Structural Engineering International

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tsei20

Structural Optimization of a Steel Arch Bridge with

Genetic Algorithm

Prof. Dr Yue Feng , Prof. Dr Chunguang Wang , Prof. Bruno Briseghella , Dr

Luigi Fenu & Prof. Dr Tobia Zordan

To cite this article: Prof. Dr Yue Feng , Prof. Dr Chunguang Wang , Prof. Bruno Briseghella , Dr
Luigi Fenu & Prof. Dr Tobia Zordan (2020): Structural Optimization of a Steel Arch Bridge with
Genetic Algorithm, Structural Engineering International

To link to this article: https://doi.org/10.1080/10168664.2020.1773373

Published online: 07 Jul 2020.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsei20
Structural Optimization of a Steel Arch Bridge with Genetic
Algorithm
Yue Feng, Prof. Dr, Department of Civil Engineering and Architecture, Hainan University, Hainan, People’s Republic of China;
Department of Civil and Environmental Engineering, ATLSS Engineering Research Center, Lehigh University, Bethlehem, PA, USA;
Chunguang Wang, Prof. Dr, School of Civil and Architectural Engineering, Shandong University of Technology, Shandong, People’s
Republic of China; Bruno Briseghella, Prof., College of Civil Engineering, Fuzhou University, Fuzhou, People’s Republic of China;
Luigi Fenu, Dr, Department of Civil Engineering, Environment Engineering, and Architecture, The University of Cagliari, Cagliari, Italy;
Tobia Zordan, Prof. Dr, BOLINA Consultant Engineering Ltd., Venice, Italy. Contact: [email protected].
DOI: 10.1080/10168664.2020.1773373

Abstract and strength as constraints. Ref. [10]

conducted the optimization at three
Structural optimization has become an important tool for structural designers. It levels of component, transverse, longi-
helps the designers to find optimal design solutions with better exploitation of tudinal configurations and overall
materials subject to various constraints. In this article, to face the critical issue system for a bridge of a given length,
of huge horizontal thrust occurred in the Calatrava bridge over the Grand width and traffic loading based on the
Canal of Venice, single-objective and multi-objective genetic algorithm (GA) Canadian and American standard.
based optimization procedures are set up, and the thickness optimization of the Ref. [11] minimized the steel weight
bridge is carried out separately with the proposed design procedures and the of composite hybrid plate girders for
optimization module implemented in ANSYS. The results are compared in highway bridges by coupling an optim-
terms of steel volume and horizontal thrust level. It shows that the GA-based ization program with a reliability
design procedures are effective tools to achieve optimal design due to the analysis program. Ref. [12] studied
global search ability of GA. Further, with these powerful tools, more the effects of component material be-
reasonable thickness distributions of steel plates and tubes of the Calatrava havior on the optimum weight of deter-
bridge can be obtained, meanwhile, the horizontal thrust can be reduced ministic truss bridge systems and
remarkably, and hence the high cost due to the large horizontal thrust of optimized truss bridges with uncertain
original design could be reduced. correlated strengths, acted upon by
Keywords: structural optimization; Genetic Algorithm; single-objective; multi- random correlated loads using a
objective; horizontal thrust reliability-based optimization approach.
Ref. [13] presented a computationally
efficient optimal design approach for
suspension bridges and performed the
layout and size optimization of a three-
Introduction optimization techniques, sizing optim- span suspension bridge with the pro-
ization is to find the optimal thickness posed method. Ref. [14] optimized the
Structural optimization is the subject of distribution of a linearly elastic plate structural members including ties, arch
achieving the best performance for a or the optimal member areas in a ribs, bottom and top bracings, floor
structure with various constraints such truss structure.6 It has revealed to be beams, and stringers of a 3-D tied-arch
as a given amount of material, limit- interesting for structural engineering bridge under AASHTO-LRFD Bridge
ation of peak stress and deflection.1,2 since it offers to the designer the possi- Design Specifications provisions. These
Based on the strong demand for light- bility to find out the most suitable published articles show that as a struc-
weight, low-cost and high-performance thickness distribution of a structure.1,2,7 tural optimization technique, sizing
structures due to the limited material The thickness distribution meets all optimization is an effective tool to ident-
resources and technological compe- specified requirements yet demands a ify the best structural solutions and
tition, the optimal structure design is minimum in terms of expenses such layouts characterized by exploitation of
becoming increasingly important.3 as weight, surface area, volume, material.
Benefit from the availability of high- stress, cost, and other factors.
speed computers and the rapid
improvements in algorithms, the struc- When sizing optimization applied to There are two classes of numerical
tural optimization has become in the bridge structures, earlier studies focus methods for solving sizing optimization
last decades an important mathemat- on identifying the optimum component problems, namely the Gradient-based
ical tool for designers and rapidly designs for individual structural types. Search Methods and Direct Search
becoming an integral part of the struc- Ref. [8] obtained the optimal design Methods. The former uses the gradi-
tural design process.4 of multi-span steel plate girders used ents of the problem functions to
in highway bridges which subjected to perform the search for the optimum
Structural optimization can be classi- moving load with the objective of mini- point, such as the Steepest Descent
fied into three categories, namely mizing the total weight. Ref. [9] opti- Method, the Conjugate Gradient
sizing, shape, and topology optimiz- mized the unstiffened and stiffened Method and the Sequential Linear
ation, each of them addresses a differ- composite hybrid plate girder with the Programming Algorithm.15 On the
ent aspect of the structural design weight of the girder as the objective contrary, Direct Search Methods do
problem.5 Among the different function and the section properties not use the gradients of the problem

