉𝑅𝑑,𝑐 = (𝐶𝑅𝑑,𝑐 ∙ 𝑘 ∙ (𝜌1 ∙ 𝑓𝑐𝑘 )3 + 𝑘1 ∙ 𝜎𝑐𝑝 ) 𝑏𝑤 ∙ 𝑑
The ACI Equations are presented in US units. The design code Eurocode 2 mandates the use of β, an
approximate strength reduction, when designing for deep beams, which accounts for the arch effect.

Fundamental period of reinforced concrete and steel structures

In structural design it is important to derive safe structures to avoid the loss of life, especially when
structures undergo earthquake excitations, thus the understanding of the dynamic behaviour of structures is
crucial during the design process. Calculating the fundamental period of structures is one of the most
important parameters that is considered when designing seismic-resistant structures (Young, 2011; Gravett
et al., 2021; van der Westhuizen et al., 2022).
The natural frequencies are determined using modal analysis which is based on the eigenvalue problem.
Eigen-frequencies are important for engineers to understand the dynamic behaviour of their structures and
which modal shapes will contribute the most during a seismic event. There are numerous methods proposed
to solve eigenvalue problems represented by Equation 4.
Kφi = λi Mφi 4
Where:
K is the stiffness matrix of the model
M is the mass of the model
φi is the vector that contains the eigenvector of the problem
λi is the corresponding eigenvalue
The method presented herein is the subspace iteration algorithm (Bathe, 1971), which is integrated
within ReConAn FEA, (2020). Validation work performed on the ability to predict the fundamental period
of structures through the use of ReConAn FEA, (2020) can be found in Markou et al., (2018), Mourlas et
al., (2019a, 2019b, 2019c), Gravett et al., (2019), and Papadrakakis and Markou (2022). This solution
algorithm finds an orthogonal basis of vectors in EK+1 , calculating in one step the required eigenvectors
when EK+1 converges to E∞ . For k = 1,2, …, iterate from Ek to Ek+1 :
KXk+1 = Mxk 5
then, find the projections of matrices K and M onto 𝐸𝑘+1 :
K k+1 = XTk+1 KXk+1 6
Mk+1 = X Tk+1 MXk+1 7
and solve for the eigensystem of the projected matrices:
𝐾𝑘+1 𝑄𝑘+1 = 𝑀𝑘+1 𝑄𝑘+1 Λ 𝑘+1 8
Thereafter, find an improved approximation to the eigenvectors:
𝑥𝑘+1 = 𝑋𝑘+1 𝑄𝑘+1 9
and then provided that the vectors X1 are not orthogonal to one of the equilibrium eigenvectors, Λ k+1 → Λ
and Xk+1 → Φ as k → ∞.
It is important to note that the convergence of this method assumes that within the iteration procedure
the vectors in Xk+1 are ordered so the ith diagonal element in Λ k+1 will always be larger than the previous
i − 1 element, i = 2, … , p. This approach ensures that the ith column in Xk+1 converges linearly to Φi
(Bathe, 1971). A detailed modal analysis is time-consuming and requires excessive computational effort,
therefore by taking into account the soil-structure interaction (SSI) phenomenon, the problem becomes
more cumbersome and computationally demanding (Gravett et al., 2021). This further highlights the
importance of developing a predictive model that will be able to compute the fundamental period of a
structure, including the SSI effect, through a minimal computational cost (van der Westhuizen et al., 2022).
In the event of a seismic excitation, the interaction between the superstructure (building) and
substructure (soil) becomes critical as it alters the distribution of stresses and strains within the
superstructure, influencing the expected results (Mourlas et al., 2020). Current building codes use empirical
equations to predict the fundamental period of structures (Jiang et al., 2020; Taljaard et al., 2021; and
Gravett et al., 2021; van der Westhuizen et al., 2022) that do not take into account the SSI effect. Therefore,
design formulae that can be found in literature and design codes omit the use of SSI that, in some cases,
significantly influences the natural frequency of buildings due to flexibility induced by soft soils. A number
of design code formulae that are investigated in the international literature for both steel and concrete
structures can be seen in Equations 10 to 13.
𝐻 𝐻 10
NEAK (RC frames) 𝑇𝑁𝐸𝐴𝐾 = 0,09 ∙ ∙√
√𝐿 𝐻 + 𝑝 ∙ 𝐿
 0,075
Eurocode 8 and SANS10160-4 𝐶𝑇 =
√𝐴𝑐
(RC Frames with concrete or
Where: 11
masonry shear walls)
𝐿𝑤𝑖 2
𝐴𝑐 = ∑ [𝐴𝑖 (0,2 + ( ) )]
𝐻
0,75
𝑇𝐸𝐶 = 𝐶𝑇 ∙ 𝐻
Where:
Eurocode 8 and SANS10160-4 𝐶𝑡 = 0.085 for moment resistant space steel frames 12
(RC and steel frames) 𝐶𝑡 = 0.075 for RC moment-resisting frames and for
eccentrically braced frames
𝐶𝑡 = 0.05 for all other buildings
𝑇1 = 0.0724(𝐻)0.8 for steel moment-resisting frames 13
ASCE 7-05 (steel frames)
𝑇1 = 0.0731(𝐻)0.75 for eccentrically braced steel frames
As stated previously and as can be seen in the formulae described above, there is no presence of
accounting for the SSI effect, while the parameters used to compute the fundamental period practically
foresee the use of the height of the structure. The only code that takes into account the width of the structure
and the RC shear wall ratio is the new Greek earthquake-resistant code (NEAK).

Deflection of horizontally curved steel I-beams

The use of curved beams has been increasing in popularity throughout the years due to the aesthetically
pleasing designs they produce in buildings as well as in industrial applications and bridges. They can be
found in applications ranging from girders in modern highway bridges, interchange facilities, as well as
industrial buildings where they are used as crawl beams. The mechanical behaviour of curved steel I-beams
is still not well understood and currently, there are no precise formulae available to determine their
deflections nor their overall mechanical response due to different types of loads. This lack of understanding
is attributed to the mechanical parameters that affect the beams’ response when they undergo gravitational
loads. These parameters include a combination of shear, flexure and torsion deformations that generate a
complicated stress-strain development at a material level, while the lack of using a 3D material model
approach does not allow the accurate prediction of the local and global mechanical response of these
structures. The design of light structures is often governed by the serviceability considerations such as
deflection (Ababu, et al., 2022), which highlights the importance of understanding the behaviour of curved
beams and their deflections when in the design phase.
By researching the international literature, one will realise that the most widely used formula for
determining the deflection of this type of beam is Castigliano’s second theorem presented in Equation 14.
This formula is not applied directly but the deflection is derived using the equation and an integration
algorithm (Dahlberg, 2004).
𝜕𝑈 14
𝛿=
𝜕𝑃
Other formulae researchers have attempted to provide, based on Castigliano’s second theorem, for the
determination of the deflection of curved beams, which are not readily available to professional civil
engineers and can only be applied to specific load cases and support conditions. Other studies conducted to
determine the deflection do not analyse the parameters that may influence the deflection and those that did
use Castigliano’s theorem (Ababu, et al., 2022).
A few of the formulae found in Dahlberg (2004) for determining the deflections of beams with different
constraints are presented below. The first case refers to a curved cantilever beam, which is curved to form
a quarter of an ellipse. The beam is clamped at one end and a force P acts at the free end. Using Castigliano’s
theorem along with other mathematical methods, such as a numerical integration method, Equation 15 is
derived.
 𝑃𝑎3 𝑃𝑎3 15
𝛿= 𝐼1 (𝛽) + 𝐼 (𝛽)
𝐸𝐼 𝐺𝐾t 2

