1 s2.0 S0141029623010726 Main

Engineering Structures 293 (2023) 116657
Contents lists available at ScienceDirect
Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct
Metamodel-assisted meta-heuristic design optimization of reinforced

concrete frame structures considering soil-structure interaction
Iván Negrin a, *, Moacir Kripka b, Víctor Yepes a
a
Institute of Concrete Science and Technology (ICITECH), Universitat Politècnica de València, 46022 Valencia, Spain
b
Graduate Program in Civil and Environmental Engineering, University of Passo Fundo, Km 292, BR 285, Passo Fundo 99052-900, RS, Brazil
A R T I C L E I N F O A B S T R A C T
Keywords: It is well known that conventional heuristic optimization is the most common approach to deal with structural
Structural optimization optimization problems. However, metamodel-assisted optimization has become a valuable strategy for
Reinforced concrete decreasing computational consumption. This paper applies conventional heuristic and Kriging-based meta-
Frame structures
heuristic optimization to minimize the CO2 emissions of spatial reinforced concrete frame structures, considering
CO2 emissions
Metamodel
an aspect usually ignored during modeling, such as the soil-structure interaction (SSI). Due to the particularities
Kriging of the formulated problem, there are better strategies than simple Kriging-based optimization to solve it. Thus, a
Soil-structure interaction meta-heuristic strategy is proposed using a Kriging-based two-phase methodology and a local search algorithm.
Three different models of structures are used in the study. Results show that including the SSI leads to different
design results than those obtained using classical supports. The foundations, usually ignored in this type of
research, also prove significant within the structural assembly. Additionally, using an appropriate coefficient of
penalization, the meta-heuristic approach can find (on average) results up to 98.24% accuracy for cohesive soils
and 98.10% for frictional ones compared with the results of the heuristic optimization, achieving computational
savings of about 90%. Therefore, considering aspects such as the SSI, the proposed metamodeling strategy allows
for dealing with high-complexity structural optimization problems.
1. Introduction excellent alternatives to solve large-scale and highly nonlinear optimi

zation problems, as is usually the case with structural optimization [10],
Design optimization is a topic widely discussed in the current context especially related to reinforced concrete (RC) structures optimization.
of structural design due to the need to minimize construction costs, For this reason, it is habitual to find structural optimization problems
material use, and decrease the construction sector’s negative impact on solved by heuristics. In the books proposed by Kaveh [11,12], many
the environment [1]. It can be achieved by strategies such as the use of examples of optimization problems in the design of civil engineering
novel building materials (e.g., low-carbon cement and clinker sub structures (mainly skeletal ones) can be found. However, more than
stitutes) [2] and recycling [3], but also through the more efficient use of these simple procedures are required to deal with real-life challenging
these materials resulting from their design optimization [4]. optimization problems. Alternatively, meta-heuristics have arisen to
As a result, more optimization objectives beyond economic ones handle such problems. They blend several simple strategies intelligently
have been emerging. A primary view of environmental assessment can to explore the solution space more efficiently.
be applied using single criteria reliably representing environmental Reinforced concrete (RC) frame building structures comprise a sig
impact. The most representative ones are CO2 emissions and embodied nificant portion of the construction sector. They are associated with
energy (EE) [5], which have proven to hold a direct relationship with substantial economic costs and environmental impacts [13]. Therefore,
economic cost [6,7,8]. it is crucial to obtain designs that minimize the adverse effects and
Structural optimization problems have two main ways to be solved. maximize the advantages of this construction type. Several authors have
On the one hand, exact methods are generally based on mathematical carried out studies to optimize these structures. Most of these studies
programming. On the other hand, heuristics consist of artificial intelli focus on essential concrete elements or two-dimensional frames. Only a
gence strategies that imitate natural processes [9]. These algorithms are few authors have applied design optimization to three-dimensional RC
* Corresponding author.
E-mail address: [email protected] (I. Negrin).
https://doi.org/10.1016/j.engstruct.2023.116657
Received 15 March 2023; Received in revised form 18 July 2023; Accepted 23 July 2023
Available online 31 July 2023
0141-0296/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
I. Negrin et al. Engineering Structures 293 (2023) 116657
frame buildings, including notable examples [14–18]. It is mainly due to dedicated to exposing and discussing the results. Finally, conclusions are
the high complexity involved in modeling these structures. Therefore, drawn in Section 4.
high computational consumption is one of the most critical challenges in
structural optimization problems. The solution for most of these engi 2. Methodology
neering problems consists of accurate and expensive numerical methods,
usually represented by partial differential or integral equations, e.g., 2.1. Problem description
structural finite element analysis (FEA), where a single function evalu
ation is usually considerably time-consuming [19]. Thus, if common In this paper, the structure used in previous work [8] is implemented
optimization problems involve thousands of single-function evaluations, as the basis of the investigation. Additionally, two other case studies
structural optimization problems generally require significant time and similar to the basic one are used to generalize the results. In these, the
computational resources. Based on this, metamodel-assisted structural span lengths (CS-2) and the number of levels (CS-3) are varied (see
design optimization (MASDO) has arisen as a valuable alternative to Fig. 1). The cross-sections of the elements are rectangular. This
deal with these complex problems. This kind of meta-heuristic approach research’s novelty lies in including shallow foundations and the corre
uses a surrogate model instead of the complex one to perform the sim sponding SSI during modeling. Note that the shallow individual footing
ulations of the optimization procedure. The most basic methodology variant is selected as it is the one that would best suit the modeled
consists of obtaining a group of design vectors inside the design space, structures. If the height of the structures is increased, there will come a
for which high-fidelity simulations (e.g., FEA) are performed. Then, the time when the base area of the foundations will be so large that they will
high-fidelity values regression or interpolation models are built and can overlap. It is where other typologies, such as combined footing or mat
be analyzed by, e.g., optimization algorithms [19]. Applying this strat foundations, would be necessary. If the height of the building continues
egy with the combination of clever optimization methodologies can be to increase, the use of piles would be required to increase the bearing
very useful in quickly obtaining good optimization results. The most capacity of the foundations, which opens the possibility of using the
common metamodels are polynomial regression, Neural Networks (NN), mat-pile combination. On the other hand, using the simpler typology
and Kriging models [20]. However, Kriging methodologies have been allows better applicability of the proposed methodology, which, once
one of the most used to aid structural optimization problems due to their validated, can be applied to other more complex structures.
simplicity and effectiveness. They create a metamodel using optimal The loads imposed on the system are shown in Table 1. Eqs. [1–4]
interpolation based on regression against observed values of the sur define the essential load combinations used, where D, L, and W are the
rounding data points, weighted according to spatial covariance values dead, live, and extreme wind loads, respectively. It is essential to
[5]. highlight that other combinations are used to design the foundations and
On the other hand, another problem usually found in the literature is to check the serviceability limit state.
the non-inclusion of modeling aspects such as the soil-structure inter
1.2D + 1.6L (1)
action (SSI). Authors usually assume structures with idealized or classic
supports (e.g., fixed), even when the assembly of soil and foundations is
1.2D + 0.8W (2)
not perfectly rigid. Support displacements (settlements) exist and in
fluence how the superstructure works, which is not the case with classic 1.2D + 1.4W + 0.5L (3)
supports. This idealization (no SSI consideration) leads to an inefficient
superstructure design. Due to the nature of settlement occurrence, it will 0.9D + 1.4W (4)
suffer an accelerated deterioration with the consequent need for extra
maintenance during its life cycle [21,22,23]. Some features are considered in this problem to obtain relevant re
Moreover, several distinctions must be considered between static sults from a practical structural engineering point of view. The variables
and dynamic SSI in the analysis being performed for the SSI consider are discrete to handle with constructive dimensioning of the elements
ation. An example of a dynamic SSI implementation can be found in and real types of concrete. Additionally, solutions of reinforcement
[24], where foundations are modeled as rigid strip footings, and the soil (longitudinal and transversal) are adapted to commercially available bar
is layered with constant material properties along its depth. An approach diameters, including longitudinal bars cut-off (detailing) and their
similar to the one used in [25] is considered to build the constitutive actual placement in the cross-sections of the elements (see Fig. 2).
model of the soil, where the shear wave velocity and the friction angle Other aspects generally ignored by designers, such as the reduction
are considered parameters of the soil layers. Otherwise, considering this of elements’ stiffness by cracking, second-order analysis of structures,
study’s static approach, SSI is considered to adopt a Winkler model, as and the after-mentioned SSI consideration, are also included. The CSi-
proposed by [26]. Thus, to relate the contact soil pressure p to the SAP2000 platform is used as a calculation engine via the Applied Pro
foundation settlement S, only one equation that includes both linear and gram Interface (API) SAP2000-MATLAB to deal with such a problem.
non-linear soil behavior is considered. Therefore, in this work, instead of For the creation of the models, the superstructure elements are consid
creating a three-dimensional model of the soil, the vertical direction is ered frame type. The foundations are modeled as shell elements. They
eliminated while considering the behavior in this direction by modeling are discretized (meshed) to improve their behavior as slab footings. A
the contact of the foundation and the soil as a discrete number of springs sensitivity analysis concluded that four divisions in each direction are
of stiffness k. This stiffness depends directly on p and S. sufficient for good results. Now, these elements are initially loose. It is
Considering the SSI during modeling, this paper applies conventional where the soil-structure interaction comes in. To simulate this phe
heuristic optimization and a Kriging-based meta-heuristic optimization nomenon, a stiffness coefficient k is calculated (see section 2.2), which
strategy to optimize spatial RC frame structures. For this reason, the two can be assigned in two ways: as an area (directly to the bottom surface of
main goals are to demonstrate the differences in optimizing the design of the element) or distributed to the nodes formed by discretizing the
a structure with idealized supports and considering the SSI using two element. In this case, the first way is used because it is faster from the
types of soils. Secondly, to implement a Kriging-based structural opti computational point of view. In either case, the results are the same. A
mization strategy to reduce the high computational consumption of the complete structure model can be seen in Fig. 3(a). The frame structure
conventional heuristic optimization trying to keep accurate optimiza design for strength and serviceability Limit States is based on the ACI
tion results. Consequently, the organization of the paper is as follows. 318–14 code.
Section 2 explains the general methodology, including the problem
description, the explanation of the SSI consideration, the formulation of
the optimization problem, and the strategies used to solve it. Section 3 is
2
Fig. 1. Three case studies (CS). CS-1 is the basic case study, CS-2 is similar to CS-1, but X and Y are 8 and 6 m respectively, and CS-3 is CS-1 with an additional level.
coordinates (Ŕ *,S) and unlimitedly approaches the asymptote p=q*br_II

