Electronic Journal of Structural Engineering

Original Article

Cite this: DOI:10.56748/ejse.233642 Machine learning-based optimum reinforced

Received Date: 22 September 2022 concrete design for progressive collapse
Accepted Date: 30 January 2023
Mohammad Javad Esfandiari a*, Homa Haghighi a and Girum Urgessa a
1443-9255 a
Department of Civil Engineering, George Mason University, VA, USA
https://ejsei.com/ejse * Corresponding Author: [email protected]
Copyright: © The Author(s).
Published by Electronic Journals for Abstract
Science and Engineering
International (EJSEI). This paper investigated progressive collapse analysis of three-dimensional (3D) reinforced concrete (RC)
This is an open access article under frames that are optimized for carrying structural loads by introducing a unique simultaneous multi-column
the CC BY license. removal using Machine Learning. The various load paths resulting from multiple-column removal are
https://creativecommons.org/license
incorporated in the optimization automatically. The investigation includes formulating an integrated
s/by/4.0/
computational framework that incorporates a self-training machine learning algorithm. The efficiency of the
algorithm is tested by using several hundreds of optimized structures. The efficiency of the computational
framework was shown by conducting a comprehensive study on the optimization and behavior of structures
considering seismic loading, alternative load path due to progressive collapse, and second order (P–delta) effects.
The results show that the proposed framework ensures that system solutions meet both structural integrity and
constructability requirements of the ACI and the Unified Facilities Criteria.

Keywords
Machine Learning, Structural Optimization, Progressive Collapse, Reinforced Concrete, Artificial intelligence

ability of robots. They concluded that integrating ML and optimization

1. Introduction significantly increases the quality of decision making and learning
capability in decision systems.
Local failure of structural members will result in increased internal As far as machine learning techniques, Jong‐Su Jeon et al. (Jeon et al.,
forces and overloading and may cause a progressive collapse of the entire 2014) proposed probabilistic joint shear strength models by implementing
or a part of a given structures. A few examples of progressive collapse (PC) ML. They used the prediction model as joint response models for evaluation
include the Murrah Building in Oklahoma City in 1995, the Twin Towers in of seismic performance and inelastic responses of frames. (Nick et al.,
New York City in 2001, and the Plasco Building in Tehran in 2017, and the 2015) have used different machine learning techniques for identifying
Hard Rock Hotel in New Orleans in 2019. PC results in economic losses and the existence and location of damage, and the type and severity of damage.
death of occupants. Hence, structures must be designed to better withstand Ni Hong-Guang and Wang Ji-Zong (Ni and Wang, 2000) used a multi-layer
progressive collapse. feed-forward neural network and presented a method for predicting 28-
The number of PC studies consisting of experimental programs and day compressive strength of concrete. Dac-Khuong Bui et al. (Bui et al.,
numerical studies have increased significantly since 2001. Byfield et al 2018) developed a model for determining the tensile strength of High-
(Byfield et al., 2014), Wang et al (Wang et al., 2014) and Qian and Li (Qian Strength Concrete. They selected neural network for their research due to
and Li, 2015) provided detailed reviews on this topic. the nonlinear relation between concrete strength and its components.
As far as numerical studies are concerned, various representative In this paper, classification techniques are used because the
models are found in literature (Bao et al., 2008; Talaat and Mosalam, 2007; algorithm needs to correctly determine the class labels for unseen instances
Buscemi and Marjanishvili, 2005) including computational FEM models for on the basis of previously observed optimum structural system data and
determining the response of structures before failure with a reasonable suggests a specific class to the optimization for further constraints
accuracy and DEM models that are more effective for moving and collision handlings. The algorithm starts to run with a few training data and as it is
between rigid bodies after failure. Marchand et al (Marchand and Stevens, applied on different structures, it can consider the result for its future
2015) argued that common structural design software (e.g., SAP2000 and training data. As the training data becomes more available, the performance
ETABS) can provide reasonable results than those determined with high- improves dramatically over time. Different classification techniques such as
fidelity physics-based software. Esfandiari et al (Esfandiari and Urgessa, decision tree, SoftMax, and nearest neighbors are presented.
2018) presented a non-linear time history pull-down of a two-span steel
frame in ETABS and showed that the numerical results were in excellent 2. Structural optimization problem for
agreement with experimental results.
Machine Learning (ML) and Artificial Intelligence (AI) applications in reinforced concrete frames
recent years have opened new opportunities for use in traditional
Esfandiari et al (Esfandiari et al., 2018b) presented a general structural
engineering problems. Machine learning mostly deals with problems where
optimization problem as shown in Equation 1.
paired examples, X → Y, exist. For this paper, X and Y can be interpreted as
features of the structure and the optimum solution, respectively. As such,
the machine learning technique attempts to map G: X → Y ̂ with the 𝑥𝑥 = [𝑥𝑥1 , … , 𝑥𝑥𝑛𝑛 ]𝑇𝑇
𝑔𝑔(𝑥𝑥) ≤ 0, 𝑘𝑘 = 1, … , 𝑚𝑚
translated domain Y ̂ distributed identically to Y. However, in the field 𝑥𝑥 𝐿𝐿 ≤ 𝑥𝑥 ≤ 𝑥𝑥 𝑈𝑈 (1)
structural engineering, the main challenge is the lack of adequate optimum
data that is needed to train the algorithm. Therefore, seeking an algorithm
that could learn to map between domains without paired input-output where x is a vector of n structural system variables, 𝑓𝑓(𝑥𝑥): 𝑅𝑅 𝑛𝑛 → 𝑅𝑅 is the
scenarios is crucial. objective function which returns a scalar value to be minimized, the vector
Zue et al. (Zhu et al., 2017) used Cycle-Consistent Adversarial Networks function 𝑔𝑔(𝑥𝑥): 𝑅𝑅 𝑛𝑛 → 𝑅𝑅 𝑚𝑚 returns a vector of length m containing the values
in image-to-image transformation for learning to map between an input and of the inequality constraints evaluated at x, and 𝑥𝑥 𝐿𝐿 , 𝑥𝑥 𝑈𝑈 are two vectors of
output image when paired examples are absent. They simultaneously length n containing the lower and upper bounds of the structural system
trained G: X → Y and another translator F: Y→ X with a cycle consistency variables, respectively. The above equation contains only inequality
loss such that F(G(x)) ≈ x and G(F(y)) ≈ y. However, in structural analysis, it constraints because equality constraints are usually not found in structural
is important to create an alternative algorithm for capturing the complex optimizations.
system behavior where few training paired data are available. Equation 2 shows a common constraint k in a structural optimization
Sra et al. (Sra et al., 2012) showed that learning from available dataset problem.
combined with optimization is applicable to a myriad of complex, dynamic,
and stochastic problems. Mosavi et al. (Mosavi and Varkonyi-Koczy, 2017) 𝑔𝑔𝑘𝑘 (𝑥𝑥) = |𝑞𝑞𝑘𝑘 (𝑥𝑥)| − 𝑞𝑞𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎,𝑘𝑘 (2)
combined machine learning with optimization to increase the learning
 where 𝑞𝑞𝑘𝑘 (𝑥𝑥) is a response measure for analysis and design x; and
𝑞𝑞𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎,𝑘𝑘 is its maximum allowable absolute value. 𝑃𝑃(𝐴𝐴)𝑃𝑃(𝐴𝐴)
𝑃𝑃(𝐵𝐵) = (7)
The objective function in structural optimization problems is generally 𝑃𝑃(𝐵𝐵)
defined by the weight or total cost of the structure (Tahmouresi et al.,
2021). When considering total cost, the concrete, steel and labor costs are
included. However, when the objective function is the total weight, only the
weight of concrete and steel are included. Equation 3 shows the resulting
objective function.