where the integrals 𝐼1 and 𝐼2 are functions of the parameter 𝛽(= 𝑏/𝑎) only
2 16
1
𝑥2 d𝑦/d𝑥 𝑥 1 d𝑦 2 𝑥
𝐼1 (𝛽) = ∫ 𝛽√1 − − (1 − ) √1 + ( ) d( )
0 𝑎2 2 𝑎 2 d𝑥 𝑎
√1 + ( d𝑦 ) √1 + ( d𝑦 )
( d𝑥 d𝑥 )

and
2 17
d𝑦
1
𝑥2 d𝑦/d𝑥 𝑥 d𝑥 d𝑦 2 𝑥
𝐼2 (𝛽) = ∫ 𝛽 √1 − 2 + (1 − ) √1 + ( ) d( )
0 𝑎 2 𝑎 2 d𝑥 𝑎
√1 + ( d𝑦 ) √1 + ( d𝑦 )
( d𝑥 d𝑥 )

where:
a is the length in the x-direction of one half-axis
b is the length in the y-direction of one half-axis
In Equations 15, 16 and 17 it can be seen how complex the derivation of such formulae can get. These
are the “basic” formulae, and more are derived in Dahlberg (2004) where the influence of bending, and
torsion are investigated. It is noteworthy to state at this point that the formulae become lengthier when
indeterminate beams are investigated.

PROPOSED PREDICTIVE MODELS DERIVED FROM AI AND ML ALGORITHMS

Predicting the shear capacity of reinforced concrete slender beams

The studies presented herein focused on the development of a single relationship that can predict the shear
capacity of slender RC beams without stirrups without using available experimental data. ML algorithms
are used to train and test models on a large database that, for the first time in international literature, was
generated using a non-linear FEA. It is important to state here that ReConAn FEA, (2020) and its ability to
predict the shear capacity of the RC slender beams was presented in Markou and Papadrakakis, (2013),
Markou et al., (2017), Mourlas et al., (2017) and Markou et al., (2021).
The first study presented here (Markou & Bakas, 2021a) comprises approximately 36,000 slender beams
without stirrups generated using a newly developed program called ReConAn FEA Multirun that generates
multiple input files and then uses ReConAn FEA to analyse them. This software reads the input file that
contains different material properties and reinforcement ratios, where it develops the finite element (FE)
models’ input files automatically (Markou & Bakas, 2021a).
A sensitivity analysis was conducted to investigate the effect of each independent variable while keeping
all other input variables constant. Through this approach, one can compare the results of each ML model
which should be as close as possible (Markou & Bakas, 2021a). Two variables that controlled the shear
prediction are the effective depth of the beam, d (Figure 1) and the width of the section, b (Figure 2) as they
were derived through the training procedure.
Figure 1: Sensitivity curves for d (Markou & Bakas, 2021a).

Figure 2: Sensitivity curves for b (Markou & Bakas, 2021a).

It must be noted here that in order to create a dataset that would have covered a wide range of beam
geometries, 95 different beam geometries were constructed, and a summary of the minimum and maximum
parameters considered can be seen in Table 1. The modelling of the beams used 20-noded isoparametric
hexahedral FE for the concrete, steel plates at the supports and 2-noded rod elements for the embedded
reinforcing, where the mesh used for discretizing the beams can be seen in Figure 3. The RC beam depicted
in this figure has a net span of 1,500 mm and a 200x300 mm section.
Table 1: RC slender beam geometry. Minimum and maximum magnitudes of the 95 beams (Markou &
Bakas, 2021a).

Variables Minimum Maximum

L [mm] 1 500 8 700
b [mm] 200 600
h [mm] 300 1350
L/h 5 11.6
h/b 1.0 4.5
Figure 3: Hexahedral and embedded rod element meshes (Markou & Bakas, 2021a).

The database is constituted of two groups, with the first group consisting of variables affecting the mesh
of the beams, namely the span (L), the beam width (b) and the effective depth (d). The second group
consisted of the material properties such as the uniaxial compressive strength in MPa (fc), concrete Young
modulus in MPa (Ec), concrete tensile strength ratio (ft) as a percentage of the uniaxial compressive strength,
the remaining shear capacity strength factor (β) as a percentage of the initial shear resistance of the material
point under investigation, the steel Young modulus in MPa (Es), steel yielding stress in MPa (fy) and the
tensile longitudinal reinforcement ratio (ρ) as a percentage of the concrete area of the beam’s section. Four
ML methods were utilized in the predictive modelling including LR, higher-order NLR, RF and GB.
Multiple error metrics were considered to quantify the ML method response. The metrics are presented in
detail in Dimopoulos et al., (2018).
A summary of the metrics obtained can be found in Markou & Bakas (2021a). The predictive formula
obtained using the higher-order NLR model can be seen in Equation 18.
𝑉𝑐 = (0.168 ∙ 𝑑) − (0.0196 ∙ 𝜌 ∙ 𝑏 ∙ 𝑑) − (1.5310−10 ∙ 𝐿3 ) 18
+(0.0197 ∙ 𝜌 ∙ 𝑓𝑐 ∙ 𝐿) − (809.8 ∙ 𝜌 2 ∙ 𝑑) + (1.55810−4 ∙ 𝜌 ∙ 𝐸𝑠 ∙ 𝑑)
+(5.7410−9 ∙ 𝑏 ∙ 𝐿2 ) − (9.6010−12 ∙ 𝐸𝑠 ∙ 𝐿2 ) − (6853.2 ∙ 𝜌 2 ∙ 𝑓𝑐 )
+(6.7810−3 ∙ 𝑓𝑡 ∙ 𝑓𝑐 ∙ 𝑑) − 36.172
For the validation of the proposed formula, 100 RC beams that were experimentally tested were used,
where the experimental results were obtained from Reineck et al., (2013). It is important to note that no
experimental data were used in the formulation of Equation 18 as opposed to what is usually performed and
presented in the international literature, where experimental results are used to train and test the proposed
predictive models. The 100 RC beams that were used during the validation procedure were selected
according to the geometric minimum and maximum values presented in Table 1. A comparison is made
between the RF trained model (RFTM), which was found to be the best predictive model in Markou and
Bakas (2021a), and Equation 1, which can be seen in Figure 4Error! Reference source not found..
According to the MAPE, the ML algorithm predictive model produced a MAPE of 11.6%, 36.4% lower
than the ACI 2019 formula. From the figures, it can also be seen that for larger sections (V c > 150kN) the
predictions do not favour safety, whereas for the ML model there is a better distribution about the diagonal
line. These beams are at higher risk in a failure scenario and the MAPE for these cases for the ML model
was 9.4%, 44% lower compared to the ACI formula. It was also found that within the experimental data
beams with the same geometry, reinforcement ratio and material properties derived different shear
capacities.
A study presented by Spijkerman et al., (2021) that was performed in parallel to the Markou and Bakas
(2021a), assumed a smaller number of RC beams for the training and testing procedure. In this study, a total
of 10,000 RC beams foresaw a rectangular section with a larger width compared to the height. The database
consisted of 20 geometrically different simply supported slender RC beams with minimum and maximum
values presented in Table 2.