Table 1
(see Fig. 4(a)).
Loads considered.
( ( q*br II ) )
Description Value p⋅S⋅ − 1
(5)
′*
R
Dead load on first floor 4.80 kN/m 2 S=
q*br II − p
Dead load on roof 5.43 kN/m2
Live load on first floor 4.00 kN/m2 Here, S is the base settlement for an acting pressure equal to the soil
Live load on roof 0.80 kN/m2
base linearity limit stress Ŕ * and q*br_II is the base bearing capacity
Wind load 0.92 kN/m2 0–5 m, 1.01 kN/m2 at 7 m,
1.13 kN/m2 at 10.5 m. pressure, based on expressions from the theory of plasticity. On the other
hand, the soil base “secant” stiffness coefficient k can be obtained easily
for the loading stage as in Eq. (6) [27].
2.2. Implementation of the static soil-structure interaction
p q*br II − p
k = tanα = = (6)
When considering the static SSI in the structural optimization S S ⋅[( q*br II ) − 1 ]
′*
R
problem, it is assumed that the structure above the underlying soil de
forms under loading–unloading, which results in a redistribution of in This stiffness coefficient is applied to the finite springs introduced in
ternal forces in the superstructure. The soil is modeled as a linearly the nodes created when meshing the foundation base. By making more
elastic half-space, considering the depth constraint of compressible divisions, the foundation will have a more realistic behavior, but the
thickness, while the foundation is considered a shallow slab footing (see model will become more complex. Fig. 3(b) shows how the elements are
Fig. 3(a)). discretized. On the other hand, Fig. 3(c) shows the non-uniform pressure
In practical calculations, the consideration of non-linearity in distribution for an exterior foundation. It is because the foundations
determining settlements in type II collapsible soils was introduced in settle irregularly, depending on the loads coming from the superstruc
[26]. However, this model can be adapted and extended to other soil ture, i.e., the springs deform according to their stiffness k depending on
types and foundation typologies [27]. In his 1969 study [26], Klepikov the configuration of the superstructure and the foundation itself. This
introduced an approximation of the relationship between acting pres model has the limitation that the springs have a linear behavior. How
sure and settlement (p vs. S) for a shallow foundation on a soil base. He ever, when applied superficially to the base of the foundation, the nodes
represented this relationship through a hyperbolic equation, denoted as formed when discretizing the element will have a coefficient according
Eq. (5). The “pressure-settlement” curve passes through the point with to their position in the element (interior, exterior, and corner). In
addition, these nodes will deform according to the load each receives,
Fig. 2. Transversal and longitudinal reinforcement bars cut-off and distribution in RC building frames, x1 and x2 are design variables, As y A’s are the real steel area
in traction and compression respectively [1].
3
Fig. 3. Concepts related to SSI modeling: (a) general considerations, (b) meshing of shell elements, and (c) non-uniform pressure distribution of a foundation base.
Fig. 4. Curve representing the acting pressure and settlement (p vs. S), (a) is an abstraction, (b) is the real p vs. S curves for an interior foundation and (c) is a zoom of
the graphic of (b). Here, Ŕ * is the linearity limit stress, S is the base settlement for an acting pressure equal to the soil base linearity limit stress Ŕ * and q*br_II is the
base bearing capacity pressure. For more information, refer to [1].
distributed non-uniformly throughout the shell element. Thus, the nodes and settlement. According to [28], this curve is acceptable for obtaining
formed (1) have different coefficients, and (2) receive different loads. soil stiffness when considering SSI. It has been proven in actual exper
Therefore, the foundation can not only settle but also rotate. Thus, the iments for different soil types [29,30,31,32].
behavior of the supports is very much in line with what happens in re It should be noted that the model proposed for considering the SSI is
ality. It is obviously not the case with rigid supports. It is important to relatively simple. Even so, and taking into account that it is based
note that several authors have validated this strategy of relating pressure fundamentally on the calculation of the actual settlements S, several
4
aspects could be indirectly considered, such as: inhomogeneity of the previous model is completed with the foundations and the springs
geological structure of the base, presence of water level, separate soil with k stiffness. The model analysis with the SSI consideration is carried
lenses and different inclusions, the possibility of soil flooding and out next, leading to new values of internal forces. The final design of the
change of its properties, particular consideration of residual and elastic whole structure is performed. Then, the CO2 emissions are obtained
settlements of the base, and more. If future research would like to according to the total volume of work. Fig. 5 represents this process.
include any of these aspects, providing more information about these
special soil conditions underlying the foundations would be beneficial. A
highly detailed explanation of the considerations for modeling SSI can 2.3. Formulation of the optimization problem
be found in [1].
The economic cost is probably the most widely implemented objec
2.2.1. Pressure-settlement curves for different type of soils tive in optimizing the design of reinforced concrete structures. However,
One of the points to highlight in this study is the inclusion of the SSI the environmental approach to optimization problems has arisen in
during modeling and its influence on the structure’s optimal design. recent years. The environmental impact of buildings can be directly
Another aspect being considered is the underlying soil’s influence. For measured using simple targets such as EE or CO2 emissions.
this purpose, two soils are used in the SSI modeling. Soil 1, as seen in Several studies have established comparisons between optimizing
Table 2, is predominantly cohesive, while soil 2 is predominantly using economic or environmental criteria. In [33], it is compared the
frictional. results of optimizing the economic cost and CO2 emissions in the design
It is essential to analyze how the soils behave during loading to un of RC structures similar to those in this study. It is concluded that these
derstand how the superstructure performs based on the soil type. It is relationships vary depending on the ductility class. A similar study [34]
achieved by using actual behavior curves of the soils for a specific type of shows that the economic cost must be moderately increased to decrease
foundation. These curves are similar to the one depicted in Fig. 4(a), emissions significantly. However, optimizing both objectives is only
which is an abstract representation. The theory behind this approach is partially counterproductive, as demonstrated by performing a multi-
that differential settlements result in a redistribution of the internal objective optimization including both criteria [35]. In [8], a compara
forces of the superstructure, causing an increase, particularly in the tive study was performed using the economic cost and the environ
bending moment. mental impact (involving both EE and CO2 emissions). It was found that
Fig. 4(b) displays the curves for an interior foundation, where the any of these objectives leads to good results as measured by the others.
thick line represents soil 2, and the thin one represents soil 1. It is However, the best option is to use CO2 emissions since the other in
observed that soil 2 has a greater bearing capacity than soil 1 and ex dicators keep good levels when using it as the optimization target. For
hibits a larger zone of non-linear behavior, unlike soil 1, where the zone this reason, this study involves the CO2 emissions single-objective
of non-linearity is minimal. It causes that underneath the foundation, optimization of the tri-dimensional frame RC structures shown in
type 1 soils generally present a linear behavior, while type 2 soils behave Fig. 1. Hence, the objective is to minimize Eq. (7).
non-linearly. Therefore, differential settlements are more significant in CO2 emissions =
∑
ei × mi (x1 , x2 , ..., xn ) (7)
type 2 soils. In Fig. 4(c), the curves are zoomed in. Suppose that dp is the i=1,n
possible difference in pressures of two adjacent foundations, e.g., an
interior one with more pressure than an exterior one. This differ Here, ei represents the unit CO2 emissions (see Table 3), mi are the
ence dp is placed in the usual working zone for each soil type, which is measurements concerning the construction units in the function of the
not precisely at the bearing capacity limit q*br_II (continuous black ver design variables (x1, x2,…,xn). Table 3 shows unit values of considered
tical line in Fig. 4(b)) but shifted to the left. The curve is a function of the materials and activities obtained from the 2016 database of the Institute
settlement S and is constructed for Limit State II. At the same time, the
geotechnical design of the foundation is made according to Limit State I,
with different safety coefficients. Thus, the real points of p vs. S are
moved. Therefore, while in soil 2, this movement keeps the points in a
zone of non-linearity, in soil 1, they move to the area of linear behavior.
The zoomed graph shows how the same difference in pressure dp causes
a larger difference dS in soil 2 (with non-linear behavior) than in soil 1
(with linear behavior). Therefore, it can be said that this type of struc
ture behaves more irregularly on type 2 soils than on type 1 soils.
2.2.2. Algorithm for SSI modelling during structural optimization