𝑓𝑓𝑂𝑂𝑂𝑂𝑂𝑂 = 𝑂𝑂𝑂𝑂𝑂𝑂𝑐𝑐 + 𝑂𝑂𝑂𝑂𝑂𝑂𝑠𝑠 + 𝑂𝑂𝑂𝑂𝑂𝑂𝑓𝑓 (3)

where 𝑂𝑂𝑂𝑂𝑂𝑂𝑐𝑐 , 𝑂𝑂𝑂𝑂𝑂𝑂𝑠𝑠 , 𝑎𝑎𝑎𝑎𝑎𝑎 𝑂𝑂𝑂𝑂𝑂𝑂𝑓𝑓 are the costs of concrete, reinforcing steel
bars, and labor, respectively.
Equations 4-6 show the costs of each component when the objective
function is the total cost of a structural frame.

𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐 𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏

𝑂𝑂𝑂𝑂𝑂𝑂𝑐𝑐 = 𝐶𝐶𝐶𝐶 �� 𝑏𝑏𝑖𝑖 . 𝑑𝑑𝑖𝑖 . 𝐿𝐿𝑛𝑛 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐,𝑖𝑖 + � 𝑏𝑏𝑤𝑤 𝑗𝑗 . ℎ𝑗𝑗 . 𝐿𝐿𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 𝑗𝑗 � (4)
𝑖𝑖=1 𝑗𝑗=1
𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐 𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏,𝑖𝑖 𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐 𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡,𝑖𝑖 PSO
𝑂𝑂𝑂𝑂𝑂𝑂𝑠𝑠 = 𝐶𝐶𝑠𝑠 . 𝛾𝛾𝑠𝑠 . �� � 𝐴𝐴𝑆𝑆𝑆𝑆 𝑗𝑗 . 𝐿𝐿𝑏𝑏𝑏𝑏𝑏𝑏 𝑗𝑗 + � � 𝐴𝐴𝑆𝑆ℎ 𝑘𝑘 . 𝐿𝐿𝑡𝑡𝑡𝑡𝑡𝑡 𝑘𝑘
𝑖𝑖=1 𝑗𝑗=1 𝑖𝑖=1 𝑘𝑘=1
𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏,𝑚𝑚

+ � � 𝐴𝐴𝑆𝑆𝑆𝑆 𝑙𝑙 . 𝐿𝐿𝑏𝑏𝑏𝑏𝑏𝑏 𝑙𝑙 (5)

𝑚𝑚=1 𝑙𝑙=1
𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐 𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡,𝑚𝑚

+ � � 𝐴𝐴𝑆𝑆ℎ 𝑛𝑛 . 𝐿𝐿𝑏𝑏𝑏𝑏𝑏𝑏 𝑛𝑛 �
𝑚𝑚=1 𝑛𝑛=1

𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐

𝑂𝑂𝑂𝑂𝑂𝑂𝑓𝑓 = 𝐶𝐶𝑓𝑓 �� �2(𝑏𝑏𝑖𝑖 + ℎ𝑖𝑖 ) . 𝐿𝐿𝑛𝑛 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐,𝑖𝑖 �

𝑖𝑖=1
𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏

+ � ��𝑏𝑏𝑤𝑤 𝑗𝑗 + 2ℎ𝑗𝑗 � . 𝐿𝐿𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 𝑗𝑗 � (6)

𝑗𝑗=1
𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐

− � 𝑏𝑏𝑘𝑘 + 𝑑𝑑𝑘𝑘 �
𝑘𝑘=1
where 𝑁𝑁𝑐𝑐𝑐𝑐𝑐𝑐 , 𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 , 𝑏𝑏, 𝑑𝑑, 𝑏𝑏𝑤𝑤 , ℎ, 𝐿𝐿, and 𝐿𝐿𝑛𝑛 are the number of columns,
the number of beams, the width of column, the depth of column, the width DM
of beam, the height of beam, the length of the members, and the length of (AI and ML)
clear span between supports, respectively; 𝐶𝐶𝐶𝐶 , 𝐶𝐶𝑓𝑓 and 𝐶𝐶𝑠𝑠 are unit cost of the
concrete, the labor and the steel, respectively; 𝐴𝐴𝑠𝑠𝑠𝑠 , 𝐿𝐿𝑏𝑏𝑏𝑏𝑏𝑏 , and 𝑁𝑁𝑏𝑏𝑏𝑏𝑏𝑏 are the
DMPSO
area, the length and the number of main rebars placed in the member while
𝐴𝐴𝑠𝑠ℎ , 𝐿𝐿𝑡𝑡𝑡𝑡𝑡𝑡 , and 𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡 are the area, the length and the number of shear
reinforcement used in the member respectively; and 𝛾𝛾𝑠𝑠 is the density of
rebars (kg/m3). Fig 1. PSO versus DMPSO

