www.nature.

com/scientificreports

OPEN STM-based symbolic regression

for strength prediction of RC deep
beams and corbels
Khaled Megahed
This study uses symbolic regression with a strut-and-tie model to predict the shear strength of
reinforced concrete deep beams (RCDBs) and corbels (RCCs). Previous studies have proposed two
distinct types of models for estimating shear capacity: explainable models based on theoretical
derivations and black-box models derived from machine learning (ML) methods. This study proposes a
hybrid model derived from the strut-and-tie model (STM), where the performance of STM is enhanced
through the ML approach using genetic programming. This model is based on a comprehensive
experimental database of 810 tests for the shear strength of RC deep beams and 371 tests for RC
corbels from various research papers. The developed STM-based symbolic regression (SR-STM)
integrates two distinct force-transferring mechanisms: the diagonal strut mechanism utilizing concrete
strength and the truss mechanism utilizing orthogonal web reinforcement. The SR-STM model is both
robust and interpretable, demonstrating high prediction accuracy with mean values of the prediction-
to-actual ratios of 0.999 and 1.004 and coefficient of determination values of 0.913 and 0.862 for
RCDBs and RCCs, respectively, while providing explainable mathematical expressions that align with
the mechanical principles of STM. The developed SR-STM model is benchmarked against several
state-of-the-art models and evaluated against the CatBoost ML technique, demonstrating acceptable
performance. The results highlight the SR-STM model’s effectiveness in providing reliable predictions
and valuable insights for practical engineering applications. Furthermore, a SHAP (Shapley Additive
Explanations) analysis was performed, and its results align with the SR-STM model, confirming the
model’s effectiveness in accurately capturing the key factors influencing the shear strength of RCDBs
and RCCs.

Keywords Machine learning, Symbolic regression, Strut-and-tie model, Deep beams, Corbels, CatBoost
model

Deep reinforced concrete (RC) elements, such as deep beams (RCDBs) and corbels (RCCs), characterized by a
small span-to-height ratio1,2, are frequently used as load-transferring components in structures, such as transfer
girders and pile caps, due to their superior shear strength compared to slender members. Despite their extensive
usage, designing these elements is challenging due to the nonlinear influence of various parameters on their
shear behavior. Multiple shear strength models have been studied, including those utilizing machine learning
methods3–8, the strut-and-tie model (STM)9–11, the compression field method12, and finite element analysis13.
Traditional design methods, such as the STM, often fail to adequately capture the complex relationships between
parameters impacting shear strength, resulting in inaccurate strength prediction. Additionally, existing code
provisions, such as ACI 31814, and EC215, along with previous mechanical models10,11, offer straightforward
procedures for calculating the shear capacity of RCDBs and RCCs. However, their conservative approach and
inconsistency with test results fail in developing a comprehensive model that can accurately estimate the shear
capacity of these elements.
Deep beams and corbels are frequently characterized by discontinuity regions, commonly called “D-region
members.” Within these D-regions, nonlinear strain distributions arise from sudden changes in geometry
or loading configurations16, making traditional design methods based on classical Bernoulli beam theory
inadequate. The strut-and-tie model (STM) has emerged as a robust and effective approach for designing deep
reinforced concrete elements. Numerous STMs have been proposed in the literature for estimating the shear
capacities of the D-regions11,17−19. However, despite the numerous STMs proposed, some models have been
criticized for their inconsistent and complex predictions of shear strength2,4,8,19.

Department of Structural Engineering, Mansoura University, PO BOX 35516, Mansoura, Egypt. email:
[email protected]

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 1

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Machine learning (ML) has recently become increasingly prominent in various engineering applications,
offering an alternative approach to traditional mechanical theories. ML algorithms, including artificial neural
networks, genetic algorithms, and ensemble learning methods, have been widely applied to predict the shear
capacity of deep RC elements3–8. For instance, Ma et al4. applied six different ML models to forecast the
shear capacity of RCDBs and compared their performance with previous state-of-the-art models. Feng et al5.
investigated four standard ensemble learning models—random forests, gradient boosting regression trees,
adaptive boosting, and extreme gradient boosting (XGBoost)—to estimate the shear capacity of RCDBs using
a dataset comprising 271 samples. Ashour et al3. utilized genetic expression programming to formulate an
empirical expression for the shear capacity of RCDBs relying on 141 test data points. In addition, Shahnewaz et
al6. and Wakjira7employed genetic algorithm for shear strength prediction of RCDBs. However, these models
are challenging to apply in practical engineering design because the purely data-driven prediction procedure
cannot be translated into a feasible mathematical expression for structural application. Consequently, data-
driven algorithms are often viewed as black-box models. Furthermore, many expressions derived from ML
techniques, such as genetic expression programming (GEP) and genetic algorithms (GA), are often criticized for
their lack of physical significance and excessive complexity3,6,7, as outlined in Table 1. In addition, recent studies
have demonstrated the potential of ML to optimize and innovate within the field of civil engineering. Inqiad
et al. used GEP for predicting the compressive strength of self-compacting concrete. Moreover, Khaled et al20.
used different ML models to predict the axial capacity of rectangular concrete-filled concrete columns, including
the Gaussian process (GPR) and the extreme gradient boosting model (XGBoost). Additionally, Moradi et al.
[46] introduced a novel experimental approach to enhance the shear capacity of RC beams by utilizing fiber-
reinforced polymer wraps. These advancements highlight the growing role of machine learning in improving the
accuracy, efficiency, and innovation of civil engineering practices.
Due to the differing prediction mechanisms of explainable models (e.g., strut-and-tie models) and black-
box models (e.g., data-driven models), these approaches have traditionally been considered independent in
resistance prediction3–8. Previous studies have favored strut-and-tie models for their interpretable mechanisms,
while black-box models have been preferred for their superior performance. This study uses a machine-learning-
based symbolic regression (SR) technique to develop an ML-aided STM for predicting the shear strength of
RCDBs and RCCs. This technique serves as a hybrid model or an intermediate solution, effectively bridging
the gap between mechanical-based and black-box models, thereby gaining popularity in recent studies21,22.
Given the robust performance of STM in estimating the shear capacity of deep RC elements, integrating the SR
technique can calibrate the STM, resulting in a hybrid model with high prediction accuracy.
This study aims to develop a symbolic regression model based on the Strut-and-Tie Model (SR-STM) to
predict the shear strength of RCDBs and RCCs precisely. The novelty of this research lies in integrating the
symbolic regression (SR) technique with the robust performance of STM. An enforced structure tree for symbolic
regression is utilized to achieve three primary objectives: (1) reduce the search space, thereby enhancing the
efficiency of the genetic programming process; (2) provide explainable expressions consistent with the Strut-
and-Tie Model; and (3) diverge from previous research that generated uninterpretable functions using SR.
The term “interpretability” in this context refers to the model’s ability to produce equations that are not only
computationally feasible but also aligned with the physical behavior of the structures, making them meaningful
and actionable for engineers. This approach ensures that the resulting models are not only computationally
feasible but also maintain interpretability and consistency with established mechanical principles. Additionally,
the developed model was calibrated against a comprehensive database of test results, including 810 deep beams
and 371 corbels, which is more extensive than those considered by state-of-the-art models11,17–19. The full list
of input features, including geometric configurations, concrete properties, reinforcement characteristics, and
loading plate dimensions, is detailed in the next section (see the Experimental Database section). Various metrics
were applied to evaluate the model’s accuracy and generality. Finally, the predictions obtained from the proposed
STM-based model were compared with those from the CatBoost ensemble machine learning technique, known
for its superior performance8, and three existing closed-form models11,17–19. This comparison highlights the
effectiveness and reliability of the proposed model in predicting shear strength while ensuring interpretability
and alignment with mechanical principles.