For the SSI consideration, it is necessary to start from some point.
Accordingly, the calculation starts by analyzing a model with idealized
supports (fixed for this case). Based on the results of the first step, the
geotechnical design is performed. Here, the dimensions of the founda
tions are calculated, and the structural design is also performed. Then,
the stiffness coefficient is calculated for each group of foundations. This
step is called the pre-design of the foundation. With these results, the
Table 2
Properties of the soils considered in the study.
Soil FI C E γ μ ø
(◦ ) (kPa) (MPa) kN/m3 (◦ )
1 8 60 12 19.0 0.40 76
2 32 10 15 17.5 0.30 56
FI: Soil friction angle C: Cohesion E: Modulus of elasticity γ: Density μ: Poisson’s ratio
Fig. 5. General algorithm for the SSI consideration during structural optimi
ø: see Fig. 6
zation process.
5
Table 3 Table 4
Unit CO2 emissions for materials and activities. Design variables and corresponding ranges for each case study.
Material Units CO2 em (kg) Variable Range
Formwork m2 2.53
CS-1 and CS-3 CS-2
Steel (G-60)* kg 3.01
Concrete 25 MPa m3 244.94 Depth of beams of first design 300 mm ≤ x1 ≤ 600 400 mm ≤ x1 ≤ 700
Concrete 30 MPa m3 279.21 group mm mm
Concrete 35 MPa m3 305.96 Width of beams of first design 200 mm ≤ x2 ≤ 300 mm
Concrete 40 MPa m3 307.06 group
Activities Depth of beams of second design 250 mm ≤ x3 ≤ 500 300 mm ≤ x3 ≤ 600
Concrete placement Beams m3 34.72 group mm mm
Width of beams of second design 200 mm ≤ x4 ≤ 300 mm
Columns m3 37.20
group
Found m3 19.84
Dimension in x-axis of interior 250 mm ≤ x5 ≤ 500 mm
Earthwork Excavation m3 3.99
columns
Refill m3 12.80
Dimension in y-axis of interior 250 mm ≤ x6 ≤ 500 mm
* fy=420 MPa, E=220 GPa
columns
Dimension in x-axis of exterior 250 mm ≤ x7 ≤ 500 mm
columns
of Construction Technology of Catalonia [36]. The CO2 emissions values Dimension in y-axis of exterior 250 mm ≤ x8 ≤ 500 mm
encompass the use of materials that involve emissions at the different columns
stages of production and placement. Consequently, the higher this value Dimension in x-axis of corner 250 mm ≤ x9 ≤ 500 mm
is the lower its sustainability. columns
Dimension in y-axis of corner 250 mm ≤ x10 ≤ 500 mm
As mentioned, all design variables are discrete and can only take
columns
relevant values from an engineering point of view. In the case of the Rectangularity of interior 0.5 ≤ x11 ≤ 2.0
frame elements, variables are related to the cross-section dimensions foundations
and are limited to multiples of 5 cm. Variables related to foundations Rectangularity of exterior 0.5 ≤ x12 ≤ 2.0
foundations
rectangularity (L/B, see Fig. 6) can handle nine possible values repre
Rectangularity of corner 0.5 ≤ x13 ≤ 2.0
senting four possibilities for each direction plus the square configura foundations
tion, e.g., [0.50, 0.63, 0.75, 0.88, 1.00, 1.25, 1.50, 1.75, 2.00]. There are fc′ in beams 25 MPa ≤ x14 ≤ 40 MPa
three variables related to this issue. Finally, there are three more vari fc′ in columns 25 MPa ≤ x15 ≤ 40 MPa
ables considering the type of concrete for each group of elements fc′ in foundations 25 MPa ≤ x16 ≤ 40 MPa
(beams, columns, and foundations), where each one can handle four
possible values. The property considered as a variable is the specified
They are sometimes referred to as state equations. These constraints
compressive strength (fc′ ). Other properties are also variables (modulus
fulfill the design limit states, i.e., defining the values the variable pa
of elasticity, for example), but they depend on fc′ , whose values vary in
rameters must meet to satisfy behavioral requirements. In structural
steps of 5 MPa. Table 4 shows the design variables and their corre
optimization, the behavioral constraints are usually set by design stan
sponding ranges. It is straightforward to note that considering the
dards. Eq. (8) represents the classical formulation of behavioral
foundations within the structural assembly (and the corresponding soil-
constraints.
structure interaction) modifies the basic formulation of the problem,
where the foundations were not considered [8]. With the addition of 4 gj (x1 , x2 , ..., xn )⩽0 (8)
more variables, and being the ones related to the foundations’ rectan
Constraints related to the strength (ultimate) limit state for RC frame
gularity very influential in the general formulation, the space of solu
elements are automatically satisfied through the API SAP2000-MATLAB
tions grows exponentially, becoming more challenging to explore and,
platform (calculus of the reinforcing steel area), which computes the
therefore, to optimize.
structural design according to the standards. The geotechnical and
On the other hand, this problem involves two fundamental types of
structural design of the foundations is implemented in a routine pro
constraints. The first group is called design (or explicit) constraints. They
grammed in MATLAB according to the Eurocode. It is essential to note
are imposed directly on the design variables and function as limits on the
that this problem is formulated with a deterministic approach, so the
movement of these variables. The intervals shown in Table 4 are explicit
design is based on the Limit States method. Therefore, aspects such as
constraints. They appear for buildability, architecture, functionality,
the variability of soil parameters are considered when applying the
and transportation.
different safety coefficients. On the other hand, constraints related to
The other group is the so-called behavioral or implicit constraints.
serviceability limit state accomplishment are deflections in beams, limit
displacement at the top of the building, or cracking of concrete mem
bers. As mentioned, these constraints are produced by applying the
standards. The way to verify these behavioral constraints is that when
any of them is not fulfilled, the value of the objective function for the
current solution is penalized so that the algorithm discards it as possible.
For more detailed information on how the deterministic approach works
and the implementation of the constraints, refer to the document [1].
2.4. Solution of the optimization problem
Two main strategies are used to solve the formulated problem. The
first presents a classical heuristic approach using a relatively new
Biogeography-based Optimization (BBO) strategy. The second relies on
kriging formulation metamodels to optimize surrogate models and avoid
Fig. 6. Typical shallow foundation and excavation scheme, p(n) is the rect costly accurate simulations, using the BBO strategy as the basic opti
angularity (variable) [1]. mization algorithm. It should be noted that the optimization of this type
6
of problem is a complex process in which it is not guaranteed to find the tests) created using the basic Kriging methodology (see next section). A
global optimum. In addition, meta-heuristic optimization based on polynomial of order 0 is used as the regression model to construct the
metamodels is proposed as an alternative to conventional optimization. metamodel. The utility (or measure used to compare the performance of
It is intended to significantly reduce computational costs while main each method) is the final value obtained in each optimization process.
taining the quality of the solutions. The figure shows that BBO is superior in both types of optimization,
especially in the real model. It also shows that using metamodels is an
2.4.1. Conventional heuristic optimization excellent alternative to perform parameter tuning processes and
Heuristics are the most implemented strategies to solve structural compare options to give a solution to a real problem using a surrogate
design optimization problems. It is precisely the case for optimization model. In this case, this metamodel (created using a polynomial of order
problems in the design of RC structures. In the optimization of simple 0, see next section) could be more precise regarding values. Even so, it
elements (beams or columns), several heuristics have been used, such as captures the fundamental characteristics of the natural phenomenon
Harmony Search (HS) [37,38] or Glowworm Swarm Algorithm [39]. and is much more computationally low-cost to evaluate. It is highly
Genetic Algorithms (GA) have been implemented for the solution of the desirable in parameter setting or comparison of alternatives to solve real
optimization of RC plane frames [40,41], as well as HS [42] or Simu problems.
lated Annealing (SA) [9,43]. On the other hand, to solve optimization The BBO algorithm represents mathematical models of how species
problems of three-dimensional concrete structures, it has been imple migrate from one island to another, how new species arise, and how
mented Particle Swarm Optimization (PSO) [14,15], the Flower Polli species become extinct. A more detailed explanation of the BBO strategy
nation Algorithm [17], the Cascade Optimization Method [18], HS [44], can be found in [1,8,48]. In order to explain why BBO seems to be a
SA [45,46], or the old bachelor algorithm [47]. proper strategy for dealing with discrete optimization problems, it is
Alternatively, there is a heuristic of relatively little diffusion called essential to highpoint the main difference between this strategy with
Biogeography-based Optimization, proposed by [48] and, although classical Evolutionary Algorithms (EAs), such as GA, for example. It is
ephemeral, has been used in problems of optimal design of civil engi associated with the recombination operator. When classical EAs
neering structures. In [49], it is implemented a version of BBO with Levy combine complete solutions, the BBO algorithm processes solutions
flight distribution (LFBBO) in the design optimization of cantilever from variable to variable (not solution to solution). That is to say, when
retaining walls. In [50], BBO is used to optimize the design of 3D and GA combines two solutions to create a new one, BBO can obtain solu
plane frame steel structures. In the case of RC structures, BBO has been tions from more than two previous solutions. In addition, the combi
successfully implemented to solve a design optimization problem. In nation and the mutation operators can affect the variable involved in the
[21,51], BBO was compared to other heuristics in these types of prob same process of getting new solutions.
lems, outperforming some classical ones such as GA, HS, or PSO. In [1], Highlight that one of the advantages of BBO over other heuristics is
the main parameters of BBO were tuned to deal with problems related to the rapid convergence to outstanding solutions very quickly. It is espe
the formulated in this paper. Finally, in [8], BBO was implemented to cially desirable in these extremely computationally expensive processes.
optimize the design of the same structure of this paper without consid One of the reasons for this fast and efficient convergence is the algo
ering foundations and SSI during modeling. rithm’s effectiveness with relatively low population sizes. For example,
As mentioned, previous works have demonstrated the superiority of in the parameter tuning process mentioned above, the optimal popula
BBO over other methodologies to solve this type of problem. However, tion size was 80 individuals, while for the other two heuristics, it was
considering the benefits of working with metamodels, an experiment is around 200. More information on this aspect can be found in [1].
carried out to select the best heuristic that will be the basis of the meta-
heuristic algorithm. In addition, it will serve as a reference to compare 2.4.2. Kriging-based optimization
and validate the results of the proposed meta-heuristic. The other two Conventional heuristic optimization processes are usually quite
heuristics tested are GA and PSO. Note that the parameters of the computationally expensive, even when using a tuned method with fast
methods are selected according to the recommendations provided in and efficient convergence, such as BBO. Alternatively, MASDO allows
[51]. Fig. 7 shows the behavior of each method in the optimization of for reducing computational consumption. The basic idea consists of
the real model (the equis represent the values obtained in three tests) and using a surrogate model (instead of the real and complex one) built from
of a metamodel (the boxplots represent the statistical analysis of twenty an initial sampling of points. This surrogate model can predict the
output data (objective response) from any input data (design variables)
in the design space.
There are three main steps to building a metamodel: (1) selection of
the initial sample of points inside the design space, also called the design
of experiment (DoE), (2) construction of the approximate mathematical
model using the initial sample of points using a metamodeling technique
and (3) validation of the constructed surrogate model. The main
objective of the metamodel construction is to obtain a model that pre
dicts the actual response with the best possible accuracy [5].
To select the initial sample of points to construct the metamodel
(DoE), the sample size and the position of these points must be consid
ered. On the one hand, the sample size (N) is directly related to the
number of variables. On the other hand, once the number of points has
been selected, they should be placed in such a way as to gather as much
information as possible. DoE can be split into two main groups. Classic
designs tend to place the sample points around the border of the design
space and only locate a few points inside it [52]. The other group, called
space-filling designs, is more suitable for building advanced meta
Fig. 7. Comparison of the performance of the three heuristics in optimizing the models. One of the popular ones is Latin Hypercube Sampling (LHS).
real model (equis) and the surrogate model (boxplots, where bottom whisker, In this work, the DoE will be performed through LHS. The effec
box bottom, middle, top, and top whisker denote the minimum, 25th percentile, tiveness of this technique has been proved in several studies [53]. LHS
median, 75th percentile and maximum utility of each process). was proposed by [54]. This method determines the N number of non-
7
overlapping intervals for each variable from several design variables (v) with a coefficient of penalization CP. This point is added to those used to
and some initial input sample points (N). Then, the design space is build the initial metamodel. If the solution is feasible, it is taken as the
distributed in Nv sections. Each point is placed in such a way as to optimal solution.
correspond to a combination of different intervals of each variable One value that defines and regulates the process’s accuracy is the
design range. Consequently, LHS designs ensure all design variables are coefficient of penalization CP applied to infeasible solutions, which is
represented with their intervals [5]. common in this type of optimization problem. Even more significantly,
Once the appropriate set of points has been selected and the corre the best solutions are often close to these “not feasible” points. There
sponding high-fidelity simulations for each point have been performed, fore, a sensitivity analysis is conducted to observe the strategy’s
the next step is to choose a metamodel and fitting strategy. As behavior for different CP values. Fig. 8(a) shows the results of studying
mentioned, NN and Kriging strategies are the most implemented tech three values: 1.00 (no penalization), 1.15, and 1.30. In this case, an
niques in structural engineering. However, NN-based models usually initial number of sampling points of N = 100 is considered. Five tests are
require several sample points and considerable computational time to performed for each penalty coefficient, and the mean of the five values is
train the network [55]. On the contrary, Kriging-based models are plotted. The figure shows that as the penalization coefficient increases,
flexible strategies and less time-consuming than NN-based techniques the accuracy of the metamodel decreases. Therefore, CP = 1 is used for
[56]. For this reason, the metamodel construction in this study is based the simple Kriging-based optimization.
on the Kriging formulation and the DACE Kriging Toolbox V 2.0 [57]. On the other hand, another essential aspect is the number of initial
Kriging models are extensive and very popular global approximation points N obtained in the DoE. A sensitivity study similar to the previous
techniques based on the work of Daniel G. Krige [58], adapted to handle one is carried out. Five values are analyzed: 20, 50, 100, 200, and 300.
geostatistical problems [59]. A more detailed explanation of the Kriging Five tests are performed for each value, and the average of these values
formulation can be found in [57,60]. is plotted. Fig. 8(b) shows that the MAPE values are outstanding. It in
dicates that the metamodels obtained from the points located by LHS are
2.4.3. Simple Kriging-based optimization pretty accurate.
The usual methodology of Kriging-based optimization starts with However, simple Kriging-based optimization often produces unsat
generating the N initial sample points using the LHS technique. It is isfactory results, as optimal values tend to be located at the extremes of
necessary to obtain the real fitness value of each point. Then, the met the variables, especially the lower ones. It frequently leads to constraints
amodel is built according to the mathematical strategies used by being violated, requiring additional points to be added to the initial
Kriging-based approaches. It is also important to highlight that each sampling from which the first metamodel was created. Without a coef
metamodel type has its associated fitting model. In this case, the ficient of penalization (CP), the algorithm may repeatedly select the
hyperparameters used to fit the model are determined by maximum same point as the optimal solution, even if it fails to meet the constraint
likelihood estimation [60]. Once the metamodel has been built, it is requirements. However, the solution found in such cases could be better,
necessary to validate it. For this purpose, measurements based on error consisting mainly of variables with extreme values. Increasing the CP
estimation are usually used. In this case, the mean absolute percentage results in a “see-saw effect”, affecting the area surrounding the infeasible
error (MAPE) is implemented, measured using ten randomly obtained point and all other values. Therefore, when optimizing the metamodel, if
points (that do not coincide with those obtained in the DoE) as shown in the obtained point violates constraints, the optimum becomes the
Eq. (9). opposite value through penalization (see-saw effect). It is usually the
∑n ∧
case for variables most affected by constraint violations, such as beam
depths or column section dimensions.
|yi − yi |
(9)
i=1 yi
MAPE = × 100%
n Thus, the points obtained by LHS produce a highly accurate meta
model only within the interior of the solution space, as illustrated in
Here, n is the number of points used for the measurement (ten in this
Fig. 9. This situation becomes more intricate for the problem since there
case), while yi and ŷi are the actual and predicted values, respectively.
are too many variables compared to the limited number of possible
This MAPE value must be lower than a previously established threshold.
values they can handle. In other words, the interior solution space is
If this criterion is not met, another ten randomly selected points (that do
minimal compared to the points at the boundaries, which causes the
not coincide with any previous point) are added. The process is repeated
“see-saw effect”. Moreover, since meaningful solutions are often found
until this condition is met, i.e., the metamodel has adequate accuracy.
at these frontier points, this explains the outcome when applying simple
Then, the optimization process proceeds. For this case, the BBO algo
Kriging-based optimization to the problem.
rithm explained above is used. This algorithm optimizes the response
surface generated by the metamodel. Once the “optimal solution” of the
2.4.4. Kriging-based meta-heuristic optimization
metamodel has been found, it is necessary to check if it is feasible using a
Considering the limitations of the simple strategy to deal with the
high-fidelity simulation. If the solution is not feasible, it is penalized
Fig. 8. Study of (a) the coefficient of penalization (CP) and (b) the initial sampling size (N).
8
Fig. 9. Example of the low effectiveness of the metamodel at the boundaries of the solution space, (a) simulated response, (b) real response and (c) error between the
simulated and the real. Coefficient of penalization equal to 1.00. The values of the other variables are randomly selected.
formulated optimization problem, a two-phase strategy based on Kriging metamodels (two-phase Kriging-based strategy). From this point, a local
metamodels is designed. It also includes concepts of online optimization search heuristic using the actual response surface is applied to improve
and a local search heuristic. The global strategy is called Kriging-based the solution obtained using the surrogate models. The steps are as fol
meta-heuristic optimization (KBMO). lows (see Fig. 10):
Using only LHS is not a good choice for this type of problem,
although it provides excellent results in the interior of the solution 1. The initial DoE is designed by obtaining 20 points using LHS, 32
space. Therefore, a DoE is designed using points obtained through LHS exterior points, and 16 corner ones for 68 initial points. Two exterior
and others arbitrarily located at the boundaries. The basic idea of this points are obtained for each variable (one upper and one lower), thus
strategy is the following: to perform a first exploration of the solution the 32 exterior points. An exterior point is considered to be one in
space to “detect” promising areas using a general metamodel; to build a which the value of a variable is fixed at one of the extremes while the
“local” metamodel in the most feasible area; to optimize the “local” other variables take random values. A corner point is considered to
metamodel and to obtain the best solution resulting from the use of be set to an extreme value. At the same time, the other variables take
Fig. 10. Flowchart of Kriging-based meta-heuristic optimization (KBMO).
9
a random value from the two extremes that correspond to each one, i. metamodel is checked with ten selected points following the same
e., all variables have extreme values. Since fewer corner points exist, criteria as point 1 until it meets the established criteria.
only 16 are considered, one for each variable. Then, the high-fidelity 4. This metamodel is optimized using the simple BBO strategy. Until
simulation is performed for each point. The metamodel is created here, this part of the process is named Kriging-based optimization
using a first-order polynomial for the regression model (see how (KBO).
Kriging works in suggested references). The metamodel is “vali 5. Considering that the computation time is still relatively low
dated” using ten randomly selected points (five interior ones ob compared to the conventional procedure, the process is finished with
tained through LHS and five exterior-corner ones) and the the local optimization algorithm (CDLIS, see Fig. 11), starting from
corresponding MAPE. If the MAPE value is not less than 6%, these 10 the point obtained using the metamodels. The final solution is
points are incorporated into the initial DoE, and the metamodel is considered the optimal one.
updated. The process is repeated until the criterion is met. Once a
good sample of initial points is obtained, another metamodel is It is essential to highlight that each of the high-fidelity simulations
created with this sample, using a polynomial of order 0 (constant). It performed in the process is stored so that if a point is repeated, it is
creates a less accurate metamodel than the one obtained with the unnecessary to perform costly extra simulations.
first-order polynomial, but the “see-saw effect” does not affect it. On the other hand, the local search algorithm is based on strategies
2. The metamodel is optimized with the simple BBO heuristic. The such as the Simplex Nelder-Mead algorithm [61] and Integer Linear
generated response surface is quite complex, i.e., each optimization Programming [62]. It consists of performing discrete unit steps and
process usually gives different results, although with much better gradually storing the results to improve the analyzed solution. The steps
quality than those obtained with the metamodels built using a first- are as follows (see Fig. 11):
order polynomial. Ten optimization procedures are performed,
which, when using the metamodel, present a little computation time 1. Each variable increases and decreases by one unit in value, starting
compared to the accurate simulations. The best solution is selected from the basic solution. Then, the corresponding high-fidelity
from the ten obtained. simulation is run, and the results are stored. If the variable’s value
3. A new local solution space is created by adding one value up (+1) is in one of the limits, it is not violated, so only one solution is
and one value down (-1) to the “optimal” values of each variable analyzed. In addition, each new solution is checked in the database
using this solution. If the value of a variable is located at the extreme, against all previous solutions to avoid repeating additional costly
two units are added or subtracted so that each variable can take three accurate simulations. If already performed, it is not necessary to run
values. Using similar criteria as in point 1, a new DoE is created (10 the FE software. Therefore, each iteration will have at most twice the
LHS points, 15 exterior, and 10 corner points, for a total of 35 number of variables.
points), and the “local metamodel” is created this time using a first- 2. Once all the results are stored, the new solution is formed, updating
order polynomial as the regression model. The accuracy of the the values of each variable that improved the basic solution. For
example, in Fig. 11, variable 1 improved the solution by decreasing
Fig. 11. Flowchart of Constrained Deterministic Local Iterative Search (CDLIS). Left: general flowchart. Right: example of an iteration in a problem with 3 variables.
10
its basic value (p(1)-1), and variable 2 did not improve with either dimension on the x-axis. Here the rectangularity is generally more sig
step. Variable 3 improved by increasing it (p(3)+1), so the basic nificant. The interior columns usually have a rectangular section with a
solution [p(1), p(2), p(3)] becomes [p(1)-1, p(2), p(3)+1]. It is called larger side in the y-axis direction. It is to increase the horizontal stiffness
new solution 1. In addition, the best individual solution is identified, of the structure in that direction, which is critical against wind action.
which is the best solution of all those obtained during the iteration. It Note that these columns do not have predominant bending due to
is the new solution 2. gravitational loads since the structures are symmetrical.
3. The high-fidelity simulation obtains the actual value of the objective On the other hand, the interior foundation, which does not receive a
function provided by new solution 1. If this value improves the value predominant bending moment due to gravitational loads, tends to be
of the basic solution, the new solution 1 becomes the basic, and a designed with a rectangular footing to deal with the horizontal wind
new iteration begins. load in its critical direction. Exterior foundations usually have square
4. If the value of solution 1 does not improve the base one, the quality of footings. Corner foundations tend to be designed with a rectangular base
the new solution 2 (already stored) is analyzed. If this new solution is with a larger dimension on the x-axis to cope with gravity loads. It can be
better than the basic one, the latter becomes the new solution 2 seen how there is a match between the columns and the corresponding
(single best solution), and a new iteration begins. If the new solution foundations. Regarding the quality of the concrete, the elements that
2 does not provide a better result, the process is ended, and the basic work primarily in bending (beams and foundation footing) require
solution is considered the optimal global solution. concrete with low compressive strength. On the other hand, since this
property (fc′ ) is significant in columns, the best choice is usually high-
3. Results and discussion strength concrete.
The results and their discussion focus on two fundamental aspects:

