Surrogate-Based Fragility Analysis and Probabilistic Optimisation of Cable-Stayed Bridges Subject To Seismic Loads

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

Engineering Structures 256 (2022) 113949

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Surrogate-based fragility analysis and probabilistic optimisation of


cable-stayed bridges subject to seismic loads
Andrea Franchini *, Wendel Sebastian, Dina D’Ayala
Department of Civil, Environmental and Geomatic Engineering, University College London, London, UK

A R T I C L E I N F O A B S T R A C T

Keywords: The lack of computationally efficient, probabilistic performance-based design strategies for cable-stayed bridges
Cable-stayed bridge hinders their optimal design in seismic regions. Thus, this paper proposes the implementation of parameterised
Seismic performance-based design fragility functions (PFFs) for the surrogate-based sensitivity analysis and performance-based optimisation of
Parameterised fragility
these structures. In particular, PFFs are exploited to define computationally efficient decision variables and four
Optimisation
Earthquake engineering
optimisation strategies that aim to optimise the probability of seismic damage and the direct losses related to the
seismic repair cost. To illustrate this overall strategy, a 542 m three-span cable-stayed bridge is considered. The
ability to predict dynamic behaviour is improved by using the measured dynamic response of the reference cable-
stayed bridge to inform the numerical model’s architecture. Then, key design variables of tower and cable cross-
sections are selected for optimisation. The sensitivity analysis shows conflicting demands placed upon design
variables by different decision variables (e.g. damage probability and the sum of construction and repair costs,
CRC) and the crucial importance of accounting for design variables interaction when making choices on design
updating. Therefore, numerical optimisation emerges as the most efficient tool to deal with these issues. With
respect to the reference structure, the proposed single-objective optimisation strategies reduced the system-level
damage probability by 3.5 times, and yielded an 88% reduction of repair cost, and a 40% decrease of CRC. A
Pareto-front of optimal design variables was also calculated to simultaneously optimise system damage proba­
bility and CRC.

1. Introduction variables, cables forces and control devices for cable-stayed bridges
subject to earthquakes [10–13]. Nonetheless, analyses and decisions
Due to their light weight and low inherent damping, cable-stayed made in the design phase are significantly affected by uncertainties, and
bridges are vulnerable to earthquakes [1] and significant damages to therefore appropriate performances of an infrastructure system may be
the deck-to-tower connections, abutments, and concrete towers have only guaranteed in a probabilistic sense [14]. Thus, in order for de­
been identified after past seismic events [2–5]. Furthermore, bridges signers and stakeholders to gain confidence that the structure will
usually represent vital links in the transport infrastructure network perform as expected when an extreme event occurs, the probabilistic
[6,7], whose disruption in seismic events might lead to substantial Performance-Based Earthquake Engineering developed by the Pacific
economic losses and delay the recovery of whole regions. Therefore, Earthquake Engineering Research Center [15–18] (PEER-PBEE herein­
even localized component failures may not be acceptable. after) is, at present, the most reliable approach.
Optimisation techniques significantly aid the achievement of eco­ Despite the better performances that can be achieved through PEER-
nomic and seismically efficient structures. However, optimising cable- PBEE, deterministic design approaches and performance criteria are still
stayed bridges is complex because of the many design variables and the most used for designing cable-stayed bridges [19]. This is mainly due
design constraints that should be considered, often nonlinear and con­ to the extensive computational effort required for seismic fragility
flicting [8]. Considerations aimed at optimising the tower geometry to assessment, which grows with the number of design variables. Conse­
avoid excessive damage deriving from the transverse deck-tower in­ quently, few studies have implemented probabilistic approaches for the
teractions were given in [9]. Moreover, several researchers imple­ seismic optimisation of these structures, mainly focusing their analysis
mented deterministic approaches to optimize geometry, cross-sectional on a limited number of design variables, such as those of mitigation

* Corresponding author.
E-mail addresses: andrea.franchini.19@ucl.ac.uk (A. Franchini), w.sebastian@ucl.ac.uk (W. Sebastian), d.dayala@ucl.ac.uk (D. D’Ayala).

https://doi.org/10.1016/j.engstruct.2022.113949
Received 19 August 2021; Received in revised form 20 December 2021; Accepted 24 January 2022
Available online 12 February 2022
0141-0296/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Franchini et al. Engineering Structures 256 (2022) 113949

devices. Wu et al. [20] argued that deterministic analysis methods, such behaviour prediction in the present study. This test data verification
as the Response Spectrum Analysis, cannot account for uncertainties due reduced the uncertainty in predictive capability, giving confidence
to record-to-record variability. Thus, they implemented system-level that the obtained results are reliable and that the proposed design
fragility functions to design the stiffness of cable restrainers for a me­ strategy is applicable to new structures.
dium span cable-stayed bridge. Zhong et al. [21] investigated the
optimal design variables of fluid viscous dampers to retrofit cable-stayed The remaining part of the manuscript is organized as follows: Section
bridges. This study used system-level fragility functions obtained using 2 details the proposed methodology for the PEER-PBEE design of cable-
the cloud approach and a joint probabilistic seismic demand model to stayed bridges; Section 3 describes a case-study cable-stayed bridge and
derive structural performance indicators. Wen et al. [22] optimized the its updating process; design space exploration and surrogate-based
parameters of viscous dampers and metallic dampers at the deck-to- performance optimisation are then performed in Section 4; finally,
pylon connection of a cable-stayed bridge by implementing a genetic conclusions are drawn in Section 5.
algorithm that aimed to minimize the repair cost ratio (RCR). The
concept of RCR was initially proposed by Xie and Jian [23] as a system- 2. Parameterised fragility functions for PEER-PBEE design and
level probabilistic performance index to quantify the damaging poten­ optimisation
tial of a selected bridge design. Zhong et al. [24] used Gaussian process
surrogate models to perform global sensitivity analysis and risk-based 2.1. Proposed methodology
optimisation of fluid viscous dampers in cable-stayed bridges. The au­
thors achieved a 72% reduction of system-level seismic risk and The procedure for PEER-PBEE assessment (referred to as forward
observed that component level objectives are generally conflicting, PBEE analysis [41]) consists in the following independent logical steps:
preventing their simultaneous optimisation. Conflicting damage prob­ 1) Hazard analysis; 2) Demand analysis through time history simula­
ability levels in different bridge components were also found by Wei tions; 3) Damage analysis to obtain fragility functions; 4) Decision var­
et al. [25]. After using fragility functions to compare typical cable- iable (DV) calculation; 5) Risk-informed decision making. The DVs
stayed bridge structural systems (rigid, longitudinally constrained and should be relevant to the stakeholder and might represent direct or in­
floating), these authors observed that optimizing multiple bridge design direct economic losses, fatalities, downtime, or similar.
variables through fragility analysis would require an excessive compu­ Using the PEER-PBEE methodology for new design entails the
tational effort. The need for further development in procedures for the inversion of the above process, leading to the inverse PBEE analysis
optimisation of cable-stayed bridges considering seismic actions and [41]. In this respect, inverting the methodology requires the designer to
resilience was also highlighted by [8]. update the design and perform the assessment iteratively. Two ap­
Machine Learning (ML) techniques might be implemented to over­ proaches are possible for design updating: trial-and-error and numerical
come the limitations posed by the high computational cost for devel­ optimisation. Based on experience and sensitivity analysis, trial-and-
oping fragility functions. In particular, ML has been used to develop error does not guarantee to find an optimal solution. On the other
seismic demand models conditioned on multiple predictors and to hand, numerical optimisation involves a significant computational effort
obtain multidimensional fragility models for portfolios of highway because of the objective function’s evaluations it requires at each iter­
bridges [26–34]. Such models are called parametrized fragility functions ation step, which results in the need to repeat the forward analysis many
(PFFs). Beyond the use for regional risk assessment, Dukes et al. [35] times, changing design variables and recording DV. Most of the
proposed their implementation in the bridge design process to quickly computational effort is required by demand analysis and damage anal­
investigate the effect of design variables on seismic performance. The ysis, which aim to obtain structural fragility functions. A seismic
methodology extension to multi-hazard fragility assessment was also fragility function FR = P[D > CLSi |IM] provides the conditional proba­
studied [36,37]. Gehl and D’Ayala [6,38] implemented Bayesian Net­ bility that the seismic demand (D) on a structural member or system
works to derive system-level fragility surfaces for multi-span simply- exceeds its capacity (C) for the considered limit state (LSi) and for a
supported bridges subject to earthquakes, ground failures and floods. given level of earthquake intensity measure (IM) [42,43].
Analytical fragility assessment has at its core the use of a numerical On the other hand, Parameterised Fragility Functions (PFFs) are
method, such as Finite Element modelling (FEM), to predict the seismic conditioned on both the IM and the n design variables Xi of the cable-
response of cable-stayed bridges. Now, in order to maximise any FEM’s stayed bridge (collected in the vector X ∈ Rn ).
ability to capture the static and dynamic behaviours of bridges which are
FR = P[D > CLSi |IM, X] = PF(IM, X) (1)
complex due to their geometry (e.g. cable-stayed) or their materials (e.g.
composites), tuning of the FEM architecture (types/layouts of elements, Therefore, both the intensity measure and the design variables can be
etc.) is often informed by comparisons with data from tests on real included as predictors in the fragility model. Generally speaking, cable-
bridges [39,40]. This approach will be adopted in the present paper. stayed bridge design variables can be divided into several categories:
The reviewed approaches using the PEER-PBEE design for cable- variables defining bridge geometry; variables defining structural mem­
stayed bridges optimised a limited number of design variables, mainly bers cross-sectional dimensions and reinforcement; variables defining
focusing on vibration control devices. Moreover, to the authors’ vibration control device properties. Nevertheless, the distinction be­
knowledge, using PFFs for the seismic fragility analysis and the tween fixed parameters and design variables is project-specific, as it
surrogate-based probabilistic optimisation of cable-stayed bridges might depend on site constraints, stakeholder preferences and budget,
cannot be found in the literature. Therefore, the main objectives of this urbanistic and architectural requirements, etc. Therefore, this paper
paper consist of: addresses the optimisation of structural members cross-sections without
loss of generality for the proposed procedure. In particular, the selected
• Exploiting PFFs for design space exploration and for developing design variables consist of towers cross-sectional dimensions, rein­
probabilistic performance-based seismic optimisation strategies for forcement ratios and cables cross-sectional area, which are all critical to
cable-stayed bridges. seismic performance. Further details are provided in Section 3.1.
• Proving the applicability and the advantages of the proposed stra­ PFFs provide closed-form equations for the damage probability.
tegies through the calculation of ’optimal combinations’ of key Therefore, their main advantage as a tool for structural optimisation
design variables for a real cable-stayed bridge, which was previously consists in the possibility of computing probabilistic DVs without per­
designed according to capacity design principles. forming additional FE analyses. Consequently, PFFs allow decoupling
• Using data from tests on the real cable-stayed bridge to inform tuning the fragility assessment from the optimisation cycle. This fact aids the
of the architecture of a FE model, which is then used for dynamic computational efficiency of the function evaluation phase in numerical