Reference Number Models: Statistical criteria

 a −0.874
Wakjira 20207 371 GP: Vth = 0.0456fc′ 0.619ρ 0.411 bw dµ = 0.82, COV = 0.305 (for RCDBs without web reinforcement)
l d

Ashour 20033 141      2  

V = bw h fc′ −4.56 + 1.68 ad ρ 2l + 2.45 + 0.1 ad − 1.16 ad + 3.12ρ t ρ l + 0.3ρ hw + 0.4ρ vw
GP:
µ = 1.11, Std =0.21 (for RCDBs)
  a 0.23  a   161   2.65
GA: Vu = bw hfc′ 2
5 − 14 d + 0.85(ρ l ρ hw ρ vw )0.1 − 35 d ρ hw ρ vw − 200 ad ρ l ρ hw ρ vw
Shahnewaz 20206 381
µ = 0.99, CoV = 0.232 (for RCDBs)
  0.044 
GA: Vu = bw hfc′ 1.74 − 2 ad + 0.5ρ 0.14 µ = 1.01, CoV = 0.257 (for RCDBs)

Table 1. Summary of previous GP and GA models for predicting shear strength of deep RC elements.

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 2

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Figure 1. The dimensions of RC deep beam and force transfer-mechanisms.

Figure 2. The dimensions of RC doubly symmetric corbel and force transfer-mechanisms.

Experimental database
The schematic diagrams of the shear mechanisms of RCDBs and RCCs are illustrated in Figs. 1 and 2, respectively.
A comprehensive dataset was collected to develop ML models, consisting of 810 RCDB and 371 RCC experiments
sourced from existing literature and a database gathered by Chetchotisak et al19. Detailed information about the
collected database is available in the supplementary data. Numerous experimental and theoretical studies9,11,17,18,
have identified that the shear capacity of deep RC elements is influenced by various factors, categorized into
six groups: (1) geometric configurations: beam height (h), effective height (d), width (bw), shear span (a) and
shear span-to-depth ratio (a/d); (2) concrete properties, i.e., concrete strength (fc’); (3) bottom longitudinal
reinforcement characteristics: reinforcement ratio (ρl), and strength (fyl); (4) web reinforcement characteristics:
vertical web reinforcement ratio (ρv) and strength (fyv), horizontal web reinforcement ratio (ρh) and strength
(fyh); (5) loading plate dimensions: top plate width (wtp) and bottom plate width (wbp); (6) horizontal force for
corbels (Nc). Table 2 summarizes statistical information for the established database.

Strut-and-tie model
Figures 1 and 2 illustrate two load-transferring mechanisms in a typical D-region for RCDB and RCC members:
the diagonal strut mechanisms (Figs. 1(a), 2(a)) and truss mechanisms (Figs. 1(b), 2(b)). In the diagonal strut

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 3

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Statistics (RCDBs) Statistics (RCCs)

Variable Symbol Type Min Max Mean Std Min Max Mean Std
Beam height h (mm) Input 160 2100 552 279 120 1143 395 185
Beam effective height d (mm) Input 140 2000 488 260 92 1059 352 171
Beam width bw (mm) Input 51 914 194 120 115 600 184 57
Shear span a (mm) Input 80 4375 618 445 50 870 184 128
Shear span-to-depth ratio a/d Input 0.27 2.5 1.29 0.54 0.11 1.69 0.59 0.35
Concrete strength fc′ (MPa) Input 11.3 120.1 40.4 21.5 15 105 43 18.4
Bottom reinforcement ratio ρ l (%) Input 0.26 11.33 2 1.11 0.21 4.93 1.24 0.76
Bottom reinforcement strength fyl (MPa) Input 267 1330 468 130 298 1480 433 131
Vertical web reinforcement ratio ρ v (%) Input 0 3.17 0.17 0.35 0 2.33 0.4 0.44
Vertical web reinforcement strength fyv (MPa) Input 0 855 211 230 0 760 270 235
Horizontal web reinforcement ratio ρ h (%) Input 0 2.86 0.29 0.37 0 1.24 0.09 0.26
Horizontal web reinforcement strength fyh (MPa) Input 0 1051 280 226 0 750 78 197
Vu
Shear strength index vn = bw hfc′
Output 0.03 0.29 0.14 0.05 0.05 0.34 0.16 0.05

Table 2. Statistic features of the experimental dataset.

mechanisms, the primary force-transferring system consists of diagonal compression struts (depicted as dashed
lines) and tension ties (depicted as solid lines), which intersect at nodes called nodal zones. While the truss
mechanism describes how the shear force is resisted by the orthogonal shear reinforcement11,23. This mechanism
is composed of two subtrusses: the horizontal subtruss, where tension forces are transferred by the horizontal
web reinforcement (red lines), and the vertical subtruss, where tension forces are transferred by the vertical web
reinforcement (blue lines). The term θ in Figs. 1 and 2 represent the angle of the diagonal concrete strut relative
to the horizontal plane and can be calculated using Eq. (1).
jd
tan θ = (1)
a

where jd = d − kd/3is the height of the moment arm11,17, with d being the effective depth and kd​the depth of
the compression zone of the member. The term kdcan be derived from the elastic bending theory11,17 as:
 
kd = (nρ l )2 + 2 (nρ l ) − (nρ l ) d(2)

where n represents the ratio of Young’s modulus of steel to concrete. For the case of RCC under a horizontal force
Nc (Fig. 2), the term ρl in Eq. (2) is assumed by previous studies17,19,23 as:
Al − Nc/fyl
ρl = (3)
bw d

where Al is the cross-sectional area of the longitudinal steel reinforcement.

Generally, an STM consisting of diagonal strut and truss mechanisms is considered to be a statically
indeterminate structure11,17,18,23. For simplicity, the shear strength of deep RC element (Vn) is assumed to be
the sum of the shear strengths provided by the diagonal concrete strut (Vc) and truss mechanisms (Vw), as
established by previous studies10,11,17, as follows:
Vn = Vc + Vw (4)

Shear strength contributed by diagonal strut mechanism

According to the STM approach, the brittle shear failure in deep RC elements is primarily governed by the
biaxial strength of the diagonal struts, which is usually less than the strength of the nodal regions18. Many
studies13,24−26 have employed the Mohr–Coulomb failure criterion for defining the biaxial strength of diagonal
struts, as follows:
σt σc
+ = 1.0(5)
ft fc′

where ft is the concrete tensile strength and σt and σc indicate the principal tensile and compressive stresses inside
the diagonal strut, respectively, as shown in Fig. 1(a). For simplicity, the stress ratio σt/ft in Eq. (5) is assumed as
the ratio of the tie tension force Tto the tensile strength of longitudinal steel bars13,19 as follows:

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 4

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Figure 3. Mutation and crossover operations in SR model.