the influence of considering the SSI and the application of the proposed Table 6
Comparison of results of optimization procedures with and without considering
metamodeling strategy. The methodology is based on performing the
soil-structure interaction during modeling (using CS-1*).
basic experiments on case study 1 and validating the results with the
other two models. Additionally, this type of structure’s behavior is Elements No SSI SSI SSI
(soil (soil
analyzed by the different conditions generated in this study.
1) 2)
CO2 Beams Steel 2917 3223 3274
3.1. Influence of SSI consideration emissions Concrete 3155 2584 2584
(kg) Formwork 352 301 301
Total 6424 6108 6159
The results’ first point focuses on including the SSI during modeling. Columns Steel 1968 2838 3024
For this purpose, the three case studies are optimized by considering Concrete 1333 1952 2345
classical (fixed) supports and the SSI using the two soil types. In addi Formwork 175 193 234
tion, the optimization with classical supports does not consider the Total 3476 4983 5603
Superstructure 9900 11 11
environmental cost of the foundations in the objective function, unlike
091 762
the optimization considering the SSI. As mentioned, the primary hy Foundations Soil Soil
pothesis is that a structure modeled with SSI will have a more stressed 1** 2**
superstructure due to the logical differential settlements between Steel 2694 1539 2613 1062
contiguous supports. Concrete 872 427 868 476
Table 5 shows the optimum values of the variables for each model. It Formwork 77 68 81 80
can be appreciated that the results are different for each support con Earthwork 5867 5014 5801 4963
Total 9510 7048 9363 6581
dition, although some aspects maintain their tendency. For example,
TOTAL 19 16 20 18
corner columns tend to be square or with little rectangularity in the 410 948 454 343
direction of predominant bending due to gravitational loads (“x” axis in
* Results of CS-2 and 3 are graphically represented in Fig. 12.
these cases). This rectangularity tends to increase in soil 2, where dif
** Foundations are designed and their environmental cost is calculated for the
ferential settlements are higher. As do corner columns, exterior columns
optimal solution (model with classic supports, no SSI).
usually have rectangular cross-sections with the most significant
Table 5
Design variables and corresponding optimal values for each case study.
Variable Optimal values
CS-1 CS-2 CS-3
No SSI Soil 1 Soil 2 No SSI Soil 1 Soil 2 No SSI Soil 1 Soil 2
x1 (m) 0.45 0.35 0.35 0.60 0.45 0.50 0.45 0.45 0.50
x2 (m) 0.20 0.20 0.20 0.25 0.25 0.25 0.20 0.25 0.25
x3 (m) 0.40 0.35 0.35 0.40 0.35 0.35 0.40 0.35 0.35
x4 (m) 0.20 0.20 0.20 0.25 0.25 0.25 0.20 0.25 0.25
x5 (m) 0.30 0.30 0.25 0.25 0.30 0.30 0.30 0.30 0.30
x6 (m) 0.25 0.40 0.35 0.45 0.40 0.45 0.45 0.40 0.45
x7 (m) 0.35 0.35 0.45 0.35 0.50 0.50 0.40 0.50 0.50
x8 (m) 0.25 0.25 0.30 0.25 0.25 0.30 0.25 0.25 0.30
x9 (m) 0.25 0.30 0.35 0.25 0.40 0.50 0.25 0.40 0.50
x10 (m) 0.25 0.30 0.40 0.25 0.35 0.35 0.25 0.35 0.35
x11 – 0.75 0.63 – 1.00 1.25 – 1.00 0.88
x12 – 1.25 1.00 – 1.00 1.00 – 1.00 1.00
x13 – 1.00 0.75 – 1.25 1.50 – 1.25 1.25
x14 (MPa) 25 25 25 25 25 25 25 25 25
x15 (MPa) 25 40 25 40 40 40 40 40 40
x16 (MPa) – 25 25 – 25 25 – 25 25
11
Table 6 highlights the differences in the designs with and without the number of levels increases, the consideration of SSI becomes even
SSI. It is constructed with the basic case study (CS-1) results. First, it is more significant. It is due to the increase in axial forces and differential
observed that the superstructure of the models optimized with SSI is settlements. Furthermore, the predominantly frictional soil is the most
much more environmentally costly than that of the model with classical influential, obeying the theories discussed in Section 2.2. The figure also
supports: 12.03% for soil 1 and 18.81% for soil 2. It does not mean that confirms the significant influence of the columns on the differences
not including the SSI is better. On the contrary, using a model with observed for each support condition. It is magnified in CS-3 due to
classical supports does not consider additional stresses introduced into increased differential settlements and bending in the superstructure.
the structure, which is why its optimal design is “less costly”. When Thus, it is confirmed that this increase in bending resulting from the SSI
modeling the system with SSI, the optimal results are more expensive phenomenon significantly influences the design of the columns.
because the superstructure is more stressed. Thus, the elements are
designed with more material. Therefore, the design of a model that does
3.2. Application of KBMO
not consider the SSI does not cause the collapse of the structure but
rather a gradual and accelerated deterioration because of the appear
Conventional optimization using FE software as a computational
ance of stresses for which it was not designed. It causes a considerable
engine for modeling, analysis, and structural design is highly costly. For
increase in the use of resources for its maintenance. It should be
these reasons, the results will be measured in terms of the accuracy of
remembered that support settlements are a slow process that could take
the metamodel-based strategy compared to conventional optimization,
years to reach its final stage. On the other hand, obtaining a more costly
highlighting the savings achieved in computational time. The element
superstructure for soil 2 proves the hypothesis in section 2.2.1 about the
considered in the analysis of KBMO is the coefficient of penalization CP.
increase in internal forces resulting from differential settlement, which
It is important to remember that optimization based on metamodels
is higher for these types of soils.
aims to reduce computational costs while maintaining good results. In
On the other hand, it can be seen that the costs of the beams do not
this section, we use the models with SSI (with both soils) to test the
vary much from one model to another, the columns being the main ones
effectiveness of the proposed methodology.
affected. Differential settlements cause an increase in the bending
moment. This increase directly affects the beams. It is also transferred to
3.2.1. Obtaining the appropriate CP
the columns due to how statically indeterminate structures work (as in
As mentioned above, an aspect of significant importance in imple
this case). Therefore, the beams’ design does not “suffer” too much from
menting metamodel-based optimization is the coefficient of penalization
this increase in bending since they are elements intended to support this
applied to infeasible solutions. Therefore, the study of applying the
stress. However, columns are intended to support axial stress funda
KBMO methodology starts with searching for the appropriate CP. Case
mentally, so this increase in bending resulting from the consideration of
study 1 is used for this purpose.
the SSI does significantly affect their design. It causes a significant in
crease in the dimensions and rectangularity of the columns’ cross-
3.2.1.1. Predominantly cohesive soil. Table 7 and Fig. 13 show how, in
sections and the need for reinforcement to deal with these internal
general, the results are better for low CPs. It is because, as proved in
forces more efficiently. Table 6 shows that for soil 1, the cost of the
section 2.4.2, more accurate predictions are obtained for low CP values.
columns in comparison with the model using classic supports increases
It occurs in both strategy steps: the Kriging-based optimization (KBO)
by 43.35% (46.55% for concrete and 44.21% for steel), while for soil 2,
and the completion with the local heuristic technique. The final results
the increase is 61.19% (75.92% for concrete and 53.66% for steel).
are the best in the case of CP=1.15, even if the KBO offers the worst. The
Fig. 12 shows the graphical representation of Table 6. In addition,
response surface generated with this value of CP presents a zone of good
the results of case studies 2 and 3 are included. Here we can see how the
local optima corresponding to an area of the real surface close to the
conclusions obtained using the CS-1 are reaffirmed. As can be seen, there
global optimum. On the other hand, it can be said that the most stable
is a tendency to increase the absolute differences in the superstructure
values were obtained for a CP=1.05 since both results (intermediate and
design. In the case of CS-2, although the absolute differences increase
final) are excellent. This configuration also offers the most significant
compared to CS-1, their ratio expressed as a percentage of the model
computational savings.
without SSI decreases due to higher total emissions. However, it can be
In Fig. 13, the final results are pretty accurate. Three CP configura
seen how in CS-3, this ratio shoots up. Here it can be concluded that as
tions have above 98% accuracy compared to CO. The cases of CP equal
Fig. 12. Graphical representation of the results referring to SSI consideration’s influence on the superstructure. Left: Results shown in Table 6 and those obtained
with CS-2 and 3. Right: Absolute differences for each case study and each soil type compared to the model without SSI.
12
Table 7
Overview on results of applying the proposed metamodeling strategy in comparison with Conventional Optimization (CO) of the model considering SSI for the
predominantly cohesive soil.
Method CP Mean results Best results
Time (s) Comp. time savings CO2 emissions Accuracy with respect to CO CO2 emissions Accuracy with respect to CO
(%) (kg) (%) (kg) (%)
KBMO 1.03 10 840 91.97 20 786 98.76 20 511 99.72