2
A. Franchini et al. Engineering Structures 256 (2022) 113949

optimisation algorithms, for which the DVs represent objective func­ unknown, and the FE analysis provides only y. Metamodeling, therefore,
tions. On the other hand, the PEER-PBEE optimisation described in consists in using statistical techniques to build a computationally effi­
Section 2.1 requires the computation of fragility functions through FE cient approximation ̂ y (with ̂y = m(x) + ε) of the relationship between y
analysis at each iteration of the searching algorithm. Such an approach and x (i.e., y ≈ ̂y = m(x) + ε) [44]. ε represents the total error ac­
is too computationally demanding for cable-stayed bridges. counting for approximation (goodness-of-fit) and measurement
Fig. 1 provides an overview of the proposed methodology. The (random) errors. Afterwards, surrogate models are implemented to
procedure starts with the generation of PFFs (detailed in Section 2.2) generate survival-failure vectors for bridge components by performing
and exploits them to aid the investigation of design alternatives (PBEE Monte Carlo simulations (MCSs), and PFFs are eventually obtained
use I, Section 2.4) and to define probabilistic optimisation strategies through logistic regression (Step 4). With reference to Fig. 1, critical
(PBEE use II, Section 2.5). Both proposed uses serve to determine the aspects of the four steps are summarized below.
decision variables, which provide the framework for decision making.
As stated above, probabilistic DVs defined through PFFs are efficient to 2.2.1. Step 1 – Experimental design
compute. The optimisation cycle (bottom left part of Fig. 1) exploits DVs The objective of experimental design is to build a design matrix
computational efficiency for rapid numerical differentiation of the per­ containing realisations of the predictors on which to apply regression
formance function over the design space. This operation serves the techniques to obtain the surrogate metamodels. Each row of the design
determination of gradients that identify the direction of minimum for matrix, an experimental design realisation, can be obtained by sampling
the successive iterative step. realisations of selected bridge parameters and randomly pairing them to
It should be noted that a cable-stayed bridge is subject to several load the IM of an earthquake belonging to the considered ground motion set.
scenarios during its service life, including self-weight, traffic, wind, and Experimental design starts with the identification of a set of n design
earthquake. This paper addresses the self-weight and earthquake load variables X = [X1 , X2 , ⋯, Xn ] which are statistically significant for the
condition, assuming that it is dominant with respect to the others. seismic response of cable-stayed bridges. A typical step to identify sig­
nificant parameters is screening [45], which would require running
2.2. Generation of parameterised fragility functions nonlinear time history analysis to record peak responses variability
induced by the change of design variables. However, considering the
To generate PFFs, this paper follows the procedure outlined by Dukes large number of parameters involved in the design of cable-stayed
et al. [35], which consists of Steps 1 to 4 depicted in Fig. 1. This figure bridges, such a procedure becomes computationally expensive even at
also highlights the “Demand analysis” and the “Damage analysis” phases its preliminary stage. A further issue regards the effect of ground motion
to ease the comparison with the conventional PEER-PBEE framework randomness on the screening process: it is not guaranteed that the sig­
described in Section 2.1. nificant parameters for different peak responses remain the same when
The “Demand analysis” phase (Steps 1 to 3 in Fig. 1) generates sur­ subject to ground motions with different properties. On the other hand,
rogate models (also called metamodels) for bridge components peak correctly accounting for ground motion randomness in the screening
responses. Consider a peak response y = f(x), which is a function of the phase would require a number of time history analyses similar to that
required to generate metamodels.
vector of predictors x = [IM, X] ∈ ℝn+1 . Usually, the relationship f is

Fig. 1. Overview of the proposed probabilistic performance-based seismic optimisation methodology for cable-stayed bridges.

3
A. Franchini et al. Engineering Structures 256 (2022) 113949

As an alternative, the selection of design variables can be based on accurate results and avoid structural members oversizing, alternative
the expertise of the analyst/designer. A preliminary analysis of the techniques should be considered in future studies that include correlated
statistical significance of all the selected design variables can then be parameters as design variables.
performed after time history analyses are completed and before building
metamodels. This second approach is adopted in this paper. First, a set of 2.2.4. Step 4 – Logistic regression
na accelerograms is selected to account for the randomness of ground Limit States (LS) should be defined to quantify damage occurrence in
motion. Then, na design variable realisations are sampled through Latin the structure. LS capacities are a quantitative measure of the capacity C
Hypercube Sampling (LHS) [46]. Eventually, each realisation is of bridge components to withstand the demands D placed upon them by
randomly paired to two ground motions from the set, resulting in a the earthquake [42]. These capacities are probabilistic variables with
design matrix of 2na rows and n + 1 columns. When using LHS, the associated uncertainty. In the spirit of PBEE, C can be thought of as a
minimum number of experiments for optimally spaced design points is level at which the bridge loses some measure of functionality. According
usually two times the number of design parameters [27], requiring to the analyst’s interest, the functionality ranges from serviceability to
na ≥ 2n. Further considerations on sample size are provided in Section collapse [49]. Typically four limit states (LS) are considered [50]: slight
4.3. (LS1), moderate (LS2), extensive (LS3), and complete damage (LS4).
Component capacities can be assumed to be lognormally distributed
2.2.2. Step 2 – Nonlinear FE time history analysis [51].
After determining the design matrix, a nonlinear FE model of the In order to build PFFs, peak response metamodels and limit state
bridge is built for each line of the design matrix, and two accelerograms capacity models are used to perform MCSs and generate binary survival-
are successively applied at the fixed nodes. Nonlinear time-history an­ failure vectors [30] of Nsim components. In this study, Nsim is determined
alyses are then performed, recording the peak responses (engineering through sensitivity analysis on the stability of fragility functions’ me­
demand parameters, EDPs) of sensitive bridge components. In partic­ dians. These vectors contain realisations of a Bernoulli random variable,
ular, this study considers the following EDPs: tower section curvatures whose distribution’s expected value can be estimated through Logistic
ductility, tower drift ratio, longitudinal displacement of bearings and Regression. A detailed description of the procedure to perform logistic
cables’ axial force ratio (AFR, defined as the ratio of the axial force regression and obtain PFFs can be found in [27,35]. The following
capacity to demand). expression is eventually obtained:
eα0 +αIM ∙ ln(IM) + α∙X
2.2.3. Step 3 - Surrogate models choice and fitting FRLSi,k (IM, X) = (3)
1 + eα0 +αIM ∙ ln(IM) + α∙X
This step aims at obtaining response surface surrogate models of each
EDP. The polynomial Response Surface (RS) method [47] represents a where α = [α1 , α2 , ⋯, αn ] and αi (i = 0, IM, 1, 2, ....n) are the logistic
natural extension to multiple dimensions of linear, one-dimensional regression coefficients, and k might refer to the k-th component or k =
probabilistic seismic response models in traditional fragility analysis, sys for system-level fragility. Different perspectives on the meaning (and
which usually condition the logarithmic value of EDPs on the logarithm corresponding mathematical representation) of system damage have
of IM. Therefore, in this study, RSs are used to predict ln(̂
y ) conditioned been described by Dueñas-Osorio and Padgett [52]. This study assumes
on the vector of predictors x = [ln(IM), X]. Four different RS metamodels that the system enters a specific damage state if any of its components
(M1, M2, M3 and M4, see Eq. (2)) are considered: they differ by the exceeds its corresponding limit state capacity, commonly referred to as
maximum polynomials order and interaction terms considered. Inter­ the “series system” assumption. Cable-stayed bridges actually behave as
action effects on peak response might be relevant for redundant struc­ mixed series–parallel systems, resulting in enhanced robustness. In this
tures such as cable-stayed bridges. In Eq. (2), k = n+1. At this stage, a respect, the series system assumption is conservative and was imple­
significance test is implemented to identify significant predictors before mented by several researchers (e.g. [53,54]). Bridge components are
building advanced metamodels. The p-values of first-order linear classified as primary or secondary according to their contribution to the
regression are adopted to test the null hypothesis that a predictor is not structural system damage level: it is assumed that secondary compo­
significant in determining peak responses. In particular, the null hy­ nents only contribute to the slight and moderate damage states. System-
pothesis is rejected if the p-value associated with one of the predictors is level binary survival-failure vectors are calculated using Monte Carlo
less than 0.05. A forward removal process is then iteratively imple­ simulations, and logistic regression is applied to obtain system-level
mented, removing one non-significant predictor at a time until all the PFFs.
predictors are significant. Because this study’s overall goal is to show how surrogate modelling
and parameterised fragility can aid the PEER-PBEE design of cable-

k
M1 : ln(̂y ) = β0 + β i xi stayed bridges, seismic parameterised fragility functions are generated
i=1 following a procedure widely used in the literature for bridges [27,35].

k ∑
k ∑
k Nevertheless, the authors acknowledge that several different (and
M2 : ln(̂y ) = β0 + β i xi + βij xi xj possibly more efficient) choices could have been made. In this respect,
i=1 i=1 j=1,i<j
(2) future research should investigate alternative statistical learning tech­

k ∑
k
niques that might result in enhanced computational efficiency and
M3 : ln(̂y ) = β0 + β i xi + βii x2i
i=1 i=1
accuracy.

k ∑
k ∑
k ∑
k
M4 : ln(̂y ) = β0 + β i xi + βii x2i + βij xi xj 2.3. Decision variables definition
i=1 i=1 i=1 j=1,i<j

The general form of an optimisation problem [55] envisages two sets


This paper addresses the optimisation of components cross-sectional
of functions: an objective function and a set of constraint functions. The
properties, which are generally characterised in terms of uncorrelated
latter consist of equations and inequalities that the design variables, or
variables: choosing a value for one variable does not constrain the range
some functions of the variables, must satisfy. Within the domain of
of possible values another variable can assume, as long as performance
definition of the design variables, usually Rn , the set of points that satisfy
requirements are satisfied. On the other hand, material properties are
all the constraints forms the feasible region to which the solution of the
usually correlated, which might result in overestimating bridge fragility
optimisation problem belongs.
[48]. From the perspective of PEER-PBEE design, overestimating
One of the main barriers to cable-stayed bridge optimisation is the
fragility leads to a conservative design. Nevertheless, to achieve more
extensive running time to calculate the objective function (DVs in PEER-

4
A. Franchini et al. Engineering Structures 256 (2022) 113949

PBEE) at every step of the selected solution algorithm. Therefore, be estimated from component-level PFFs. Eventually, the total repair
recalling that PFFs provide a closed-form equation for the probability of cost RC(X, PGV) is computed as the sum of the contributions of each
exceeding a selected damage state, the surrogate-based design optimi­ component.
sation methodology proposed in this paper exploits PFFs to determine Because the RC depends on the PGV, in previous studies, Repair Cost
DVs as functions of the design variables X. Because it does not require to Ratio (RCR) surfaces were obtained as a function of peak responses;
perform FE simulations at each iteration, such an approach dramatically then, optimal design variables were identified at different IM levels by
aids computational efficiency and represents a point of novelty in this means of additional time history analyses [22]; alternatively, peak re­
paper. In that respect, to the best of the authors’ knowledge, surrogate- sponses were set as the objective function, and then RCR was used to
based probabilistic seismic performance optimisation strategies for assess the validity of the achieved solutions [23]. However, to achieve
cable-stayed bridges have not yet been proposed. an overall optimal value of design variables and directly implement RC
It is also acknowledged that, beyond being adopted as objective as the objective function, the dependency of RC on the intensity measure
functions, some DVs might represent desirable constraint functions for should be removed. To this end, in the current study, the distribution of
the studied problem. Considering the computational burdens of the RC|PGV is calculated by randomly sampling realisations of PGV at the
conventional PEER-PBEE, this option is made possible by using PFFs. In construction site and computing the corresponding RC. This procedure is
particular, this paper implements DVs related to seismic damage to treat computationally efficient thanks to the closed-form equation of PFFs,
the uncertainties affecting constraint conditions, yielding an additional which can be quickly implemented to compute the damage probabilities
point of novelty. in Eq. (5). After obtaining the RC distribution, its 95th percentile
(defined as RC95 (X)) is proposed in this study as an objective function.
2.3.1. Average median peak ground velocity (AMPGV) The probability density function of PGV at the construction site is
Previous works on performance-based analysis and optimisation of assumed as lognormal distributed and, consequently, also the distribu­
bridges [21,24,25,56] considered the IM’s median value of each damage tion of RC results lognormal. Parameters of the PGV lognormal distri­
state’s fragility curve as a separate performance indicator of system- bution can be obtained through seismic hazard analysis at the
level performance. The median IM (named IM) represents the earth­ construction site.
quake intensity with a 50% probability of causing the structure to enter
the relevant damage state and is therefore regarded as an appropriate 2.3.4. Construction and repair cost (CRC)
indicator of damage probability. As discussed further in Section 4.1, this This study defines the Construction and Repair Cost, CRC(X), as the
study adopts the peak ground velocity (PGV) as IM. To account for the linear combination of C0 (X) and RC95 (X), which were defined above.
four system-level damage states with one single function, the first DV is This DV allows a designer to balance the initial capital investment and
defined as the average value of the median PGVs for the four defined the cost the asset owner should bear in case of earthquake-induced
damage states (AMPGV(X)). damage. In Sections 2.4 and 2.5, the decision variables described
above are implemented for design space exploration and to develop
2.3.2. Construction materials cost (C0) optimisation strategies.
A second decision variable is the construction materials cost for the
cable-stayed bridge, which can be computed as follows: 2.4. PBEE use I: Design space exploration
C0 (X) = Cdeck + Cbearings + C(X) (4)
The closed-form expression of the DVs defined in Section 2.3 can be
The details of the deck are not considered for optimisation. Thus, its used for design space exploration, providing insightful information
cost is assumed to be independent of the selected design variables. The about the effects of single and pairwise design variables change on the
cost of the bearings is roughly assumed as 10% of the cost of a leg in the overall performance of the structure.
pylon: previous researchers [22,23] argued that the replacement cost of Consider a generic decision variable DV(X), a function of the design
bearings (accounting for direct and indirect losses) is about 15–25% of variables X. Let X0 be the initial design configuration and Xk a new
the total bridge construction cost, whereas pylons account for 70%. design configuration with the same components as X 0 except for the k-th
Therefore, the assumption made in this study is a reasonable estimation one, which is a 10% perturbation of X0,k . In this study, ΔDV k (Xk ) is
of bearings’ construction cost. Furthermore, the cost of the foundations defined as the percentual variation of the decision variable resulting
is neglected. C(X) is the portion of bridge cost which directly depends on from perturbation of the k-th design variable. The analysis of ΔDV k
design variables, and can be estimated by summing the products of identifies which parameters are more influential in determining the
materials quantities and unit costs. The unit costs were taken as 3.5 and value of the selected performance measure and whether their variation
7.5 €/kg for towers and cables, respectively [10]. As for the reinforce­ affects the bridge’s seismic performance. A similar approach is imple­
ment, 1.1 €/kg was considered a figure representative of current market mented in [35] to investigate the effect of design parameters on fragility
values. estimation.
A cable-stayed bridge is a complex, hyperstatic system in which
2.3.3. Repair cost (RC) many components interact to determine the overall structural response.
The repair cost (RC), also considered in this study as a decision In turn, this results in the importance of understanding the combined
variable, quantifies the material repair cost in case of earthquake effect of design variables on system-level response. To that end, this
occurrence. The RC calculation follows the procedure outlined in [23] study builds matrices of contour plots [45] that present the variation of
for the definition of repair cost ratio, which was applied to optimise
PGV LSi (easily computable from PFFs) as a function of two design vari­
vibration control devices in cable-stayed bridges by [22]. The following
ables while keeping the others set to their average value. As discussed in
equation is derived for the k-th component:
Section 2.3, previous studies recognized the median PGV of system-level
( )
∑ 4 fragility functions as an efficient indicator of bridge performance
RCk (X, PGV) = pi,k (X, PGV)∙di,k Ck (X) (5) [21,24,25,56].
i=1