(a) (b)
Figure 4: Experimental shear strength vs Prediction with (a) ACI (2019) and (b) Random forests
(Markou & Bakas, 2021a).

Table 2: RC slender beam geometry. Minimum and maximum magnitudes of the 20 beams (Spijkerman et
al., 2022).

Variables Minimum Maximum

L [mm] 1 500 2 700
b [mm] 300 750
h [mm] 150 350
L/h 6 10.8
h/b 0.33 0.83

Each of the 20 RC geometrically unique beams were analysed 500 times for different material and
reinforcement ratio combinations forming the 10 000 results used to train and test the predictive model.
The combinations consisted of different magnitude variations of the fc (kPa), Ec (kPa), ft, β, Es (kPa), fy
(kPa) and As (m2), similar to the study conducted by Markou & Bakas (2021a) presented above.
The formula developed, was then used to predict the shear strength of experimental data found in the
international literature. In total, three formulae were produced to investigate the robustness and accuracy,
whereas the formula presented herein in Equation 19 was the one that had the minimum MAPE (5.1%). As
it can be seen, the proposed formula has 10-features and was compared to both Eurocode 2 and ACI 318-
19.
𝑉𝑐 = (0.253939 ∙ 𝜌 ∙ 𝑑 2 ) − (0.0571682 ∙ 𝑑 ∙ 𝐿 ∙ 𝑝) + (0.0967954 ∙ 𝜌 ∙ 𝑏 ∙ 𝑑) 19
+(19.3334 ∙ 10−9 ∙ 𝐸𝑆 ∙ 𝑓𝑐 ∙ 𝑑) − (0.234107 ∙ 10−6 ∙ 𝑓𝑐 ∙ 𝐿2 ) + (0.003624 ∙ 𝜌 ∙ 𝐿2 )
−(0.00809799 ∙ 𝜌 ∙ 𝐿 ∙ 𝑏) + (5.38644 ∙ 10−6 ∙ 𝑓𝑐 ∙ 𝑏 ∙ 𝑑) + (343643 ∙ 𝜌 ∙ 𝑓𝑡 ∙ 𝛽)
−(387.761 ∙ 𝜌 2 ∙ 𝑑) + 0.950862
Figure 5: Comparison between experimental and ACI prediction (Spijkerman et al., 2022).

Figure 6: Comparison between experimental and EC2 prediction (Spijkerman et al., 2022).

Figure 7: Comparison between experimental and proposed formula (Spijkerman et al., 2022).

Figure 5, 6 and 7 contain a comparison between the experimental results and Equations 1, 3 and 19,
respectively. The results in all three figures look quite similar, however, for ultimate forces below 50 kN
there is a closer and more accurate prediction for the proposed formula. At forces larger than 50 kN the
proposed formula becomes less accurate and is more conservative. A validation database was selected from
experiments found in the international literature, where 36 beams were selected with properties within the
range of the limits presented in Table 2. The proposed formula was found to outperform the existing design
code formulae, where the MAPE of the proposed formula was 47.03%, and for the cases of the design
codes, it was found that the ACI318-19 derived a respective MAPE of 145.51% and from Eurocode 2 the
MAPE was 112.98%. This finding, which states that the proposed formula derives an almost 3 times lower
error compared to the two codes, highlights the importance of further investigating the proposed approach.

Predicting the shear capacity of FRP reinforced concrete deep beams

As it was presented in AlHamaydeh et al., (2022), the aim of that research work was to model and interpret
the behaviour of FRP-bar RC deep beams, while forming a dataset that was going to allow the development
of a predictive model that would be able to overperform the current design code formulae in computing the
shear capacity of these structural members.
A sensitivity analysis was conducted, and it was found that the shear span has an inverse relationship
with the shear strength, and this is attributed to the arch action phenomenon experienced by deep beams.
This relationship can be seen in Figure 8. Similar to the sensitivity analysis performed on slender beams,
the geometric characteristics of the beams (b and d) have a significant influence on the shear capacity. A
total of 93 models representing various possible FRP-reinforced deep beams were analysed through
ReConAn FEA, where a summary of the minimum and maximum parameters used to develop the dataset
can be seen in Table 3.

Figure 8: Sensitivity analysis for alpha (AlHamaydeh et al., 2022).

Table 3: FRP deep beam geometry. minimum and maximum magnitudes of the 93 beams (AlHamaydeh et
al., 2022)

Variables Minimum Maximum

b [mm] 200 400
Ec/fc (Concrete) 701 940
Ec/fc (FRP bars) 71 75
a/h 0.5 2
ρ [%] 0.2991 2.595
 (b)
(a)
Figure 9: RC beam (a) 20-noded hexahedral mesh and (b) embedded rebar elements (AlHamaydeh et al.,
2022).

For the needs of this research work, the modelling of the RC deep beams, 20-noded hexahedral FE were
used and the FRP reinforcement was modelled through the use of rod elements. The mesh of one of RC
deep beams can be seen in Figure 9, where the 20-noded hexahedral elements and the embedded rebar
macro-elements can be depicted. It is noteworthy to state here that the beams were tested under four-point
bending. The study also included an investigation of the effect of different factors on the shear capacity and
the crucial parameters were highlighted. Regarding the validation of the adopted software in capturing the
shear capacity of deep RC beams with FRP-bars can refer to Markou and AlHamaydeh, (2018).
After successfully constructing the database through nonlinear analysis, AlHamaydeh et al., (2022)
proposed a 5-feature formula that was generated using a higher-order stepwise regression (HOSR)
algorithm, where the resulting relationship can be seen in Equation 20.
VRk,cp = (1.2 ∙ 10−4 ∙ fc ∙ d ∙ b) − (1.1 ∙ 10−5 ∙ a2 ∙ fc ) + (6.33 ∙ 10−6 ∙ ffrp∙ fc ∙ b) 20
−(1.47 ∙ 10−6 ∙ a ∙ b ∙ d) + (9.4210−8 ∙ a2 ∙ L) + 115
The correlation between the numerically generated shear strength and those developed from the formula
can be seen in Figure 10. The dotted lines in Figure 10 represent a +/-95% confidence intervals and an
acceptable range of shear values to check the reliability and validity of Equation 20. So as to investigate
the models’ predictive capabilities, 43 experimentally tested RC deep beams were used. A comparison of
the MAPE of the predictive model as well as the various design code formulae can be seen in Table 4.

Figure 10: Correlation between numerically predicted and formula predicted shear strength
(AlHamaydeh et al., 2022).

Table 4: Comparison of MAPE of formulae (AlHamaydeh et al., 2022).

AI ACI 440 ACI 440 (Modified) Eurocode (without β) Eurocode (with β)

MAPE 57 77 83 91 34
 It can be seen in Table 4 that Eurocode 2 (using the β parameter) had the minimum MAPE among the
design codes presented. The proposed formula was also found to outperform the two design code formulae
considered in that research study, where further investigation is deemed necessary in improving the
predictive capabilities of the proposed model. It is also important to be noted that the Eurocode equation is
originally developed for steel and not FRP reinforcement, which is something that is accounted for in ACI
440.