Structural Engineering International 2020 Scientific Paper 1

functions and only the function values iteration. In this study, the individuals constraint violations.19
are used in the search process, such as are sorted based on their fitness
Powell Method and Nature-inspired values, and scaled values are assigned F = OBJ × (1 + C) (1)
Search Methods.16 GA belongs to according to their rankings, for
Nature-inspired Search Methods and instance, an individual
√ with rank k where F is fitness function, OBJ is the
is known as its global search ability. has scaled value 1/ k. Then the selec- objective function to be minimized, C
tion moves along a line that is laid out is constraint violation function. The
In this paper, to face the critical issue of
by sections of the length proportional following expression for constraint vio-
huge horizontal thrust occurred due to
to the scaled values and allocates the lation function is used:
the high rise-to-span ratio of the Cala-
individual from the section it lands
trava bridge which main structure was
on, the process is shown in Fig. 1.
constructed with steel, design pro-

Nl

Nl
cedures coupling numerical analysis Crossover is a process of combining C= Cis + Cid (2)
software MATLAB17 with finite and mixing two different designs into i=1 i=1
element software ANSYS18 is formed, the population, it introduces variation
then the thickness optimization is into a population. There are many where Cis , Cid , and Nl are constraint
carried out separately and the results methods for performing crossover, violation for stress, constraint violation
are compared in terms of steel such as scattered, single point, and for deflection, and number of load
volume and horizontal thrust level. uniform crossover. In this study, a scat- cases, respectively. The constraint vio-
tered crossover scheme is used. It lation Ci is defined as
creates a random binary vector with
Genetic Algorithm and ⎧
the same length of the bit string of ⎨0 ai ≤ 0
Design Procedure parent individuals and used as an oper- Ci = a1/2 0 , ai ≤ 1 (3)
ator to generate the new child individ- ⎩ i2
Genetic Algorithm ai ai . 1
ual. The bit in first parent individual is
GA is a search algorithm that is based chosen if the element value in the
where αi is defined according to state
on Darwin’s theory of natural. Differ- vector is 1; otherwise, the bit is
variables that is the stress level and
ent from traditional optimization chosen from the second parent
deflection should be lower than an
methods working with variables them- individual.
allowable value.
selves, it works with a coded set of
Mutation is a process to make small
the variables and searches from a
population of design variables rather
random changes in the individuals in |Si |
the population to provide genetic ai = −1 (4)
than improving a single design vari- |Sai |
diversity and enable the genetic algor-
able. It uses objective function infor-
ithm to search a broader space. In this
mation and does not require gradients where Si and Sai are the state variables
study, a Gaussian mutation function
of constraint functions that make GA and allowable state variables in terms
which adds a random number taken
more general than gradient-based of stress and deflection level in the ith
from a Gaussian distribution is used
methods.19 It starts with coding a set load case.
as an operator. If the random number
of design variables as well as popu-
is less than the mutation probability, In summary, the GA based design pro-
lation. The initial population is gener-
the value of the bit is switched from 0 cedure of single objective optimization,
ated randomly with a uniform
to 1 or vice versa. which involves interaction between
distribution. Among the coding
numerical analysis software
methods, the binary coding is the
MATLAB and finite element software
most common which implies the indi- Design Procedure of Single
ANSYS for solving the sizing optimiz-
viduals in the population with 0 and 1 Objective Optimization ation problems is given as follows:
bit strings and is used in this study.
An important step in single-objective
From the generated population, the GA is defining the fitness function 1. Setup the main procedure in
next generation of designs is evolved according to the objective to be mini- MATLAB. Input the GA control
by performing three operations, mized. The fitness function assigns a parameters include population
namely selection, crossover, and fitness value to each design variable, size, crossover probability, mutation
mutation.16 Selection is a process of and a higher fitness value implies a probability, maximum generation,
selecting a set of designs with higher better design. In practice, the fitness fitness function, and function
fitness values from the current popu- function is generally defined by pena- tolerance.
lation and carrying them into the next lizing the objective function with 2. Generate the initial population ran-
domly. Decode the binary design
variables into decimal values and
write these values into an input file
for ANSYS.
3. Read the input file of design vari-
ables in ANSYS and perform finite
element analysis. Calculate the
state variables and write them into
an input file for MATLAB.
4. Read the input file of state variables
Fig. 1: Process for selection of designs in GA in MATLAB and calculate the

