Structural Optimization Using Genetic Algorithms

2022 International Conference on INnovations in Intelligent SysTems and Applications (INISTA) | 978-1-6654-9810-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/INISTA55318.2022.9894243

1st Alexandra Băicoianu 2nd Alexandru Garofide

Department of Mathematics and Informatics Siemens Industry Software
Transilvania University of Braşov Braşov, Romania
Braşov, Romania [Link]@[Link]
[Link]@[Link]

3rd Roberta-Iuliana Luca 4th Mihai Vlădărean

Faculty of Mathematics and Informatics SoundStack
Transilvania University of Braşov Braşov, Romania
Braşov, Romania vladareanmihai@[Link]
[Link]@[Link]

Abstract—One of the most discussed topics at present is related more in their life. Applied together or individually, these
to optimization processes. We are living in a world in which techniques provide solutions for different problems. The prob-
everything gets more and more expensive every day. So, to avoid lem is to optimize a structure, so the answer came quickly
a significant crisis, we need to re-evaluate the cost of everything
that surrounds us, including the materials used in construction. consisting of a GA development.
The problem that comes is not related to the materials themselves, This research aim is to successfully implement a GA that,
but the quantities used. For each structure, we can deduce two starting from an input structure with different restrictions, the
requirements: practicability and resistance. To meet these, the evolutionary algorithm can optimize it. It must determine this
tendency is to create a structure with more support than it needs structure considering some constraints and the stress to which
just to guarantee safety. Unfortunately, this implies a greater
weight that the structure must support and also greater costs it is subjected. For particular inputs, the chosen methodology
for building it. The object of this paper is to properly develop did not provide the optimal solution due to the way the
a Genetic Algorithm based on the methods of mathematical algorithm performs, the study will, however, provide a result
programming for this class of structural optimization problems that is as close as possible to the ideal one. This is a sacrifice
in the presence of multiple design constraints. This state-of- that must be taken when using GAs, as the brute force
the-art heuristic solution, alongside other related works, should
demonstrate the utility of Genetic Algorithms in these types of approach could take billions of years to provide a solution,
problems, and improve the time necessary for the implementation while our method takes only a few seconds. Therefore, we
of an optimal architecture. choose to take the chance of not providing the best solution
Index Terms—Genetic Algorithm, Finite Element Method, Fi- possible, but gaining massively on the computational time side.
nite Element Analysis, Structural Optimization, Fitness Function This study should demonstrate the utility of GAs in real
optimization problems and give a solution for the automation
I. P ROBLEM S TATEMENT of this process. In this way, we can let people be involved
in those activities that cannot be realized without human
In manufacturing, it is often desirable to minimize the intervention.
amount of material used to package a product with a certain Moreover, the proposed solution is more efficient than the
volume. Our study decided on the approach of using Genetic classical mathematical optimization methods [9], [11] that are
Algorithms (GAs) [8] to optimize this quantity of material in used on large scale for this subject.
some construction.
The current broadly used approach is represented by man- II. BACKGROUND AND R ELATED W ORK
ual calculations. The architects and the engineers need to The problem of the time-consuming manual structural op-
reconsider again and again different scenarios to find the timization has been observed since the beginning of the
best solution. This whole process finishes, in general, with field. Structural optimization generates a component design
a compromise because it is extremely hard for them to create that exhibits maximum structural utility subject to a set of
both a resistant and low-priced structure. Another disadvantage functional requirements and constraints on the component’s
of this method is the constant need for at least one person who structural behavior [6]. Because of this, multiple studies have
can do all the mathematical operations and deduce different been actively carried out to propose different solutions to this
formulas. optimization problem. The use of GAs has also been studied
In the last few years, Artificial Intelligence techniques on multiple occasions [1], [2], [5], [7], [12]. The difference
showed remarkable improvements, helping people more and between these studies consists of a different gene encoding,
978-1-6654-9810-4/22/$31.00 ©2022 IEEE different selection and crossover methods, different fitness