3. Proposed DMPSO-ML algorithm The probability of geometry and loading for a given member
specification can be determined from previously analyzed optimum
3.1 Overview structures. Accordingly, DM implements machine learning to find the most
probable structural member according to the geometry, loading condition
Figure 1 shows the basic concept of the particle movement in the and location of the member. However, for incorporating dependencies,
traditional Particle Swarm Optimization method (PSO) versus the proposed Bayes’ theorem needs fundamental assumptions about dependence and
Decision-Making Particle Swarm Optimization method (DMPSO). As shown independence between system variables, and determining the marginal in
in the figure, each particle in PSO only searches for the best solution the Bayes’ theorem is computationally expensive. For that reason,
according to its own best experience, and the best solution is determined alternative machine learning methods can be cost effective and are
by all particles (Kennedy and Eberhart, 1995). However, in structural investigated in this paper.
optimization, it is important to ensure that the structure is stable and safe. The decision maker algorithm can also disregard a solution at any given
An experienced structural engineer can decide the sort of alterations in the time of the computation process when it determines that a better fitness
system variables that could lead to a preferable solution. As an example, if cannot be obtained. The principle of pruning from AI was adopted in this
the demand to capacity ratio of a member is greater than one, the paper, which allows the DM algorithm to ignore portions of the search space
acceptable solution for addressing this issue may be changing the cross- or analysis that make no difference to the final choice. The heuristic
sectional sections of the member or the rebar reinforcement ratio of the evaluation function allows us to estimate the objective function without
section. However, a human decision maker cannot be available and actively doing a complete analysis. When pruning is applied to a standard search
participate in the solution process and direct it according to the preferences tree, it returns the same move as a search would, but it prunes away
in the entire process of an optimization. The decision maker (DM) algorithm branches that cannot possibly influence the final decision. This requires
acts similar to an experienced structural engineer. examining first the successors that are likely to be the optimum solution.
To equip the DM with machine learning power, different ML methods
More details regarding how the DM is formulated and operates, and
were investigated in this paper.
how it is fused with PSO, are discussed in the following subsections.
The first step was the collection and preparation of the training data
set. Since there was no training data available at the beginning, a small sub-
3.2 DMPSO Algorithm Enhanced with ML set of structures was selected, and the optimization algorithm was used for
producing the training set. The small sub-set of structures was later used to
DMPSO uses an informed strategy and the knowledge beyond the produce more complex training sets. Then, the behavior of 640 more
definition of the problem itself, to empower PSO optimization algorithm complex structures were considered as the training data. For each
and accelerate convergence toward the optimum solution. The DM structure, the structure was separately optimized for 10 random column
formulation in this paper, which is inspired by Bayes’ theorem, seeks the removal scenarios for progressive collapse analysis. The goal was to classify
probability of a member not failing given the geometry and loading the best cross section for the elements under different loading conditions.
application. Bayes’ theorem is stated mathematically by Equation 7. For this purpose, three separate machine learning (ML) models were
trained. Table 1 provides the specifics of these ML models.

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

2
Table 1. Detail of the ML models For the beams, cross sectional dimensions and the location of the rebar
were considered as variables. The corresponding reinforcement amount
Model 1 Model 2 was calculated and placed later based on strength requirements.
Number of bays in The section restrictions did not affect the final results because the initial
Element type guess from ML is fed to the optimization algorithm with no restriction on its
each direction
variable. In other words, these initial guesses show the preference of the DM to
Maximum bay span Maximum adjacent bays length at the
in each direction element the optimization algorithm and guide it to find a better solution.
Number of stories Moment and shear of the element For identifying the best ML technique that will be integrated with the
Features optimization algorithm, the following classifiers were investigated.
Dead load Top and bottom connected beams,
The number of stories above the
Live load
element
3.3 K-Nearest Neighbor (KNN)
The number of stories below the The K-nearest-neighbor (KNN) density estimation method (Paya et al.,
Seismic parameters
element. 2008) was applied to each class followed by employing Bayes’ theorem.
Initial randomized Consider a data set comprising Nk members’ cross section in class Ck with
Optimum separate elements based on
Usage section at the a total of N points. To classify a new structural element x, a sphere centered
conventional loading
beginning on x can be drawn in the feature space, precisely holding K elements’ cross
section regardless of their class. Suppose this sphere contains Kk member
Model 1 related the final result of the average size of optimum elements from class Ck. The posterior probability of selecting a structural element’s
in each story to the overall geometric feature of the structure, including cross section can be obtained by applying the Bayes’ theorem as shown in
number of bays in each direction, maximum bay span in each direction, Equation 8.
number of stories, dead load, live load and seismic parameters. This model
was only used in the first iteration to generate initial randomized sections. 𝑃𝑃(𝐶𝐶𝑘𝑘 )𝑝𝑝(𝐶𝐶𝑘𝑘 ) 𝐾𝐾𝑘𝑘
The second and third models connected the output of the optimum 𝑝𝑝(𝐶𝐶𝑘𝑘 ) = = (8)
𝑝𝑝(𝑥𝑥) 𝐾𝐾
separate elements to its learning feature for the whole structure and
progressive collapse removal scenario cases, respectively. The learning
features considered include element type, actual bay length, moment and
shear of the element before and after removal scenarios, bottom and top In the KNN formulation, the cross section with the largest posterior
connected beams, and the number of stories above and below the removed probability should be assigned to element x to minimize the probability of
element. misclassification. When the algorithm needs to find an appropriate cross
To streamline the problem for ML and to avoid overfitting, the class of section for a new element, it can identify the K nearest similar members
the sections for columns and beams were restricted to 8 and 6 sections, from the training data set of optimum structures and then assign a cross
respectively as shown in Fig. 2 and Fig. 3. Rectangular cross-sections were section with the average of variables from the KNN. It is important to find
deliberately considered with 100mm difference in width and height to have the best K for the problem. Small Ks result in many small regions of each
the best arrangement of the classes covering most of the practical results. class and make the model more biased. On the other hand, large Ks led to
fewer larger regions which may affect the final result.