2 Scientific Paper Structural Engineering International 2020

Fig. 2: Design procedure of single objective
optimization

penalty function and fitness for

every individual of the population.
5. Check the convergence criteria, ter- Fig. 3: Design procedure of multi-objective optimization
minate the design process if the
maximum generation is reached or MATLAB is adopted. To maintain 3. Create an offspring population Qt
the function tolerance is satisfied. the diversity among population of size Np through selection, cross-
Otherwise, generate the next gener- members, a crowd-comparison over, and mutation.
ation through selection, crossover, process to get an estimate of the 4. Form a combined population of Pt
and mutation. density of solutions surrounding a par- + Qt. Perform non-dominated
6. Repeat steps 2 to 5 until an optimal ticular solution in the population is sorting on the combined
solution is obtained. introduced.25 The process assigns a population.
value named crowded distance to 5. Form the next parent population
All previous steps are summarized in each design solution. In the same Pt+1 based on the rank and
the flow chart shown in Fig. 2. front, the higher crowded distance of crowded distance of each solution
a solution, the higher probability to in the combined population.
be selected. The design procedure of 6. Check the convergence criteria, ter-
Design Procedure of Multi- multi-objective optimization coupling minate the design process if the
Objective Optimization MATLAB with ANSYS for solving maximum generation is reached or
Structural design optimization pro- the sizing optimization problems is the function tolerance is satisfied.
blems in the field of civil engineering given as follows: Otherwise, generate the next off-
are usually multi-criteria. It makes spring population Qt+1 through
these problems to have not only one 1. Setup the main procedure in selection, crossover, and mutation.
but a set of solutions which represent MATLAB. Input the GA control 7. Repeat steps 4 to 5 until the Pareto-
the best possible trade-offs among the parameters include population size optimal front is obtained.
different objectives to be optimized. Np, crossover probability, mutation
Among the available computational probability, maximum generation, All previous steps are summarized in
methods that can be used to solve fitness function, and function the flow chart shown in Fig. 3.
multi-objective optimization problems, tolerance.
multi-objective GAs are by far the 2. Generate the parent population Pt
most well-known metaheuristic algor- of size Np randomly. Perform non- Calatrava Bridge
ithms and have been widely used in dominated sorting on the parent
population. Assign a rank to each General Situation
real-world structural design pro-
blems.20–24 In this study, the multi- solution according to its non-domi- Calatrava Bridge, the fourth bridge
objective GA implemented in nated level. spanning the Grand Canal of Venice,

Structural Engineering International 2020 Scientific Paper 3

 Fig. 7: FE model in ANSYS

of the bridge and existing soil con-

dition, a rigid, heavy, and deep foun-
dation is built on each side and a
system of hydraulic jacks at the con-
Fig. 4: Ponte della Costituzione—Santiago Calatrava, 2008, Venice (Italy) (Photo: the nection between the abutments and
authors) the superstructure is installed in the
real bridge. The foundation is a 22 m
deep diaphragm wall caisson consisting
of two parallel 22.5 m diaphragms with
two 9 m long front diaphragm walls
that stiffened with internal partitions
and a solid upper plate. It leads to an
extremely high total cost of the
bridge, approximately 10 million euros.
The fourth Bridge over the Grand
Canal is a steel arch footbridge with a
main span of 80.8 m. It is 93.8 m in
total while count from two end steps.
4.67 m rise gives the bridge a 1/16
rise-to-span ratio. The width varies
from a minimum of 5.58 m at the foot
to a maximum of 9.38 m in the middle
Fig. 5: General arrangement of Calatrava Bridge of the bridge. The height is 3.2 m at
the foot at the steps, reaching 9.28 m
at the highest point at the center. The
later has its official name “Ponte della importance not only in Venice city general arrangement is shown in Fig. 5.
Costituzione”, opened to the public but at national and international level
It was designed as an arch bridge with a
on September 11, 2008. After a public too, and it has also taken on a historical
radius of 180 m. The bridge consists of
process, Spanish architect Santiago significance. Immediately following the
a central arch, two side arches, and two
Calatrava was asked to design the completion of the newest bridge, there
lower arches. The axes of the central
new bridge in November 1999.26,27 are a lot of comments including criti-
arch and two lower arches are two
Calatrava selected to add new cism and praise. However, due to the
elements and designed a steel arch dimension curves in the vertical
lack of wheelchair access, lack of
bridge with a radius of 180 m, as plane. Two lower arches are symmetri-
necessity, and its modern appearance,
shown in Fig. 4. The construction of cally placed and 1.85 m away from the
criticism received heatedly. Moreover,
the fourth bridge has assumed great central arch in the transverse direction.
considering the huge horizontal force
The axes of side arches are 3 dimension
curves, placed symmetrically and a
2.56 m minimum at the foot to a
4.46 m maximum in the middle away
from the central arch.
Five arches are connected by girders,
the latter placed perpendicular to the
arches and joined them together. Hori-
zontal girders connect side arches and
central arch together, while vertical
girders connect lower arches and
central arch together. The girders
consist of steel tubes and plates,
which form closed section boxes. The
typical cross section as well as open
star-shaped cross section is shown in
Fig. 6. The maximum width is 9 m at
Fig. 6: Typical cross-section (right half, units: mm)