where di,k is the damage ratio of the k-th component at the i-th
2.5. PBEE use II: Surrogate-based design optimisation
damage state; Ck is the replacement cost of the k-th component, which is
assumed to equal the construction material cost for the considered Four optimisation strategies are considered based on the computa­
component; pi,k is the damage probability for each damage state that can tionally efficient DVs defined in Section 2.3 (Table 1). The DVs are used

5
A. Franchini et al. Engineering Structures 256 (2022) 113949

Table 1
Optimisation strategies.
Definition Objective Problem formulation

Strategy #1 Maximise the average of the median PGVs (AMPGV) of the bridge system ⎪ C0 (X* ) ≤ C0,ref


max AMPGV(X), subject to LB ≤ X* ≤ UB
*


⎩ PGV LS1 (X ) ≥ PGVsite
σcr (X* )/σ(X* ) ≥ 3

Strategy #2 Minimise the Repair Cost ⎪ C0 (X* ) ≤ C0,ref



⎨ LB ≤ X* ≤ UB
min RC95 (X), subject to PGVLS1 (X* ) ≥ PGVsite



⎪ AMPGV(X* ) ≥ AMPGVref

σcr (X* )/σ(X* ) ≥ 3

Strategy #3 Minimise the Construction and Repair Cost ⎪ LB ≤ X* ≤ UB


PGVLS1 (X* ) ≥ PGVsite
min CRC(X), subject to *
⎪ AMPGV(X ) ≥ AMPGVref


[ ] σcr (X* )/σ(X* ) ≥ 3
Strategy #4 (multi- Maximise the average of median PGVs and simultaneously minimise the Construction and − AMPGV(X)
min ,
objective) Repair Cost. CRC(X)

⎨ LB ≤ X* ≤ UB
subject to PGVLS1 (X* ) ≥ PGVsite

σcr (X* )/σ(X* ) ≥ 3

to define either or both the objective function and the boundary con­ maker is required to select an optimal solution based on his/her sub­
ditions of the problem. The solution vector of design variables is labelled jective criteria of preference.
X* . Additional boundary conditions are set for all the proposed strate­
The objective of the first optimisation strategy (Strategy #1) is to gies. First, each optimisation strategy is such that it searches the solution
maximize AMPGV(X). Intuitively, a higher value of AMPGV is associated within selected lower and upper bounds, i.e. LB ≤ X ≤ UB. Details on
with a structure with a lower probability of damage when subject to an the boundaries selected in this study are provided for the selected case
earthquake. The advantage of implementing AMPGV consists in gaining study in Section 4.6. A second consideration regards the control of
an overall description of system-level performance. However, using the buckling failure, which might result from the reduction of the towers’
average of the medians as a DV might lead to a stiff and ductile bridge cross-sectional area to minimise cost-related DVs. In this respect,
(which could enter LS1 for a low value of EDP and then reach LS4 for a constraint equations are set to guarantee an adequate safety margin
substantial value of EDP) having the same AMPGV as another optimal against instability. The buckling load σcr is computed according to
solution which entails more bunched-up fragility functions, with high Eulerian buckling theory and a safety factor γ cr = σcr /σ ≥ 3 is required
flexibility and low ductility. Although extensive ductility might be for each optimised configuration X* . In terms of damage control per­
beneficial in terms of ultimate behaviour, triggering of non-linearity for formance, it was argued in Section 2.5 that controlling AMPGV might
low values of EDPs might be undesirable for serviceability damage result in high probabilities of exceeding low damage states, which might
states. The selected DVs for optimisation intrinsically account for such be undesirable in the occurrence of low-intensity earthquakes. There­
effects. Additionally, if specific performance goals are desirable, fore, it was chosen to constrain PGV LS1 to be no less than the median
consistent constraint conditions can be set. To avoid excessive capital PGV at the construction site, PGV site . A different choice could be made
investment in the attempt of increasing system safety, a constraint according to the desired performance objective.
equation is used to require that the initial cost C0 (X) of the optimized In the remainder of the paper, the proposed methodology is applied
structure does not exceed the cost of the reference structure. Details on to a case study bridge.
the reference structure are provided in Section 3. In an actual design
situation, C0 might be substituted by a threshold agreed with the asset 3. Case study bridge
owner or by the estimated cost of preliminary design.
Strategy #2 and #3 target the cost of the structure. Specifically, they 3.1. Bridge description and finite element model
aim at minimizing RC95 and CRC, respectively. The AMPGV was previ­
ously defined as an overall measure of system-level performance. To The Quincy Bayview Bridge [58], located in Illinois, USA, is chosen
guarantee that the mentioned costs are minimized while maintaining an as a case study to demonstrate the proposed procedure. As shown in
acceptable damage probability, PFFs are applied to define nonlinear Fig. 2a, the structure consists of a 274 m main span and two 134 m
boundary conditions. More in detail, the minimum acceptable value of symmetric side spans. Two H-shaped reinforced concrete towers of 71 m
AMPGV(X* ) is limited to be no less than the value computed for the height from the waterline support 56 stay cables arranged on two planes
reference bridge AMPGV ref . In Strategy #2, C0 is also considered as a and in a semi-fan configuration. Cables represented in Fig. 2a with the
constraint. same line type have the same cross-sectional area.
In the framework of this study, the AMPGV(X) and the CRC(X) prove Each tower (Fig. 2b) is formed by two concrete legs, a lower strut
to be conflicting objectives. Therefore, Strategy #4 is developed as a (Sec D-D) supporting the deck and an upper strut (Sec C-C) connecting
multi-objective optimisation problem. In the presence of trade-offs be­ the legs at the top. The legs present two sections along their length
tween conflicting objectives, usually, no solution exists which minimizes (section A-A, section B-B), with the section change occurring at the deck-
all of them simultaneously. Accordingly, solving a multi-objective to-tower connection. Details on the cross sections can be found in
optimisation problem involves finding a set of Pareto optimal solutions Fig. 2c-f. In the real structure, the section change takes place 2.5 m
that entail different trade-offs between objective functions. In particular, above the deck, but this does not significantly affect the dynamic
a solution is defined as ’Pareto optimal’ if one objective cannot be behaviour, as shown in Section 3.2. An additional simplification in this
improved without degrading at least one other objective decision vari­ study, also not relevant in terms of dynamic behaviour, consists in the
able [57]. The set of Pareto optimal solutions is called the Pareto front. A assumption that the thickness of the walls in section B-B is the same
Pareto front is calculated, which optimises the trade-off between along the two horizontal dimensions. The deck (Fig. 2g) is formed by a
AMPGV(X) and the CRC(X). After the Pareto front is found, the decision- steel–concrete composite section, with two main girders at the sides and

6
A. Franchini et al. Engineering Structures 256 (2022) 113949

West East

(a)

Section A-A
5.21
Traffic barrier

0.23
2.21
Z’

1.93
Y’

0.48
Girder Stringer
4-4
0.46 13.26 0.46

(c) (g)
Section B-B Section C-C Section D-D

3-3 Z’ Z’
2-2 Z’
Y’ Y’ Y’

1-1

(b) (d) (e) (f)


Fig. 2. Quincy Bayview Bridge: (a) longitudinal view; (b) tower section; (c-f) main cross sections; (g) composite deck. Dimensions in metres [m].

five internal stringers in the longitudinal direction. reinforcement choices as highlighted by [19]. For the shear reinforce­
As discussed in Section 2.1, and without loss of generality for the ment volumetric ratio, a range of 0.5%-1.5% was considered.
applicability of the proposed PEER-PBEE framework, this paper ad­ A 3D Finite Element (FE) model of the bridge (Fig. 3) was developed
dresses the optimisation of towers cross-sectional dimensions (X1 to X7), in the OpenSees platform [63], which allows performing detailed time
reinforcement ratios (X8 to X11) and cables cross-sectional area (X12 to history analysis considering materials and geometric nonlinearities.
X15). Consistently, the selected design variables Xi are defined in Moreover, OpenSees easily allows the parametrisation of bridge design
Table 2. The “Reference” column lists design variables values in the variables for optimisation. In [58], the finite element model was built
reference structure [58]. Since details on reinforcement were not pro­ using linear elastic elements, and nonlinear static and dynamic features
vided, commonly selected values are assumed. For the experimental were neglected. Such an approach provided modal characteristics which
design process, DVs were assigned a uniform distribution with upper and were close to those observed from the ambient vibration test on the real
lower bounds defined in the relevant columns of Table 2. The choice of a structure [64], and therefore, the authors considered the elastic
uniform distribution reflects that any combination of code-compliant assumption acceptable. However, a significant nonlinear behaviour of
DVs is accepted as long as it ensures a satisfactory structural cable-stayed bridge materials is expected in strong earthquakes [1],
performance. requiring nonlinear analysis. Moreover, several studies highlighted that
A ± 30% variation from the reference value was allowed for the gross the dynamic analysis of cable-stayed bridges should assume the
dimensions of the tower leg cross-sections (X1, X2), gross dimensions of deformed geometric configuration obtained through nonlinear static
the tower struts (X4, X5) and the cables’ cross-sectional area (X12 to analysis under dead loads [65,66]. The authors of this paper also
X15). Such a range of variation is in good agreement with the boundaries observed a better representation of modal properties when considering
allowed in previous studies on the cross-sectional optimisation of cable- the nonlinear behaviour induced by cables sagging. For these reasons,
stayed bridges (e.g. [11,59]) and with proportions observed in con­ the present study explicitly considers geometric and materials
structed bridges [60]. The thickness-to-short side ratio in the hollow nonlinearities.
sections of tower legs (X3) is allowed to vary in the range 0.25–0.4: the As depicted in Fig. 3a, the part of the leg from the pile caps to the
same lower bound was adopted by Santos et al. [59] and yields a min­ upper strut is modelled using distributed plasticity fibre elements based
imum wall thickness of 0.40 m, in line with the values computed by on the force formulation, while the remaining parts are assumed to
Camara [60] for anchorage zones and with optimal pylon dimensions remain elastic. The same assumption about linear behaviour was drawn
suggested by Hassan et al. [61]. Meanwhile, the upper bound of 0.4 for the upper and the lower strut beams (sections C-C and D-D in Fig. 2).
guarantees that the hollow section remains such in each design variable The choice of distributed plasticity elements aims at capturing the ef­
realisation. Because of the lack of data on typical struts thickness in fects of axial force-moment interaction and material nonlinearity under
concrete cable-stayed bridges, the same bounds as tower legs sections all load conditions. Moreover, such elements intrinsically account for
were considered for thickness-to-short side ratio in struts (X6, X7). The ductility reduction induced by high compression force and high rein­
flexural reinforcement ratio varies between 1.5% and 4%, a range that forcement ratios. Fibre section details are provided in Fig. 3b, while
respects the requirements of Eurocode 2 [62] and reflects common Fig. 4 depicts materials constitutive laws. Both the core and the cover