Figure 4. (a) Flow charts of Symbolic regression. (b) The optimal tree-based individuals for RC deep beam
strength, where operator O, in the forced structure part, can be any arbitrary binary operator while, in the
optimized structure part, variable C is constant.

σt T
≈ (6)
ft (ρ l bw d) fyl

As depicted in Fig. 1(a), the tension force T at the CCT node (node surrounded by two struts and one tie) is
calculated through the balance of horizontal forces as follows:
T = Vc/tan θ (7)

While the diagonal compressive stress σ c acting on the concrete strut is given by:
Cc Vc
σc = = (8)
Astr Astr sin θ

where Astr is diagonal strut area as illustrated in Figs. 1(a) and 2(a) by the following expression:

Astr = bw (kd)2 + (wl )2(9)

where wlis the projected horizontal width of the CCC node (node surrounded by three struts)11,17,19,23,27. For
simplicity in this study, wl is assumed to be equal to the width of the top plate for deep beams and wc/2 for doubly
symmetric corbels, where wc is the column width (Fig. 2(a)).
By substituting Eqs. (6) and (8) into Eq. (5), the shear capacity​of the diagonal strut mechanism (Vc) can be
expressed as:
1 1 1
= + (10)
Vc Vs Vt

where Vs and Vt define the strength of the diagonal strut and tension tie, respectively, given by:
Vs = fc′ Astr sin θ , Vt = ρ l bw dfyl tan θ (11)

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 5

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Parameters Value Parameters Value

Number of generations 200 Allowed Binary operators -, +, *, ^, /
Total number of populations 20 Loss function Algorithm 1
Population size 50 Constraints {‘^’:(–1,10)}(a)
Maximum length of expressions (total number of nodes) 30 Nested constraints ‘^’:{‘^’:0,’/’:1}(b)
Parsimony (factor controls the expression complexity) 0.02 model_selection Accuracy

Table 3. The parameters of the SR model used in generating expressions. (a)The ‘^’:(–1,10) constraint means
that the left argument of the power function can exhibit any level of complexity, whereas the right argument is
restricted to a maximum complexity of 10 nodes. (b)Nested constraints govern how operators can be combined
or nested. The constraint ‘^’:{‘^’:0,’/’:1} specifies that ‘^’ operator cannot be used inside another ‘^’ operator,
but ‘/’ operator can be nested once in ‘^’ operator.

Previous models Formulas


Hwang and Lee18 VHwang = Kξ Astr sin θ , ξ = 3.35/ fc′ ≤ 0.52 K is factor accounting for web reinforcement
VRusso = kc (kχ fc′ sin θ + khρ hfyhtan θ + kv a/dρ v fyv ) bw d, χ = 0.74r3 − 1.28r2 + 0.22r + 0.87, r = fc′ /105
Russo et al11,17. For deep beams, kc = 0.76, kh = 0.25, kv = 0.35
For corbels, kc = 0.8, kh = 0.65, kv =0
VM IST = Vc + Fhtan θ + Fv
Vc = 1 +1 1 ≤ 0.85fc′ Astr sin θ , Fh = khAhfyh, Fv = kv Av fyv ,
Vstr Vtie
  
Chetchotisak et al. (MIST)19 Vstr = λ (fc′ )β Astr sinθ , Vtie = α Al fyt + 0.5 fc′ wtbw tanθ
For deep beams, λ = 1.864, β = 0.704, α = 3.512, kh = 0.161, kv = 0.153
For corbels, λ = 1.039, β = 0.847, α = 2.973, kh = 0.389, kv = 0.215

with Astr = bw (kd)2 + (wl )2 , wt = 2 (h − d)
 
Vn = 1 +
1
1 + vhtan θ + vv bw dfc′
vs sin θ vt tan θ

Current study
For deep beams, vs = 0.8αd ks1.1, vt = 1.6(ρel)0.7, vh = 0.14ρeh, vv = 0.31ρev
For corbels, vs = 0.8α c ks0.78, vh = 0.17(ρ eh)0.83, vv = 0.27ρ ev
vt = 3.0ρ el,
      
with α d = fc′ /30 − 0.15, α c = fc′ /60 + 3.1 , ks = k 2 + (wl /d)2, ρ el = ρ l ff′l , ρ eh = ρ h ffh′ , ρ ev = ρ v ffv′
c c c

Table 4. Summary of previous mechanical models in predicting shear strength. where K is strut-and-tie index
accounting for the influence of the web reinforcement., wt is the widths the tie.

Equation (10) expresses the strength contributions of the concrete strut Vs and the tension tie Vt to the overall
shear capacity. Rewriting Eq. (10) in a dimensionless form results in:
1 1 1
= + e (12)
vc kssin θ ρl tan θ

where
  
Vc fyl
vc = , ks = (k)2 + (wl /d)2, ρel = ρl (13)
bw dfc′ fc′

The term ks defines the ratio of strut width to beam depth.

Shear strength contributed by truss mechanism

According to the force equilibrium in the truss mechanism18,27 in Figs. 1(b) and 2(b), the shear strength Vw can
be written as:
Vw = Fhtan θ + Fv (14)

where Fh and Fv refer to the average tensile forces carried by the horizontal and vertical web reinforcements,
respectively, in the D-regions. Accordingly, in a manner similar to Eq. (12), the shear strength of the truss
mechanism Vw can be rewritten in dimensionless form as follows:

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 6

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

   
Vw fyh fyv
vw = = ρehtan θ + ρev , with ρeh = ρh , ρev = ρv (15)
bw dfc′ fc′ fc′

From Eqs. (12) and (15), the overall contribution of the diagonal strut and truss mechanisms can be formulated
as:
 