Determining the fundamental period of RC structures

An initial study was conducted (Gravett et al., 2019) to investigate and then propose an approach that will
alleviate numerical constraints regarding performing modal analysis of large-scale models through 3D
detailed modeling. The numerical limitations are numerical instabilities that occur when the opening and
closing of cracks occur and the excessive computational demand required when dealing with small
numerical models. This issue can be solved by performing modal analyses on structures discretized with
the hybrid model (HYMOD) approach (Gravett et al., 2019) without compromising the accuracy of the
model. The HYMOD approach was adopted for the development of the structures presented herein as
described by Markou and Papadrakakis (2015). The HYMOD approach combines hexahedral and beam
column finite elements as it is depicted in Figure 11 where it is shown how the coupling between the
hexahedral and beam-column elements through kinematic constraints is performed.
Extending the Gravett et al., (2019) research work, a relatively large dataset was developed and
presented in Gravett et al., (2021) that foresaw the analysis of multistorey RC buildings initially constructed
and investigated in Markou et al. (2018). Initially, 2-, 4-, 6-, 8-, 10 and 20-storey models without and with
shear walls were created as shown in Figure 12.
Each storey height was 3 m tall, while each frame consists of 3 bays with an 8.9 m total building width.
Furthermore, two different foundation types were considered, namely a raft slab and an isolated footing so
as to decide which type of foundation to use according to the soil characteristics, since the SSI effect was
also considered in this study (Gravett et al., 2021). In order to account for the SSI effect, two soil types
were considered. The isolated footings were used in models with a medium solid (E = 700 MPa), and a raft
slab foundation was adopted for the case of soft soil (E = 65 MPa). A total of 475 numerical results were
used in this study as well as 60 out-of-sample modal results for the validation of the proposed formula.
Figure 13 shows the first 3 modes for a 2-storey as well as a 20-storey RC building.
The numerical models analysed in Gravett et al., (2019) were modified accordingly by Gravett et al.,
(2021) to include a soil domain discretized with hexahedral finite elements. A total of 475 modal results
were generated considering the minimum and maximum values in terms of geometry and soil material
properties as seen in Table 5. Figure 14 shows different models that foresee the discretization of the soil
domain through the use of hexahedral elements.

Figure 11: Hybrid model showing the interface between the elements (Markou & Papadrakakis, 2015).
 (a)

(b)

Figure 12: Initial models (a) without shear walls, (b) with shear walls (Gravette et al., 2019, 2021).

(a)
 (b)

Figure 13: First 3 modes of (a) 2-storey and (b) 20-storey building without shear walls (Gravette et al.,
2019, 2021).

Three design formulae were generated using the NLR ML algorithm (Gravett et al., 2021). It is important
to note that the formulae are, at present, only valid for the values as seen in Table 5. The 20-feature formula
as seen in Equation 21 is discussed because it derived the minimum MAPE (8.5%) and the highest
correlation (91.5%) when used on the 60 out-of-sample validation dataset.

(a) (b)
Figure 14: Soil discretized with hexahedral meshes of 10-storey RC models for (a) raft slab foundation
and (b) isolated footings (Gravett et al., 2021).
Table 5: Minimum and maximum magnitudes of the RC Frames (Gravette et al., 2019).

Variables Minimum Maximum

Soil Depth [m] 1 60
Soil E [kPa] 65 000 700 000
H [m] 6 30
L [m] 6.25 18.25
B [m] 6.25 18.25
ρ(%) 0 82.9

𝑇 = (0.032867&7 ∙ H) − (143.475 × 10−6 ⋅ 𝜌 ⋅ H) − (79.921 × 10−8 ⋅ L ⋅ B 2 ) 21

−(10.458 × 10−14 ⋅ 𝜌 ⋅ L ⋅ 𝐸𝑠 ) − (55.061 × 10−6 ⋅ 𝜌 ⋅ L) − (71.214 × 10−12 ⋅ 𝐸𝑠 ⋅ D2 )
+(5.016 × 10−6 ⋅ L ⋅ H ⋅ D) − (2.93 × 10−18 ⋅ B ⋅ 𝐸𝑠2 ) + (7.121 × 10−10 ⋅ 𝐵 ⋅ 𝐸𝑠 )
+(8.702 × 10−6 ⋅ 𝐿 ⋅ 𝐻 2 ) + (93.621 × 10−8 ⋅ 𝐿 ⋅ 𝜌 2 ) − (6.093 × 10−6 ⋅ 𝐿 ⋅ 𝐻 ⋅ 𝜌)
+(3.351 × 10−6 𝜌 ⋅ 𝐻 2 ) + (60.549 × 10−12 ⋅ 𝐵 ⋅ 𝐸𝑠 ⋅ 𝐷) + (16.982 × 10−5 ⋅ 𝐵2 )
−(64.89 × 10−6 ⋅ 𝐻 2 ) − (9.776 × 10−12 ⋅ 𝐻 ⋅ 𝐸𝑠 ⋅ 𝐷) + (1.477 × 10−17 ⋅ 𝐻 ⋅ 𝐸𝑠2 )
−(44.111 × 10−10 ⋅ 𝐻 ⋅ 𝐸𝑠 ) − (2.891 × 10−6 ⋅ 𝐻 ⋅ 𝐷 ⋅ 𝐵) + 3.8476 × 10−4
The validation dataset consisted of 60 additional modal results from buildings with geometries that differ
from those seen in Table 5 and used for the training and testing of the predictive models. The correlation
was found to be high for all three formulae, but as stated above only the 20-feature formula is discussed
herein due to its superior predictive capabilities illustrated in Figure 15. It is also noteworthy to state here
that the corresponding MAPE obtained by Eurocode 8 and other proposed formulae found in the
international literature was larger than 30% compared to the 5.68% obtained from the proposed formula.

0.6
Formulae Predicted Period (s)

0.5

0.4
y = 0.9105x + 0.031
0.3 R² = 0.9115

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6
Numerically Predicted Period (s)
Figure 15: Correlation of 20-feature formula on validation data (Taljaard et al., 2021).

Comparing the performance of the proposed fundamental period formulae, Figure 16 was developed
which compares the 20-feature formula with four other formulae proposed in Hatzigeorgiou and Kanapitsas
(2013), Chopra and Goel (2000), Amanat and Hoque (2006) and Asteris et al. (2015) . Table 6 shows the
MAPE for the different formulae as they resulted when implemented on the validation dataset of the 60
out-of-sample data. It is easy to observe that the proposed fundamental period formulae (Gravett et al.,
2021) exhibit superior accuracy when used to predict the fundamental periods of the under-study structures.
 1.6
20-Feature Formula
Goel and Chopra (2020)
1.4
Formula Predicted Period (s)

Amanat and Hoque (2006)

1.2 Hatzigeorgiou and Kanapitsas (2013), without infill walls
Hatzigeorgiou and Kanapitsas (2013), with infill walls
1 Asteris et al. (2015)

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1 1.2
Numerically Predicted Period (s)

Figure 16: Comparison between the 20-feature formula predictions and other fundamental period
formulae proposed in the international literature (adapted from Gravett et al., 2021).