7
A. Franchini et al. Engineering Structures 256 (2022) 113949

Table 2 calibrated to match the measured modal behaviour to the closest extent.
Bridge design variables definition, reference design values and variability range Further details are provided in Section 3.2.
for optimisation. Boundary conditions are chosen according to the description and
Design variables Values design drawings provided in [58]. Since the piers are founded on
Notation Description Units Reference Lower Upper
densely piled foundations driven to the bedrock, soil-structure interac­
bound bound tion effects are neglected, and the tower bases are fixed. At the eastern
and western ends of the deck, the vertical and transversal displacements
X1 Section A-A and B-B m 4.42 3.094 5.746
gross dimension: Y’ side of the girders are fixed, while longitudinal displacement and rotations
X2 Section A-A and B-B m 2.13 1.491 2.769 around all three axes are free.
gross dimension: Z’ side Despite the presence of sliding bearings at the two towers, for dy­
X3 Section B-B thickness-to- – 0.3521 0.25 0.4 namic analysis, Wilson and Gravelle [58] suggested fixing the deck at
short side ratio
X4 Section C-C and D-D m 3.05 2.135 3.965
both towers and all points of connection in terms of translational degrees
dimension: Y’ side of freedom. However, the constrain conditions under quasistatic thermal
X5 Section C-C, D-D m 4.11 2.877 5.343 and wind loads in the real structure are different, as the deck longitu­
dimension: side Z’ side dinal translational degree of freedom at the east tower is released.
X6 Section C-C thickness-to- 0.2 0.25 0.4

Because of the good matching with the ambient vibration test [64], but
short side ratio
X7 Section D-D thickness- – 0.1705 0.25 0.4 also recognizing the possible, significant contribution of bearings to the
to-short side ratio bridge system vulnerability, as highlighted by previous studies [20], the
X8 Longitudinal steel % 2.75 1.5 4 modelling approach shown in Fig. 3c for the deck-to-tower connection is
reinforcement ratio of adopted in this study. The two main deck girders are vertically sup­
Section A-A
X9 Longitudinal steel % 2.75 1.5 4
ported at the lower strut beams by means of rigid links representing the
reinforcement ratio of sliding bearings, whose longitudinal behaviour (X direction) is modelled
Section B-B through elastic-perfectly plastic springs (Fig. 4d). The transversal
X10 Volumetric ratio of % 1 0.5 1.5 displacement (Z direction) at the strut level and at the deck slab level is
transverse steel
constrained to the tower nodes, as it is also the longitudinal displace­
reinforcement of Section
A-A ment at the deck-to-tower connection level. These boundary conditions,
X11 Volumetric ratio of % 1 0.5 1.5 selected after comparing the available choices, are validated in Section
transverse steel 3.2 throughout the comparison with ambient vibration tests.
reinforcement of Section
B-B
X12 Cables’ area – type 1 cm2 89.7 62.79 116.61
3.2. FE model updating and validation
X13 Cables’ area – type 2 cm2 67.7 47.39 88.01
X14 Cables’ area – type 3 cm2 53.7 37.59 69.81 FE modelling provides an approximate representation of the (espe­
X15 Cables’ area – type 4 cm2 34.8 24.36 45.24 cially material and geometrically nonlinear) dynamic behaviour of
complex structures such as cable-stayed bridges, so it is important to
concrete are modelled with the Concrete01 material (Fig. 4a), which tune the model’s architecture using data from tests on the real bridge.
implements a Kent-Scott-Park model and allows to take into account the Moreover, since modal analysis is performed after (material and
confinement effect. Reinforcement bars are represented through the geometrically) nonlinear gravity analysis, the dynamic behaviour of this
bilinear constitutive law with linear strain-hardening shown in Fig. 4b model might differ from the elastic behaviour studied in [58]. Therefore,
(Steel01 material in OpenSees). model updating was conducted to validate the assumptions on boundary
The apparent axial stiffness of cables is greatly influenced by sagging, conditions and calibrate uncertain parameters. For the considered
which depends on their state of tension: if cable tension increases, sag­ bridge, results of ambient vibration measurements were provided by
ging reduces, and the apparent axial stiffness increases consequently [64]. From a quantitative point of view, two favoured tools for
[67]. A possible approach to model this behaviour consists in consid­ comparing modal properties are the percentual frequency error and the
ering equivalent, straight, tension-only truss elements with an equiva­ Modal Assurance Criterion (MAC) [69].
lent elastic modulus that accounts for the mentioned nonlinear effects The weighted sum of these two quantities can be used to build an
[68]. According to this approach, the apparent stiffness of a sagging objective function for model updating [70]. On these bases, an optimi­
cable can be expressed as [67]: sation problem was built in this study to update the model parameters
listed in Table 4 and match the dynamic behaviour of the real structure
AEeq =
AE
[ s,cab ] (6) to the closest extent. The optimisation problem was solved through the
(ρL)2 A∙Es,cab ‘fmincon’ function in MATLAB [71], which performed modal analysis
1+ 12T 3 through the OpenSees software [63] at each function evaluation step.
Complete information regarding natural frequency and modal shapes
where A is the area of the cable cross-section; Es,cab and ρ are the was provided only for the first 15 modes [64]. Therefore, model
elastic modulus and the mass density of the cable material; T is the updating and validation were performed for them. They include vertical
tensile force in the cable, which is calculated for dead loads through an (flexural) modes, denoted as ’Vi’, and transverse-torsional modes,
iterative procedure and also provides the value for the prestress σprestress denoted as ’TTi’. Here, i represents the mode number. Firstly, a pa­
in the cable’s constitutive law. The trilinear constitutive law presented rameters sensitivity study was performed by investigating the effect of a
in Fig. 4c [21] is assigned to each cable element. With reference to Fig. 4, 10% variation of each updating parameter on bridge eigenvalues,
the properties implemented to define the constitutive laws of materials revealing that all the selected parameters significantly affect at least one
are collated in Table 3. of the provided experimental modes (Fig. 5a). It can also be observed
The deck is expected to remain in the elastic range when subject to that parameters might oppositely affect different modes, increasing or
earthquake; therefore, as shown in Fig. 3a, it is modelled through a decreasing the eigenvalues.
central spine of 3D elastic beam-column elements [58]. Rigid links Fig. 5b presents the frequency error in the initial model, updated
connect the spine to cables anchors and support lumped masses to model, and model built in [58]. The updated model presents a frequency
correctly account for the torsional inertia induced by girders, stringers error that is consistently within an acceptable range, with a maximum
and traffic barriers. The deck’s stiffness is assigned to the spine and error of 10.7% for V3. MAC comparison is presented in Fig. 5c for this

8
A. Franchini et al. Engineering Structures 256 (2022) 113949

Lumped Elastic beam-column


Rigid link
mass element

Sec B-B
Tension-only
truss element

Rigid link

Sec A-A
Deck spine
Shell
elements
Nonlinear beam-column
Y element with fibre section
X Fixed
Z
(a)
Section A-A Section B-B Equal disp.
Steel X, Z
bar
tBB Rigid link
Z’ Z’ Beam element
Y’ Cover Y’ Deck node
X1

concrete Y Tower node


Z
X1

Equal disp.
Core
concrete Z Y, Z;
X bearing
X2 X2
spring X

(b) (c)
Fig. 3. Finite element model: (a) FE frame model; (b) fibre section modelling of Section A-A and B-B; (c) deck-to-tower connection modelling. XYZ absolute reference
system. X’Y’Z’ relative reference system.

[-]

(a) (b) (c) (d)


Fig. 4. Material constitutive laws: (a) confined and unconfined concrete; (b) steel reinforcement bars; (c) cables; (d) sliding bearing (X direction).

study’s initial and update model only, as [58] did not provide this each mode, the upper diagram presents the normalized vertical
measure. All modes in the updated model have a MAC number of at least displacement of the two external nodes of the deck, while the bottom
0.6. This value was regarded as acceptable, considering the available box refers to the transversal displacement of the central node. A satis­
information. The negligible effect of assuming the same thickness in factory agreement between experimental and analytical modal shapes
both directions for the hollow section B-B is also validated by consid­ can be observed, providing further confidence that the model re­
ering the frequency error and MAC numbers as shown in Fig. 5b-c. produces the real structure’s behaviour well.
Finally, to evaluate the effect of the chosen boundary conditions and
stiffness distribution on the dynamic behaviour of the structure, Fig. 6
compares the obtained modal shapes for the first three vertical modes
(V1, V2, V3) and the first three torsional modes (TT1, TT2, TT3). For

9
A. Franchini et al. Engineering Structures 256 (2022) 113949

Table 3
Nonlinear material properties.
Definition Notation Value Units

Cover concrete compressive strength f’c 30.00 MPa


Cover concrete strain at the compressive strength εc0 0.002 –
Cover concrete ultimate stress fcu 0.00 MPa
Cover concrete ultimate strain εcu 0.004 –
Confinement factor K 1 + ρs,h fyh /f’c –
Confined concrete compressive strength f’c,C K∙f’c MPa
Confined concrete strain at the compressive strength εc0,C Kεc0 –
Confined concrete ultimate stress fcu,C 0.2f’c,C MPa
Confined concrete ultimate strain εcu,C 0.004 + 0.9ρs,h fyh /300 –
Longitudinal reinforcement yielding stress fys 414 MPa
Longitudinal reinforcement elastic modulus Es 200 000 MPa
Longitudinal reinforcement strain hardening ratio shr 0.0018 –
Hoops yielding stress fyh 420 MPa
Cable steel yielding stress fy,cab 1600 MPa

Table 4
Model updating parameters.
Definition Notation Units Initial value LB factor UB factor Updated value

Elastic modulus of concrete Econc kN/m2 3.08E+07 0.8 1.2 2.67E+07


Shear modulus of concrete Gconc kN/m2 1.23E+07 0.8 1.2 1.15E+07
Elastic modulus of deck spine Espine kN/m2 2.07E+08 0.9 1.1 2.13E+08
Density of steel ρsteel kg/m3 7849.8 0.9 1.1 7452.5
Density of concrete ρcon kg/m3 2403.0 0.9 1.1 2390.2
Area of deck spine Aspine m2 0.8344 0.8 1.2 0.9671
Flexural inertia of deck spine - horizontal axis Iz,spine m4 0.3409 0.7 1.3 0.3659
Flexural inertia of deck spine - vertical axis Iy,spine m4 19.76 0.7 1.3 16.32
Torsional inertial of deck spine Jspine m4 0.0261 0.7 1.3 0.0311
Elastic modulus of cables material Ecab kN/m2 2.00E+08 0.85 1.1 1.76E+08

4. Bridge optimisation using metamodels and PFFs be translated into peak responses of symmetric components having the
same distribution. Therefore, for the sake of safety for design and in line
4.1. Input ground motions with the standard approach of designing for the worst possible scenario,
for each group of symmetric components, the EDP distribution with the
This study considers the ground-motion suite developed by Baker highest average value is selected and considered representative of the
et al. [72] to account for earthquakes’ randomness and develop para­ group. In this way, 24 peak responses datasets are obtained, and the four
meterised fragility functions. The suite includes 160 ground motion surrogate metamodels described in Section 2.2.3 are developed for each
records selected to create a “structure- and site-independent” suite. of them.
Therefore, the presented results have general validity. Limit states thresholds for the selected EDPs are assumed to be
A more accurate and site-specific accelerograms selection should be lognormal distributed with median seismic capacity Sc and coefficients
performed when implementing this procedure for actual structural of variation CV shown in Table 5. These values are chosen following
design; however, this lies beyond the scope of this work. PGV is selected previous literature on the fragility analysis of cable-stayed bridges
as the intensity measure because several studies have shown a better [53,54].
correlation between this intensity measure and peak responses of cable- The tower’s cross sections damage is quantified in terms of curvature
stayed bridges (see, for instance, [73;74]). ductility. The yielding curvature is obtained for each design row through
moment–curvature analysis of a fibre section; then, curvature ductility is
calculated by dividing the peak curvature response by the yielding
4.2. Engineering demand parameters and damage states curvature. An additional demand parameter for the towers is the drift
ratio, defined as the modulus of the peak displacement at the top of the
To generate the design matrix, 160 realisations of the 15 design tower divided by the tower height.
variables are sampled through LHS and randomly paired with two Regarding sliding bearings, the slight damage state relates to the
ground motions from the selected set. Therefore, 320 experimental upper pad reaching the design displacement of 100 mm, while the
design rows are obtained. Nonlinear time-history analyses for each moderate damage state involves bearing unseating. For system-level
design row are performed, recording the vulnerable components’ peak fragility, bearings are considered as secondary components, as dis­
responses (EDPs) indicators. These include: the maximum longitudinal cussed in Section 2.2.4, and therefore their EDPs only contribute to LS1
(Z) and transversal (X) curvature ductility μ of the towers T1 and T2 at and LS2. The damage in the cables is quantified in terms of the Axial
the four sensitive locations identified in Fig. 2b as 1–1, 2–2, 3–3, 4–4; the Force Ratio (AFR), that is, the ratio between the peak tensile force in the
drift ratio of the towers; the longitudinal displacement of the four sliding considered cable (over the time history) and the yielding tensile force.
bearings; the axial force ratio (AFR) in the 56 cables. Since the structure The CVs of different damage states are assumed as 0.35 for all compo­
is symmetric with respect to its centre, the response of symmetric nents but cables, for which a value of 0.11 is considered [54]. The lower
components is expected to be the same. In a probabilistic sense, this can