1
Vn = 1 1 + ρhtan θ + ρv bw dfc′ (16)
e e

ks sin θ + ρe tan θ
l

Symbolic regression
Symbolic regression (SR)28,29is a genetic programming technique30designed to discover simple and interpretable
analytic equations that best fit a given problem through exploring a predefined space of mathematical expressions
and functions. SR is approached as a multi-objective optimization problem, balancing predictive accuracy and
model complexity. Genetic programming techniques, including the principles of natural selection and evolution,
are commonly used in SR to iteratively refine candidate mathematical expressions until satisfactory expressions
are achieved. This research employs a Python library called PySR31 to identify interpretable, simple expressions
for the shear strength of RCDBs and RCCs.
The SR algorithm starts with creating an initial population composed of a random combination of operational
symbols (e.g., +, -, *,/, ^, etc.) and terminals, such as input variables and constants. Each individual in the
population is represented by a tree-like expression. Selection is then performed probabilistically, favoring those
individuals that demonstrate superior performance. To prevent the generation of overly complex individuals by
SR, a complexity limit of 30 is set, meaning the total number of operators, constants, and variables in the equations
cannot exceed this value. In addition, overly complex expressions are excluded from the SR expressions such as
high exponential terms, e.g. (•)^(•^•). The selected individuals undergo mutation (Fig. 3(a, b)) or crossover
(Fig. 3(c)) to produce a new generation of populations. Figure 4(a) presents the core steps of the SR approach.
This evolutionary process employs a fitness function, defined in Eq. (17)31, to evaluate and identify the most
optimal individuals in each generation, ensuring the progressive refinement of solutions.
l (E) = lpred (E) .exp (frecency [C (E)])(17)

where lpred(E) represents the model prediction error, C(E) denotes the expression complexity E, quantified by
the total number of nodes in the expression. The term frecency [C(E)] accounts for the frequency and recency
of the expression E occurring at complexity C(E) within the population. This measure is crucial for avoiding
the overcomplication and redundancy of the generated expressions, ensuring a balance between minimizing
error and maintaining simplicity. Details of the SR parameters used for generating expressions in this study are
summarized in Table 3.
The process of finding the best expression requires many iterations and a detailed assessment of each one. Each
equation generated through these iterations is subjected to thorough evaluation and refinement, considering
factors such as the complexity of the equations, their accuracy, and their ease of interpretation.
The process of finding the best expression requires many iterations, where different parameters, such as the
number of generations, total population, and population size, are varied to generate distinct equations. Each
equation is then subjected to a detailed assessment and refinement, considering complexity, accuracy, and ease
of interpretation factors. This iterative approach ensures that the final model balances interpretability with
predictive accuracy.

STM-based symbolic regression development

The primary objective of this section is to utilize symbolic regression (SR) to refine and calibrate the strut-and-
tie model (SR-STM) for predicting the shear strength of reinforced concrete deep beams (RCDBs) and corbels
(RCCs). Unlike purely data-driven approaches where machine learning models determine prediction outcomes
directly from data, the SR-based calibration method presented here fundamentally integrates the mechanical
principles of STM with machine learning techniques to enhance prediction accuracy. This hybrid approach
merges the physical insights of STM with the optimization capabilities of symbolic regression to achieve precise
and meaningful prediction outcomes.
Numerous expressions for predicting the shear strength of D-regions have been developed based on STM
concepts, varying in complexity and accuracy, as summarized in Table 4. For instance, Russo et al11,17. proposed
a nonlinear contribution factor, χ (fc′ ) fc′ , to account for concrete strength in the strut capacity. Similarly,
Hwang and Lee18, and Chetchotisak et al19. developed nonlinear expressions for strut capacity. For the remaining
elements, such as the tie and web reinforcement contributions, they assumed linear partial contributions with
coefficients less than 1.0 for each component. This partial contribution arises from the fact that the stresses in
the web reinforcements may not reach their yield strengths simultaneously.
In this study, symbolic regression is employed to optimize the coefficient expressions for various components
of the STM: the strut (ks), tie ( ρel), vertical web reinforcement ( ρev ), and horizontal web reinforcement ( ρeh) as
defined in Eq. (16). The expression employed for the optimization of the STM-based symbolic regression process
is as follows:

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 7

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

 
1
Vn =  1 1 + vh (ρeh) tan θ + vv (ρev ) bw dfc′ (18)
vs (f ′ , ks )sin θ + v
c ( e
t ρl )tan θ

where the coefficient expressions for each component of the STM that need to be optimized by SR technique are:
vs = vs (fc′ , ks) , vt = vt (ρel) , vh = vh (ρeh) , vv = vh (ρev )(19)
      
f f f
with: ks = k 2 + (wl /d)2, ρel = ρl fyl′ , ρeh = ρh fyh′ , ρev = ρv fyv′
c c c

Symbolic regression is employed to optimize the four expressions in Eq. (19) using the PySR library. This
process involves developing a custom objective loss function designed explicitly for symbolic regression, which
predefines the form of the shear strength equation with the four expressions. The symbolic trees accepted in this
process must have of a structure of ((T1 ◦ T2) ◦ (T3 ◦ T4)), where subtrees T1, T2, T3, and T4 define the four

Figure 5. Distribution of the databases and the relationships between different parameters.

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 8

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Figure 5. (continued)

functions vs, vt, vh, and vv, respectively, and operator ∘ could be any arbitrary binary operator. This structure is
referred to as the enforced structure of the tree generated using the developed objective function, as illustrated
in Fig. 4(b). The details of the objective function are illustrated in Algorithm 1. The objective function enforces
constraints on the structure of the symbolic expressions and penalizes undesirable characteristics.
As explained in Algorithm 1, the degree of the tree head is checked, and if it is not equal to two, meaning it
should combine two subtrees, TL = (T1 ◦ T2) and TR = (T3 ◦ T4), a significant penalty of 1000 is imposed. The
left subtree (TL) and right subtree (TR) are also checked for a degree of 2, with a smaller penalty of 100 imposed
if they do not meet this criterion. The left child of the left subtree (TLL or T1) represents vs (strut contribution)
and must be a function of fc’ (concrete strength) and the ratio ks. If vs contains invalid features or negative values,
penalties are added in proportion to the number of violations. Similarly, the right child of the left subtree (TLR
or T2) is processed as vt (tie contribution), which should only be a function of ρ el. The contributions from
the left subtree (TLL, TLR) are combined to derive the concrete contribution. Similarly, the right subtree (TR)
should be branched into two children (TRL, TRR) or (T3, T4), represent the contributions of horizontal and
vertical web reinforcement, respectively. It should be noted that the penalty term increases progressively by how

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 9

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Figure 6. Comparison between proposed equations and previous models.

Training data Testing data All data

Metrics CatBoost SR-STM CatBoost SR-STM CatBoost SR-STM Hwang18 Russo11 Chetchotisak (MIST)19
Mean µ 0.999 0.993 1.002 1.024 0.999 0.999 1.056 0.979 0.989
%CoV 4.56 14.11 14.80 15.86 7.79 14.55 18.75 18.48 15.24
R2 0.994 0.919 0.908 0.879 0.979 0.913 0.827 0.870 0.906
MAPE 3.295 10.89 10.67 13.84 4.770 11.48 16.51 14.58 11.95
RMSE(kN) 0.213 0.913 0.758 0.957 0.389 0.922 1.300 1.128 0.956
a20-index 0.998 0.86 0.833 0.772 0.965 0.842 0.684 0.731 0.828

Table 5. Comparison of the developed ML models for RC deep beams.