Table 6: Absolute average errors for different formulae on the validation dataset (Gravett et al., 2021).
Mean Absolute
Percentage Error
3-Feature Formula 14.4%
5-Feature Formula 10.26%
20-Feature Formula 5.68%
Eurocode 8 (2004) 45.24%
NEAK (2000) 51.67%
Hatzigeorgiou and Kanapitsas, (2013), without infill
75.67%
walls, W = 0
Hatzigeorgiou and Kanapitsas, (2013), with infill
30.88%
walls, W = 1
Asteris et al. (2015) without infill walls, aW = 1, Et = 0 34.03%
Amanat and Hoque (2006) 56.98%
Goel and Chopra (2000) 54.05%

Furthermore, in their study, Gravett et al., (2021) developed and implemented shallow and deep
neural networks to train, test and validate predictive models by using the developed dataset.
Specifically, a 50 neurons network was utilized for the approximation and training of the model with
the algorithm defined in Bakas et al. (2020). So as to select 50 neurons, the researchers implemented
the recommendation that foresees the use of a 10% on the number of dataset values. Figure 17 shows
the structure of the ANN used to train, test, and validate the corresponding predictive model as
presented by Gravett et al., (2021).
Figure 17: Artificial Neural Network Architecture (Gravett et al., 2021).
According to the numerical investigation involving the ANN model, it was found that it exhibited an R²
= 0.9709, which was close to the 5-feature formula, (R² = 0.9594), as well as a corresponding MAPE =
10.71%, which is similar to that obtained from the 5-feature formula (MAPE = 10.26%). Nevertheless, the
MAPE on the validation dataset of the 60 out-of-sample modal results was only 3%, deriving a more
accurate prediction compared to the 20-feature formula, demonstrating a significantly more accurate
prediction compared to all the formulae investigated by Gravett et al., (2021). Similar results were reported
for the cased of the deep learning algorithm that was used to train a similar predictive model. As was
presented in Gravett et al., (2021), the MAPE of the predictive model obtained from the deep learning
algorithm was found to be slightly better than that of the 20 terms nonlinear formula (5.68%). It was also
stated that the advantage of using the non-linear regression approach is that of a closed-form solution which
is achieved, in contrast with the "black-box" when the ANN is used to develop the predictive model.
The research work performed by Gravett et al., (2021) was extended by Carstens et al., (2022) where
the 475 numerical results were increased by constructing and analysing additional buildings with and
without accounting for the SSI effect. Table 7 summarises the new minimum and maximum values used in
this study.
Table 7: Minimum and maximum magnitudes of the new RC Frames (Carstens et al., 2022).
Variables Minimum Maximum
Soil Depth [m] 1 60
Soil E [kPa] 65 000 700 000
H [m] 3 30
L [m] 3.4 34.4
B [m] 3.4 34.4
ρ(%) 0 85.29

Once more, three formulae were developed (Carstens et al., 2022) and the most accurate proposed
formula is presented herein. As it was reported in their study, the most accurate formula was the 20-feature
formula which can be seen in Equation 22. The MAPE obtained from this formula was 1.49%, which is a
7% improvement from the Gravett et al., (2021) study. This result was obtained when using the 60 out-of-
sample results that were implemented during the validation of the newly proposed formula, whereas on the
other hand, the 3-feature formula (Equation 23) resulted in the highest correlation. Figure 18 shows the
correlation between the validation dataset and the prediction of the proposed fundamental period formula
of Equation 22, as it was presented by Carstens et al., (2022).
𝑇 = (0.0292939 ⋅ H) − (0.150825 ⋅ 10−3 ∙ 𝜌 ⋅ H) 22
+(5.82242 ∙ 10−6 ⋅ H ⋅ B 2 ) + (3.30369 ∙ 10−6 ⋅ 𝜌 ⋅ 𝐻 2 )
+(0.000215881 ⋅ H ⋅ L) − (1.89375 × 10−15 𝐸𝑠2 ⋅ D)
+(3.23855 ⋅ 10−6 ∙ L ⋅ H ⋅ D) − (6.46154 ∙ 10−6 ⋅ 𝜌 ⋅ B ⋅ H)
 −(92.5478 ∙ 10−12 ⋅ 𝐻 ⋅ 𝐸𝑠 ⋅ 𝐷) − (40.6192 ∙ 10−12 ⋅ 𝜌 ⋅ 𝐸𝑠 ⋅ 𝐷)
+(194.394 ∙ 10−9 ⋅ 𝐷 ⋅ 𝜌 2 ) + (0.0037148 ⋅ 𝐵)
+(35.8861 ∙ 10−9 ⋅ 𝜌 ⋅ 𝐻 ⋅ 𝐷) + (66.2381 ∙ 10−12 ⋅ 𝐸𝑠 ⋅ 𝐷 2 )
−(0.278639 ∙ 10−6 ⋅ 𝐷 3 ) − (0.113737 ∙ 10−9 ⋅ 𝐿 ⋅ 𝐸𝑠 ⋅ 𝐷)
−(1.6727 ∙ 10−6 ⋅ 𝐵3 ) + (30.9934 ∙ 10−6 ⋅ 𝐿 ⋅ 𝐷)
−(0.00178654 ⋅ 𝐿) + (0.645744 ∙ 10−6 ⋅ 𝐿3 ) + 0.00239996
𝑇 = (0.0310197 ⋅ 𝐻) − (0.11254 ⋅ 10−3 ∙ 𝜌 ⋅ 𝐻) + (12.9093 ⋅ 10−6 𝐻 ⋅ 𝐵2 ) + 0.0110165 23

1.2
Formula Predicted Period [s]

0.8

0.6

0.4 y = 0.9873x + 0.041

R² = 0.9789
0.2

0
0 0.2 0.4 0.6 0.8 1 1.2
Numerically Predicted Period [s]
Figure 18: Correlation of the 3-feature formula on the validation set (Carstens et al., 2022).

Determining the fundamental period of steel structures

In a similar study (van der Westhuizen et al., 2022) the development of fundamental period formulae was
conducted for steel structures. The models were created using a varying number of stories, bays, and
superstructural geometry conditions. Initially, a model with a single-storey, single bay with a storey height
of 3.5 m and a raft foundation assuming a fixed base (see Figure 19) was constructed.

Figure 19: Initial Model (van der Westhuizen et al., 2022).

 (a) (b)
Figure 20: Initial model modified to have (a) two bays and (b) three bays.