10
A. Franchini et al. Engineering Structures 256 (2022) 113949

Fig. 5. Finite element model updating: (a) sensitivity of eigenvalues to a 10% variation of updating parameters; (b) frequency error; (c) MAC number.

covariance value for cables reflects the reduced uncertainties in pre­ the next step for each peak response. It can be observed that, in most
dicting steel and tension-only component response with respect to cases, the metamodel M4 (second-order polynomial with interaction
concrete and concrete members. terms) provides the highest Adjusted R2 and is also associated with the
lowest values of RMSE.
4.3. Peak response metamodels The obtained values of goodness-of-fit measures are in line with
previous literature (e.g. [27]) and were regarded as appropriate for the
The Response Surface methodology is applied in this section to aim of this study. Meanwhile, using more advanced metamodels such as
generate metamodels of the recorded peak responses. Confidence that Gaussian Process Regression [75] might provide a more accurate pre­
the sampled design space reflects the feature of geometrical design diction of cable-stayed bridge peak responses and should therefore be
variables not being correlated is gained by analysing correlation co­ considered for future research.
efficients. The results of such check are omitted for the sake of brevity The lack of correlation between design variables in the design space
but can be found in the Supplementary Material. and the obtained values of goodness-of-fit measures confirm that the
The significance test described in Section 2.2.3 is implemented to generated sample size is appropriate for this study. However, in the
identify predictors which are not significant and can therefore be absence of these conditions, further sensitivity analyses on sample size
removed before building higher-order metamodels. Such a process should be performed to get accurate simulation results.
resulted in eliminating the predictors X7 and X10. These parameters
correspond to the thickness of the lower strut (section D-D) and the 4.4. Parameterised fragility functions
transverse reinforcement of section A-A (see Fig. 2), respectively. After
two removal steps, all the predictors are significant for at least one peak To generate PFFs, Nsim = 106 realisations of design variables are
response, and they are therefore kept to build more refined metamodels. sampled from their uniform distribution and paired to a PGV realisation.
Fig. 8 compares the Adjusted R2 and Root Mean Square Error (RMSE) The selected number of realisations derives from a sensitivity study
statistics for the generated meatamodels. They differ in terms of the aimed at identifying the number of MCS that make the fragility functions
polynomial degree (first or second-order) and for the presence of medians converge to a stable value. The PGV is also sampled from a
interaction terms with first-order predictors. The metamodel with the uniform distribution within the range PGV ∈ [0.01,2.5]m/s. Monte Carlo
highest Adjusted R2 is selected and adopted to sample peak responses at simulation is then performed to build survival-failure vectors for bridge

11
A. Franchini et al. Engineering Structures 256 (2022) 113949

1 1

0 Model-Deck side 1 0
Model-Deck side 2 Model-Deck side 1
-1 Test-Deck side 1 -1 Model-Deck side 2
Test-Deck side 2 Test-Deck side 1
0 100 200 300 400 500 0 100 200 300 400 Test-Deck
500 side 2

1 1
Model-Deck center Model-Deck center
Test-Deck center Test-Deck center

0 0

-1 -1
0 100 200 300 400 500 0 100 200 300 400 500

( a) (b)

1 1

0 0

-1 Model-Deck side 1 -1
Model-Deck side 1
Model-Deck side 2 Model-Deck side 2
0 100 200 300 400 Test-Deck
500 side 1 0 100 200 300 400 500 side 1
Test-Deck
Test-Deck side 2 Test-Deck side 2

1 1
Model-Deck center Model-Deck center
Test-Deck center Test-Deck center

0 0

-1 -1
0 100 200 300 400 500 0 100 200 300 400 500

( c) (d)

-1 Model-Deck side 1
Model-Deck side 2
Test-Deck side 1
0 100 200 300 400 500
Test-Deck side 2

1
Model-Deck center
Test-Deck center

-1
0 100 200 300 400 500

(e) (f)
Fig. 6. Comparison of experimental and FEM modes. (a-c) first three vertical (V) modes; (d-f) first three transverse-torsional (TT) modes.

components: metamodels are applied to calculate Nsim realisations of figure aids the identification of the most vulnerable components. With
seismic EDPs, which are compared to realisations of limit state capac­ respect to longitudinal curvature ductility demand, the most vulnerable
ities sampled from their lognormal distributions. location in the towers is section 1–1 at the base (Z-11), whereas high
Vulnerable components are identified in Fig. 9a, whereas Fig. 9b transversal curvatures demands are observed at sections 3–3 and 4–4 (Y-
shows the sampled demands (D) and capacities (C) distributions. This 33, Y-44). The tower drift is limited to lower damage states. Due to the

12
A. Franchini et al. Engineering Structures 256 (2022) 113949

Table 5
Definition of limit state capacity for different bridge components.
Component Demand parameter Limit states thresholds

LS1-Slight LS2-Moderate LS3-Extensive LS4-Collapse

Sc CV Sc CV Sc CV Sc CV

Tower Cross section curvature ductility 1 0.35 2 0.35 4 0.35 7 0.35


Tower Drift Ratio [%] 0.7 0.35 1.5 0.35 2.5 0.35 5 0.35
Bearing Long. displacement [mm] 100 0.35 600 0.35 – – – –
Cable Axial force ratio (AFR) 0.4 0.11 0.6 0.11 0.85 0.11 1 0.11

101
0.3

100 Ground motion suite


0.25 Median PGV

0.2
10-1

0.15
10-2
0.1
-3
10
0.05

10-4 0
10-2 10-1 100 101 0 0.5 1 1.5 2 2.5

(a) (b)
Fig. 7. Ground motion suite details: (a) Response spectra; (b) PGV distribution.

1 1
M1
M2
M3
M4
0.5 0.5
M1
M2
M3
M4
0 0
Curv Y-11
Curv Z-11
Curv Z-22
Curv Z-33
Curv Z-44
Curv Y-22
Curv Y-33
Curv Y-44

Cab 1-28-29-56
Cab 2-27-30-55
Cab 3-26-31-54
Cab 4-25-32-53
Cab 5-24-33-52
Cab 6-23-34-51
Cab 7-22-35-50
Cab 8-21-36-49
Cab 9-20-37-48
Bearing
Drift

Cab 10-19-38-47
Cab 11-18-39-46
Cab 12-17-40-45
Cab 13-16-41-44
Cab 14-15-42-43

Curv Y-11
Curv Y-22
Curv Y-33
Curv Y-44
Drift
Curv Z-11
Curv Z-22
Curv Z-33
Curv Z-44

Bearing
Cab 1-28-29-56
Cab 2-27-30-55
Cab 3-26-31-54
Cab 4-25-32-53
Cab 5-24-33-52
Cab 6-23-34-51
Cab 7-22-35-50
Cab 8-21-36-49
Cab 9-20-37-48
Cab 10-19-38-47
Cab 11-18-39-46
Cab 12-17-40-45
Cab 13-16-41-44
Cab 14-15-42-43

(a) (b)
Fig. 8. Goodness-of-fit measures (Adjusted R2 and RMSE) for the 24 peak responses.

peculiar deck-to-tower connection of this bridge that prevents the X15. The reason why cables might easily enter the slight damage state in
relative displacement of the deck with respect to tower legs at some case of an earthquake is that they are prestressed to a level close to the
distance above the lower strut, the bearings do not represent vulnerable correspondent limit state threshold.
components in this specific case. Indeed, Fig. 9b shows that the bearings By comparing each realisation of capacity and demand, a survival-
are likely only to enter the slight damage state. On the other hand, such a failure vector is built for each component and each damage state, and
design choice places higher demand on the tower legs where the deck logistic regression is applied to estimate the coefficients of the corre­
displacement is constrained. sponding fragility curve. Furthermore, based on the series assumption
Looking at the plots for cables’ AFR in Fig. 9b, it is clear that cables for system-level fragility, a system-level survival-failure vector is also
(1, 28, 29, 56) exhibit a low probability of entering even the slight built, and logistic regression is applied to estimate the relevant fragility
damage state. On the contrary, cables (5, 24, 33, 52), (6,23,34,51), (9, functions. Regression coefficients can be found in the Supplementary
20, 37, 48) and (10, 19, 38, 47) are more vulnerable, although they also Material.
show a low probability of entering higher damage states. The cross- Conventional components’ fragility functions, conditioned on the IM
sectional area of all these cables corresponds to the design variable only and typically used for assessment, can be obtained by setting all the

13
A. Franchini et al. Engineering Structures 256 (2022) 113949

Fig. 9. Vulnerable components in the bridge: (a) identification; (b) sampled distributions of demands (D) and capacity (C) for peak responses and different limit
states (LS). AFR=Axial Force Ratio. For curvature ductility, Z=longitudinal, Y=transversal.

bridge design variables to their reference value and calculating the is also characterised by a low probability over the IM domain. Similar
probability of failure at increasing levels of PGV. The resulting curves behaviour is observed for cables at moderate, extensive and collapse
are plotted in Fig. 10. Section 1-1 is identified as the most vulnerable damage states. On the other hand, these components reveal to enter the
location for longitudinal curvature, whereas the same consideration for slight damage state easily; as discussed above, this is due to the prestress
section 3–3 and 4–4 is relevant for the transversal curvature. As ex­ level.
pected from the peak responses’ distributions in Fig. 9b, the bearings Fig. 11 presents system-level fragility functions for the Quincy Bay­
exhibit a low probability of damage; drift-induced damage to the towers view bridge. The high probability of the system entering the slight

14
A. Franchini et al. Engineering Structures 256 (2022) 113949

1 1
Curv Z-11 - LS1 Curv Z-33 - LS1 Curv Y-11 - LS1 Curv Y-33 - LS1
Curv Z-11 - LS2 Curv Z-33 - LS2 Curv Y-11 - LS2 Curv Y-33 - LS2
Curv Z-11 - LS3 Curv Z-33 - LS3 Curv Y-11 - LS3 Curv Y-33 - LS3
0.8 0.8 Curv Y-11 - LS4 Curv Y-33 - LS4
Curv Z-11 - LS4 Curv Z-33 - LS4
Curv Z-22 - LS1 Curv Z-44 - LS1 Curv Y-22 - LS1 Curv Y-44 - LS1
Curv Z-22 - LS2 Curv Z-44 - LS2 Curv Y-22 - LS2 Curv Y-44 - LS2
Curv Y-22 - LS3 Curv Y-44 - LS3
0.6 Curv Z-22 - LS3 Curv Z-44 - LS3 0.6 Curv Y-22 - LS4 Curv Y-44 - LS4
Curv Z-22 - LS4 Curv Z-44 - LS4

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

(a) (b)
1 1
Drift - LS1 Bearing - LS1
Drift - LS2 Bearing - LS2 Cab 1-28-29-56 - LS1 Cab 8-21-36-49 - LS1
Drift - LS3 Bearing - LS3 Cab 2-27-30-55 - LS1 Cab 9-20-37-48 - LS1
0.8 Drift - LS4 Bearing - LS4 0.8 Cab 3-26-31-54 - LS1 Cab 10-19-38-47 - LS1
Cab 4-25-32-53 - LS1 Cab 11-18-39-46 - LS1
Cab 5-24-33-52 - LS1 Cab 12-17-40-45 - LS1
Cab 6-23-34-51 - LS1 Cab 13-16-41-44 - LS1
0.6 0.6 Cab 7-22-35-50 - LS1 Cab 14-15-42-43 - LS1