Training data Testing data All data

Metrics CatBoost SR-STM CatBoost SR-STM CatBoost SR-STM Hwang18 Russo17 Chetchotisak (MIST)19
Mean µ 0.999 1.013 1.000 0.967 0.999 1.004 1.322 1.010 1.016
%CoV 4.71 14.19 13.01 17.36 7.19 14.95 18.68 18.45 16.18
R2 0.989 0.876 0.873 0.780 0.974 0.862 0.448 0.799 0.807
MAPE 3.465 11.776 10.00 14.16 4.769 12.25 34.03 14.69 13.18
RMSE(kN) 0.303 1.185 0.843 1.338 0.464 1.217 2.43 1.468 1.437
a20-index 0.997 0.815 0.865 0.676 0.97 0.787 0.329 0.714 0.757

Table 6. Comparison of the developed ML models for RC corbels.

far it deviates from the constraints, effectively guiding the genetic algorithm towards the correct factorization.
Finally, the loss function combines the penalty terms and the Mean Absolute Percentage Error (MAPE) of the
predictions.
The optimal tree-based individuals (Fig. 4(b)) fitting the training experimental database for RC deep beams is

vs = 0.8α d ks1.1, vt = 1.6(ρel)0.7, vh = 0.14ρeh, vv = 0.31ρev (20)

where α d = fc′ /30 − 0.15.

The resulting training for RC corbel beams is

vs = 0.8αc ks0.78, vt = 3.0ρel, vh = 0.17(ρeh)0.83, vv = 0.27ρev (21)

where α c = fc′ /60 + 3.1.

The optimized coefficients (e.g., vs, vt, vh, vt) reflect the interplay between these factors and the overall shear
strength. For instance, the strut contribution coefficient expression (vs) is proportional to the ratio of strut width
to beam depth (ks). In addition, vsis inversely proportional to the concrete compressive strength due to the brittle
behavior of high-strength concrete. This result agrees with the expression provided by Russo et al11,17. in Table 4,
as the variable χ is inversely proportional to the concrete compressive strength. The remaining coefficients are
proportional to their influencing variable. The tie contribution coefficient vt is proportional to tie reinforcement
( ρ el), contribution coefficient vv is proportional to vertical web reinforcement ( ρ ev ), and contribution coefficient
vh is proportional to horizontal web reinforcement ( ρ eh), which aligns with the expected physical behavior
in STM-based design. In addition, the positive impact of the diagonal strut angle θaligns with experimental

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 10

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

findings by Kani32, which show that beams exhibit higher shear resistance at higher angles θ (lower a/d ratios).
Design examples illustrating the above equations can be found in the supplementary data.
The developed expressions for the SR-STM model are not only simple and robust but also carry physical
significance, in contrast to the GEP and GA models from previous studies (see Table 1). Moreover, the enforced
structure tree for symbolic regression reduces the search space and ensures that the resulting expressions are
both explainable and consistent with the Strut-and-Tie Model (STM).

Performance and results of the SR-STM model

In this study, the min-max scaling technique is applied to normalize the data and minimize the negative impacts
of multidimensionality. The normalized datasets are then randomly split into two sets, with 80% used for training
the model and 20% reserved for testing. The statistical distributions of these databases and relationships between
different parameters are presented in Fig. 5.
The performance of proposed SR-STM model is assessed by comparing it with the ML-based CatBoost
model and several mechanical models, including those developed by Hwang and Lee18, Russo et al11,17. and
Chetchotisak et al19.. The design equations for these three mechanical models are listed in Table 4.
Figure 6 shows scatter plots that compare experimental and predicted results for the SR-STM, CatBoost, and
mechanical models. For both the SR-STM and CatBoost models, most data points are closely aligned along the
diagonal line, indicating a high level of accuracy and reliability in the predictions made by these models. This
close correlation between the predicted and actual results highlights the strong performance of the SR-STM and
CatBoost models in accurately estimating the shear strength of RCDBs and RCCs. Detailed evaluation metrics
used to assess the performance of these models are presented in Tables 5 and 6. These metrics include the mean
(µ), coefficient of variance (CoV), coefficient of determination (R2), root mean squared error (RMSE), mean
absolute percentage error (MAPE), and a20-index, as defined below:

Algorithm 1. Objective function.

function objective_function(tree, data):
Initialize penalty term P = 0
if tree.degree ≠ 2 then P + = 1000
else
TL = tree.l #left tree
if TL.degree ≠ 2 then P + = 1000
else
TLL = TL.l # left subtree of left tree
P + = 100 * NOVs(TLL ≠ f (fc′ , ks ), TLL < 0)**
vs= eval_tree(TLL, data.x)
TLR = TL.r # right subtree of left tree
P + = 100 * NOVs(TLR ≠ f (ρ el ) ,TLR < 0)
vt= eval_tree(TLR, data.x)
1.0
vc = 1.0 1.0
+ v ∗data.x[tanθ
# concrete contribution
vs ∗data.x[sinθ ] t ]
end if
TR = tree.l #left tree
if TR.degree ≠ 2 then P + = 1000
else
TRL = TR.l # left subtree of right tree
P + = 100 * NOVs(TRL ≠ f (ρ eh ), TRL < 0)
vh= eval_tree(TRL, data.x)
TRR = TR.r # right subtree of right tree
P + = 100 * NOVs(TRR ≠ f( ρ ev ), TRR < 0)
vv = eval_tree(TRR, data.x)
vhv = vh*data.x [tanθ ] + vv # web reinforcement
contribution
end if
r = dataset.y/( vc + vhv ) # True-to-pred ratio
 
sum(abs.(r−1))
return +P # Return MAPE penalty term
length(r) +
end function

**: The expression NOV(TLL ≠ f (fc′ , ks) && TLL < 0) quantifies the number of violations where the TLL
subtree is not a function of the features fc′ and ks, or where it includes other features, or when its value is
negative.


n n n    n
y i − y i )2
i=1 ( 1  yi 100%   yi   
 , RM SE =  1
R2 = 1 −  2 , µ = , M AP E = − 1 yi − yi)2(22)
(
n − n y
i=1 i
n i=1  yi  n i=1
i=1 y i − y

In
−
this context, yi refers to the predicted value for the i-th specimen, yi is the corresponding actual output value,
y ​is the average of the actual observations, and n indicates the total number of samples in the database. The a20-
index33 assesses the percentage of specimens where the ratio yi/yi falls within the range of 0.80 to 1.20.
As demonstrated in Tables 5 and 6, both SR-STM and CATB models exhibit high accuracy, with mean µ,
R2, and a20-index values approaching 1.0, along with low CoV, MAPE, and RMSE. Specifically, the CATBoost
model achieves MAPE values of approximately 3.3% and 10.7% for RC deep beams and 3.5% and 10% for RC

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 11

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

Figure 7. Prediction errors of previous models and established ML models.