1.05 10 706 92.07 20 902 98.20 20 717 98.71
1.10 15 930 88.20 21 243 96.53 20 727 98.67
1.15 15 255 88.70 20 684 99.26 20 454 100.00
1.20 14 351 89.37 21 495 95.31 21 346 95.64
Conv. Opt. 135 000 20 532 20 454
(CO)
Fig. 13. Graphic comparison of results shown in Table 7. Left: simple box plots (extreme values and mean) of three test performed for each configuration. Right:
Accuracy of metamodel-based optimization results (in comparison to CO) for different CPs (left y-axis) and corresponding computational savings (right y-axis).
to 1.03 and 1.05 are pretty outstanding, considering not only the final 3.2.1.2. Predominantly frictional soil. Table 8 and Fig. 14 show that the
results but also those obtained with the KBO. Note that one of the three intermediate (KBO) and final results (KBMO) are less efficient. It is
tests with CP=1.03 obtained a result accuracy of 99.72% compared to because this model using soil 1 is more “unstable”, where there are
the best solution obtained by the CO, being this value even better than certainly more constraint violations in a more deformed and stressed
one of the three results obtained by the conventional method. The best superstructure. It makes the actual response surface more difficult to
value found with CP=1.05 has an accuracy of 98.71%, another excellent predict. The results with the lowest CP (1.03) are remarkable, which are
result. On the other hand, this figure shows the computational savings poor compared to those obtained for soil 1. It is due to the previous
obtained. This saving lies almost entirely in substituting high-fidelity approach to the number of constraint violations, where this low penalty
simulations with metamodels. The number of high-fidelity simulations seems less efficient in predicting the real phenomenon. Again, the results
is determined by the number of simulations needed to obtain a first with CP= 1.15 lead to reasonably good final results since both models,
surface with a MAPE lower than 6% (step 1) and by the number of it even with different soils, should have similar overall behavior. It should
erations the local search algorithm must do to find the best solution. The be noted that the best solution was obtained with CP=1.05, with an
latter is influenced by the quality of the solution obtained at the end of accuracy of 99.24%. The computational savings were also lower due to
the KBO. Low CPs offer much more computational savings because of obtaining less accurate metamodels. Overall, the results are still quite
their more accurate surrogate response surfaces and better intermediate satisfactory.
solutions. In general, using low CPs (1.03 and 1.05) saves, on average,
almost 92% of computational time, with very accurate results compared 3.2.2. Application of KBMO with CP=1.05 to case studies 2 and 3
to CO. Having concluded that the best option for penalizing infeasible so
lutions is to apply a CP of 1.05, it is tested the proposed strategy with
case studies 2 and 3. Fig. 15 shows the results for the three cases. Part (a)
Table 8
Overview on results of applying the proposed metamodeling strategy in comparison Conventional Optimization (CO) of the model considering SSI for the predom
inantly frictional soil.
Method CP Mean results Best results
Time (s) Comp. time savings CO2 emissions Accuracy with respect to CO CO2 emissions Accuracy with respect to CO
(%) (kg) (%) (kg) (%)
KBMO 1.03 12 877 90.17 19 002 96.59 18 715 97.97