Authorized licensed use limited to: Biruni Universitesi. Downloaded on November 14,2022 at 04:02:51 UTC from IEEE Xplore. Restrictions apply.
functions etc., each of them trying to reach the best solution A GA generates a set number of possible solutions. The
for the considered constraints of the structural component. total of solutions is called the population, while the solution
Actually, [1] presents a way in which the parameters for itself is called a chromosome, or simply an individual. The
a classifier are adjusted using GAs. Closer to the problem number of individuals in a generation is preserved throughout
discussed in this study are [7] and [2]. The research presented the entirety of the program. Each individual is evaluated in
in [7] focuses on a particular type of arch bridge. While in order to determine their fitness and a hierarchy is then formed.
our approach, a chromosome is represented by a structure, in A predetermined percentage of individuals are determined
their proposal, a chromosome can be anything related to the subpar and cast aside, while the others are considered to have
geometrical parameters, such as lengths, widths, diameters or a higher probability to generate a suitable answer, and thus
some material properties. Moreover, even though the purpose are sent to mate and produce offspring. The phase consisting
is the same - structural optimization, the approaches are quite of crossing and mutating occurs randomly. Although now that
different because, in this article, the focus is only on a well- there are only fit individuals, the offspring’s odds of being fit
determined type of bridge, while our idea can be used for are increased. This process is then reiterated a set number of
any type of structure. The paper [2] is very similar to the times and the solution is determined to be the fittest individual
one described in [7], the main purpose being the optimization of the last generation.
of some geometrical parameters which describe the arch There are a lot of advantages when we are talking about
of a bridge. Compared to our methodology, they use finite using the approach mentioned above over classical optimiza-
element analysis, exactly like we do. However, in our case, tion. An evolutionary algorithm is able to deal with various and
the algorithm has much more freedom to drive the shape of complex types of optimizations, allowing the objective fitness
the resulting structure. function to take different formats even if we are thinking
The usage of evolutionary algorithms is argued on their about stationarity, linearity, or continuity. Another noticeable
heuristic approach. When it comes to optimization problems, advantage of GAs is parallelism. The solution to the problem
which this is, gaining an optimal solution in a shorter amount is searched in multiple directions simultaneously because the
of time, rather than spending way more time in an attempt to offspring of the population act as independent individuals.
obtain the best possible outcome, is the preferred way. This On the other hand, there are also some disadvantages. It is
is where GAs excel, as mixing better and better solutions not easy to find the appropriate fitness function, the termina-
will cause the quality of solution to raise faster and faster, tion criteria, or the values for some important parameters such
until reaching a balance, where any solution contained in a as the rate of crossover, mutation, or the population size [15].
generation of solutions is already satisfactory, at which point, B. Finite Element Analysis
the raise gets slower, but the solutions barely change.
Finite Element Analysis (FEA) is widely used in engineer-
One big advantage to this type of algorithms is that they
ing problems to find solutions for issues with high complexity
are very versatile and can be used in structural optimization
and it is considered one of the most important instruments an
for a lot of problems, and therefore, for different components.
engineer can possess. There are a lot of engineers saying that
Moreover, even if the GA has certain fixed components, these
without FEA we wouldn’t have had so many things around
can also be improved and adapted for the problem at hand.
us.
Considering these advantages, we chose to use GAs in order
FEA is the process of analyzing the behavior of a single
to explore how they influence the process of minimizing the
part or a group of elements from an assembly having different
quantities of materials used in a structure.
constraints. This process is actually the simulation of the
III. T HEORETICAL A SPECTS AND F RAMEWORKS mathematical equations deduced using the Finite Element
A. Genetic Algorithms Method (FEM) [10]. This whole idea is considered to be the
center of engineering but not only. It is also widely used for
Genetic Algorithms are heuristic solution-search or opti-
software simulations.
mization methods [13]. Intuitively, they work in a similar
Even though the finite element method appeared officially
way to how natural evolution works. Their fundamentals were
only in 1960, the concept was discovered centuries ago and
developed based on the same ideas as the Darwinian principle
used in a natural, usual way. The main idea is to find the
of natural selection.
solution to a complex problem by replacing it with another
They were developed by John Holland, who introduced the
one for which the process of solving is already known.
concept of population individuals and defined the mandatory A significant application of those concepts is structural
operations for a GA - selection, crossover, and mutation optimization which is used to minimize or maximize the
[4]. They represent one of the Model Evolution Calculus objective fitness function taking into consideration different
branches, being used for solving optimization problems [2]. restrictions. This purpose fits the aim of the current research,
The Model Evolution Calculus is inspired by Darwin’s Theory so we will use these methods in our simulations [3].
of Evolution and genetics - the science of heredity introduced
by G. Mendel. Unlike other approaches, these solutions use C. Project Chrono
a population of elements instead of a single value to find the Project Chrono [20] is an open-source C++ library for
optimal solution for a problem. physics simulations. It can be embedded in different software