4 Scientific Paper Structural Engineering International 2020

 is considered. A uniformly distributed
Load case Loading condition(s) Loading area(s) load of pedestrians of 6 kN/m2 is sup-
1 Full bridge full width posed, with 7 different load cases, as
presented in Table 1. For the wind
2 Full bridge half width load, according to the technical
report,28 a simplified pressure of
3 2.5 kN/m2 is applied to an area with
3.7 m high and 80.8 m long. These
4 Half bridge full width load cases are not only used to calcu-
late the static behavior of the bridge
5 with the original parameter but also
used over and over again during optim-
6 Two diagonal areas of half width ization looping.

7
Mechanical Behavior
Table 1: The uniformly distributed load of pedestrians is 6 kN/m2 From the structural point of view, the
results of finite element analysis show
that the structure is not reasonable.
the center and with approximately displacement Du is obtained from The structural response of the foot-
2.1 m height. There are 38 cross sec- bridge monitoring, while the horizontal bridge is more like a girder rather
tions for each half span and 75 cross thrust T is from the finite element than an arch. The axial forces are
sections in total. results presented in the referenced taken up by the central arch and two
paper.26,27 lower arches. However, the utilization
of an open star-shaped cross section
Finite Element Model The steel grade is Fe510DD according
makes the stiffness of these main arch
to the Italian code. Its strength is equiv-
The bridge is modeled with the finite alent to S355 in Eurocode 3. Therefore,
ribs small, large bending deflection
element analysis software ANSYS occurred under symmetric and asym-
the yield stress is 355 MPa for steel
(Fig. 7). Shell element SHELL93 is metric loads. Maximum is 0.060 m
thickness no higher than 40 mm and
chosen for all the steel structure when loads add on full bridge full
is 335 MPa for steel thickness ranging
except two lower arches and steel width. The deflection shape is illus-
from 40 to 80 mm. The value of the
tubes of girders, which are simulated trated in Fig. 8.
modulus of elasticity is assumed to be
with BEAM188 element. There are 210 GPa. Poisson’s ratio and material Huge horizontal thrust occurred due to
110,869 nodes and 44,783 elements in density are set to 0.3 and 7850 kg/m3 the 1/16 rise-to-span ratio, maximum
total. The transverse constraint (DY) respectively. 6511 kN under full bridge full width
and vertical constraint (DZ) of central load case and 16,311 kN under an
arch and two lower arches at two abut- Secondary dead load including the
envelope combination of self-weight,
ments are fixed, but each horizontal weight of handrails (2.5 kN/m2 on the
secondary dead load, and pedestrian
constraint (DX) at two abutments is side arches), the weight of glass deck
load. Therefore, a system of hydraulic
simulated through one spring (10.5 kN/m2 on the central arch), and
jacks at the connection between the
element, whose stiffening coefficient the weight of Istria stone deck
abutments and the superstructure is
K = T/Du is 8.38E8 N/m. The (12 kN/m2 on the horizontal girders)
installed in the real bridge. Jacks are
intervened when horizontal displace-
ments become higher than 20 mm on
each abutment in order to restore the
initial geometry of the footbridge.

Structural Optimization

Fig. 8: Deflection shape under full bridge full width load case Optimization Formulation
Design Variables: The thickness of all
steel plates and tubes are assumed as
design variables with values ranging
between 12 and 60 mm. These vari-
ables are integer-valued take into
account the manufacturing limitation.
Eventually, due to the symmetry of
the typical cross section, 12 design vari-
ables are defined in total named as T1
to T12 as shown in Fig. 9. Among
them, T1, T2, and T3 are the central
arch, side arches, and bottom arches
Fig. 9: Design variables of thickness optimization respectively, which are circled with