3.4 SoftMax Classifier

SoftMax classifier (Duan et al., 2003) is the generalization of binary
Logistic Regression classifier to multiple classes. Softmax classifier uses a
linear classifier for mapping and generating scores as the unnormalized log
probabilities with cross-entropy loss having the form shown in Equation 9.

𝑒𝑒 𝑓𝑓𝑦𝑦𝑦𝑦
𝐿𝐿𝑖𝑖 = log � � (9)
∑𝑗𝑗 𝑒𝑒 𝑓𝑓𝑗𝑗

Where fj is the j-th element of the vector of class scores f. The use of the
exponential scores gives the unnormalized probabilities, and the division
for normalization purpose. This will ensure that the sum of the probabilities
is one. The stochastic gradient descent was used for training. Here the best
section that has the highest probability for the corresponding element was
sought. The data was trained in 16 mini batches.
Fig 2. Different section classes for columns in the ML
classification problem 3.5 Decision Tree Classifier
Decision tree classification algorithms (Safavian and Landgrebe, 1991)
have a significant potential for a variety of problems and have been used in
civil engineering applications. There are different measures that can be
utilized to determine the best way to split between classes. Gini index and
entropy were used for selecting the best split based on the degree of
impurity of the child nodes. Binary decision tree with 5 and 6 depths were
tested.

4. Metrics and performance evaluation

The data set used here was divided into 3 different groups: training,
validation, and testing. The training dataset was used to train models with
various hyper parameter values. Then the validation dataset was used to
identify the best working parameters. To validate the model, a 5-fold cross
validation was used. After that, the training and validation datasets were
used to train the final model.
Evaluation presented a major challenge. Consider a single column
under pure compression loading. This column can be designed with
different cross sections that can satisfy the stability requirements of the
structure. Since the optimization does not warrant the global optimum in a
complex system, it accepted near-optimum solutions in the evaluation.
Moreover, the element classes of the final result did not exactly match the
output of the machine learning classes. For the machine learning outputs,
Fig 3. Different section classes for beams in the ML classification only 8 classes were considered for the columns and 6 classes for the beams,
problem while the final result of the optimum structure does not have any restriction
and the dimensions of the elements might change through the optimization
process. Therefore, if the structural requirements were only checked, all

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

3
over designed solutions would pass the evaluation criteria. On the other control the exploration capabilities of the swarm. Vector 𝑥𝑥 𝑃𝑃𝑃𝑃,𝑗𝑗 denotes the
hand, if the optimum solution was only checked, the accuracy would be very personal best position which is registered by particle 𝑗𝑗, vector 𝑥𝑥 𝐺𝐺𝐺𝐺,𝑗𝑗 is the
low. This issue was addressed by finding the nearest neighbor of the actual global best position attained by the entire swarm up to the current iteration,
sections of the structure with class samples of one increment threshold for and vector 𝑥𝑥 𝐷𝐷𝐷𝐷,𝑗𝑗 indicates the position of preference of the decision maker
accepting the result. in the search space. The acceleration coefficients 𝑐𝑐1 and 𝑐𝑐2 , and 𝑐𝑐3 rule the
Finding the final optimum solution was not the goal of this initial step impact of the particle's own experiences, the other particles' experiences,
but rather keeping the variables within an acceptable range of initial and the decision maker’s preference on the trajectory of each particle,
guesses. Later the algorithm would find the best optimum solution. respectively. 𝑟𝑟1 , 𝑟𝑟2 , 𝑟𝑟3 are three random vectors with uniformly distributed
Therefore, different hyperparameters and methods were investigated, and numbers in the interval [0, 1]. The symbol “⨀” is the element-wise product
the best parameters were selected for integrating it with the optimization of two vectors. The two acceleration coefficients (c1 and c2) and the
algorithm. The comparison of the results is shown in Fig. 4. smallest and largest value of inertial factor (wmin and wmax) were taken
as 2.025, 2.025 and 0.4, 0.9, respectively (Alam et al., 2015; Eberhart and
Shi, 2000). The acceleration coefficient for DM, c3, was initially taken as
2.025 (the same as c1 and c2) and decreased over iterations. Therefore, for
the first few iterations, the algorithm mostly relies on the DM. To ensure the
functionality, if the demand to capacity ratio of a member is not within 50%
in the optimization process, the DM algorithm suggests its preference, such
as increasing or decreasing a relevant parameter, to the DMPSO algorithm.
The DM algorithm gathers statistics from a database of previously analyzed
structures to determine members most often lead to an optimum structure. In the
early iterations, there were a few choices among the large number of possible
variables. Thus, the DM commentary based on past structures has a higher
impact on DMPSO. Usually after the first 100 iterations, the DMPSO algorithm
mostly relies on optimization rather than the DM preference.

5. Incorporating progressive collapse in

DMPSO-ML
In progressive collapse analysis, multiple scenarios of removing critical
Fig 4. Box plot of different hyper parameters members should be considered, which drives the structural system and
cross-section selection to be tedious and costly. Therefore, investigating the
The number and size of rebars in each direction of the concrete section formulation of a computational framework is important for producing d
presented a challenge because specific required rebar area could be placed cost-effective solutions. This was achieved by a series of steps. First, a finite
in several arrangements. As an example, 450 mm2 of required rebar area element model capable of accurately modeling new load paths to
can be satisfied by using 6-D10, 4-D12, or 2-D25 bars. The results can be progressive collapse analysis was developed. Then, the finite element
improved by only considering the cross-sectional dimension parameters model was integrated with DMPSO to automatically evaluate structural
and the overall required area of the rebar for the section with 10 percent response for progressive collapse.
threshold for the columns and the beams. Later, when this section is ready There are two groups of constraints needed for expanding the
to be fed to the optimization algorithm, the overall required area would be optimization problem and incorporating progressive collapse. These
converted to the best arrangement for the rebars size and numbers in each include general concrete structural system constraints and progressive
direction, for that specific section. This approach dramatically improved the collapse (UFC) constraints. The first group includes parameters typical of
findings obtained as shown in Fig. 5. structural systems subjected to traditional lateral loads such as plastic
rotations.
The second group includes those related to progressive collapse as
defined in UFC (Gsa, 2003; Defense, 2005) and GSA. These constraints
ensure that the structure is capable of bridging over critical vertical load-
carrying elements that are eliminated during a progressive collapse
scenario such as redundancy requirements.
The integrated framework of the DMPSO algorithm that is empowered by
ML is shown schematically in Fig. 6 and Fig. 7.