Figure 8. Summary plot for inputs influencing shear strength index Vn/bwd fc’ .

corbels in the training and testing sets, respectively, which are the lowest among the compared models. The
SR-STM model, while slightly less accurate with MAPE values of 10.84% and 11.76% for RC deep beams and
13.84% and 14.16% for RC corbels, still provides a strong performance. Furthermore, the introduced SR-STM
model yields µ values of 0.999 and 1.004, R2 values of 0.913 and 0.862, and CoV values of 14.55% and 14.95% for
RC deep beams and RC corbels, respectively. Although the SR-derived formulas show slightly lower accuracy
compared to the CATBoost (CATB) model, they are more accessible and easier to interpret, which significantly
enhances their practical utility in engineering applications. While the CATBoost model demonstrates superior
prediction accuracy, its black-box nature restricts its practical application in engineering design. In contrast,
the SR-STM model provides a more interpretable approach that bridges the gap between theoretical insight and
practical use.

Comparisons with closed-form models

In this section, the proposed equations are compared with three existing closed-form models listed in Table 4,
including the models by Hwang and Lee18, Russo et al11,17., and Chetchotisak et al. (MIST)19 models. Tables 5
and 6 provide statistical information on the predictive capabilities of these models for RC deep beams (RCDBs)
and RC corbels (RCCs), respectively. The proposed equations yield values of µand CoV of 0.999 and 14.55%
for RCDBs and 1.004 and 14.95% for RCCs. In contrast, the existing closed-form models display CoV greater
than 18%, except for Chetchotisak et al19., which shows CoV values of 15.24% and 16.18% for RCDBs and
RCCs, respectively. This indicates that the proposed equations exhibit better predictive stability and robustness
compared to the existing models.
The models introduced by Russo11,17and Chetchotisak19show mean values with slight deviations from 1.0. In
contrast, the Hwang and Lee model18 shows the highest mean deviation for RCCs, with a mean value of 1.322.
Figure 6introduces scatter plots illustrating the relationship between experimental and predicted results across
the entire database using the proposed equations and the three closed-form models. The SR-STM proposed in
this study displays prediction-to-test ratios concentrated around unity, achieving the highest accuracy among the
previous models. Furthermore, the CATBoost model demonstrates superior performance, with CoV values of
7.79% for RCDBs and 7.19% for RCCs, highlighting the effectiveness of using ML techniques for shear strength

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 12

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

prediction. On the other hand, the prediction results of closed-form models reveal deficiencies in their prediction
mechanisms. In particular, the Hwang expression18 for RCCs shows the most significant deviation between
experimental and predicted values, with a distribution skewed over the diagonal line, indicating a tendency
toward conservative predictions. The improvement of SR-STM reflects not only higher prediction accuracy
compared to closed-form models but also in providing more specific physical significance and mathematical
equations compared to purely data-driven models e.g., the CATB model.
Figure 7 displays the prediction errors of both existing closed-form expressions and the developed ML
models. In Fig. 7(a), the CATB and ST-STM models exhibit high precision, with over 83% of testing samples
within a 20% error range. In contrast, the Hwang and Russo formulas have almost 70% of samples within the
same range. The MIST model also performs well, capturing 80% of samples within the 20% error range. In
Fig. 7(b), the ST-STM and MIST formulas for RCCs show similar performance, with a slight advantage for
the ST-STM. While the results of the proposed equations and the MIST formula are comparable, the proposed
equations are more straightforward to implement. In addition, the ST-STM significantly outperforms the Hwang
model by having nearly three times the number of test samples within the same error ranges for RC corbels.
Moreover, all performance metrics for the introduced SR-STM, as detailed in Tables 5 and 6, exceed those of the
previously introduced mechanical models. These results highlight the superior performance of employing ML
techniques, such as the CATB and ST-STM models, in predicting the shear strength of RC deep beams and RC
corbels.

Feature importance analysis

Assessing the impact of input parameters on the shear strength is crucial for designing RCDBs and RCCs. This
study utilizes the Shapley Additive Explanation (SHAP) method to identify and highlight the most significant
parameters influencing the shear strength index, Vn/bwd fc’34. As depicted in Fig. 8, the summary plot reveals the
effect of each feature on model predictions and ranks the features according to their relative importance on the
axial strength index. The feature with the highest absolute SHAP value is considered the most significant. The
span-to-depth ratio (a/d), and effective longitudinal reinforcement ratio ( ρ el) emerge as the most significant
parameters for both RCDBs and RCCs. Additionally, the analysis shows that the effective vertical and horizontal
web reinforcement ratios ( ρ ev , ρ eh) rank as the third most important feature for RCDB and RCC databases,
respectively. The significance of the remaining features is ranked in descending order.
Figure 9 illustrates the SHAP feature importance for each input variable in the RCDBs and RCCs databases.
A feature importance value greater than zero indicates a positive correlation with the strength index, whereas
a value less than zero suggests a negative impact. The length and color of the bars in Fig. 9 represent the
significance and direction (positive or negative) of each feature, respectively. It is evident that, except for the
a/d ratio, concrete strength fc’, and height h, the remaining input features have a positive and mixed influence
on the shear strength index. Increasing effective reinforcement ratios ( ρ el, ρ ev , ρ eh), top plate width for RCDBs
wb, and column width for RCCs wc, will improve the shear strength index. Conversely, the a/d ratio, concrete
strength fc’, and height h negatively impact the shear strength index. The negative impact of the a/dratio aligns
with experimental findings by Kani32, which show that beams exhibit higher shear resistance at lower a/d values.
Moreover, increasing the beam height h, reduces the shear resistance, as deeper beams weaken shear transfer
strength through aggregate interlock along critical shear cracks, resulting in higher energy release and further
reduction in shear resistance35.
The equation extracted from the SR-STM model aligns well with the results obtained from the SHAP
analysis, reinforcing the reliability and interpretability of the SR-STM model. As shown in Eqs. (18), (20), and
(21) of the SR-STM model, the concrete strength fc’ has a negative impact on the shear strength index Vn/bwd
fc’, consistent with the SHAP analysis results. Additionally, the SR-STM model demonstrates that increasing
the effective reinforcement ratios ( ρ el, ρ ev , ρ eh), top plate width for RCDBs wb, column width for RCCs wc,
and angle θ (which is proportional to the inverse of the span-to-depth ratio a/d) enhances the shear strength

Figure 9. Features importance for inputs influencing shear strength index Vn/bwd fc’.

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 13

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

index. This positive effect is also reflected in the SHAP analysis, where these parameters increase the predicted
shear strength. Therefore, the proposed equation’s behavior is consistent with the SHAP analysis, highlighting
the model’s ability to accurately capture the relationships between the input features and the shear strength of
RCDBs and RCCs.