Authorized licensed use limited to: Biruni Universitesi. Downloaded on November 14,2022 at 04:02:51 UTC from IEEE Xplore. Restrictions apply.
projects to simulate robots, mechatronic systems, virtual re- More details about this whole process can be viewed in
ality, augmented reality, vehicle dynamics, and other physics- IV-C2.
based phenomena. Even more, Chrono allows us to use finite 3) The Selection Step: Selection is the step in which some
elements and model flexible parts such as beams, cables, or parents are selected to generate offspring. These resulted in
hexahedrons. In this way, we can customize our own meshes, children becoming part of the next generation. For this, it was
change the materials and add forces. used roulette wheel selection [14], a stochastic method of se-
Using multibody dynamics, we can run simulations of rigid lection where the fitness value of an individual is proportional
bodies, add restrictions to different parts, and add motors, to its probability of selection.
dampers, or linear actuators. Chrono can be used for large- For this step, the fitness function is essential. The chromo-
scale simulations, collision detection, and vehicle dynamics. somes with the best fitness values are most likely to be chosen
There are a lot of advantages when we are thinking about in this step. Being based on probabilities, there are smaller
using Chrono. Starting from the wide range of usage areas, the chances for candidates with worse function values to appear
possibility to have visual results, stability, and ending complex in the next generation.
documentation, Chrono represents a real source of inspiration. For each individual, has been computed its cumulative
In the following sections, Project Chrono was used for probability of selection that is used next to decide whether
testing the evolutionary algorithm. Simulations for different that chromosome will be chosen for the crossover step or not.
scenarios were executed, meshes were created and custom 4) The Crossover Step: The crossover operator matches
operations for these were implemented. the selected set of individuals to produce offspring. For those
IV. M ETHODS AND A NALYSIS chromosomes that were selected to participate in the crossover
process, we used the single-point crossover to interchange
The purpose of this section is to interpret and describe the
parts of the genes of the parents. Changing the genes means
significance of our findings in relation to the research problem
automatically adding or removing some material from the
being investigated and to explain any new understanding or
structures corresponding with the individuals.
insights that emerged as a result of our research.
5) The Mutation Step: The mutation operator is used on
A. Genetic Algorithm Overview a limited number of genes of the offspring. It alters the
1) The Genetic Code: According to Charles Darwin, de- value of the gene in order to diversify the new generations
spite appearances, “it is not the strongest of the species that and minimize the chances of premature convergence. For our
survives, nor the most intelligent, but the one most responsive problem, mutation means adding or removing elements from
to change”. the mesh based on the mutation probability. A mesh, in our
In order to find the optimal solution, our GA simulates case, is formed by finite elements represented by cubes, those
the evolution of individuals using natural selections. There elements being held together using the nodes. Bit-flip mutation
are two main processes. The first one is the selection of the was considered for this step, meaning that the selected gene
best individuals used to produce the next generation. And will be replaced with its opposite. Having values of 0 and 1,
the second one is the mechanism that processes the selected gene mutation means changing its value with the other one.
individuals using crossover and mutations to create offspring.
According to natural selection, only those individuals who are B. Building the Next Optimized Generation
able to adapt to the changes in the environment will survive The suggested GA gathers all the steps mentioned above
and help in creating the next optimal generation. which can be also visualized in Fig. 1. Firstly, the population
The proposed GA implements all these steps, specializing was initialized. Even though in this step the algorithm usually
all the operations for the problem discussed in. generates a random population of chromosomes that can be
2) The Fitness Function: The fitness function is the main solutions to the problem, we decided that it would be better to
component in a GA because it says how “fit” the solution for start from some pre-build structures. This helps in guiding the
the exposed problem is. Unfortunately, there is no specific set algorithm to preserve some empty spaces in the construction.
of rules for determining this function, so the best way to find For each individual, the fitness value was computed and then
the appropriate one for a specific problem is by using “trial and the three main processes were executed: selection, crossover
error” method. The algorithm was tested for multiple objective and mutation. For all these processes it was used the binary
functions and we concluded that the best one is the following: encoding for each individual. It has been created a vector of
F = R2 ∗ (maxS − maxSS ) (1) genes containing 0 and 1 values where 1 represents that the
corresponding part of the construction is present and 0 that
where: it was removed. This idea is illustrated in Fig. 2. All these
F - the fitness function steps are repeated until the termination condition is met. The
R - the number of the removed elements algorithm was stopped when we arrived at a specific number
maxS - the maximum stress allowed of generations or find an individual that is ”‘good enough”’
maxSS - the maximum stress on the structure for the measurements. In the end, an optimized structure was
achieved.