Fig 5. Box plot when only cross-sectional dimensions were

considered as variables
The KNN-5 method resulted in the highest accuracy compared to all the
techniques investigated. The DMPSO then modifies its velocity and position
according to its experience, neighboring particles’ experience, and the
preference of the decision maker (DM) by employing Equation 10 and
Equation 11 (Esfandiari et al., 2018a).

𝑣𝑣 𝑗𝑗 (𝑡𝑡 + 1) = 𝑤𝑤𝑣𝑣 𝑗𝑗 (𝑡𝑡) + 𝑐𝑐1𝑟𝑟1⨀ �𝑥𝑥 𝑃𝑃𝑃𝑃,𝑗𝑗 − 𝑥𝑥 𝑗𝑗 (𝑡𝑡)� + 𝑐𝑐2𝑟𝑟2 ⊙ �𝑥𝑥 𝐺𝐺𝐺𝐺,𝑗𝑗 − 𝑥𝑥 𝑗𝑗 (𝑡𝑡)�
+ 𝑐𝑐3𝑟𝑟3 ⊙ �𝑥𝑥 𝐷𝐷𝐷𝐷,𝑗𝑗 − 𝑥𝑥 𝑗𝑗 (𝑡𝑡)�
(10)

𝑥𝑥 𝑗𝑗 (𝑡𝑡 + 1) = 𝑥𝑥 𝑗𝑗 (𝑡𝑡) + 𝑣𝑣 𝑗𝑗 (𝑡𝑡 + 1) (11)

Where 𝑣𝑣𝑗𝑗 (𝑡𝑡) 𝑎𝑎𝑎𝑎𝑎𝑎 𝑥𝑥 𝑗𝑗 (𝑡𝑡) represent the velocity and the position vectors Fig 6. Modified DMPSO for Progressive Collapse
of particle 𝑗𝑗 at time 𝑡𝑡, respectively. The term 𝑤𝑤 is a modifier employed to

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

4
 6. Practical example on progressive
collapse analysis of 3D optimized RC
frames.
The proposed DMPSO-ML framework was used to analyze a given 7-
story RC frame with three spans whose geometry, grouping details, and
removal scenarios are shown in Fig. 8. A total of sixteen unique column
removal scenarios were implemented in the alternate path investigation of
this example frame. Although a symmetric plan was used, sections A and C
differ in size. This implies that one set of eight removal scenarios, as shown
in Fig. 8, is considered in each of these sections. These columns eliminated
are located at the corner and middle of each direction of the members in the
first story above the grade, the story directly below the roof, and the story
above the location of a column splice at every other floor. The objective
function here is the construction weight similar to a baseline optimized
structure for conventional loadings without considering progressive
collapse (Esfandiari et al., 2018b). The frame includes 180 members, 84
beams and 96 columns, which were arranged into 42 groups: 28 groups for
beams and 14 groups for columns. It contains 574 system variables, 504
were for beams and 70 were for columns. Beams and columns were
grouped to satisfy the uniformity of members and having similar behaviors
according to their place in the frame and loading conditions. To ensure best
results for the stochastic decline, a population size of 150 was selected.
Table 2 shows the optimal frame systems from the present progressive
collapse optimization, considering all of the given removal scenarios Both
linear static (LS) and nonlinear dynamic (ND) methods were considered.
Furthermore, the result of the optimal frame analysis without including
progressive collapse is also presented as the baseline to depict the changes
made in the structural member and reinforcing steel sizes of optimal results
when compared to current integrated progressive collapse analysis.
Fig. 9 shows DMPSO-ML algorithm evolutions for obtaining the
solutions. DMPSO did not confine local values and carried on
converging. As validated, the baseline structure without considering the
progressive collapse requirements converged to the optimum results in
smaller number of structural analysis and flatten out within less
generations. This is expected because for progressive collapse analysis
DMPSO has to confirm the constraints in two steps. In the first step,
constraints related to traditional lateral systems (e.g., seismic)
requirements were checked (Randall W. Poston and Basile G. Rabbat,
Fig 7. Flowchart of decision maker empowered by Machine 2011). If the criteria are not met here, it proceeds to the Decision Maker
Learning (ML) (DM) portion of the algorithm for adjusting parameters. Otherwise, it
proceeds to the second step for checking progressive collapse
requirements. The results are not registered until every constraint
consisting of both lateral and progressive collapse requirements are met.