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

(c) (d)
1 1
Cab 1-28-29-56 - LS2 Cab 8-21-36-49 - LS2 Cab 1-28-29-56 - LS3 Cab 8-21-36-49 - LS3
Cab 2-27-30-55 - LS2 Cab 9-20-37-48 - LS2 Cab 1-28-29-56 - LS4 Cab 8-21-36-49 - LS4
Cab 3-26-31-54 - LS2 Cab 10-19-38-47 - LS2 Cab 2-27-30-55 - LS3 Cab 9-20-37-48 - LS3
0.8 Cab 4-25-32-53 - LS2 Cab 11-18-39-46 - LS2 0.8 Cab 2-27-30-55 - LS4 Cab 9-20-37-48 - LS4
Cab 5-24-33-52 - LS2 Cab 12-17-40-45 - LS2 Cab 3-26-31-54 - LS3 Cab 10-19-38-47 - LS3
Cab 6-23-34-51 - LS2 Cab 13-16-41-44 - LS2 Cab 3-26-31-54 - LS4 Cab 10-19-38-47 - LS4
Cab 7-22-35-50 - LS2 Cab 14-15-42-43 - LS2 Cab 4-25-32-53 - LS3 Cab 11-18-39-46 - LS3
0.6 0.6
Cab 4-25-32-53 - LS4 Cab 11-18-39-46 - LS4
Cab 5-24-33-52 - LS3 Cab 12-17-40-45 - LS3
Cab 5-24-33-52 - LS4 Cab 12-17-40-45 - LS4
0.4 0.4 Cab 6-23-34-51 - LS3 Cab 13-16-41-44 - LS3
Cab 6-23-34-51 - LS4 Cab 13-16-41-44 - LS4
Cab 7-22-35-50 - LS3 Cab 14-15-42-43 - LS3
Cab 7-22-35-50 - LS4 Cab 14-15-42-43 - LS4
0.2 0.2

0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

(e) (f)
Fig. 10. Fragility functions for reference Quincy Bayview Bridge components: (a) longitudinal curvature ductility; (b) transversal curvature ductility; (c) tower drift
and bearing displacement; (d) cables at slight damage state; (e) cables at moderate damage state; (f) cables at extensive and collapse damage state.

damage state is due to the combined effect of a high probability of 4.5. Design space exploration
damage in terms of both curvature ductility (section 1–1, section 3–3)
and cables AFR. In this respect, the optimisation strategies proposed in 4.5.1. Effect of single design variable variation on damage probability
Section 2.5 allow a designer to control the probability of a system Following the procedure outlined in Section 2.4, this section in­
entering selected damage states. vestigates the effect of a 10% variation of each significant design vari­
able Xi on the DVs. Xi are varied one at a time while keeping the others at
their reference value, and the percentage variation of the considered DV
is calculated.

15
A. Franchini et al. Engineering Structures 256 (2022) 113949

1 their cross-section (design variable X12) makes the structure stiffer and
constrains the top displacement of the tower, attracting more force in
this component. Consequently, reducing their size also reduces the
0.8 damage probability at tower sections. An increase in the struts’ depth
LS1 - Slight
LS2 - Moderate (design variable X4) improves system-level performance at all damage
0.6 LS3 - Extensive states because of the higher stiffness it yields to the towers in the
LS4 - Collapse transversal direction. At lower damage states, such an effect is the
opposite with respect to the design variables X1 and X2 because the
0.4 struts sections C-C and D-D are required to remain in the elastic regime.
It is also interesting to investigate the effect of the struts’ thickness
(design variable X6), which determines their torsional stiffness. This
0.2 parameter shows the highest effect on the AMPGV, which increases (less
vulnerable structure) if X6 reduces. The struts’ torsional stiffness,
0 particularly that of the upper strut (section C-C), controls the coupling
0 0.5 1 1.5 2 2.5 between the two legs and their dynamic flexural properties, affecting the
vibrational periods of the overall structure. More in detail, a greater
thickness makes the tower stiffer and the flexural vibration period
Fig. 11. System-level fragility functions for Quincy Bayview Bridge (refer­
smaller, with corresponding greater amplification of the seismic input
ence case).
and therefore greater demands at the critical sections of the legs.
An overall look at Fig. 12a reveals that increasing the value of certain
Table 6 design variables might have opposite effects on system-level fragility at
Damage ratios for the calculation of the repair cost distribution [22]. the different damage states. For instance, regarding the cross-sectional
dimensions of the towers (design variables X1, X2), an increase in
Component Damage state
their size increases the likelihood of exceeding the slight and moderate
DS1 - Slight DS2 - Moderate DS3 - Extensive DS4 - Collapse limit state (LS1, LS2), whose thresholds are close (see Table 5): this
Bearing 0.04 0.10 0.50 1.00 higher vulnerability for low damage states is due to the above consid­
Pylon 0.03 0.08 0.25 1.00 erations on the yielding curvature. However, such an increase makes the
Cable 0.03 0.08 0.25 1.00
towers (and the overall structure) stiffer, reducing the seismic demand
placed upon other components at the higher limit states. Therefore, the
The damage ratios implemented to calculate the probability distri­ probability of the system entering higher damage states reduces.
bution of the repair cost RC are provided in Table 6 [22]. To calculate The above considerations confirm the complex response of highly
such distribution, the probability density of the PGV at the construction redundant structures such as cable-stayed bridges; indeed, changing the
site is required. Without loss of generality, the PGVs of the selected properties of a member can affect the whole structural performance.
accelerograms suite are assumed to represent the occurrence of earth­ Furthermore, the process through which this occurs is not easily gen­
quakes at the construction site. A lognormal distribution is fitted to these eralisable and depends on the considered performance measure. Thus,
values and used to sample 1000 realisations of PGV. The probability the outcomes presented in Fig. 12a are powerful tools to aid the de­
distribution of RC is then obtained by calculating its value for each PGV signer’s understanding of the role that different design variables play in
realisation. Eventually, its 95th percentile RC95 is obtained. reducing the probability of the bridge system exceeding a limit state.
Results of the sensitivity study are reported in Fig. 12a-b. Fig. 12c Fig. 12b depicts the effect of single-parameters variation on different
provides the layout of the investigated design variables. Each bar rep­ decision variables which refer to the cost. As expected, the construction
resents the effect of the variation of the design variable listed on the material cost C0 decreases with any design variable doing the same; the
vertical axis on the left: blue bars refer to 10% decreased design vari­ highest contribution relates to concrete sections dimensions.
ables, while red bars to 10% increased variables. Conversely, cables exhibit a negligible contribution. When RC95 is
Concerning Fig. 12a, the probability of the structure exceeding a considered, it can be observed that the design variables X1, X2, X3, X6,
damage state decreases if the PGV moves rightwards, whereas the X12, X14, X15 have an effect consistent with what was observed for the
opposite happens for a leftwards variation. The yielding curvature of a initial cost C0: their reduction also reduces the repair cost. However, it is
RC cross-section is inversely proportional to its size perpendicular to the interesting to observe that the increment of the parameters X4, X5, X8
neutral axis. Accordingly, reducing the depth of sections A-A and B-B and X13, albeit increasing C0, reduces the repair cost. This effect relates
(design variables X1, X2) reduces the stiffness of the towers and hence to those parameters reducing the damage probability for the bridge
structure and competes with the requirements to control the initial
increases the value of PGV for slight damage state, and, given its sig­
material cost. Such observation reveals the necessity to solve a trade-off
nificant contribution to system-level fragility (see Fig. 10a), the proba­
in controlling C0 and RC. Alternatively, the construction and repair
bility of the system entering the same damage state.
costs, CRC, can be considered in a single optimisation function. C0
The other components with significant contribution to the first
dominates this DV because of its higher value with respect to the repair
system-level damage state are the cables. Fig. 10d shows that the most
cost (as shown in Section 4.6 below).
vulnerable cables are the symmetric groups including cables (5,6,9,10).
On the other hand, the competing demands between C0 and RC95 for
The cross-sectional area of all these cables corresponds to the design
some design variables make them less influential for the CRC, as can be
variable X15, whose variation has the most significant influence on LS1.
observed for the cross-sectional dimensions of section C-C (X4 and X5).
In particular, increasing the area of these cables reduces their AFR,
It is also relevant to note that, despite the negligible contribution cables’
making it more difficult to enter the slight damage state. Fig. 10e-f also
cross-section provide to the initial material cost of this structure, they
reveals that cables contribution to higher damage states reduces; how­
appear among the most contributing parameters for RC95. Fig. 12b also
ever, Fig. 12a shows that they remain among the most critical compo­
highlights competing demands placed on design variables by AMPGV
nents. This fact can be explained by considering that cable-stayed
and CRC. In particular, it is possible to observe that changing some of the
bridges are highly redundant structures in which each cable affects the
design variables to maximize AMPGV would also increase the value of
distribution of forces within the other components.
CRC, as shown by design variables from X1, X3, X4 and X5.
Backstay cables are more critical for higher limit states. Increasing
Overall, Fig. 12 highlights the conflicting contribution of the selected

16
A. Franchini et al. Engineering Structures 256 (2022) 113949

X15
X14
X13
X12
X11
X9
X8
X6
X5
X4
X3
X2
X1
Xi decreased by 10%
Xi increased by 10%

-50 0 50 -50 0 50 -50 0 50 -50 0 50 -10 0 10

(a)

X15
X14
X13
X12
X11
X9
X8
X6
X5
X4
X3
X2
X1
Xi decreased by 10%
Xi increased by 10%

-5 0 5 -20 0 20 -5 0 5

(b) (c )
Fig. 12. Effect of single design variable variation on performance measures: (a) on the median PGV (PGV) of the bridge at different damage states and on the average
median PGV (AMPGV); (b) on cost metrics: construction material cost C0, 95th percentile of repair cost distribution RC95, and Construction and Repair Cost CRC; (c)
identification of design variables.

design variables in determining different system-level performance • In line with Fig. 12, cables parameters are the most affecting pa­
measures, in the sense that increasing the value of a design variable rameters for the slight damage state; in particular, the interaction
might positively affect a performance objective while worsening others. with the shorter cables (cross-sectional area X15) is the most
In other words, design variables reveal a conflicting influence when influencing.
their variation positively affects one or more DVs while deteriorating • At higher damage states (LS3 observed in Fig. 13b), the contribution
other DVs. Consistently, design optimisation represents the most effec­ of other design variables becomes significant, and the tile contour
tive way to find a combination of design variables that achieves the lines variation is more dynamic than at lower damage states. The
desired performances while satisfying short- and long-term economic shorter cables (X15) and backstays (X12) exhibit impactful interac­
constraints. tion with the other design variables; the thickness of the upper strut
beam (X6) and section A-A longitudinal reinforcement (X8) also
4.5.2. Pairwise effect of design variables stand out. A unique ’best’ choice for each of these parameters (i.e.
To further investigate the conflicting or conflating effects of design maximizing or minimizing them) cannot be identified and depends
variables on the system performance, tile plots (see Section 2.4) are on the choice taken for all the other parameters.
produced for pair of variables. In particular, the change in median peak • Conflicting trends with respect to the single-design variable sensi­
ground velocity for system-level slight and extensive damage states tivity study (Fig. 12) can be identified, highlighting that the design
(PGV LS1 ,PGV LS3 ) is studied. The results are reported in Fig. 13. Each tile updating process must also consider parameters interaction. For
represents the variation of PGV LSi versus the two statistically significant example, Fig. 12a suggests that, to increase PGV LS3 , X6 (struts
design variables listed on the vertical and horizontal overall axes. All thickness) should be decreased and X8 (section A-A’s longitudinal
plots are set to the same PGV scale. Vertical and horizontal gradients reinforcement) increased. The opposite requirement can be identi­
mean poor interaction between the variables, positive slope shows in­ fied from the X6-X8 tile in Fig. 13b. This effect can be explained by
verse interaction, and negative slope shows direct interaction. observing that an increase in the flexural reinforcement X8 balances
Among the wide range of information that might be obtained from the effect of the increase of strut’s thickness X6, causing higher
these plots, the following overall observations are drawn: flexural demand on the critical tower sections, by increasing the
ultimate bending moment capacity of the same sections and

17
A. Franchini et al. Engineering Structures 256 (2022) 113949

[m/s]
X2 45 0.15

X3
40
X4

X15 [cm2]
0.1

X5 35

X6 0.05
30

X8
25
X9 40 45 50 55 60 65
0

X14 [cm2]
X11

X12
X13
X14
X15
X1 X2 X3 X4 X5 X6 X8 X9 X11 X12 X13 X14
(a)

[m/s]
45 2.8

X2 2.6

2.4
40
X3 2.2
X15 [cm2]

2
X4 35
1.8

X5 1.6
30
1.4
X6
1.2
25
X8 1
40 45 50 55 60 65

X9 X14 [cm2]

X11

X12

X13

X14

X15
X1 X2 X3 X4 X5 X6 X8 X9 X11 X12 X13 X14

Fig. 13. Median peak ground velocity PGV landscape for: (a) system-level slight damage state (LS1); (b) system-level extensive damage state (LS3).