The initial model was then expanded by altering the number of storeys, the geometry of the structure,
depth of soil as well as the soil’s Young modulus and orientation of columns. The models foresaw 2-, 4-,
6-, 8- and 10 stories, each with a 3.5 m height. The models were further expanded to include double and
triple spans along the x-axis as shown in Figure 20. The smallest plan was 5x3 m and the largest foresaw a
15x6 m plan view. The models were further modified to include the SSI effect where a soil domain was
discretized with depths of 1, 5, 12.5, 22.5 and 37.5 m. The superstructure was discretized through the use
of Natural Beam-Column Flexibility-based (NBCFB) finite elements (Markou & Papadrakakis, 2015),
where the raft and soil domains were discretized through 8-noded isoparametric hexahedral finite elements.
A total of 1,152 numerical results were produced for the final dataset that was used to train and test the
closed-form solutions.
The machine-learning algorithm was used to train and test the database and construct a fundamental
period formula for steel structures taking into account the SSI effect. The fundamental period 40-feature
formula obtained can be seen in Equation 24.
𝑇 = (0.194630 ∙ 𝑙𝐻2 ) + (0.0580556 ∙ 𝐶𝑂2 ∙ 𝐵) − (9.39027 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝐼𝑛𝑣𝐵 ∙ 𝑙𝐵) 24
−(8.49213 ∙ 𝐼𝑛𝑣𝐿 ∙ 𝐶𝑂 ∙ 𝐻) − (41.8498 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝑙𝐿 ∙ 𝐻) − (8.14564 ∙ 𝐼𝑛𝑣𝐸 ∙ 𝐸 ∙ 𝐻)
−(0.800465 ∙ 𝐶𝑂 ∙ 𝐵 ∙ 𝐻) + (114.808 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝐼𝑛𝑣𝐵 ∙ 𝐻) + (46.6778 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝐼𝑛𝑣𝐵2 )
+(0.0631499 ∙ 𝐵2 ∙ 𝐻) + (4.20803 ∙ 𝑙𝐵 ∙ 𝐶𝑂 ∙ 𝐻) − (0.144945 ∙ 𝑙𝐿 ∙ 𝐻 ∙ 𝐿)
+(0.847694 ∙ 𝐵 ∙ 𝐻 ∙ 𝐼𝑛𝑣𝐿) + (9.37930 ∙ 𝐼𝑛𝑣𝐿2 ∙ 𝐻) − (1.08930 ∙ 𝐼𝑛𝑣𝐶𝑂2 ∙ 𝐿)
+(4.04342 ∙ 𝐼𝑛𝑣𝐿) − (0.251627 ∙ 𝐼𝑛𝑣𝐿 ∙ 𝐶𝑂 ∙ 𝐵) − (0.00783561 ∙ 𝐼𝑛𝑣𝐵 ∙ 𝑙𝐶𝑂 ∙ 𝑙𝐸 )
+(0.523388 ∙ 𝑙𝐿2 ∙ 𝐼𝑛𝑣𝐶𝑂) + (0.0947335 ∙ 𝐼𝑛𝑣𝐻 ∙ 𝑙𝐻 ∙ 𝐿) + (46.8309 ∙ 𝐼𝑛𝑣𝐸 ∙ 𝐻 ∙ 𝑙𝐷𝑠)
+(0.00764850 ∙ 𝑙𝐻 ∗ 𝐵) + (0.000161108 ∙ 𝑙𝐿 ∙ 𝐿 ∙ 𝑙𝐸) − (20.5554 ∙ 𝐼𝑛𝑣𝐸 ∙ 𝐶𝑂 ∙ 𝐷𝑠)
−(0.00474725 ∙ 𝐼𝑛𝑣𝐿2 ∙ 𝐼𝑛𝑣𝐷𝑠) + (2.73101 ∙ 𝐼𝑛𝑣𝐿 ∙ 𝐼𝑛𝑣𝐻 ∙ 𝐶𝑂) + (0.403996 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝑙𝐵 ∙ 𝐿)
−(0.0105914 ∙ 𝑙𝐿 ∙ 𝐿 ∙ 𝐵) − (0.228100 ∙ 𝑙𝐵2 ∙ 𝐶𝑂) + (0.00265642 ∙ 𝐼𝑛𝑣𝐿 ∙ 𝐻2 )
−(2.58386 ∙ 𝐼𝑛𝑣𝐵 ∙ 𝐼𝑛𝑣𝐻 ∙ 𝐶𝑂) + (5.84142 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝐻 ∙ 𝐿) + (29.5168 ∙ 𝐼𝑛𝑣𝐶𝑂 ∙ 𝐻)
+(0.849560 ∙ 𝐼𝑛𝑣𝐿 ∙ 𝑙𝐻 ∙ 𝐶𝑂) − (2.14776 ∙ 𝐼𝑛𝑣𝐵 ∙ 𝑙𝐻 ∗ 𝑙𝐶𝑂) + (1.34222 ∙ 𝑙𝐵 ∙ 𝑙𝐻 ∙ 𝐼𝑛𝑣𝐻)
−(0.00333495 ∙ 𝑙𝐸 ∙ 𝐿 ∙ 𝐼𝑛𝑣𝐻) − (2.64111 ∙ 𝐼𝑛𝑣𝐵2 ∙ 𝐼𝑛𝑣𝐻) + (71.1358 ∙ 𝐼𝑛𝑣𝐻 ∙ 𝐷𝑠 ∙ 𝐼𝑛𝑣𝐸)
−(17.9194 ∙ 𝐼𝑛𝑣𝐸 ∙ 𝑙𝐸 ∙ 𝑙𝐿) − 1.16636
Where:
𝑇 is the fundamental period (s)
𝐷𝑠 is the depth of soil (m)
𝐸 is the soils Young’s Modulus (kPa)
𝐻 is the building height (m)
𝐿 is the length of the building parallel to the oscillating direction (m)?
𝐵 is the width of the building perpendicular to the oscillating direction (m)
𝐶𝑂 is the orientation of the columns (either a 1 or 2)
 𝑙𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟 is ln(𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟 + 1) i.e., 𝑙𝐷𝑠 = ln(𝐷𝑠 + 1)
1 1
𝐼𝑛𝑣𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒r is 𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟+1 i.e., 𝐼𝑛𝑣𝐷𝑠 = 𝐷 +1
𝑠
The correlation between the numerically predicted and formula predicted fundamental period can be
seen in Figure 21.

Figure 21: Correlation between numerically predicted and formula predicted fundamental period (van der
Westhuizen et al., 2022).

According to the validation phase performed in van der Westhuizen et al., (2022), it was found that the
formula had a mean absolute error of 2.8%, compared to Eurocode 8 which had an error of 76%. It can be
seen that the proposed formula yields a more accurate result while accounting for the all-important SSI
effect.

Determining the deflection of horizontally curved steel I-beams

The aim of this work (Ababu et al., 2022) was to derive an easy-to-apply predictive formula for the
determination of the mid-span deflection of curved steel I-beams. It is important to note that Castigliano’s’
theorem was not used in this case, where machine learning was utilised to determine an explicit closed-
form equation.
In order to generate beams using a finite element method, it is important to note what mesh size to adopt
to accurately model the beam behaviour as this affects the overall numerical results Models with varying
mesh sizes were created and the deflections were compared to a curve obtained from an experimental study
conducted by Shanmugam et al. (1995). It was found that the 20-noded isoparametric hexahedral elements
best represented the mechanical response of the under study curved steel I-beams. Figure 22 shows an
example beam developed using a 20-noded hexahedral mesh.
Figure 22: (Left) General beam, undeformed mesh and (Right) Deformed shape and von Mises stress
contour for a total deflection of 17 mm (Ababu et al., 2022).

A total of 270 beams were created that were fixed at one end with roller support at the other (also seen
in Figure 22) as this represents the boundary conditions of crawl beams. The load was applied at the mid-
span and was divided into 10 load increments where the beams were loaded until failure. A load-deflection
curve was then developed which led to a total of 1 890 unique data points that were used to train the ML
algorithm. It is important to note that the selected points foresaw a maximum of 50% load level of the
maximum load in avoiding including deflections that resulted in load levels that might generate significant
geometrical nonlinearities.
Specifically, 8 formulae were developed in this case where 5 to 40 features were considered. The
correlation remained the same no matter the number of features used but the MAPE was optimum when
the 10-feature formula was used (4.11%). The 10-feature formula, which yielded the minimum average
error for the validation data, can be seen in Equation 25. The MAPE was derived by using a validation
dataset of 15 out-of-sample data points. This pilot project demonstrates the potential of the proposed
approach, where further analysis and experimentation were deemed necessary (Ababu et al., 2022).
Deflection = (1.08128 ∙ 10−1 ∙ Q ∙ L) − (8.41124 ∙ 10−6 ∙ Q ∙ L ∙ A) 25
−(9.81969 ∙ 10−6 ∙ Q ∙ E ∙ R) + (1.82604 ∙ 10−5 ∙ Q ∙ R ∙ R)
−(2.71029 ∙ 10−4 ∙ Q ∙ R ∙ L) + (6.22991 ∙ 10−3 ∙ Q ∙ L ∙ L)
+(2.55963 ∙ 10−6 ∙ Q ∙ L ∙ Ixx ) + (2.99150 ∙ 10−7 ∙ Q ∙ fy ∙ fy )
−(1.07754 ∙ 10−4 ∙ Q ∙ E ∙ L) − 8.95792 ∙ 10−2
Where:
𝐸 is the Young’s modulus (GPa)
𝐿 is the curved length of the beam (m)
𝐴 is the section area (mm2)
𝑓𝑦 is the yielding stress
𝐼𝑥𝑥 is the second moment of area about the strong axis
𝑓𝑦 is the yielding stress (mm4)
𝑄 is the percentage of ultimate load applied to the beam as a number (50%=50)