Limitations and future work

This section addresses the limitations of the SR-STM model and highlights potential areas for future research.
While the SR-STM model shows strong predictive capabilities, its applicability is inherently limited by the scope
of the data used for its development. The model is tailored to specific geometries and material properties within
the dataset, as detailed in Table 2. For instance, the model has been developed based on beams with a/d ratios
ranging from 0.27 to 2.5 for RCDBs and 0.11 to 1.69 for RCCs and on concrete strength between 11.3 MPa and
120.1 MPa for RCDBs and 15 MPa to 105 MPa for RCCs. The reinforcement ratios and yield strengths also cover
specific ranges, which may not cover all real-world scenarios. Future work could focus on expanding the dataset
to include a broader range of geometries and material properties. As a result, applying the SR-STM model for
beams or corbels with geometries or material properties outside these ranges may lead to inaccurate predictions,
as it would involve extrapolation beyond the model’s trained data.
The evolving field of energy-harvesting concrete marks a significant step toward smart, sustainable
infrastructure. This technology integrates materials that allow concrete structures to harvest and store energy,
creating multifunctional, self-sustaining elements36–39. Machine learning, particularly symbolic regression,
can optimize the performance of these materials, enhancing the efficiency of energy-harvesting concrete. By
incorporating energy-harvesting considerations into the design, we can develop concrete structures that are
both structurally sound and energy-efficient. This approach could lead to self-powered roadways, buildings, and
bridges, advancing sustainable construction practices.

Conclusions
In conclusion, this study compiled a comprehensive database of 810 experimental tests for the shear strength of
RC deep beams (RCDBs) and 371 RC corbels (RCCs) tests from various research papers. It employed symbolic
regression (SR) techniques to refine and calibrate the Strut-and-Tie Model (SR-STM). From the evaluation
results, the following conclusions can be drawn:

• The integration of symbolic regression with the Strut-and-Tie Model successfully enhances prediction ac-
curacy while maintaining the interpretability and consistency of the models with established mechanical
principles.
• The SR-STM model yields µ values of 0.999 and 1.004, R² values of 0.913 and 0.862, and CoV values of 14.55%
and 14.95% for RC deep beams and RC corbels, respectively, indicating high predictive stability and robust-
ness.
• Compared to existing closed-form models by Hwang and Lee18, Russo et al11,17., and Chetchotisak et al.
(MIST)19, the SR-STM model shows better predictive performance, with improved CoV values and concen-
trated prediction-to-test ratios around unity.
• While the CATBoost model demonstrates superior performance with CoV values of 7.79% for RCDBs and
7.19% for RCCs, its black-box nature limits practical application in engineering design, highlighting the need
for more interpretable models like SR-STM.
• The SR-STM model significantly outperforms the Hwang model by having nearly three times the number of
test samples within the 20% error range for RC corbels, and it surpasses all performance metrics compared to
previously introduced mechanical models.
• The alignment between the SR-STM model’s equations and the SHAP analysis confirms the model’s effective-
ness in accurately capturing the key factors influencing the shear strength of RCDBs and RCCs.

The SR-STM not only achieves significant improvements in shear strength prediction but also effectively
combines the advantages of white-box and black-box models. Compared to purely data-driven approaches, the
SR-STM, with its explicit mathematical equations, is more accessible and reliable for engineers to utilize in
practical applications. While the SR-STM model demonstrates strong predictive capabilities, its effectiveness
is limited by the dataset’s scope, necessitating future research to broaden the range of geometries and material
properties to enhance model applicability and accuracy. In summary, integrating the ML-based approach
presents a promising method for accurately predicting the shear strength of RC deep elements, providing
valuable insights for engineering applications.

Data availability
All data generated or analyzed during this study are included in this published article and available in a public
repository: https://github.com/kmegahed/SR-STM.

Received: 8 July 2024; Accepted: 30 September 2024

References
1. MacGregor, J. G., Wight, J. K., Teng, S. & Irawan, P. Reinforced concrete: Mechanics and design, vol. 3. Prentice Hall Upper Saddle
River, NJ, (1997).
2. Kassem, W. Strength prediction of corbels using strut-and-tie Model Analysis. Int. J. Concr Struct. Mater. 9 (2), 255–266. https://
doi.org/10.1007/s40069-015-0102-y (2015).