Authorized licensed use limited to: Biruni Universitesi. Downloaded on November 14,2022 at 04:02:51 UTC from IEEE Xplore. Restrictions apply.
 1) First Tests: The first approach used as an objective
function a combination of the number of removed elements
and the stress to which the structure was subjected:

F = 2R ∗ (maxS − maxSS ) (2)

This layer of tests showed that the structure was optimized

but not as much as we would like. The results can be visualized
in Fig. 4. It’s obvious that the quantity of materials used was
Fig. 1. Key steps in a Genetic Algorithm.
minimized but if we are talking about structural optimization,
these results are not very optimistic.
Apparently, using this function as our fitness function,
determined the algorithm to oscillate between less than 10
individuals because of the exponential part of the function.
The number of removed elements has to be important, but the
results show that the weight accorded to it is too high.

Fig. 2. Encoded genes for individuals and their graphic representation.

C. Discussions and Results

Fig. 4. Resulted structure after the first tests.
In order to evaluate and see the behavior of the structures
optimized by the GA, using Project Chrono [21] some simu-
2) Second Tests: After the first results, it was obvious that
lations were executed.
the objective function needed some changes. The algorithm
There are three main steps for setting up a simulation. The was tested with different fitness functions to select which one
first one, and probably the most important, is choosing the produces the best results. It was considered a small structure
system. Chrono offers support for two types of systems, one with 10 cubes on the Ox axis, 5 on the Oy axis and 3 on the
of them being used for non-smooth contacts and the other one Oz axis. Having an initial population of 100 individuals and
for smooth contacts. In our case, considering that we do not different fitness functions, Table I shows how many iterations
have collisions in our simulations, there is no difference in the algorithm executed until the population got stabilized. The
using one or the other system. In the next step, we had to population is considered stable when a small percentage of
choose a time-stepper, also known as a time integrator. We individuals cover the majority of chances of being selected for
used Euler-Implicit-Linearized, especially because it supports crossover. More specifically, we considered that the population
FEA well and it is also fast. The last step is selecting a solver is stable when 10% of the individuals cover 90% of the
which must also work with finite elements. Our choice was chances of being selected for the next steps.
to use Minres solver because of its good convergence and its Considering the notation:
support in working with FEA.
Fig. 3 presents all these steps, showing also some of the S = maxS − maxSS (3)
possibilities for each component that must be set. The chosen
options are specified in bold.
TABLE I
T HE NUMBER OF ITERATIONS FOR EACH FITNESS FUNCTION

No. Fitness Function Iterations

(1) 2R *S 4
(2) R0.5 *S 31
(3) R*S 40
(4) R2 *S 40

After multiple tests and after analyzing the evolution of

the four fitness functions shown in Fig. 5, we concluded
Fig. 3. Steps for setting up a simulation. that the last function is the best for our purpose. The main
improvement that we can observe in the resulting structures is