FUTURE RESEARCH
This chapter highlighted the use of AI and ML algorithms in developing formulae that successfully
determined the mechanical properties and response of structures and structural members. The predictability
of these models, however, depends on the datasets that are constructed and used to train in each case. In
some instances, it is necessary to expand the datasets to come up with formulae that can predict properties
in a larger range. For example, Equation 25 can only be used to predict accurately the fundamental period
of a structure that falls within the dataset’s limits and requires expansion if it were to be applied to a larger
geometrical set of structures. A current research project is being completed in which the dataset of
determining the fundamental period of steel structures is being expanded to include an odd number of
stories as well as more Young’s modulus values. It is expected that the dataset will contain 100,000
numerical results obtained in a similar way as discussed in the chapter through the use of a high-
performance computer that facilitates a large number of numerical modelling and ML training.
With the successful use of AI and ML algorithms in the determination of mechanical behaviour of
structural elements of buildings, Civil Engineers will be able to incorporate the proposed formulae during
their design achieving additional accuracy compared to the current design codes. This method, as stated,
will be also implemented in the research of other structural types and configurations, such as the case of
predicting of the horizontal displacement of piles embedded in clay material. Furthermore, RC beams with
stirrups will be investigated including deep and slender beams which are reinforced with standard and FRP
bars. RC structural columns and shear walls will also be analysed thereafter.
In expanding the curved beams dataset, other geometries, boundary conditions and different steel yield
strengths are currently being explored, where the rotation of the section of the beams is being investigated.
This research work will also be expanded to include curved RC beams that are also found to be of significant
interest. Since there are no formulae current available like the one presented herein, it is necessary to
compare the results obtained to that of experimental results, which is currently a research project that is
being performed at the laboratories of the University of Pretoria at the Civil Engineering department.
Through the presented research work, it has been established that time can be saved in terms of the
development and analysis of models using FEA software, however, the ML algorithm may produce
equations that are long and difficult to make use of when manual calculations are performed. The
development of an automated ML software is currently being developed that will be able to analyse a dataset
and perform these calculations for the user, in addition to the ability to use excel sheets or programming
calculator apps that incorporate the developed predictive models. This would make the analysis of the
results significantly easier. The development of such an application is now being finalised, thus providing
with the ability to reduce the required time developing graphs that further help to understand the ML
algorithm’s output. This will also allow more engineers to be able to use the developed technology,
therefore, train their own predictive models.

CONCLUSION
While there are, in some cases, formulae available to design engineers this does not necessarily mean that
they are the most accurate, easily accessible, or easy to apply. The chapter presented through this research
work summarized the use of AI and ML in developing predictive models for the design of structural
elements and full-scale structures through an innovative approach that combines advanced FEA for the
construction of datasets that are then used to train new predictive models. It has been shown that the
predictive models are more accurate compared to the current design code formulae, which highlights the
prospects of the proposed approach in developing more accurate design codes leading to increased safety,
sustainability, and reduced cost. Finally, further numerical investigation and experimentations are needed
to establish objective and accurate design formulae that will derive from the proposed methodology.