Structural Engineering International 2020 Scientific Paper 5

dashed lines. T8 and T9 are the side
plates, T11 and T12 are the top plates. 2. Given the horizontal thrusts of the
bridge under secondary dead load
State Variables: The design constraints
(F2) and maximum pedestrian load
related to stress and deflection level
(F3) are the same kind of objective,
of bridge summarized as following are
the optimization problem is formu-
considered in the optimization process.
lated as having two different objec-
tive functions: minimize the
simax − smax ≤ 0 i = 1, 2, . . . , LC horizontal thrusts of the bridge
 under self-weight (F1) and second-
355/gM0 t ≤ 40 mm
smax = ary dead load (F2).
335/gM0 40 mm , t ≤ 80 mm
(7) min F = (F1 , F2 ) (10) Fig. 11: Convergence history of minimizing
horizontal thrust under envelope
where simax is the maximum von Mises combination
stress of steel plates and tubes for ith
pedestrian load case combined with Optimal Results of Single
self-weight, secondary dead load, and Objective Optimization the optimal design obtained is used
wind load, smax is the maximum allow- Optimal Result with Gradient-Based as the starting point to repeat the
able stress of steel plates and tubes, Method optimization process. The results
gM0 is the partial factor for the resist- show the values of the objective func-
ance and equals to 1 according to To compare the results obtained with tion at different design steps. It can be
Eurocode 3. the Gradient-based Method, the seen that the consistent reduction of
optimization is carried out with the the objective function is effective in
design optimization tool implemented
dimax − dmax ≤ 0 i = 1, 2, . . . , LC (8) in ANSYS. The solver first transforms
Fig. 9a. The objective function is
reduced from 16,311 kN in the orig-
the constrained problem into an inal design to 13,543 kN, namely
where dimax is the maximum vertical unconstrained one via penalty func-
deflection of the deck for ith pedestrian 17.0% reduction. The horizontal
tions, and then the derivatives of the thrusts of the bridge under self-
load case, dmax is the maximum allow- objective function are evaluated at
able vertical deflection of the deck weight (F1), secondary dead load
some point in the design space and (F2) and maximum pedestrian load
which is taken as L/1000, L is the used to determine the best search
length of the bridge. (F3) are reduced from 6319, 3481,
direction from that point, namely the and 6511 kN to 5406, 2835.1, and
Objective Function: Considering that direction of steepest descent. Follow- 5301.5 kN, namely 14.4%, 18.6%,
the horizontal thrusts of the bridge ing that, a line search is conducted and 18.6% reduction, respectively.
under self-weight (F1), secondary along the search direction to locate Meanwhile, the total volume of steel
dead load (F2) and maximum ped- the minimum point along this line. is increased from 46.54 to 49.23 m3,
estrian load (F3) are different single The derivatives are evaluated again at namely 5.8% rising.
objective functions, the optimum the minimum point just found to get a
design is found with two different new search direction, and another line
Optimal Result with GA
formulations: search is conducted. The procedure is
continued until a sufficiently small In order to solve this problem, the
1. Combines all criteria into a single change in the value of the function is population size, crossover probability,
scalar parameterized objective found.18,15 and mutation probability are set as 12,
function by using the same weight, 0.8 and 0.2 in the design procedure. It
The iterative process of the bridge is
the optimization problem is formu- means that there are 12 individuals in
illustrated in Fig. 10. Since the
lated as minimizing the horizontal each generation, and 80% of individuals
optimum solution depending on the
thrust F of the bridge under an in the next generation are produced by
initial values, different starting points
envelope combination. Hence, the crossover. The design process will ter-
are adopted to carry out the optimiz-
objective function is given by minate if the maximum generation 30
ation process. At first, the optimal
is reached or the average change in
design is obtained with the original
F = F1 + F2 + F3 (9) the fitness function value is less than
design as starting point. Thereafter,
function tolerance 1e-6. There is not
any serious attempt made to find the
best parameter setting for GA in this
study due to the high computational
effort, which takes around 10 min to
carry out a single run of finite element
analysis, and 60 h in total with an 8-
core, 128 GB of memory computer.
Hence, there is no absolute guarantee
that a global optimal solution has been
obtained.
Fig. 10: Convergence iteration of optimization process with ANSYS. (a) Starting point is the Convergence history which shows the
original design, (b) Starting point is the optimal design solution obtained best fitness values at different

6 Scientific Paper Structural Engineering International 2020

 Self- Secondary dead Full length full Full length half Half length full Two diagonal areas of
Load Cases weight load width width width half width
Original Fx (kN) 6319 3481 6511 3256 3245 3245
Model
Dmax 0.056 0.032 0.060 0.047 0.035 0.033
(m)
Optimized Fx (kN) 5087 2768 5170 2585 2577 2577
Dmax 0.056 0.031 0.058 0.040 0.033 0.031
(m)
Reduction Fx (%) 19.5 20.5 20.6 20.6 20.6 20.6
Dmax 0.0 3.1 3.3 14.9 5.7 6.1
(%)
Table 2: Horizontal thrust and deflection under load cases before and after optimization