Fig 8. Geometry, member classification, and removal scenarios

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

5
Thus, the need for evolutionary generation to handle these two types of Bao Y, Kunnath SK, El-Tawil S, et al. (2008) Macromodel-Based
constraints is evident in the progressive analysis case. Simulation of Progressive Collapse: RC Frame Structures. Journal of
Fig. 10 illustrates the demand to capacity ratio (DCR). The demand Structural Engineering 134: 1079-1091.
capacities for LS and ND were calculated for regular loading after the Bui D-K, Nguyen T, Chou J-S, et al. (2018) A modified firefly algorithm-
structure was totally designed for progressive collapse. This implies that artificial neural network expert system for predicting compressive and
after the process of removing elements are completed and the optimum tensile strength of high-performance concrete. Construction and Building
result was obtained, the DCR is calculated using those sections. Bigger Materials 180: 320-333.
sections were used in LS and ND when compared to the structure Buscemi N and Marjanishvili S. (2005) SDOF Model for Progressive
considered without progressive collapse. As a result, the DCR of elements Collapse Analysis. Structures Congress 2005. 1-12.
in LS is the least among all the methods. Nevertheless, all the DCRs obtained Byfield M, Mudalige W, Morison C, et al. (2014) A review of progressive
were above 0.61, which shows that the algorithm can obtain acceptable collapse research and regulations. Proceedings of the Institution of Civil
results. Engineers - Structures and Buildings 167: 447-456.
Note that in the ML algorithm, the results are obtained instantaneously Defense Do. (2005) Unified Facilities Criteria: Design of Buildings to
given data is used to train the algorithm and the pretrained model is used Resist Progressive Collapse (UFC 4-023-03). Department of Defense
for prediction. The optimization piece requires close to 12 hours to run Washington, DC, USA.
because each iteration takes around 10 seconds. Duan K, Keerthi SS, Chu W, et al. (2003) Multi-category Classification by
Soft-Max Combination of Binary Classifiers, Berlin, Heidelberg: Springer
Berlin Heidelberg.
Eberhart RC and Shi Y. (2000) Comparing inertia weights and
constriction factors in particle swarm optimization. Proceedings of the 2000
Congress on Evolutionary Computation. CEC00 (Cat. No.00TH8512). 84-88
Esfandiari M and Urgessa G. (2018) A Pull-down Dynamic Analysis of
Two-Span Steel Frames Subjected to Progressive Collapse.
Esfandiari MJ, Urgessa GS, Sheikholarefin S, et al. (2018a) Optimization
of reinforced concrete frames subjected to historical time-history loadings
using DMPSO algorithm. Structural and Multidisciplinary Optimization 58:
2119-2134.
Esfandiari MJ, Urgessa GS, Sheikholarefin S, et al. (2018b) Optimum
design of 3D reinforced concrete frames using DMPSO algorithm. Advances
in Engineering Software 115: 149-160.
Gsa U. (2003) Progressive collapse analysis and design guidelines for
new federal office buildings and major modernization projects. Washington,
DC.
Fig 9. Convergence rate from DMPSO, based on LS, ND compared Jeon J-S, Shafieezadeh A and DesRoches R. (2014) Statistical models for
to the system without considering progressive collapse. shear strength of RC beam-column joints using machine-learning
techniques. Earthquake Engineering & Structural Dynamics 43: 2075-2095.
Kennedy J and Eberhart R. (1995) Particle swarm optimization.
Proceedings of ICNN'95 - International Conference on Neural Networks.
Perth, WA, Australia: IEEE, 1942-1948
Marchand KA and Stevens DJ. (2015) Progressive Collapse Criteria and
Design Approaches Improvement. Journal of Performance of Constructed
Facilities 29: B4015004.
Mosavi A and Varkonyi-Koczy AR. (2017) Integration of Machine
Learning and Optimization for Robot Learning. Advances in Intelligent
Systems and Computing Recent Global Research and Education:
Technological Challenges, 519: 349-355.
Ni H-G and Wang J-Z. (2000) Prediction of compressive strength of
concrete by neural networks. Cement and Concrete Research 30: 1245-1250.
Nick W, Asamene K, Bullock G, et al. (2015) A study of machine learning
techniques for detecting and classifying structural damage. International
Fig 10. Maximum DCR of members for each analysis method Journal of Machine Learning and Computing 5: 313.
Paya I, Yepes V, González-Vidosa F, et al. (2008) Multiobjective
Optimization of Concrete Frames by Simulated Annealing. Computer-Aided
Civil and Infrastructure Engineering 23: 596-610.
7. Conclusion Qian K and Li B. (2015) Research Advances in Design of Structures to
Resist Progressive Collapse. Journal of Performance of Constructed Facilities
This paper presented the integration of optimization and progressive 29: B4014007.
collapse analysis computational framework empowered by machine Randall W. Poston and Basile G. Rabbat. (2011) Building Code
learning. The main objective was to evaluate the behavior of reinforced Requirements for Structural Concrete (ACI 318-11): American Concrete
concrete structures while satisfying the limitations and specifications of the Institute.
American Concrete Institute 318 code and Unified Facilities Criteria Safavian SR and Landgrebe D. (1991) A survey of decision tree classifier
progressive collapse requirements. Three machine learning algorithms, K methodology. IEEE Transactions on Systems, Man, and Cybernetics 21: 660-
nearest neighbor, SoftMax, and decision tree classifiers were evaluated. The 674.
KNN machine learning algorithm provided better performance to alleviate Sra S, Nowozin S and Wright SJ. (2012) Optimization for machine
the computational challenge for the structural optimization problem learning, Cambridge, Massachusetts, London, England: Mit Press.
involving progressive collapse. Tahmouresi, A., Robati, A. et al. (2021) A Combined Genetic Algorithm-
Using a case study, the analysis was shown to enhance load Artificial Neural Network Optimization Method for Mix Design of Self
redistribution capability of the structure by considering the alternate path Consolidating Concrete, International Journal of Structural and Civil
criteria through finding appropriate structural member sizes. Meanwhile, Engineering Research, Vol. 10, No. 3, doi: 10.18178/ijscer.10.3.106-112.
the cost of the example frame analyzed reduced substantially through the Talaat M and Mosalam KM. (2007) Towards Modeling Progressive
optimization process. The case study showed the capability of the DMPSO- Collapse in Reinforced Concrete Buildings. Research Frontiers at Structures
ML algorithm to accelerate convergence toward the optimum system Congress Long Beach, California, United States: American Society of Civil
solutions while reducing computational effort. Engineers, 1-16.
Wang H, Zhang A, Li Y, et al. (2014) A review on progressive collapse of
building structures. The Open Civil Engineering Journal 8: 183-192.
References Zhu J-Y, Park T, Isola P, et al. (2017) Unpaired Image-to-Image
Translation using Cycle-Consistent Adversarial Networks. (Accessed March
Alam MN, Das B and Pant V. (2015) A comparative study of 01, 2017).
metaheuristic optimization approaches for directional overcurrent relays
coordination. Electric Power Systems Research 128: 39-52.