Authorized licensed use limited to: Biruni Universitesi. Downloaded on November 14,2022 at 04:02:51 UTC from IEEE Xplore. Restrictions apply.
the fact that the algorithm eliminated a lot of elements from solution, we tested the behavior of the genetic algorithm in
the second layer. Some of the results are exposed in Fig. 6. multiple scenarios.
The implementation and all the tests per- The exposed numerical tests presented demonstrate the
formed are available on a private repository, computational advantages of the discussed approach, which
[Link] that can be become more pronounced in large-scale structural optimization
accessed on request. problems. The proposed evolutionary algorithm proved to be
a robust and efficient method for structural optimization.
R EFERENCES
[1] E. D. Carvalho, R. R. V. Silva, F. H. D. Araújo, R. A. L. Rabelo and A.
O. de Carvalho Filho, “An approach to the classification of COVID-19
based on CT scans using convolutional features and genetic algorithms”,
Comput Biol Med., September 2021.
[2] M. Z. A. Elrehim, M. A. Eid and M. G. Sayed, “Structural optimization
of concrete arch bridges using Genetic Algorithms”, Ain Shams Engi-
neering Journal, vol. 10, Issue 3, September 2019, pp. 507-516, ISSN
2090-4479.
[3] L. Haldurai, T. Madhubala and R. Rajalakshmi, “A Study on Genetic
Algorithm and its Applications”, International Journal of Computer
Sciences and Engineering, vol. 4, Issue 10, October 2016, ISSN 2347-
2693.
[4] J. Holland, “Genetic Algorithms”, Scientific American, vol. 267, no. 1,
Fig. 5. Evolution of fitness values for each objective function. 1992, pp. 66–73.
[5] W. M. Jenkins, “Towards structural optimization via the genetic algo-
rithm”, Computers & Structures, vol. 40, Issue 5, 1991, pp. 1321-1327,
ISSN 0045-7949.
[6] U. Kirsch, “Structural Optimization: Fundamentals and Applications”,
Springer, 1993, ISBN 3540559191.
[7] K. Korus, M. Salamak and M. Jasiński, “Optimization of geometric
parameters of arch bridges using visual programming FEM components
and genetic algorithm”, Engineering Structures, vol. 241, August 2021,
ISSN 0141-0296.
[8] A. Lambora, K. Gupta and K. Chopra, “Genetic Algorithm- A Literature
Review,” 2019 International Conference on Machine Learning, Big Data,
Cloud and Parallel Computing (COMITCon), 2019, ISBN 978-1-7281-
Fig. 6. Final results. 0211-5.
[9] M. M. Meerschaert, “Mathematical Modeling”, 3rd ed., Elsevier, 2007,
ISBN 9780123708571.
V. C ONCLUSIONS AND F UTURE W ORK [10] S. S. Rao, “The Finite Element Method in Engineering”, 6th ed.,
Butterworth-Heinemann, 2018, ISBN 9780128117682.
The intention of this paper was to find a sense for using [11] R. A. Sarker and C. S. Newton, “Optimization Modelling: A Practical
GAs in structure optimizations. The algorithm proved to be Approach”, 1st ed., CRC Press, 2008, ISBN 9781420043105.
[12] Y. Sun, B. Xue, M. Zhang and G. G. Yen, “Evolving Deep Convolutional
the best fit for tackling this issue. In addition, the way it is Neural Networks for Image Classification”, arXiv, October 2017.
designed, by considering the solution as an abstract notion that [13] M. D. Vose, “The simple genetic algorithm : foundations and theory”,
has little to no correlation to its meaning, allows for working MIT Press, 1999, ISBN 026222058X.
[14] S. L. Yadav and A. Sohal, “Comparative Study of Different Selection
with an enormous number of solutions without crippling the Techniques in Genetic Algorithm”, International Journal of Engineering,
performance. Even though this study uses the standard steps Science and Mathematics, vol. 6, Issue 3, July 2017.
of a GA, the applicability domain is rather new and not so [15] X. S. Yang, “Nature-Inspired Optimization Algorithms”, 2nd ed., Aca-
demic Press, 2021, pp. 91-100, ISBN 9780128219867.
exploited. Moreover, the representation of the chromosomes [16] ACE OCP, Structural Design Optimization Software. [Online].
and the subject represent this study’s innovative component. [Link]
There already exist some well-known frameworks such as [17] Allplan, BIM Software. [Online].
[Link]
Allplan [17], ACE OCP [16], OptiStructure [19] or Altair [18] Altair, Topology Optimization Software. [Online].
[18] which are used in structural optimizations, but our goal [Link]
was to investigate how the whole process works and how [19] OptiStructure. [Online].
[Link]
can a GA be integrated. The purpose of this paper was [20] Project Chrono. [Online].
successfully achieved, but some feasible comparisons with all [Link]
these exposed frameworks can be done. The proposed model [21] Simulation System. [Online].
[Link]
can be effortlessly exported and easily compared with the
existing classical models.
During the evolution of this study, we encountered some
challenges. Chromosomes defining, fitness function identifica-
tion, generating a scene that can highlight the obtained results
and the individual validators are some of the most significant
steps that required our attention. In order to find the best

Authorized licensed use limited to: Biruni Universitesi. Downloaded on November 14,2022 at 04:02:51 UTC from IEEE Xplore. Restrictions apply.