the horizontal thrust and deflection arch plays an important role in

under all load cases are decreased bearing the structural forces and the
after optimization. In contrast to stresses are obtained under the envel-
deflection decreasing, the reductions ope of all live loads.
of horizontal force are more remark-
Considering the vibration of the bridge
able and are around 20% under all
is a critical issue, the modal analyses
load cases.
are carried out before and after the
The stress state ratio, understood as optimization. The frequencies of the
the ratio between the maximum stress first four order vibration mode of the
and the allowable stress, is calculated original design are 1.19, 1.36, 2.04,
along the longitudinal axis of the and 2.85 respectively. In the best
bridge for the points 1, 2 and 3 (see design solution, higher values 1.23,
Fig. 9) and plotted in Fig. 12. All of 1.38, 2.33, and 2.87 are obtained. The
Fig. 12: The stress state ratio at representa- the points are at the top plates. From results show that the stiffness of the
tive points of optimal design the figure it can be seen that, the bridge is slightly increased after
maximum stress of point 1 (central optimization.
arch) is reached to the limit and has
Eigenvalue buckling analyses of the
the highest value 1, whereas the
generations is illustrated in Fig. 11. The original design and optimized design
largest stress state ratios of point 2
best design solution is found after 30 are also carried out. The first 30 buck-
and point 3 are around 0.85 and 0.75,
generations. The results show that the ling eigenvalues of two designs are
respectively. The results show that the
reduction of the objective is well obtained, and all of them are local
optimal solution exploits the materials
appreciable. The horizontal thrust buckled modes but none of them is
of central arch more effectively than
under envelope combination of the overall buckled mode. The first eigen-
the components at point 2 and point
best design solution is reduced from value is 1.287 in the original design,
3. Moreover, the stress of point 1 has
16,311 kN in the original design to whereas it increased to 1.395 in the
an overall higher value than point 2
13,025 kN, namely 20.1% reduction. optimized design. The results show
and point 3. The result may be
The horizontal thrusts of the bridge that the potential of local buckling
explained by the fact that the central
under self-weight (F1), secondary still exists but the behavior is improved
dead load (F2) and maximum ped-
estrian load (F3) are reduced from
6319, 3481, and 6511 kN to 5087.4,
2768, and 5169.8 kN, namely 19.5%,
20.5%, and 20.6% reduction, respect-
ively. Meanwhile, the total volume of
steel is slightly higher than the original
design, 2.75% increased from 46.54 to
47.82 m3.
To compare with the original model
without thickness optimization, the
static behaviors of the optimum
design under the same load cases are
calculated. The results of horizontal
thrust and deflection under load cases Fig. 13: Results of multi-objective optimization (a) evolution of Pareto-optimal front, and (b)
before and after optimization are final set of Pareto-optimal designs with original design and optimal design obtained with
listed in Table 2. It can be seen that ANSYS

Structural Engineering International 2020 Scientific Paper 7

Design T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 V (m3) F1 (kN) F2 (kN) F3 (kN)
Original 25 20 60 16 30 30 30 12 25 30 16 25 46.54 6319 3481 6511
Optimal with ANSYS 20 14 36 26 51 24 53 51 20 21 39 38 49.23 5406 2835.1 5301.5
Optimal with single-objective GA 21 14 32 33 53 26 48 30 22 41 18 29 47.82 5087.4 2768 5169.8
A 12 26 15 21 59 33 15 34 31 42 23 14 43.1 4766.6 2895.9 5386.4
B 36 12 12 60 60 52 60 44 46 60 25 50 64.8 6187.8 2472.3 4624.4
Table 3: Thickness distributions, steel volume and horizontal thrust before and after optimization (unit: mm)

Optimal Result of Multi-Objective

Optimization
The multi-objective optimization is
performed with the same parameters
as in single-objective GA except using
a population size of 30 individuals.
Fig. 13a shows the evolution of the
Pareto-optimal front at generations 1,
5, 10, 20 and 30, and Fig. 13b presents
the final set of Pareto-optimal designs
with original design and optimal
design obtained with ANSYS. Simi-
larly, there is no absolute guarantee
that a global Pareto-optimal solution
has been obtained, but from Fig. 13 it
can be seen that a good spread of sol-
utions is found even at a local Pareto-
optimal front.
Three of the designs on the Pareto-
optimal front, including designs
labeled A, B, and optimal design
obtained with single-objective GA,
are singled out for further investi-
gation. Together with the original
design and optimal design obtained
Fig. 14: Steel volume and deflection for the Pareto-optimal designs at different generations with ANSYS, the design variables,
plotted as a function of horizontal thrust, (a–b) steel volume, (c) maximum deflection under steel volume and corresponding
pedestrian load, (d) maximum deflection under self-weight objective function values for the
chosen designs are listed in Table 3.
Compared with the original design,
the horizontal thrusts of optimized
after the optimization. The results can as the lower bound of these com-
designs are reduced greatly. Among
be explained by the fact that a rela- ponents, which is the lowest thickness
the optimized designs, design A has
tively high thickness of 12 mm is set adopted in the original design.
the lowest F1 as well as steel volume
but the highest F2 and F3. In contrast,
design B has the highest F1 as well as
steel volume but the lowest F2 and
F3. Whereas the optimal design
obtained with ANSYS is dominated
by the optimal design obtained with
single-objective GA.