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

6
Appendix
Table 2. Results of optimum systems based on LS; ND compared to the system without considering progressive collapse.
LS ND DMPSO
Sectional Sectional Sectional
Reinforcements Reinforcements Reinforcements
dimensions dimensions dimensions
Height

Width

Height

Width

Height

Width
Top -rebar
Type Top-rebar Bot-rebar Bot -rebar Top -rebar Bot -rebar

5-D16 6-D16 6-D16 3-D12 5-D16 5-D12 5-D12 3-D12 4-D14 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T1 350 350 3-D12
3-D14
4-D12 4-D12 1-D16 3-D12 350 350 4-D12
3-D14
3-D14 3-D14 4-D16 3-D14 300 300 3-D14
3-D12
3-D14 3-D14 1-D16 3-D14
3-D16 5-D16 3-D16 4-D16 5-D12 5-D12 4-D12 3-D14 3-D14
T2 400 350 3-D16
3-D14
5-D16 3-D16
6-D14
4-D16 350 350 5-D12
3-D14
5-D12
5-D14
4-D14
5-D14 350 350 3-D14
3-D14
3-D14
3-D14 3-D14 3-D14
4-D14 4-D14 5-D14 3-D12 6-D14 4-D16 3-D16 2-D12 3-D12 3-D12 3-D12
T3 350 350 4-D12
3-D12
4-D12 4-D12 1-D16 4-D12 350 300 2-D14
3-D12
3-D14
3-D12
2-D12
3-D12 300 300 4-D16
3-D12
4-D16
3-D12
1-D16
3-D12
5-D16 4-D16 2-D16 2-D16 4-D14 3-D14 5-D14 3-D14 3-D14
T4 350 350 4-D14
3-D14
3-D14 2-D14
5-D14
3-D14 350 350 4-D14
3-D14
3-D14
5-D14
5-D12
4-D14 350 350 2-D22
3-D14
2-D22
4-D14 4-D14 4-D14
5-D16 5-D16 3-D16 3-D12 2-D16 5-D14 5-D14 3-D12 4-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T5 400 300 4-D14
3-D12
4-D14 1-D14 1-D16 2-D14 350 350 4-D12
3-D12
4-D12 2-D16 3-D12 3-D14 300 300 3-D20
3-D12
3-D20 1-D14 1-D16 1-D14
4-D16 5-D16 3-D16 3-D16 5-D14 4-D14 5-D16 4-D14 3-D14 3-D14
T6 400 350 5-D12
3-D14
3-D12 2-D12
4-D14
4-D12 400 350 5-D14
3-D14
4-D14 3-D12 2-D12
6-D16 350 350 2-D22
3-D14
2-D22
4-D14 4-D14 4-D14
9-D14 5-D14 5-D14 3-D12 4-D14 4-D16 3-D16 4-D16 3-D12 3-D12 3-D12
T7 450 350 9-D14
3-D14
5-D14 3-D12 1-D16 4-D12 400 300 3-D14
3-D12
3-D14
6-D12
4-D16
5-D12 300 300 2-D22
3-D12
2-D22
3-D12
1-D16
3-D12
4-D12 5-D12 5-D12 3-D12 5-D12 5-D12 4-D12 3-D12 2-D14 3-D12 3-D12 3-D12 3-D12 3-D12
T8 400 300 5-D14
3-D14
6-D14 2-D14 4-D16 2-D14 350 300 5-D12
3-D12
4-D12 1-D14 3-D14
4-D14 300 300 1-D16
3-D12
1-D16 1-D10 4-D16 1-D10
3-D16 3-D16 3-D16 3-D14 3-D16 3-D14 3-D14 3-D16 3-D14 3-D14 3-D14
T9 300 300 2-D12
3-D14
2-D12 2-D12 1-D12 2-D12 300 300 2-D12
3-D14
2-D12
3-D14
3-D14
3-D14 300 300 3-D22
3-D14
3-D22
3-D14
1-D12
3-D14
3-D16 4-D16 2-D16 3-D12 4-D16 2-D12 2-D12 3-D12 2-D16 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T10 350 300 2-D14
3-D12
4-D14 2-D14 3-D18 3-D14 300 300 3-D16
3-D12
3-D16 2-D14 2-D12 2-D14 300 300 3-D24
3-D12
3-D24 3-D12 3-D18 3-D12
3-D16 3-D16 3-D16 3-D12 2-D16 4-D14 4-D14 3-D12 2-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T11 300 300 3-D14
3-D12
3-D14 2-D14 3-D18 2-D14 300 300 3-D12
3-D12
3-D12 2-D20 2-D14 2-D10 300 300 3-D25
3-D12
3-D25 2-D18 3-D18 2-D18
4-D16 5-D16 4-D16 3-D12 5-D16 2-D16 2-D16 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T12 350 350 4-D14
3-D12
4-D14 3-D14 3-D18 4-D14 300 300 2-D16
3-D12
2-D16 3-D16 3-D12 3-D16 300 300 4-D22
3-D12
4-D22 3-D16 3-D18 3-D16
3-D14 4-D14 3-D14 3-D14 2-D14 4-D12 4-D12 3-D14 4-D14 3-D14 3-D14 3-D14 3-D14 3-D14 3-D14
T13 350 350 4-D16
3-D12
4-D16 3-D16 3-D16 3-D16 350 350 4-D14
3-D14
5-D14 2-D12 4-D16 2-D12 350 350 3-D25
3-D14
3-D20 2-D12 3-D16 2-D12
5-D16 8-D16 6-D16 3-D12 5-D16 5-D12 4-D12 3-D12 3-D14 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T14 400 350 3-D12
3-D12
5-D12 4-D12 3-D18 3-D12 350 300 5-D12
3-D12
4-D12 2-D12 3-D14 2-D12 300 300 3-D24
3-D12
3-D24 2-D12 3-D18 2-D12
3-D12 4-D12 2-D12 3-D12 3-D12 5-D12 5-D12 3-D12 3-D12 3-D12 3-D12
T15 350 350 5-D16
3-D12
6-D16 3-D16 1-D16 4-D16 350 350 5-D12
3-D12
5-D12
3-D12
3-D14
3-D12 300 300 2-D16
3-D12
2-D16
3-D12
1-D16
3-D12
5-D16 3-D16 2-D16 3-D12 2-D16 4-D12 3-D12 5-D12 3-D12 3-D12 3-D12
T16 400 350 5-D16
3-D14
3-D16 1-D16 1-D16 1-D16 350 350 5-D14
3-D12
4-D14
5-D12
4-D12
6-D12 300 300 3-D16
3-D12
3-D16
3-D12
1-D16
3-D12
7-D16 4-D16 5-D16 3-D12 4-D16 3-D14 3-D14 2-D14 3-D12 3-D12 3-D12 3-D12
T17 350 350 4-D12
3-D12
3-D12 3-D12 1-D12 3-D12 350 300 3-D14
3-D12
3-D14
3-D14
3-D14 1-D14 350 300 4-D16
3-D12
4-D16
3-D12
1-D12
3-D12
3-D16 3-D16 3-D16 3-D12 3-D16 5-D14 5-D14 3-D12 4-D14 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T18 350 350 3-D16
3-D12
3-D16 3-D16 2-D16 3-D16 350 350 4-D12
3-D12
4-D12 1-D16 4-D14 1-D16 350 300 3-D20
3-D12
3-D20 1-D16 2-D16 1-D16
4-D14 6-D14 3-D14 3-D14 4-D14 5-D12 4-D12 3-D14 6-D14 3-D14 3-D14 3-D14 3-D14 3-D14 3-D14
T19 450 300 4-D14
3-D14
6-D14 3-D14 1-D12 4-D14 450 350 6-D14
3-D14
6-D14 3-D16 2-D14 2-D16 400 300 3-D20
3-D14
3-D20 2-D16 1-D12 2-D16
5-D14 4-D14 3-D14 3-D12 2-D14 4-D14 4-D14 3-D14 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T20 400 350 6-D16
3-D14
4-D16 4-D16 1-D16 5-D16 400 350 4-D14
3-D12
4-D14 3-D12 3-D12
6-D12 300 300 2-D25
3-D12
2-D25 1-D16 1-D16 1-D16
5-D14 5-D14 3-D14 2-D16 5-D14 5-D14 5-D14 2-D16 5-D12 2-D16 2-D16 2-D16 2-D16 2-D16 2-D16
T21 450 350 3-D16
2-D16
3-D14 7-D14 3-D14 5-D14 400 300 4-D12
2-D16
4-D12 2-D14 5-D12 2-D14 350 300 2-D22
2-D16
2-D22 3-D14 3-D14 3-D14
5-D14 5-D14 5-D14 5-D14 4-D16 4-D16 3-D14 5-D12
T22 500 400 1-D16
3-D20
1-D16 2-D14
3-D20
5-D14 500 350 4-D16
2-D20
4-D16 2-D12 1-D12
2-D20 450 300 2-D20 2-D20 2-D20 2-D20 2-D20 2-D20
3-D16 3-D16 2-D16 3-D12 2-D16 2-D14 2-D14 2-D12 3-D12 3-D12 3-D12
T23 300 300 3-D14
3-D12
3-D14 2-D14 2-D24 2-D14 300 300 3-D16
3-D12
3-D16
3-D12
3-D14
3-D12 300 300 4-D20
3-D12
4-D20
3-D12
2-D24
3-D12
4-D12 4-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T24 350 300 5-D14
3-D12
5-D14 3-D14 2-D24 4-D14 300 300 3-D14
3-D12
3-D14 1-D16 2-D14 2-D16 300 300 3-D25
3-D12
3-D25 3-D14 2-D24 3-D14
6-D16 8-D16 8-D16 3-D12 7-D16 5-D16 4-D16 3-D12 4-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T25 450 350 4-D12
3-D14
5-D12 5-D12 2-D24 5-D12 400 300 3-D12
3-D12
3-D12 2-D20 3-D14 2-D20 300 300 5-D20
3-D12
5-D20 2-D20 2-D24 2-D20
6-D12 5-D12 3-D12 3-D12 5-D12 2-D12 2-D12 3-D12 2-D14 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T26 350 350 3-D14
3-D12
3-D14 4-D14 2-D24 7-D14 300 300 3-D16
3-D12
3-D16 2-D20 2-D14 2-D20 300 300 5-D20
3-D12
4-D20 2-D20 2-D24 2-D20
6-D16 4-D16 3-D16 3-D12 4-D16 4-D16 4-D16 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T27 350 350 5-D14
3-D12
4-D14 2-D14 2-D24 3-D14 350 350 3-D12
3-D12
3-D12 2-D20 4-D16 2-D20 300 300 5-D20
3-D12
5-D20 2-D20 2-D24 2-D20
6-D12 7-D12 7-D12 3-D12 7-D12 4-D16 3-D16 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12 3-D12
T28 400 350 6-D12
3-D14
7-D12 7-D12 2-D20 7-D12 350 300 3-D12
3-D12
2-D12 2-D16 3-D14 2-D16 300 300 3-D20
3-D12
3-D20 2-D16 2-D20 2-D16
T29 350 350 4-D16 4-D16 350 350 3D-14 4D-14 300 350 2D-14 4D-14