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 14

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

3. Ashour, A. F., Alvarez, L. F. & Toropov, V. V. Empirical modelling of shear strength of RC deep beams by genetic programming.
Comput. Struct. 81 (5), 331–338. https://doi.org/10.1016/S0045-7949(02)00437-6 (2003).
4. Ma, C. et al. Prediction of shear strength of RC deep beams based on interpretable machine learning, Constr. Build. Mater., vol. 387,
no. July p. 131640, 2023, doi: (2022). https://doi.org/10.1016/j.conbuildmat.2023.131640
5. Feng, D. C., Wang, W. J., Mangalathu, S., Hu, G. & Wu, T. Implementing ensemble learning methods to predict the shear strength
of RC deep beams with/without web reinforcements. Eng. Struct. 235, 111979. https://doi.org/10.1016/j.engstruct.2021.111979
(2021).
6. Shahnewaz, M., Rteil, A. & Alam, M. S. Shear strength of reinforced concrete deep beams – A review with improved model by
genetic algorithm and reliability analysis. Structures. 23, 494–508. https://doi.org/10.1016/j.istruc.2019.09.006 (2020).
7. Wakjira, T., Ibrahim, M., Sajjad, B. & Ebead, U. Shear capacity of reinforced concrete deep beams using genetic algorithm. IOP
Conf. Ser. Mater. Sci. Eng. 910 (1). https://doi.org/10.1088/1757-899X/910/1/012002 (2020).
8. Megahed, K. Prediction and reliability analysis of shear strength of RC deep beams. Sci. Rep. 14 (1), 14590. https://doi.org/10.1038/
s41598-024-64386-w (2024).
9. Park, J. & Kuchma, D. Strut-and-tie model analysis for strength prediction of deep beams. ACI Struct. J. 104, 657–666 (2007).
10. Matamoros, A. B. & Wong, K. H. Design of Simply Supported Deep Beams Using Strut-and-Tie Models, ACI Struct. J., vol. 100, no.
6, pp. 704–712, [Online]. Available: (2003). https://www.scopus.com/inward/record.uri?eid=2-s2.0-0346243623&partnerID=40&
md5=3f8071840dea1f601ddc0f176bdb303b
11. Russo, G., Pauletta, M. & Venir, R. Reinforced concrete deep beams- Shear Strength Model and Design Formula. ACI Struct. J., 102,
3, https://doi.org/10.14359/14414
12. Vecchio, F. J. & Collins, M. P. The Modified Compression-Field Theory for Reinforced Concrete Elements Subjected to Shear,
318Reference, vol. 19, no. 16.
13. Tang, C. Y. & Tan, K. H. Interactive Mechanical Model for Shear Strength of Deep Beams, J. Struct. Eng. - J STRUCT ENG-ASCE,
vol. 130, doi: 10.1061/(ASCE)0733-9445(2004)130:10(1534). (2004).
14. A. C. I. C. 318, Building code requirements for structural concrete: (ACI 318 – 19) ; and commentary (ACI 318R-19). Farmington
Hills, MI: American Concrete Institute, [2019].
15. Hendy, C. R. & Smith, D. A. Designers’ Guide to EN 1992-2: Eurocode 2: Design of Concrete Structures: Part 2: Concrete Bridgesvol.
17 (Thomas Telford, 2007).
16. Schlaich, J., Schaefer, K. & Jennewein, M. Toward a consistent design of structural concrete. PCI J. 32 (3), 74–150. https://doi.
org/10.15554/pcij.05011987.74.150 (1987).
17. Russo, G., Venir, R., Pauletta, M. & Somma, G. Reinforced concrete corbels - Shear strength model and design formula, ACI Struct.
J., vol. 103, no. 1, pp. 3–10, [Online]. Available: (2006). https://www.scopus.com/inward/record.uri?eid=2-s2.0-32544444278&par
tnerID=40&md5=5ea8be15e672cd267a9207886dc8e81d
18. Shyh-Jiann, H. & Hung-Jen, L. Strength Prediction for Discontinuity Regions by Softened Strut-and-Tie Model, J. Struct. Eng., vol.
128, no. 12, pp. 1519–1526, Dec. doi: 10.1061/(ASCE)0733-9445(2002)128:12(1519). (2002).
19. Chetchotisak, P., Teerawong, J. & Yindeesuk, S. Modified interactive strut-and-tie modeling of reinforced concrete deep beams and
corbels, Structures, vol. 45, no. September, pp. 284–298, doi: (2022). https://doi.org/10.1016/j.istruc.2022.08.116
20. Megahed, K., Mahmoud, N. S. & Abd-Rabou, S. E. M. Application of machine learning models in the capacity prediction of RCFST
columns. Sci. Rep. 13 (1), 20878. https://doi.org/10.1038/s41598-023-48044-1 (2023).
21. Ben Chaabene, W. & Nehdi, M. L. Genetic programming based symbolic regression for shear capacity prediction of SFRC beams.
Constr. Build. Mater. 280, 122523. https://doi.org/10.1016/j.conbuildmat.2021.122523 (2021).
22. Megahed, K., Mahmoud, N. S. & Abd-Rabou, S. E. M. Prediction of the axial compression capacity of stub CFST columns using
machine learning techniques. Sci. Rep. 14 (1), 2885. https://doi.org/10.1038/s41598-024-53352-1 (2024).
23. Lu, W. Y., Lin, I. J. & Hwang, S. J. Shear strength of reinforced concrete corbels. Mag Concr Res. 61 (10), 807–813 (2009).
24. Zhang, N. & Tan, K. H. Direct strut-and-tie model for single span and continuous deep beams. Eng. Struct. 29 (11), 2987–3001.
https://doi.org/10.1016/j.engstruct.2007.02.004 (2007).
25. Hanoon, A. N., Jaafar, M. S., Al Zaidee, S. R., Hejazi, F. & Aziz, F. N. A. A. Effectiveness factor of the strut-and-tie model for
reinforced concrete deep beams strengthened with CFRP sheet, J. Build. Eng., vol. 12, no. June pp. 8–16, 2017, doi: (2016). https://
doi.org/10.1016/j.jobe.2017.05.001
26. Li, R., Deng, M., Zhang, Y. & Wei, D. Shear strengthening of reinforced concrete deep beams with highly ductile fiber-reinforced
concrete jacket, J. Build. Eng., vol. 48, no. December p. 103957, 2022, doi: (2021). https://doi.org/10.1016/j.jobe.2021.103957
27. Hwang, S. J., Lu, W. Y. & Lee, H. J. Shear strength prediction for deep beams, ACI Struct. J., vol. 97, no. 3, pp. 367–376, [Online].
Available: (2000). https://www.scopus.com/inward/record.uri?eid=2-s2.0-0034191709&partnerID=40&md5=740d3368c45f408c
b6472d616e219e3e
28. Koza, J. R. Genetic programming as a means for programming computers by natural selection. Stat. Comput. 4 (2), 87–112. https://
doi.org/10.1007/BF00175355 (1994).
29. Udrescu, S. M., Tegmark, M., Feynman, A. I. & A physics-inspired method for symbolic regression. Sci. Adv. 6 (16), eaay2631
(2020).
30. Goldberg, D. E. & Holland, J. H. Genetic algorithms and machine learning. Mach. Learn. 3 (2), 95–99. https://doi.
org/10.1023/A:1022602019183 (1988).
31. Cranmer, M. Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl, 2023, [Online]. Available: http://
arxiv.org/abs/2305.01582
32. Kani, G. How safe are our large reinforced concrete beams? in Journal Proceedings, vol. 64, no. 3, pp. 128–141. (1967).
33. Asteris, P. G. & Mokos, V. G. Concrete compressive strength using artificial neural networks. Neural Comput. Appl. 32, 11807–
11826. https://doi.org/10.1007/s00521-019-04663-2 (2020).
34. Wang, J., Lu, R. & Cheng, M. Application of ensemble model in capacity prediction of the CCFST columns under axial and
eccentric loading. Sci. Rep. 13 (1), 9488. https://doi.org/10.1038/s41598-023-36576-5 (2023).
35. Chen, H., Yi, W. J. & Ma, Z. J. Shear size effect in simply supported RC deep beams. Eng. Struct. 182, 268–278. https://doi.
org/10.1016/j.engstruct.2018.12.062 (2019).
36. Sakalauskas, K. & Kaklauskas, G. Pure shear model for crack width analysis of reinforced concrete members. Sci. Rep. 13 (1),
13883. https://doi.org/10.1038/s41598-023-41080-x (2023).
37. Gerges, N. N. et al. Flexural capacity of eco-friendly reinforced concrete beams. Sci. Rep. 13 (1), 20142. https://doi.org/10.1038/
s41598-023-47283-6 (2023).
38. Ahmad, I. & Shokouhian, M. Promoting Sustainable Green Infrastructure: Experimental and Numerical Investigation of Concrete
Reinforced with Recycled Steel Fibers, Arch. Adv. Eng. Sci., no. SE-Research Articles, pp. 1–13, Jun. doi: (2024). https://doi.
org/10.47852/bonviewAAES42022837
39. Bin Inqiad, W., Ali Raza, M. & Asim, M. Predicting 28-Day Compressive Strength of Self-Compacting Concrete (SCC) Using
Gene Expression Programming (GEP), Arch. Adv. Eng. Sci., no. SE-Research Articles, pp. 1–13, Nov. doi: (2023). https://doi.
org/10.47852/bonviewAAES32021606

Acknowledgements
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 15

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/

profit sectors.

Author contributions
K.M is repsonsable for all manscript.

Funding
Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooper-
ation with The Egyptian Knowledge Bank (EKB).

Competing interests
The author declares no competing interests.

Additional information
Supplementary Information The online version contains supplementary material available at https://doi.
org/10.1038/s41598-024-74803-9.
Correspondence and requests for materials should be addressed to K.M.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and
indicate if changes were made. The images or other third party material in this article are included in the article’s
Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included
in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or
exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy
of this licence, visit http://creativecommons.org/licenses/by/4.0/.

© The Author(s) 2024

Scientific Reports | (2024) 14:25066 | https://doi.org/10.1038/s41598-024-74803-9 16

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]