1.05 14 764 88.73 18 646 98.53 18 483 99.24
1.10 15 629 88.07 18 981 96.71 18 885 97.05
1.15 13 716 89.53 18 963 96.80 18 623 98.47
1.20 14 842 88.67 19 321 94.86 19 026 96.28
Conv. Opt. 131 000 18 376 18 343
(CO)
13
Fig. 14. Graphic comparison of results shown in Table 8. Left: simple box plots (extreme values and mean) of three test performed for each configuration. Right:
Accuracy of metamodel-based optimization results (in comparison to CO) for different CPs (left y-axis) and corresponding computational savings (right y-axis).
Fig. 15. (a) Simple box plots of the three tests performed (conventional optimization and KBMO) for each case study for both soil types. (b) Achieved accuracy of
KBMO compared to CO for each case. Note that the differences between the means and the minimum value obtained from the three tests are represented.
shows that the model for structure 2 is the most difficult to optimize, CS-3, the results are slightly higher: 90.5 and 89.4 %. In general, these
especially for soil 2, as shown in part (b). It is because this structure is results are similar to those obtained in CS-1.
probably more prone to obtain infeasible solutions. In this case, as the
length of the beams increases, this infeasibility is given by the stiffness 3.2.3. Exploring the systems under investigation
constraint of these elements (i.e., violation of the deflection constraint at One of the benefits of using metamodels is that thanks to the
the center of the span). This situation, as mentioned, becomes more cheapness of the simulations, several experiments can be performed, and
critical in soil 2, causing a more stressed superstructure. This theory the solution space can be explored better. It allows a deeper under
about the inversely proportional relationship between the number of standing of the systems under analysis. Considering that several solu
infeasible solutions and the accuracy of metamodel-assisted optimiza tions were obtained in the study of the appropriate CP using CS-1, these
tion strategies is verified by obtaining less accurate results for case are used to go inside into the structure’s behavior. Other more gener
studies 2 and 3. Additionally, this is proven by the fact that the struc alized conclusions are also accepted by analyzing the solutions obtained
tures modeled on soil 2 are more challenging to optimize using the with the other two case studies.
metamodel strategy. In Fig. 15(b), the results for model 1 suggest the Hypotheses on the behavior of this type of structure, such as
opposite, but this is true for this case of CP = 1.05 specifically. In the obtaining more stressed superstructures when modeling the SSI, have
previous section, it was found that, in general, the results are more stable been previously discussed. Several of these hypotheses can be deduced
for soil 1. from Fig. 16. Perhaps the most notable case is the relationship between
In summary, it can be stated that the results are pretty satisfactory the cost of beams and columns. It is often the case that as the cost of
since, on average, solutions above 98% accuracy are achieved compared concrete in beams decreases, there is an increase in the cost of columns,
to conventional optimization. In some specific cases, even 99% is significantly reinforcing steel, as can be seen in Fig. 16(c) comparing the
exceeded. Regarding computational savings, a similar analysis is per cases of CP=1.05 with CO or with CP=1.10. The same occurs in Fig. 16
formed for CS-2 and 3 as for CS-1 (tables 7 and 8). For CS-2, these (d) with CP=1.03 compared to CP=1.05 or CP=1.10 with CP=1.15. It is
savings, on average, are 90% and 89 % for soils 1 and 2, respectively. For because stiffer (more expensive) beams cause the columns to be
14
Fig. 16. Breakdown of each of the best solutions obtained by each configuration for CS-1, (a) and (b) represent the emissions broken down by elements, while (c) and
(d) denote each element broken down by the main components. Additionally, (a) and (c) are related to the soil 1, while (b) and (d) to the soil 2.
subjected to fewer bending moment and a corresponding reduction in 31, and 36%. For CS-2, this distribution is 34–25-41 for soil 1 and 40–31-
reinforcing steel. 29 for soil 2. In CS-3, the optimal distribution would be 32–28-40 for soil
It is essential to note the differences between the solutions for each 1 and 36–38-26 for soil 2. Regarding the components analyzed, it can be
soil type. Fig. 16(a) and (b) show that the columns of models with soil 2 seen that the formwork has very little influence, unlike the earthwork.
are more expensive, although the foundations are much cheaper due to On the other hand, the optimal (environmental) cost ratio of steel
the higher bearing capacity of this type of soil. Considering that foun –concrete is always helpful in reinforced concrete structures. For the CS-
dations are usually ignored in this type of study, a few lines are dedi 1 with soil 1, this ratio is 1.25, 1.45, and 3.01 for beams, columns, and
cated to their influence on the structural assembly. In CS-1, foundations foundations, respectively. For soil 2, the ratios are 1.25, 1.29, and 2.23.
represent 46 and 36% of the emissions of the structural assembly for The overall ratios are 1.61 for soil 1 and 1.37 for soil 2. For the second
soils 1 and 2, respectively. In CS-2, they represent 41 and 29 %, while in case study, these ratios are 1.08, 1.57, and 3.47 (1.60 in general) for soil
CS-3, 40 and 26 %, respectively. Therefore, foundations are very influ 1, and 0.96, 1.31, and 2.42 (1.20 in general) for soil 2. CS-3′s optimal
ential in the final output of this type of structure. In addition, they distribution is 1.12, 1.20, and 3.39 (1.47 in general) for soil 1, and 1.05,
regulate how the superstructure works with their geometry and the 1.51, and 2.35 (1.33 in general) for soil 2.
interaction with the soil. It is also essential to emphasize the significant
impact on CO2 emissions of earthwork activities. In CS-1, they represent 4. Concluding remarks and future work
62% of the total emissions of foundations (28% of global emissions) for
soil 1 and 75% (27% of the total) for soil 2. In CS-2, this ratio is 51% Although conventional heuristic optimization has been successfully
(21% of the total) for soil 1 and 69% (20% of the total) for soil 2. For CS- implemented for years in structural optimization, this strategy may be
3, these values are 51% (20%) and 71% (18%), respectively. insufficient to deal with more complex problems. This increase in
Other significant aspects are the optimal emission ratios between complexity is due to the development of more accurate models, even
elements and components. For CS-1 with soil 1, the distribution is though modeling aspects such as soil-structure interaction (SSI) are
approximately 30, 24, and 46% between beams, columns, and founda usually not considered. This paper has evaluated the use of metamodel-
tions. In the case of soil 2, this distribution is more equitable, leaving 33, assisted optimization to minimize the CO2 emissions of spatial
15
reinforced concrete frame buildings, taking into account the SSI, which Kripka: Conceptualization, Writing – review & editing, Supervision,
is implemented using a Winkler model. The surrogate models are built Project administration. Víctor Yepes: Conceptualization, Resources,
using a Kriging-based methodology. Supervision, Project administration, Writing – review & editing, Fund
Due to the particularities of the formulated discrete optimization ing acquisition.
problem, there are better alternatives than simple Kriging-based opti
mization. Thus, a meta-heuristic strategy is proposed using a Kriging- Declaration of Competing Interest
based two-phase methodology to build the surrogate model with a
local search algorithm. The authors declare that they have no known competing financial
Results show that the SSI should be addressed during modeling. interests or personal relationships that could have appeared to influence
When considering this aspect, the superstructure design provides the work reported in this paper.
different results than when it is not considered, i.e., the SSI consider
ation leads to more stressed superstructures, especially the columns. Data availability
This phenomenon is more evident in predominantly frictional soils than
in predominantly cohesive ones. Considering the foundations within the Data will be made available on request.
structural assembly is another aspect that should be addressed. They are
very influential both in the general outputs of the structure and in the Acknowledgments
way in which the system distributes the loads. On the other hand,
applying the Kriging-based meta-heuristic optimization allows for This work was supported by the grant PID2020-117056RB-I00,
obtaining very accurate results compared with those obtained with the which was funded by MCIN/AEI/10.13039/501100011033 and by
conventional heuristic optimization, with computational savings of “ERDF A way of making Europe”. Grant PRE2021-097197 funded by
about 90%. This accuracy varies depending on the coefficient of MCIN/AEI/10.13039/501100011033 and by FSE+.
penalization. Generally, better results are obtained for low coefficients
(1.03–1.05). References
Promising lines of research in this field should focus on implement
ing surrogate-based optimization to solve optimization problems with [1] Negrin I, Roose D, Chagoyén E, Lombaert G. Biogeography-Based Optimization of
RC structures including static soil structure interaction. Struct Eng Mech 2021;80
more encompassing formulations, considering environmental, social, (3):285–300. https://doi.org/10.12989/sem.2021.80.3.285.
and constructive objectives. Moreover, these objectives should not only [2] Gartner E. Industrially interesting approaches to “low-CO2 cements”. Cem Concr
be limited to mere optimization up to the design stage, i.e., they should Res 2004;34(9):1489–98. https://doi.org/10.1016/j.cemconres.2004.01.021.
[3] Thormark C. A low energy building in a life cycle-its embodied energy, energy need
also include the Life-Cycle Analysis of the structures. Another interesting for operation, and recycling potential. Build Environ 2002;37(4):429–35. https://
point of view would be identifying the least influential variables within doi.org/10.1016/S0360-1323(01)00033-6.
the optimization results and converting them into constant parameters [4] Yeo D, Gabbai R. Sustainable design of reinforced concrete structures through
embodied energy optimization. Energ Buildings 2001;43(8):2028–33. https://doi.
by assigning them values that have proven optimal from previous org/10.1016/j.enbuild.2011.04.014.
studies. It would make the optimization problem easier to solve, [5] Penadés-Plà V, García-Segura T, Yepes V. Accelerated optimization method for
considerably reducing computation time. Alternatively, considering that low-embodied energy concrete boxgirder bridge design. Eng Struct 2019;179:
556–65. https://doi.org/10.1016/j.engstruct.2018.11.015.
soil-structure interaction is a novel aspect in this research, the influence
[6] Camp CV, Huq F. CO2 and cost optimization of reinforced concrete frames using a
of soil conditions underlying the foundations, such as the presence of a big bang-big crunch algorithm. Eng Struct 2013;48:363–72. https://doi.org/
water table, more than one stratum, or drainage conditions, could be 10.1016/j.engstruct.2012.09.004.
investigated. Other ways of modeling this interaction can also be [7] Yepes V, Martí JV, García-Segura T. Cost and CO2 emission optimization of
precastprestressed concrete U-beam road bridges by a hybrid glowworm swarm
explored. An interesting alternative would be implementing other algorithm. Autom Constr 2015;49:123–34. https://doi.org/10.1016/j.
models beyond the proposed linear springs, such as using dashpots with autcon.2014.10.013.
a Kelvin-type or a similar approach. It would be very effective also for [8] Negrin IA, Chagoyén E. Economic and environmental design optimisation of
reinforced concrete frame buildings: A comparative study. Structures 2022;38:
dynamic analysis. Therefore, further refining the proposed methodology 64–75. https://doi.org/10.1016/j.istruc.2022.01.090.
with other more complex methods and checking if it is worthwhile [9] Paya I, Yepes V, González-Vidosa F, Hospitaler A. On the Weibull cost estimation of