therefore reducing the likelihood of these sections exceeding higher the curvature demand at section 1–1 and the AFR for cables (6, 23, 34,
damage states. 51) with respect to section B-B depth and thickness (X1, X3). Different
combinations of these variables lead to different structural behaviours,
The pairwise effect analysis reveals that the competing demand in which either curvature ductility or cable stresses govern structural
placed on design variables also depends on variables interaction. failure. As proved in the next section, numerical optimisation can indi­
Therefore, numerical optimisation is advised to identify the optimal rectly identify the ideal structural behaviour for the chosen design
combination of design variables that also account for interaction effects. objectives.
Further support to this statement is lent by the observation that
seismic input in complex structures as cable-stayed bridges leads to
nonconvex or even nonconnected behaviour. In this respect, the varia­ 4.6. Surrogate-based optimisation
tion of the selected design variables might cause a shift from a structure
with higher curvature ductility demand placed on tower sections to The design space exploration performed in Section 4.5 revealed
another structure with higher stress in cables. For instance, Fig. 14 competing demands placed upon the same design variable by different
(obtained for a PGV = 1.0 m/s) shows a nonconvex response surface of decision variables and the crucial role that design variables interaction
plays in determining them. Thus, the optimisation strategies

18
A. Franchini et al. Engineering Structures 256 (2022) 113949

that the bridge system presents a high probability of entering the slight
damage state even for low levels of PGV (PGV LS1 = 0.03m/s). The
analysis of Fig. 9b and Fig. 10d, as well as the sensitivity study of Section
4.5, revealed that such low value is mainly due to cables prestressing
level. As discussed in Section 2.5, it was chosen to constrain PGV LS1 to be
no less than 0.25 m/s, value which approximates the median PGV at the
construction site (see Fig. 7b). When the AMPGV is not the object of
optimisation, its value is constrained to be no less than the AMPGV of the
performance of the bridge designed using the reference values in Table 2
(AMPGV ref = 2.29m/s). A similar consideration was made for C0 : in
Strategy #1 and #2, its value is constrained not to exceed the con­
struction material cost of the reference structure C0,ref = 39€M. Finally,
the most critical locations for buckling check are sections 3–3 and 4–4 in
Fig. 2b, for which appropriate boundary conditions were set.
The formulated problems are solved through the optimisation
toolbox of the MATLAB software [71] as constrained nonlinear multi­
Fig. 14. Peak response surfaces. variable problems. More in detail, the function ‘fmincon’, which im­
plements a sequential quadratic programming method, is used for
summarized in Table 1 are applied in this section to compute optimal single-objective strategies (#1 to #3). ‘paretosearch’ is instead imple­
design variable combinations. Case study-specific constraint conditions mented to solve the multi-objective optimisation problem (Strategy #4)
are described before presenting the optimisation results. and obtain the Pareto front through pattern search.
The constraint LB ≤ X ≤ UB for the space of design variables listed Table 7 and Fig. 15a summarize the optimal values X* of the design
in Table 2, which was adopted for experimental design, are held for variables for each single-objective optimisation strategy. Results are
optimisation. In terms of damage control performances, Fig. 11 shows

Table 7
Ratios between optimal design values and Quincy (reference) values.
Design Variable X1 X2 X3 X4 X5 X6 X8 X9 X11 X12 X13 X14 X15

Case Optimal value Xi*/Reference value Xireference


Upper bound 1.30 1.30 1.14 1.30 1.30 2.00 1.45 1.45 1.50 1.30 1.30 1.30 1.30
Lower bound 0.70 0.70 0.71 0.70 0.70 1.25 0.55 0.55 0.50 0.70 0.70 0.70 0.70
Quincy, reference 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Strategy #1 0.70 0.70 1.14 1.30 1.30 1.25 1.45 0.55 1.50 0.70 1.30 1.30 0.70
Strategy #2 0.70 0.70 0.71 1.30 1.23 1.25 1.45 1.45 1.44 0.70 1.23 0.88 0.84
Strategy #3 0.70 0.70 0.71 0.70 0.70 1.25 1.45 0.94 1.50 0.70 1.24 1.25 1.00

2 10 40
C_0 [M €]

8 30
6
20
1.5 4
2 10
0 0
Ref. #1 #2 #3 Ref. #1 #2 #3
1

3
40
RC_95 [M €]

CRC [M €]

0.5 2 30
Upper and Lower Bound Strategy #2
Quincy (reference) Strategy #3 20
1
0 Strategy #1 10
0 0
X1
X2
X3
X4
X5
X6
X8
X9
X11
X12
X13
X14
X15

Ref. #1 #2 #3 Ref. #1 #2 #3 Ref.

(a) ( b)
Fig. 15. Single-objective optimisation results: (a) optimal design variables; (b) optimal value of decision variables.

Table 8
Single-objective optimisation results. Values in brackets represent the percentual variation of the decision variable with respect to the reference structure.
Case Objective Decision Variables

PGVLS1 [m/s] AMPGV [m/s] C0 [€ M] RC95 [€ M] CRC [€ M]

Quincy (reference) – 0.03 2.29 38.88 2.51 41.39


Strategy #1 Maximise AMPGV 0.25 10.28 (+348.15%) 32.27 (-17.00%) 0.64 (-74.56%) 32.91 (-20.49%)
Strategy #2 Minimise Repair Cost RC95 0.25 6.79 (+195.92%) 29.42 (-24.34%) 0.31 (-87.81%) 29.72 (-28.19%)
Strategy #3 Minimise Construction and Repair cost CRC 0.25 2.29 (-0.00%) 23.97 (-38.35%) 0.78 (-68.83%) 24.75 (-40.20%)

19
A. Franchini et al. Engineering Structures 256 (2022) 113949

34
2 Feasible region
Pareto Front (Strategy #4)
32 Strategy #1
Strategy #2
1.5
Strategy #3

CRC [M €]
30
1
28

0.5
Upper and Lower Bound 26
Quincy (reference)
Pareto optimal solutions
0
24
X1
X2
X3
X4
X5
X6
X8
X9
X11
X12
X13
X14
X15
2 4 6 8 10

( a) (b)
Fig. 16. Results of Strategy #4 for multi-objective optimisation: (a) Pareto optimal solution design variables; (c) Pareto front and comparison with single-objective
optimisation strategies.

presented as ratios of the reference values. The obtained DVs are instead using the minimization of repair cost (Strategy #3), albeit not yielding a
listed in Table 8, with the difference from the reference case tabulated in dominant solution in terms of CRC and AMPGV, gives a good trade-off
percentage. The DVs are also plotted in Fig. 15b. between safety and cost.
From Fig. 15a, it can be observed that the three strategies led to the Overall, the multi-objective approach provides a broader set of
lowest values of the external dimensions of the tower leg cross-sections optimal parameters configurations, among which a decision-maker
(design variables X1, X2). This can be explained considering that their might select the one which suits most based on subjective criteria.
reduction decreases cost-related DVs (Fig. 12b) while increasing PGV LS1 A general look at the results of Fig. 15 and Fig. 16 proves that nu­
and having limited contribution to AMPGV (Fig. 12a). In addition, merical optimisation algorithms represent an effective way to deal with
Strategy #1 and #2 increased the strusts’ gross section dimensions (X4, competing demands placed upon cable-stayed bridges’ design variables
X5) to control the considered damage probabilities better. Conversely, by different performance metrics when considering seismic design. In
X4 and X5 were reduced in Strategy #3, which is consistent with their addition, surrogate metamodeling and PFFs aided computational effi­
observed effect from Fig. 12b. More variability can be observed in ciency, making the proposed approach valuable for optimising these
optimal cables’ cross-sections (design variables X12 to X15). structures.
Strategy #1 achieved the highest system-level safety, increasing
AMPGV by 3.5 times with respect to the reference structure; meanwhile, 5. Conclusions
the approach entailed the highest construction material cost and highest
construction and repair cost among the optimised configurations. This paper proposed a computationally efficient methodology for the
Strategy #2 reduced the direct economic losses in case of an earthquake probabilistic optimisation of cable-stayed bridges in earthquake-prone
(RC95 ) by 87.81% while containing the initial investment within the set regions. In particular, it demonstrated how surrogate models and par­
budget. Finally, Strategy #3 decreased CRC by 40.20%. ameterised fragility functions can be implemented in the PEER-PBEE
Fig. 12 revealed competing demands placed on design variables by framework for design space exploration and optimisation. The ach­
AMGV and CRC. Therefore, a multi-objective optimisation with pattern ieved results yield the following conclusions:
search was implemented (Strategy #4) to identify optimal design vari­
able configurations and the Pareto front depicted in Fig. 16. Linear • Parameterised fragility functions represent a computationally effi­
interpolation between optimal points is assumed. The same figure also cient tool to define probabilistic decision variables that can be used
reports the performance metrics points resulting from single-objective for design space exploration and optimisation of cable-stayed
optimisation strategies. bridges. Four optimisation strategies were proposed to minimize
Comparing Fig. 15a and Fig. 16a, it emerges that all the strategies damage probability, construction cost and seismic damage-induced
minimize the gross section dimensions of the tower leg (X1, X2) and the repair cost.
area of backstay cables (X12). Such a combination minimises the cost • As a case study, a 542 m three-span cable-stayed bridge was
while reducing damage probability at the tower section (see discussion considered, and its modelling parameters were updated to match
in Section 4.5.1). Still regarding cables, the sensitivity study of Section values of dynamic properties measured by an experimental
4.5 revealed that the cross-sectional areas X12 and X15 are critical for campaign. This enabled improvement of the FE model’s architecture,
system-level performance and that their simultaneous reduction vastly thereby enhancing confidence that the FE predictions forming the
increase higher damage states’ medians (Fig. 13). Consistently, all core of the fragility study are reliable and that the proposed strate­
optimal solutions reduced the cross-sections of these two cables to gies can be applied to design new structures.
maximise AMPGV. However, Pareto optimal solutions exhibit a higher • Surrogate-based sensitivity analysis revealed conflicting demands
variability for the other variables, uncovering solutions that couldn’t be placed on geometrical design variables (towers’ cross-section, rein­
found through single-objective optimisation strategies. forcement, cables’ area) by different DVs and the crucial importance
In Fig. 16b, single-objective optimisation results from Strategy #1 of accounting for design variables interaction when making choices
and #3 lay at the extremes of the Pareto front and were used as initial on design updating.
points. On the other hand, the results of Strategy #2, whose purpose was
to minimize the 95th percentile of repair cost distribution, lay close to
the front and falls within the feasible region. Thus, Fig. 16b shows that