Discussion
The steel volume and deflection of the
bridge for the Pareto-optimal designs
at different generations are plotted as
a function of the horizontal thrust in
Fig. 14. In Fig. 14a–b there is a clear
trend of increasing in steel volume as
Fig. 15: Design variables for the Pareto-optimal designs plotted as a function of horizontal F1 rising, whereas F2 decreases as the
thrust, (a) T1-T3, (b) T4, T5, and T10 steel volume rising. The former trend

8 Scientific Paper Structural Engineering International 2020

is obvious as F1 is the resultant of self- figure, it can be seen that the vast Funding
weight as well as the total steel volume. majority of optimal designs have T4
The latter trend may be explained by ranging from 30 to 60 mm. This work was supported by Key Research
and Development Program of Shandong
the fact that the optimized designs
Province: [grant number 2019GHY112076];
have more reasonable thickness distri- Foundation of Hainan University: [grant
butions which changed the forces Conclusions number kyqd1632]; Science and Technology
taken up by the central arch and the Development Project of Zibo: [grant
In this paper, to face the critical issue of
lower arches. Since the rise of the number 2017KJ040017]; Shandong Provin-
huge horizontal thrust occurred due to cial Natural Science Foundation: [grant
central arch is higher than two lower
the high rise-to-span ratio of the Cala- number ZR2017LEE031]; Hainan Provin-
arches, if the forces taken up by two
trava bridge, the optimization pro- cial Natural Science Foundation: [grant
lower arches are higher, the horizontal number 518QN216].
cesses of steel members of this
force under secondary dead load is
structure are carried out. The following
lower.
conclusions could be drawn from the
In Fig. 14c there is a clear trend of results: References
decreasing in maximum deflection
under pedestrian load as F1 rising. It 1. With the optimization tool [1] Briseghella B, Fenu L, Feng Y, Mazzarolo E,
confirms that the optimization process implemented in ANSYS, the hori- Zordan T. Topology optimization of bridges sup-
found designs with more reasonable zontal thrusts of the bridge under ported by a concrete shell. Struct. Eng. Int. 2013;
thickness distributions as well as high 23(3): 285–294. https://doi.org/10.2749/
self-weight (F1), secondary dead
101686613X13363929988214
static stiffness of the bridge for a load (F2) and maximum pedestrian
given amount of material. However, load (F3) are reduced from 6319, [2] Briseghella B, Fenu L, Lan C, Mazzarolo E,
in Fig. 14d there is no trend of decreas- 3481, and 6511 kN to 5406, 2835.1, Zordan T. Application of topological optimiz-
ation to bridge design. J. Brid. Eng. 2013; 18(8):
ing or increasing in maximum deflec- and 5301.5 kN, namely 14.4%, 790–800. https://doi.org/10.1061/(ASCE)BE.
tion under self-weight. The result may 18.6%, and 18.6% reduction, 1943-5592.0000416
be explained by the fact that the steel respectively. Meanwhile, the total
[3] Huang X, Xie M. Evolutionary Topology
weight and the static stiffness of the volume of steel is increased from
Optimization of Continuum Structures: Methods
bridge increase simultaneously as the 46.54 to 49.23 m3, namely 5.8% and Applications, Wiley, New York, 2010.
total steel volume rising. The former rising.
results in higher deflection under self- 2. With the single-objective GA, the [4] Christensen PW, Klarbring A. An
Introduction to Structural Optimization,
weight, while the latter leads to a sig- horizontal thrusts of the bridge Springer, Berlin, 2009.
nificant reduction of deflection under under self-weight (F1), secondary
all load cases. dead load (F2) and maximum ped- [5] Bendsoe MP, Sigmund O. Topology
Optimization: Theory, Methods and
estrian load (F3) are reduced to
Design variables for the final set of Applications, Springer Verlag, Berlin, 2003.
5087.4, 2768, and 5169.8 kN,
Pareto-optimal designs are plotted as [6] Rozvany GIN. A critical review of estab-
namely 19.5%, 20.5%, and 20.6%
a function of horizontal thrust in lished methods of structural topology optimiz-
reduction, respectively. Meanwhile,
Fig. 15. The results indicate that the ation. Struct. Multidiscipl. Optim. 2009; 37(3):
the total volume of steel has 217–237. https://doi.org/10.1007/s00158-007-
optimal thickness distribution is
2.75% rising. 0217-0
obtained according to the structural
3. With the multi-objective GA, a
behavior of steel components. As men- [7] Briseghella B, Fenu L, Feng Y, Lan, C,
fairly well defined front of non-
tioned before, the axial forces are Mazzarolo, E, Zordan T. Optimization indexes
dominated designs is found. The to identify the optimal design solution of shell-
taken up by the central arch and two
optimal thickness distributions are supported bridges. J. Brid. Eng. 2015; 21(3):
lower arches. On the other hand, the
obtained according to the structural 04015067. https://doi.org/10.1061/(ASCE)BE.
role of two side arches is to maintain 1943-5592.0000838
behavior of steel components.
the bridge stable. It can be seen from
4. By means of sizing optimization, [8] Adeli H, Mak K. Interactive optimization
Fig. 15a, the vast majority of optimal
more reasonable thickness distri- of plate girder bridges subjected to
designs have T1 ranging from 12 to
butions of steel members can be moving loads. Comp.-Aid. Design. 1990; 22(6):
40 mm. Similarly, T3 is ranged from 368–376. https://doi.org/10.1016/0010-4485
obtained, and the horizontal
12 to 35 mm, whereas T2 ranged from (90)90087-S
thrust could be reduced remark-
12 to 25 mm.
ably, meanwhile, the high cost due [9] Dhillon BS, Kuo C-H. Optimum design
Due to the utilization of open star- to the large horizontal thrust of of composite hybrid plate girders. J. Struct.
shaped cross section, star-shaped steel the original design could be Eng. 1991; 117(7): 2088–2098. https://
doi.org/10.1061/(ASCE)0733-9445(1991)117:7
plates not only withstand shearing reduced. (2088)
force of the main arch but also bear 5. Results also show that a better
local bending moment, so some parts optimum could had been obtained [10] Cohn MZ, Lounis Z. Optimal design of
structural concrete bridge systems.
of vertical girders are under great if more design variables would had
J. Struct. Eng. 1994; 120(9): 2653–2674. https://
strain. As a result, T5 and T10 in the been considered to enable a vari- doi.org/10.1061/(ASCE)0733-9445(1994)120:9
optimized designs are in the range of ation of the thickness of the plates (2653)
40–60 mm. Horizontal girders bear along the longitudinal axis of the
[11] Hendawi S, Frangopol DM. Design of com-
large bending moment owing to the bridge. This was not possible in posite hybrid plate girder bridges based on
maximum 9 m width of star-shaped this work because of the associated reliability and optimization. Struct. Safety. 1994;
cross section which corresponds to a computational cost, but can be the 15(1–2): 149–165. https://doi.org/10.1016/0167-
4.5 m cantilever beam. From the object of future research. 4730(94)90057-4