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

7
T30 350 350 4-D16 3-D16 350 350 4D-16 3D-16 300 350 2D-20 2D-20
T31 350 350 4-D20 4-D20 350 350 4D-20 4D-20 300 350 2D-25 2D-25
T32 350 350 4-D20 4-D20 350 350 3D-25 2D-25 300 350 3D-25 2D-25
T33 350 400 3-D20 4-D20 350 350 2D-25 2D-25 300 350 2D-25 2D-25
T34 400 400 5-D16 5-D16 350 400 3D-16 3D-16 350 400 3D-16 3D-16
T35 400 400 3-D20 5-D20 350 400 4D-16 4D-16 400 400 2D-16 6D-16
T36 350 400 5-D20 3-D20 350 350 4D-20 3D-20 300 300 4D-20 2D-20
T37 350 400 3-D20 5-D20 350 350 2D-20 2D-20 300 300 2D-20 2D-20
T38 400 400 4-D20 4-D20 400 400 3D-16 5D-16 400 300 2D-16 6D-16
T39 450 400 5-D16 4-D20 450 400 2D-20 4D-20 450 300 2D-16 6D-16
T40 450 400 4-D25 4-D25 450 400 2D-25 2D-25 450 350 2D-25 2D-25
T41 450 450 4-D25 5-D25 450 450 3D-25 3D-25 450 400 2D-25 2D-25
T42 500 500 6-D25 4-D25 500 500 5D-25 5D-25 450 400 2D-25 6D-25

Electronic Journal of Structural Engineering, 2023, Vol 23, No. 2

8