REFERENCES
A. C. I. Committee, “ACI 440.1 R-15: guide for the design and construction of structural concrete
reinforced with FRP bars,” American Concrete Institute, Farmington Hills, 2015.
A.C. Institute, Building Code Requirements for Structural Concrete (ACI 318-19): Commentary on
Building Code Requirements for Structural Concrete (ACI 318R-19): an ACI Report, American Concrete
Institute. ACI, 2019.
Ababu, E. M., Markou, G., & Bakas, N.P. (2022). Using Machine Learning and Finite Element Modelling
to Develop a Formula to Determine the Deflection of Horizontally Curved Steel I-beams [Paper
presentation]. ICAART, Portugal.
Abed, F., El-Chabib, H., & AlHamaydeh, M. (2012). Shear characteristics of GFRP-reinforced concrete
deep beams without web reinforcement. Journal of Reinforced Plastics and Composites, 31(16), 1063-
1073.
AlHamaydeh. M., Markou, G., Bakas, P.N. & Papadrakakis, M. (2022), AI-Based Shear Capacity of FRP-
Reinforced Concrete Deep Beams without Stirrups, Engineering Structures, 264.
Amanat, K. M., & Hoque, E. (2006). A rationale for determining the natural period of RC building frames
having infill. Engineering Structures, 28(4), 495-502.
Asteris, P. G., Repapis, C. C., Cavaleri, L., Sarhosis, V., & Athanasopoulou, A. (2015). On the fundamental
period of infilled RC frame buildings. Structural Engineering and Mechanics, 54(6), 1175-1200.
Bakas, N.P., Langousis, A., Nicolaou, M. & Chatzichristofis, S.A. (2020), A Gradient Free Neural Network
Framework Based on Universal Approximation Theorem. arXiv preprint arXiv:1909.13563.
Bakas, N.P., Markou, G., Charmpis, D., & Hadjiyiannakou, K. (2021). Performance and scalability of deep
learning models trained on a hybrid supercomputer: Application in the prediction of the shear strength of
reinforced concrete slender beams without stirrups [Paper presentation]. COMPDYN, Greece.
Bathe, K-J. (1971). Solution Methods for Large Generalized Eigenvalue Problems in Structural
Engineering, Report UC SESM 71-20. Civil Engineering Department, UC Berkeley.
CEN. Eurocode 8: Design of structures for earthquake resistance. Part 1: general rules, seismic actions and
rules for buildings. European Standard EN 1998-1:2004, Comite Europeen de Normalisation, Brussels,
Belgium, 2004.
Chopra, A. K., & Goel, R. K. (2000). Building period formulas for estimating seismic displacements.
Earthquake Spectra, 16(2), 533-536.
Dahlberg, T. (2004). Procedure to calculate deflections of curved beams. International journal of
engineering education, 20(3), 503-513.
Dhahir, M. K. (2017). Shear strength of FRP reinforced deep beams without web reinforcement. Composite
Structures, 165, 223-232.
Dhahir, M. K. (2018). Strut and tie modelling of deep beams shear strengthened with FRP laminates.
Composite Structures, 193, 247-259.
Dimopoulos, T., Tyralis, H., Bakas, N.P. & Hadjimitsis, D. (2018) Accuracy measurement of Random
Forests and Linear Regression for mass appraisal models that estimate the prices of residential apartments
in Nicosia, Cyprus. Advances in Geosciences, 45, 377-382.
El-Sayed, A.K., El-Salakawy, E.F. & Benmokrane, B. (2005). Shear strength of concrete beams reinforced
with FRP bars: design method. 7th International Symposium on Fiber-Reinforced (FRP) Polymer
Reinforcement for Concrete Structures, 230, 955-974.
El-Sayed, A.K., El-Salakawy, E.F. & Benmokrane, B. (2006). Shear capacity of high-strength concrete
beams reinforced with FRP bars. ACI Materials Journal, 103(3), 383.
EN1992-1-1, Eurocode 2 (2004), Design of concrete structures.
European Committee for Standardization (2015). Eurocode 2: Design of concrete structures: Part 1-1:
General rules and rules for buildings: European Committee for Standardization.
Gravett, D. Z., Mourlas, C., Markou, G. & Papadrakakis, M. (2019) Numerical performance of a new
algorithm for performing modal analysis of full-scale reinforced concrete structures that are discretized
with the HYMOD approach [Paper presentation]. COMPDYN, Streamed from Athens.
Gravett, D. Z., Mourlas, C., Taljaard, V-L., Bakas, N., Markou, G., & Papadrakakis, M. (2021). New
fundamental period formulae for soil-reinforced concrete structures interaction using machine learning
algorithms and ANNs. Soil Dynamics and Earthquake Engineering, 144, 106656.
Hatzigeorgiou, G. D., & Kanapitsas, G. (2013). Evaluation of fundamental period of low‐rise and mid‐rise
reinforced concrete buildings. Earthquake engineering & structural dynamics, 42(11), 1599-1616.
Markou, G., & Papadrakakis, M. (2013). Computationally efficient 3D finite element modelling of RC
structures. Computers and concrete, 12(4), 443-498.
Markou, G., & Papadrakakis, M. (2015). A simplified and efficient hybrid finite element model (HYMOD)
for non-linear 3D simulation of RC structures. Engineering Computations, 32(5), 1477-1524.
Markou, G., Mourlas, C., & Papadrakakis, M. (2017). Cyclic nonlinear analysis of large-scale finite element
meshes through the use of hybrid modelling (HYMOD). International Journal of Mechanics, 11(2017),
218-225.
Markou, G., & AlHamaydeh, M. (2018). 3D finite element modelling of GFRP-reinforced concrete deep
beams without shear reinforcement. International Journal of Computational Methods, 15(02), 1850001.
Markou, G., Mourlas, C., Bark, H., & Papadrakakis, M. (2018). Simplified HYMOD non-linear simulations
of a full-scale multistory retrofitted RC structure that undergoes multiple cyclic excitations–An infill RC
wall retrofitting study. Engineering Structures, 176, 892-916.
Markou, G., Garcia, R., Mourlas, C., Guadagnini, M., Pilakoutas, K., & Papadrakakis, M. (2021). A new
damage factor for seismic assessment of deficient bare and FRP-retrofitted RC structures. Engineering
Structures, 248, 113152.
Markou, G. & Bakas P.N. (2019), AI reinforced concrete: Algorithms predict concrete strength without
training on experimental data, Innovate, 14, 8-10.
Markou, G. & Bakas, N. P. (2021a). Prediction of the Shear Capacity of Reinforced Concrete Slender
Beams without Stirrups by Applying Artificial Intelligence Algorithms in a Big Database of Beams
Generated by 3D Nonlinear Finite Element Analysis. Computers and Concrete, 28(6), 433-447.
Markou, G. & Bakas P.N. (2021b), Developing reinforced concrete structures through AI algorithms and
large-scale modelling. Innovate, 16, 38-40.
Mourlas, C., Papadrakakis, M., & Markou, G. (2017). A computationally efficient model for the cyclic
behavior of reinforced concrete structural members. Engineering Structures, 141, 97-125.
Mourlas, C., Gravett, D. Z., Markou, G., & Papadrakakis, M. (2019a). Investigation of the soil structure
interaction effect on the dynamic behavior of multistorey RC buildings. [Paper presentation]. VIII
international conference on computational methods for coupled problems in science and engineering
(coupled problems), Spain.
Mourlas, C., Markou, G., & Papadrakakis, M. (2019b). Accurate and computationally efficient nonlinear
static and dynamic analysis of reinforced concrete structures considering damage factors. Engineering
Structures, 178, 258-285.
Mourlas, C., Markou, G., & Papadrakakis, M. (2019c). 3D Detailed Modeling of Reinforced Concrete
Frames Considering accumulated damage during static cyclic and dynamic analysis–new validation case
studies. [Paper presentation]. COMPDYN, Greece.
Mourlas, C., Khabele, N., Bark, H. A., Karamitros, D., Taddei, F., Markou, G., & Papadrakakis, M. (2020).
Effect of Soil–Structure Interaction on Nonlinear Dynamic Response of Reinforced Concrete Structures.
International Journal of Structural Stability and Dynamics, 20(13), 2041013.
New Greek Seismic Code (NEAK), Athens. 2000.
Papadrakakis, M. & Markou, G. (2022). Assessment of the Structural Response of Steel Reinforced and
Steel-Fibre Reinforced Concrete Structures with 3D Detailed Modeling: Limitations and Remedies. In
Current Trends and Open Problems in Computational Mechanics (pp. 395-405). Springer.
Razaqpur, A. G. & Isgor, O. B. (2006). Proposed shear design method for FRP-reinforced concrete
members without stirrups. ACI Structural Journal, 103(1), 93.
Reconan FEA v2.00, User’s manual. (2020).
https://www.researchgate.net/publication/342361609_ReConAn_v200_Finite_Element_Analysis_Softwa
re_User's_Manual
Reineck, K-H., Bentz, E.C., Fitik, B., Kuchma, D.A. & Bayrak, O. (2013), ACI-DAFSTB Database of shear
tests on slender reinforced concrete beams without stirrups, ACI Structural Journal, 110 (5).
SABS Standards Division, South African National Standard (SANS), Edition 1 Part 4: Seismic actions and
general requirements for building, 2009.
Spijkerman Z., Bakas P.N., Markou G. & Papadrakakis. M. (2021) Predicting the shear capacity of
reinforced concrete slender beams without stirrups by applying artificial intelligence algorithms [Paper
presentation]. COMPDYN 2021, Greece.
Taljaard, V-L., Gravett, D. Z., Mourlas, C., Markou, G., Bakas, N., & Papadrakakis, M. (2021)
Development of a new fundamental period formula by considering soil-structure interaction with the use
of machine learning algorithms [Paper presentation]. COMPDYN 2021, Streamed from Athens.
Thai, H-T. (2022). Machine learning for structural engineering: A state-of-the-art review. Structures, 38,
448-491.
Tureyen, A.K., & Frosch, R.J. (2002). Shear tests of FRP-reinforced concrete beams without stirrups. ACI
Structural Journal, 99(4), 427-434.
van der Westhuizen, A.M., Markou, G. & Bakas, N.P. (2022) Development of a New Fundamental Period
Formula for Steel Structures Considering the Soil-structure Interaction with the Use of Machine Learning
Algorithms. [Paper presentation]. ICAART, Portugal.
Yang, K-H., Chung, H-S., Lee, E-T., & Eun, H-C. (2003). Shear characteristics of high-strength concrete
deep beams without shear reinforcements. Engineering Structures, 25(10), 1343-1352.
 Young, K. C. (2011). An Investigation of the Fundamental Period of Vibration of Irregular Steel Structures
[Unpublished Thesis]. The Ohio State University, Columbus.

View publication stats