Structural Engineering International 2020 Scientific Paper 9

[12] Nakib R. Deterministic and reliability- [19] Pezeshk S, Camp C, Chen D. Design of non- Optim. 2016; 53(3): 545–566. https://doi.org/10.
based optimization of truss bridges. Comput. linear framed structures using genetic optimization. 1007/s00158-015-1291-3
Struct. 1997; 65(5): 767–775. https://doi.org/10. J. Struct. Eng. 2000; 126(3): 382–388. https://doi.org/
[24] Zavala GR, Nebro AJ, Luna F, Coello CAC.
1016/S0045-7949(94)E0289-E 10.1061/(ASCE)0733-9445(2000)126:3(382)
A survey of multi-objective metaheuristics
[13] Cao H, Qian X, Chen Z, Zhu, H. Layout and [20] Konak A, Coit DW, Smith AE. Multi- applied to structural optimization. Struct.
size optimization of suspension bridges based on objective optimization using genetic algorithms: Multidiscipl. Optim. 2014; 49(4): 537–558.
coupled modelling approach and enhanced particle a tutorial. Reliab. Eng. Syst. Saf. 2006; 91(9): https://doi.org/10.1007/s00158-013-0996-4
swarm optimization. Eng. Struct. 2017; 146: 170– 992–1007. https://doi.org/10.1016/j.ress.2005.11.
[25] Deb K. Multi-Objective Optimization Using
183. https://doi.org/10.1016/j.engstruct.2017.05.048 018
Evolutionary Algorithms, John Wiley & Sons,
[14] Latif M, Saka M. Optimum design of tied- [21] Poirier JD, Vel SS, Caccese V. Multi- Chichester, 2001.
arch bridges under code requirements using objective optimization of laser-welded steel
[26] Zordan T, Briseghella B, Mazzarolo E. Bridge
enhanced artificial bee colony algorithm. Adv. sandwich panels for static loads using a genetic
structural optimization through step-by-step evol-
Eng. Softw. 2019; 135: 102685. https://doi.org/10. algorithm. Eng. Struct. 2013; 49:508–524. https://
utionary process. Struct. Eng. Int. 2010; 20(1): 72–
1016/j.advengsoft.2019.102685 doi.org/10.1016/j.engstruct.2012.10.033
78. https://doi.org/10.2749/101686610791555586
[15] Rothwell, A. Optimization Methods in [22] Talaslioglu T. Optimal dome design
[27] Zordan T, Briseghella B, Siviero E. The
Structural Design. Springer, Berlin, 2017. considering member-related design constraints.
fourth bridge over the Grand Canal in Venice:
Frontiers of Struct. Civil. Eng. 2019; 13(5):
[16] Arora, J. Introduction to Optimum Design, from idea to analysis and construction. Struct.
1150–1170. https://doi.org/10.1007/s11709-019-
Academic Press, Waltham, MA, USA, 2004. Eng. Int. 2010; 20(1): 6–12. https://doi.org/10.
0543-5
2749/101686610791555667
[17] Documentation M. The MathWorks Inc,
[23] Zavala G, Nebro AJ, Luna F, Coello CAC.
2005. [28] Siviero, E. Independent check design of the
Structural design using multi-objective meta-
IVth bridge over the Grand Canal. Final
[18] Ansys R. 11.0 Documentation for ANSYS. heuristics. Comparative study and application
Report. Padova, 2007.
ANSYS Inc, Canonsburg, PA, 2007. to a real-world problem. Struct. Multidiscipl.

10 Scientific Paper Structural Engineering International 2020