would be fascinating. In addition, other artificial intelligence techniques building frames designed by simulated annealing. Meccanica 2010;45:693–704.
https://doi.org/10.1007/s11012-010-9285-0.
(e.g., Neural Networks) could be considered to deal with the high [10] Molina-Moreno F, García-Segura T, Martí JV, Yepes V. Optimization of buttressed
computational costs. Furthermore, once the importance of modeling SSI earth-retaining walls using hybrid harmony search algorithms. Eng Struct 2017;
in frame structures has been demonstrated, other typologies should be 134:205–16. https://doi.org/10.1016/j.engstruct.2016.12.042.
[11] Kaveh A. Advances in Metaheuristic Algorithms for Optimal Design of Structures,
tested to increase the sustainability of this type of structure based on the Springer International Publishing, Switzerland, 3rd edition, 2021. 10.1007/978-3-
methodologies proposed in this work. Once more complex structures are 319-05549-7.
implemented, other foundation alternatives such as combined footings, [12] Kaveh A. Applications of Metaheuristic Optimization Algorithms in Civil
Engineering, Springer. Switzerland 2017. https://doi.org/10.1007/978-3-319-
mat foundations, or mat-piles combinations can be used. An exciting
48012-1.
research to develop would be to analyze the sustainability of each of [13] Olivier J, Janssens-Maenhout G, Muntean M, Peters J, Trends in Global CO2
them and their relation with the superstructure design. Regarding the Emissions,. Report. Joint Research Centre, Hague: European Commission; 2015.
latter, variable section beams, “I” or “T” sections, slab-column typol [14] Lagaros N. A general purpose real-world structural design optimization computing
platform. Struct Multidiscip Optim 2013. https://doi.org/10.1007/s00158-013-
ogies (without beams), hybrid elements, and others can be alternatives 1027-1.
to be explored to improve the economic, environmental, constructive, [15] Esfandiari M, Urgessa G, Sheikholarefin S, Dehghan Manshadi G. Optimum design
social, and durability indexes of this type of construction. Another of 3D reinforced concrete frames using DMPSO algorithm. Adv in Eng Soft 2018;
115:149–60. https://doi.org/10.1016/j.advengsoft.2017.09.007.
promising line of research is focused on the type of optimization [16] Martins A, Simões L, Negrão J, Lopes A. Sensitivity analysis and optimum design of
formulation. Performance-based, reliability-based, or robust designs are reinforced concrete frames according to Eurocode 2. Eng Opt 2019. https://doi.
interesting and efficient approaches when considering other aspects, org/10.1080/0305215X.2019.1693554.
[17] Mergos P. Optimum design of 3D reinforced concrete building frames with the
such as dynamic analysis or uncertainty. flower pollination algorithm. J Build Eng 2021;44. https://doi.org/10.1016/j.
jobe.2021.102935.
CRediT authorship contribution statement [18] Salimi P, Bondarabadi H, Kaveh A. Optimal Design of Reinforced Concrete Frame
Structures Using Cascade Optimization Method. Period Polytech-Civ Eng 2022;66
(4):1220–33. https://doi.org/10.3311/PPci.20868.
Iván Negrin: Conceptualization, Methodology, Software, Investi
gation, Formal analysis, Visualization, Writing – original draft. Moacir
16
[19] Filomeno-Coelho R. Metamodels for mixed variables based on moving least Mead. Obras y Proyectos 2019;26:74–86. https://doi.org/10.4067/S0718-
squares. Application to the structural analysis of a rigid frame. Optim Eng 2013. 28132019000200074.
https://doi.org/10.1007/s11081-013-9216-8. [42] Triches-Boscardin J, Yepes V, Kripka M. Optimization of reinforced concrete
[20] Bäckryd RD, Ryberg AB, Nilsson L. Multidisciplinary design optimisation methods building frames with automated grouping of columns. Autom Constr 2019;104:
for automotive structures. Int. J. Automot. Mech. Eng., 2017, 14(1), pp. 4050-67. 331–40. https://doi.org/10.1016/j.autcon.2019.04.024.
10.15282/ijame.14.1.2017.17.0327. [43] Paya I, Yepes V, Hospitaler A, González-Vidosa F. CO2-optimization of reinforced
[21] Negrin I, Chagoyén E, Negrin A. Parameter tuning in the process of optimization of concrete frames by simulated annealing. Eng Struct 2009;31:1501–2158. https://
reinforced concrete structures. DYNA 2021;88(216):87–95. http://doi.or doi.org/10.1016/j.engstruct.2009.02.034.
g/10.15446/dyna.v88n216.87169. [44] García-Segura T, Penadés-Plá V, Yepes V. Sustainable bridge design by metamodel-
[22] Navarro I, Yepes V, Martí JV, González-Vidosa F. Life cycle impact assessment of assisted multi-objective optimization and decision-making under uncertainty.
corrosion preventive designs applied to prestressed concrete bridge decks. J Clean J Clean Prod 2018;202:904–15. https://doi.org/10.1016/j.jclepro.2018.08.177.
Prod 2018;196:698–713. https://doi.org/10.1016/j.jclepro.2018.06.110. [45] Penadés-Plà V, Yepes V, García-Segura T. Robust decision-making design for
[23] Pons JJ, Penadés-Plà V, Yepes V, Martí JV. Life cycle assessment of earth-retaining sustainable pedestrian concrete bridges. Eng Struct 2020. https://doi.org/
walls: An environmental comparison. J Clean Prod 2018;192:411–20. https://doi. 10.1016/j.engstruct.2019.109968.
org/10.1016/j.jclepro.2018.04.268. [46] Mathern A, Penadés Plà V, Armesto-Barros J, Yepes V. Practical metamodel-
[24] Khatibinia M, Salajegheh E, Salajegheh J, Fadaee M. Reliability based design assisted multi-objective design optimization for improved sustainability and
optimization of reinforced concrete structures including soil structure interaction buildability of wind turbine foundations. Struct Multidisc Optim 2022. https://doi.
using a discrete gravitational search algorithm and a proposed metamodel. Eng org/10.1007/s00158-021-03154-0.
Optim 2013;45(10):1147–65. https://doi.org/10.1080/0305215X.2012.725051. [47] Martínez-Martín FJ, Yepes V, González-Vidosa F, Hospitaler A. Optimization
[25] Zhang Y, et al. Two-dimensional nonlinear earthquake response analysis of a Design of RC Elevated Water Tanks under Seismic Loads. Appl Sci 2022;12(5635).
bridge-foundation-ground system. Earthq Spectra 2018;24:343–86. https://doi. https://doi.org/10.3390/app12115635.
org/10.1193/1.2923925. [48] Simon D. Biogeography-Based Optimization. IEEE Transactions on evolutionary
[26] Klepikov SN. General solution for beams and plates on elastically deforming bases computation, 2008, 12(6.), pp. 702-13. 10.1109.
with varying stiffnesses. In: Bases, foundations and soil mechanics research. [49] Aydogdu I. Cost optimization of reinforced concrete cantilever retaining walls
Budivielnik: Kiev; 1969. p. 37–47. under sismic loading using a boiogeograpy-based optimization algorithm with
[27] Klepikov SN, Tregub AS, Matveev IV. Calculation of buildings and works on Levy flights. Eng Opt 2017;49(3):381–400. https://doi.org/10.1080/
collapsible soils. Kiev: Budivielnik; 1987. 0305215X.2016.1191837.
[28] Bao T, Lui Z. Evaluation of Winkler Model and Pasternak Model for Dynamic Soil- [50] Shallan O, Maaly HM, Sagiroglu M, Hamdy O. Design optimization of semi-rigid
Structure Interaction Analysis of Structures Partially Embedded in Soils. Int. J. space steel frames with semi-rigid bases using biogeography-based optimization
Geomech., Vol. 20, No. 2. 10.1061/(ASCE)GM.1943-5622.0001519. and genetic algorithms. Struct. Eng. Mech., 2019, 70(2), pp. 221-31. 10.12989/
[29] Matlock, H. Correlations for design of laterally loaded piles in soft clay. In Vol. sem.2019.70.2.221.
1204 of Proc., Offshore Technology in Civil Engineering’s Hall of Fame Papers [51] Negrin I, Roose D, Chagoyén E. Parameter tuning strategies for metaheuristic
from the Early Years, 1970, 77–94. Houston: Offshore Technology Conference. methods applied to discrete optimization of structural design. Available at: Revista
10.4043/1204-ms. Inv Op 2022;43(2):241–58. https://rev-inv-ope.pantheonsorbonne.fr/sites/de
[30] Cox W., Reese L. C., Grubbs B. R. Field testing of laterally loaded piles in sand. In fault/files/inline-files/43222-08.pdf.
Proc., Offshore Technology Conf., 1974, 1–14. Houston: Offshore Technology [52] Myers R, Montgomery D, Anderson-Cook C. Response surface methodology:
Conference. 10.4043/2079-MS. process and product optimization using designed experiments 1995. https://doi.
[31] Reese LC, Welch RC. Lateral loading of deep foundations in stiff clay. J Geotech org/10.2307/1270613.
Eng Div 1975;101(7):633–49. https://doi.org/10.1061/AJGEB6.0000177. [53] Chuang CH, et al. Multidisciplinary design optimization on vehicle tailor rolled
[32] Taghavi A, Muraleetharan KK. Analysis of laterally loaded pile groups in improved blank design. Struct Multidiscip Opt 2008;31:551–60. https://doi.org/10.1007/
soft clay”. Int J Geomech 2016;17(4):1–13. https://doi.org/10.1061/(ASCE) s00158-007-0152-0.
GM.1943-5622.0000795. [54] McKay M, Beckman RJ, Conover WJ. Comparison of Three Methods for Selecting
[33] Mottaghi L, Izadifard RA, Kaveh A. Factors in the Relationship Between Optimal Values of Input Variables in the Analysis of Output from a Computer Code.
CO2 Emission and Optimal Cost of the RC Frames. Period Polytech-Civ Eng 2021; Technometrics 1979;21(2):239–45. https://doi.org/10.1080/
65(1):1–14. https://doi.org/10.3311/PPci.16790. 00401706.1979.10489755.
[34] Kaveh A, Izadifard RA, Mottaghi L. Optimal design of planar RC frames considering [55] Forrester AIJ, Keane AJ. Recent advances in surrogate-based optimization. Prog
CO2 emissions using ECBO, EVPS and PSO metaheuristic algorithms. J Build Eng Aerosp Sci 2009;45:50–79. https://doi.org/10.1016/J.PAEROSCI.2008.11.001.
2020;28. https://doi.org/10.1016/j.jobe.2019.101014. [56] Li YF, Ng SH, Xie M, Goh TN. A systematic comparison of metamodeling
[35] Kaveh A. Cost and CO2 Emission Optimization of Reinforced Concrete Frames techniques for simulation optimization in decision support systems. Appl Soft
Using Enhanced Colliding Bodies Optimization Algorithm. In: Applications of Comput 2010;10:1257–73. https://doi.org/10.1016/J.ASOC.2009.11.034.
Metaheuristic Optimization Algorithms in Civil Engineering. Cham: Springer; [57] Lophaven SN, Nielsen HB, Søndergaard J. DACE: A Matlab Kriging toolbox. Version
2017. https://doi.org/10.1007/978-3-319-48012-1_17. 2.0., Kongens Lyngby, Denmark: Informatics and Mathematical Modelling,
[36] Catalonia Institute of Construction Technology. BEDEC PR/PCT ITEC material Technical University of Denmark, 2002. 10.2307/1270613.
database; 2016. [58] Krige DG. A statistical approach to some mine valuation and allied problems on the
[37] Medeiros GF, Kripka M. Optimization of reinforced concrete columns according to Witwatersrand. University of the Witwatersrand, South Africa; 1951. Master’s
different environmental impact assessment parameters. Eng Struct 2014;59: thesis,Available at:.
185–94. https://doi.org/10.1016/j.engstruct.2013.10.045. [59] Cressie NAC. Statistics for spatial data. New York, NY, USA: John Wiley & Sons;
[38] Medeiros GF, Kripka M. Modified harmony search and its application to cost 1993.
minimization of RC columns. Adv. Comput. Des., 2017, 2(1), pp. 1-13. 10.12989/ [60] Lophaven SN, Nielsen HB, Søndergaard J. Aspects of the Matlab toolbox DACE,
acd.2017.2.1.001. Kongens Lyngby. Denmark: Informatics and Mathematical Modelling, Technical
[39] Yepes V, García-Segura T, Martí JV, Alcalá J. Optimization of concrete I-beams University of Denmark; 2002. Available at: https://citeseerx.ist.psu.edu/viewdoc/
using a new hybrid glowworm swarm algorithm. Lat Am J Solids Struct 2014; download?doi=10.1.1.69.7029&rep=rep1&type=pdf.
Volumen 11:1190–205. https://doi.org/10.1590/S1679-78252014000700007. [61] Nelder JA, Mead R. A simplex method for function minimization. Comput J 1965;
[40] Serpik I, Mironenko I, Averchenkov V. Algorithm for Evolutionary Optimization of 7:308–13. https://doi.org/10.1093/comjnl/7.4.308.
Reinforced Concrete Frames Subject to Nonlinear Material Deformation- [62] Dantzig GB. Linear Programming and Extensions. Santa Monica, California:
International Conference on Industrial Engineering, ICIE 2016. Procedia Princeton University Press, 1963. Dantzig GB. Linear Programming and Extensions.
Engineering, Issue 150, pp. 1311-16. 10.1016/j.proeng.2016.07.304. Santa Monica, California: Princeton University Press, 1963. 10.7249/R366.
[41] Negrin I, Negrin MA, Chagoyén E. Optimización de pórticos planos de hormigón
armado utilizando una hibridación de algoritmos genéticos y el algoritmo Nelder-
17