20
A. Franchini et al. Engineering Structures 256 (2022) 113949

• Numerical optimisation strategies are recommended to better deal [17] Moehle, J. and G.G. Deierlein. A framework methodology for performance-based
earthquake engineering. in 13th World Conference on Earthquake Engineering.
with conflicting demands and variables interaction. Through single-
2004. Vancouver, B.C., Canada.
objective optimisation (Strategy #1 to #3), a 348% increment of the [18] Günay S, Mosalam KM. PEER performance-based earthquake engineering
AMPGV, an 88% reduction of repair cost, and a 40% decrease of CRC methodology, revisited. J Earthquake Eng 2013;17(6):829–58.
were achieved. On the other hand, Strategy #4 was used to obtain a [19] Shama A, Jones M. Seismic Performance-Based Design of Cable-Supported Bridges:
State of Practice in the United States. J Bridge Eng 2020;25(12):04020101.
set of Pareto optimal design variable combinations that simulta­ [20] Wu W, Li L, Shao X. Seismic assessment of medium-span concrete cable-stayed
neously optimize AMPGV and CRC. The minimization of repair cost bridges using the component and system fragility functions. J Bridge Eng 2016;21
yields a good trade-off between safety and cost DVs. (6):04016027.
[21] Zhong J, et al. System-based probabilistic optimization of fluid viscous dampers
equipped in cable-stayed bridges. Adv Struct Eng 2018;21(12):1815–25.
This study addressed the self-weight and earthquake load case, [22] Wen J, et al. Performance-based seismic design and optimization of damper
which was assumed to dominate other load cases. Meanwhile, the devices for cable-stayed bridge. Eng Struct 2021;237:112043.
[23] Xie Y, Zhang J. Optimal design of seismic protective devices for highway bridges
complete design of cable-stayed bridges should also consider additional using performance-based methodology and multiobjective genetic optimization.
constraint conditions to control components serviceability and guar­ J Bridge Eng 2017;22(3):04016129.
antee safety for other loads. Thus, the proposed procedure sets the basis [24] Zhong J, et al. Risk-informed sensitivity analysis and optimization of seismic
mitigation strategy using Gaussian process surrogate model. Soil Dyn Earthquake
for further and more accurate developments in the framework of multi- Eng 2020;138:106284.
hazard design. [25] Wei B, et al. System-based probabilistic evaluation of longitudinal seismic control
for a cable-stayed bridge with three super-tall towers. Eng Struct 2021;229:
111586.
Declaration of Competing Interest [26] Seo J, Linzell DG. Horizontally curved steel bridge seismic vulnerability
assessment. Eng Struct 2012;34:21–32.
[27] Ghosh J, Padgett JE, Dueñas-Osorio L. Surrogate modeling and failure surface
The authors declare that they have no known competing financial visualization for efficient seismic vulnerability assessment of highway bridges.
interests or personal relationships that could have appeared to influence Probab Eng Mech 2013;34:189–99.
the work reported in this paper. [28] Dukes, J.D., Application of bridge specific fragility analysis in the seismic design process
of bridges in California. 2013, (Doctoral Dissertation, Georgia Institute of
Technology).
Acknowledgements [29] Ghosh J, et al. Seismic reliability assessment of aging highway bridge networks
with field instrumentation data and correlated failures, I: Methodology.
Earthquake Spectra 2014;30(2):795–817.
The financial support of the Maurice Franses Memorial Trust and [30] Rokneddin K, et al. Seismic reliability assessment of aging highway bridge
that of UCL’s Department of Civil, Environmental and Geomatic Engi­ networks with field instrumentation data and correlated failures, II: Application.
Earthquake Spectra 2014;30(2):819–43.
neering are greatly acknowledged. The support of UCL high-
[31] Mangalathu S, Heo G, Jeon J-S. Artificial neural network based multi-dimensional
performance computing power is also appreciated. fragility development of skewed concrete bridge classes. Eng Struct 2018;162:
166–76.
[32] Mangalathu S, Jeon JS, DesRoches R. Critical uncertainty parameters influencing
Appendix A. Supplementary data seismic performance of bridges using Lasso regression. Earthquake Eng Struct Dyn
2018;47(3):784–801.
Supplementary data to this article can be found online at https://doi. [33] Mangalathu S, Jeon JS. Stripe-based fragility analysis of multispan concrete bridge
classes using machine learning techniques. Earthquake Eng Struct Dyn 2019;48
org/10.1016/j.engstruct.2022.113949.
(11):1238–55.
[34] Jeon J-S, et al. Parameterized seismic fragility curves for curved multi-frame
References concrete box-girder bridges using Bayesian parameter estimation. J Earthquake
Eng 2019;23(6):954–79.
[35] Dukes J, et al. Development of a bridge-specific fragility methodology to improve
[1] Camara A. Seismic behaviour of cable-stayed bridges: A review. Medcrave Online
the seismic resilience of bridges. Earthquakes and Structures 2018;15(3):253–61.
Journal of Civil Engineering 2018;4(3):161–9.
[36] Kameshwar S, Padgett JE. Multi-hazard risk assessment of highway bridges
[2] Filiatrault A, Tinawi R, Massicotte B. Damage to cable-stayed bridge during 1988
subjected to earthquake and hurricane hazards. Eng Struct 2014;78:154–66.
Saguenay earthquake. I: Pseudostatic analysis. J Struct Eng 1993;119(5):1432–49.
[37] Kameshwar S, Padgett JE. Effect of vehicle bridge interaction on seismic response
[3] Chang K-C, et al. Lessons learned from the damaged Chi-Lu cable-stayed bridge.
and fragility of bridges. Earthquake Eng Struct Dyn 2018;47(3):697–713.
J Bridge Eng 2004;9(4):343–52.
[38] Gehl, P. and D. D’Ayala. Integrated multi-hazard framework for the fragility
[4] Siringoringo DM, Fujino Y. Observed dynamic performance of the Yokohama-Bay
analysis of roadway bridges. in Proceedings of the 12th International Conference
Bridge from system identification using seismic records. Structural Control and
on Applications of Statistics and Probability in Civil Engineering (ICASP12). 2015.
Health Monitoring: The Official Journal of the International Association for
Vancouver, Canada.
Structural Control and Monitoring and of the European Association for the Control
[39] Kildashti K, et al. Drive-by-bridge inspection for damage identification in a cable-
of Structures 2006;13(1):226–44.
stayed bridge: Numerical investigations. Eng Struct 2020;223:110891.
[5] Bruneau M, Wilson JC, Tremblay R. Performance of steel bridges during the 1995
[40] Sebastian WM, Johnson M. Interpretation of sensor data from in situ tests on a
Hyogo-ken Nanbu (Kobe, Japan) earthquake. Can J Civ Eng 1996;23(3):678–713.
transversely bonded fibre-reinforced polymer road bridge. Structural Health
[6] Gehl P, D’Ayala D. Development of Bayesian Networks for the multi-hazard
Monitoring 2019;18(4):1074–91.
fragility assessment of bridge systems. Struct Saf 2016;60:37–46.
[41] Li Y, Conte JP, Gill PE. Probabilistic performance-based optimum seismic design
[7] Gehl P, D’Ayala D. System loss assessment of bridge networks accounting for multi-
framework: illustration and validation. Computer Modeling in Engineering &
hazard interactions. Struct Infrastruct Eng 2018;14(10):1355–71.
Sciences 2019;120(3):517–43.
[8] Martins AMB, Simões LMC, Negrão JHJO. Optimization of cable-stayed bridges: A
[42] Padgett JE, DesRoches R. Methodology for the development of analytical fragility
literature survey. Adv Eng Softw 2020;149:102829.
curves for retrofitted bridges. Earthquake Eng Struct Dyn 2008;37(8):1157–74.
[9] Camara A, Efthymiou E. Deck–tower interaction in the transverse seismic response
[43] D’Ayala, D., et al., Guidelines for analytical vulnerability assessment of low/mid-rise
of cable-stayed bridges and optimum configurations. Eng Struct 2016;124:
Buildings—Methodology. Vulnerability Global Component project. 2014, GEM
494–506.
Technical Report.
[10] Ferreira F, Simões L. Synthesis of three dimensional controlled cable-stayed bridges
[44] Simpson TW, et al. Metamodels for computer-based engineering design: survey and
subject to seismic loading. Comput Struct 2020;226:106137.
recommendations. Engineering with Computers 2001;17(2):129–50.
[11] Ferreira FLS, Simoes LMC. Optimum design of a controlled cable stayed bridge
[45] Forrester, A., A. Sobester, and A. Keane, Engineering design via surrogate modelling: a
subject to earthquakes. Struct Multidiscip Optim 2011;44(4):517–28.
practical guide. 2008: John Wiley & Sons, Ltd.
[12] Simões LMC, Negrão JHJO. Optimization of cable-stayed bridges subjected to
[46] McKay MD, Beckman RJ, Conover WJ. A comparison of three methods for selecting
earthquakes with non-linear behaviour. Eng Optim 1999;31(4):457–78.
values of input variables in the analysis of output from a computer code.
[13] Martins AMB, Simões LMC, Negrão JHJO. Optimization of concrete cable-stayed
Technometrics 2000;42(1):55–61.
bridges under seismic action. Comput Struct 2019;222:36–47.
[47] Box GEP, Wilson KB. On the experimental attainment of optimum conditions. J Roy
[14] Ang A-S. Life-cycle considerations in risk-informed decisions for design of civil
Stat Soc: Ser B (Methodol) 1951;13(1):1–38.
infrastructures. Struct Infrastruct Eng 2011;7(1–2):3–9.
[48] Li H, et al. Seismic fragility assessment framework for highway bridges based on an
[15] SEAOC, Vision 2000: Performance-Based Seismic Engineering of Buildings. 1995:
improved uniform design-response surface model methodology. Bull Earthq Eng
Structural Engineers Association of California, Sacramento (CA).
2020;18(5):2329–53.
[16] Mackie, K. and B. Stojadinovic. Fragility curves for reinforced concrete highway
[49] Nielson BG, DesRoches R. Seismic fragility methodology for highway bridges using
overpass bridges. in 13th World conference on earthquake engineering. Vancouver BC,
a component level approach. Earthquake Eng Struct Dyn 2007;36(6):823–39.
Canada.

21
A. Franchini et al. Engineering Structures 256 (2022) 113949

[50] Kircher CA, Whitman RV, Holmes WT. HAZUS earthquake loss estimation [63] Mazzoni S, et al. OpenSees Command Language Manual. 2006.
methods. Nat Hazard Rev 2006;7(2):45–59. [64] Wilson JC, Liu T. Ambient vibration measurements on a cable-stayed bridge.
[51] Cornell CA, et al. Probabilistic basis for 2000 SAC federal emergency management Earthquake Eng Struct Dyn 1991;20(8):723–47.
agency steel moment frame guidelines. J Struct Eng 2002;128(4):526–33. [65] Nazmy AS, Abdel-Ghaffar AM. Non-linear earthquake-response analysis of long-
[52] Dueñas-Osorio L, Padgett JE. Seismic reliability assessment of bridges with user- span cable-stayed bridges: Theory. Earthquake Eng Struct Dyn 1990;19(1):45–62.
defined system failure events. J Eng Mech 2011;137(10):680–90. [66] Abdel-Ghaffar AM, Nazmy AS. 3-D nonlinear seismic behavior of cable-stayed
[53] Pang Y, et al. Seismic fragility analysis of cable-stayed bridges considering different bridges. J Struct Eng 1991;117(11):3456–76.
sources of uncertainties. J Bridge Eng 2014;19(4):04013015. [67] Nazmy AS, Abdel-Ghaffar AM. Three-dimensional nonlinear static analysis of
[54] Li C, et al. Seismic fragility analyses of sea-crossing cable-stayed bridges subjected cable-stayed bridges. Comput Struct 1990;34(2):257–71.
to multi-support ground motions on offshore sites. Eng Struct 2018;165:441–56. [68] Ernst, J.H., Der E-Modul von Seilen unter berucksichtigung des Durchhanges. Der
[55] Nocedal, J. and S. Wright, Numerical optimization. 2006: Springer Science & bauingenieur, 1965. 40(2): p. 52-55.
Business Media. [69] Pastor M, Binda M, Harčarik T. Modal assurance criterion. Procedia Eng 2012;48:
[56] Pang Y, He W, Zhong J. Risk-based design and optimization of shape memory alloy 543–8.
restrained sliding bearings for highway bridges under near-fault ground motions. [70] Simoen E, De Roeck G, Lombaert G. Dealing with uncertainty in model updating
Eng Struct 2021;241:112421. for damage assessment: A review. Mech Syst Sig Process 2015;56:123–49.
[57] Awad, M. and R. Khanna, Efficient learning machines: theories, concepts, and [71] The Mathworks, I., MATLAB version 9.5.0.1298439 (R2018b) Update 7. 2018:
applications for engineers and system designers. 2015: Springer nature. Natick, Massachusetts.
[58] Wilson JC, Gravelle W. Modelling of a cable-stayed bridge for dynamic analysis. [72] Baker, J.W., et al., New ground motion selection procedures and selected motions for the
Earthquake Eng Struct Dyn 1991;20(8):707–21. PEER transportation research program, in PEER report 3. 2011, Pacific Earthquake
[59] Santos CAN, et al. Structural optimization of two-girder composite cable-stayed Engineering Research Center.
bridges under dead and live loads. Can J Civ Eng 2020;47(8):939–53. [73] Zhong J, et al. Optimal intensity measures in probabilistic seismic demand models
[60] Casado, A.C., Seismic behaviour of cable-stayed bridges: design, analysis and seismic of cable-stayed bridges subjected to pulse-like ground motions. J Bridge Eng 2019;
devices. 2011, (Doctoral dissertation, Universidad Politécnica de Madrid). 24(2):04018118.
[61] Hassan MM, El Damatty AA, Nassef AO. Database for the optimum design of semi- [74] Guo J, et al. Optimal intensity measures for probabilistic seismic demand models of
fan composite cable-stayed bridges based on genetic algorithms. Struct Infrastruct a cable-stayed bridge based on generalized linear regression models. Soil Dyn
Eng 2015;11(8):1054–68. Earthquake Eng 2020;131:106024.
[62] CEN, EN 1992-1-1 Eurocode 2: Design of concrete structures - Part 1-1: General [75] Williams, C.K. and C.E. Rasmussen, Gaussian processes for machine learning. Vol. 2.
rules and rules for buildings. 2005: Comité Européen de Normalisation (Brussels). 2006: MIT press Cambridge, MA.
p. 230.

22

You might also like