1 s2.0 S0141029623010726 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S0141029623010726 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0141029623010726 Main

Uploaded by

Copyright:

Available Formats

Engineering Structures 293 (2023) 116657

Contents lists available at ScienceDirect

Metamodel-assisted meta-heuristic design optimization of reinforced

1. Introduction excellent alternatives to solve large-scale and highly nonlinear optimi

coordinates (Ŕ ,S) and unlimitedly approaches the asymptote p=qbr_II

2.2.2. Algorithm for SSI modelling during structural optimization

(◦ ) (kPa) (MPa) kN/m3 (◦ )

2.4. Solution of the optimization problem

Fig. 10. Flowchart of Kriging-based meta-heuristic optimization (KBMO).

The results and their discussion focus on two fundamental aspects:

CS-1 CS-2 CS-3

No SSI Soil 1 Soil 2 No SSI Soil 1 Soil 2 No SSI Soil 1 Soil 2

KBMO 1.03 10 840 91.97 20 786 98.76 20 511 99.72

KBMO 1.03 12 877 90.17 19 002 96.59 18 715 97.97

You might also like

1 s2.0 S0141029623010726 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S0141029623010726 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0141029623010726 Main

Uploaded by

Copyright:

Available Formats

Engineering Structures 293 (2023) 116657

Contents lists available at ScienceDirect

Metamodel-assisted meta-heuristic design optimization of reinforced

1. Introduction excellent alternatives to solve large-scale and highly nonlinear optimi­

coordinates (Ŕ *,S) and unlimitedly approaches the asymptote p=q*br_II

2.2.2. Algorithm for SSI modelling during structural optimization

(◦ ) (kPa) (MPa) kN/m3 (◦ )

2.4. Solution of the optimization problem

Fig. 10. Flowchart of Kriging-based meta-heuristic optimization (KBMO).

The results and their discussion focus on two fundamental aspects:

CS-1 CS-2 CS-3

No SSI Soil 1 Soil 2 No SSI Soil 1 Soil 2 No SSI Soil 1 Soil 2

KBMO 1.03 10 840 91.97 20 786 98.76 20 511 99.72

KBMO 1.03 12 877 90.17 19 002 96.59 18 715 97.97

You might also like

1. Introduction excellent alternatives to solve large-scale and highly nonlinear optimi

coordinates (Ŕ ,S) and unlimitedly approaches the asymptote p=qbr_II