Professional Documents
Culture Documents
Full Title: Clustering-Based Active Learning Reliability of FRP Strengthened RC Beams with
Random Field Nonlinear Finite Element to Model Spatial Variability
Abstract: This paper presents a framework for assessing the reliability of fiber-reinforced polymer
(FRP) strengthened reinforced concrete (RC) beams in flexure using stochastic
nonlinear finite element (SNFE) analysis and k-w-means clustered active learning
Kriging Monte Carlo Simulation (AK-MCS), where the spatial variation in the concrete
and bond material properties are considered. A computer algorithm was developed to
augment commercially available non-linear finite element analysis software and
automate the process for conducting the SNFE-Clustered AK-MCS analysis. The k-w-
means clustering is based on the U learning function to provide multi-point enrichment
to improve convergence of the stopping criteria by allowing parallel computation of the
SNFE models. Parametric analysis indicated the accuracy of the reliability prediction of
the examined member and proved the efficiency of the proposed analysis in reducing
the number of calls to SNFE models compared with existing literature, when using
probability-based stopping criteria.
Additional Information:
Question Response
The journal requires that all submissions The paper presents a novel computational technique to assess the reliability of FRP
fall within its aims and scope, explained strengthened structures by considering the spatial distribution of the material properties
here. Please explain how your submission in non-linear finite element simulation. The technique utilizes active learning Kriging,
fits the journal's aims and scope. computer clustering, and stochastic nonlinear finite element simulation.
Papers published in ASCE Journals must Stochastic nonlinear finite element simulation offers significant opportunities in real-life
make a contribution to the core body of practice to assess the reliability of degrading structures strengthened using FRP.
knowledge and to the advancement of the However, the computational cost is generally prohibitive. In this research, the efficiency
field. Authors must consider how their of performing such analysis is significantly improved by utilizing adaptive Kriging (form
new knowledge and/or innovations add of AI), and clustering (computer parallelization). The framework is proven to be efficient
value to the state of the art and/or state of and can be applied to a variety of FRP strengthening configurations.
the practice. Please outline the specific
contributions of this research in the
comments box.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
secured? If yes, please explain in the
comment box.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
manuscript pages. Papers that exceed the
limits must be justified. Grossly over-
length papers may be returned without
review. Does this paper exceed the ASCE
length limitations? If yes, please provide
justification in the comments box below.
Please provide the previous manuscript The paper was submitted to the ASCE Journal of Structural Engineering (MS STENG-
number and explain what you have 13041), and an initial rejection was made. The comments were focused on improving
changed in this current version in the the quality of presenting the paper contribution and expanding on the literature review.
comments box below. You may upload a Both aspects were improved in the revised submission:
separate response to reviewers if your a. The research significance was re-written as can be seen in the revised "practical
comments are extensive. consideration" section of the paper. The paper contribution is focused on applying
clustered AK analysis for FRP strengthening. This is the first study to apply clustered
as follow-up to "Was this paper AK analysis for stochastic nonlinear finite element modeling of FRP strengthened
previously declined or withdrawn from this beams. Verifying the accuracy and efficiency of AK analysis for FRP design using
or another ASCE journal? If so, please stochastic FE is needed to validate the method of analysis for future use.
provide the previous manuscript number b. A complete paragraph was included in the introduction section to provide a state of
and explain what you have changed in the art review of current work related to AK parallelization as requested by the
this current version in the comments box reviewers.
below. You may upload a separate
response to reviewers if your comments The authors believe that the paper's focus fits better with the scope of ASCE Journal of
are extensive." Composite for Construction and thus, it was decided not to rebut the decision by ASCE
Journal of Structural Engineering.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
authors and manuscript number (if
available) for the companion papers along
with your detailed justification for the
editor in the comments box below. If there
is no justification provided, or if there is
insufficient justification, the papers will be
returned without review.
When submitting a manuscript, authors d. Some or all data, models, or code generated or used during the study are
must include a section heading titled proprietary or confidential in nature and may only be provided with restrictions.
“Data Availability Statement” before the
“Acknowledgments” section or after the
“Conclusion." Within the section, authors
will include one or more of the following
statements, as well as all citations to data,
code, or models. You can read more
about the Data Availability Statement
policy here.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
Please select one or more of the
statements below that apply to your
manuscript. The statement(s) listed in
your manuscript should match those you
select in your response to this question.
Please describe which data are The computer code developed for the paper is a proprietary.
proprietary or confidential and what
restrictions apply.
as follow-up to "When submitting a
manuscript, authors must include a
section heading titled “Data Availability
Statement” before the “Acknowledgments”
section or after the “Conclusion." Within
the section, authors will include one or
more of the following statements, as well
as all citations to data, code, or models.
You can read more about the Data
Availability Statement policy here.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
ASCE is a signatory of the Sustainable Industry, Innovation and Infrastructure - - Build resilient infrastructure, promote
Development Goals (SDG) Publishers inclusive and sustainable industrialization and foster innovation
Compact, supporting the U.N. Sustainable
Development Goals listed below. You can
read more about our commitment here. If
your submission addresses any of these
goals, please check up to five of the
relevant boxes below.
I. Authorship Responsibility
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
journal article or 500 words from a book,
written permission from the copyright
holder must be obtained and supplied with
the submission. Each author has a
responsibility to identify materials that
require permission by including a citation
in the figure or table caption or in
extracted text.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
formats and media of expression now
known or later developed, including
electronic, to the American Society of Civil
Engineers subject to the following:
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
• Authors may post the final draft of their
work on open, unrestricted Internet sites
12 months after publication in an ASCE
Journal, provided the draft contains a link
to the published version at
ascelibrary.org.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
exemption in its entirety. If at least one
author is not a Crown Government
employee, copyright must be transferred
to ASCE by that author.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
I, the corresponding author, confirm that
the content, figures, drawings, charts,
photographs, and tables in the submitted
work are either original work created by
the authors listed on the manuscript or
work for which permission to re- use has
been obtained from the creator.
as follow-up to "COPYRIGHT
TRANSFER AGREEMENT
I. Authorship Responsibility
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
holder to reproduce any material that (1)
they did not create themselves and/or (2)
has been previously published, to include
the authors’ own work for which copyright
was transferred to an entity other than
ASCE. For any figures, tables, or text
blocks exceeding 100 words from a
journal article or 500 words from a book,
written permission from the copyright
holder must be obtained and supplied with
the submission. Each author has a
responsibility to identify materials that
require permission by including a citation
in the figure or table caption or in
extracted text.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
the exclusive copyright interest in the
above-cited manuscript (subsequently
called the “work”) in this and all
subsequent editions of the work (to
include closures and errata), and in
derivatives, translations, or ancillaries, in
English and in foreign translations, in all
formats and media of expression now
known or later developed, including
electronic, to the American Society of Civil
Engineers subject to the following:
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
version submitted to ASCE after peer
review and prior to copyediting or other
ASCE production activities; it does not
include the copyedited version, the page
proof, a PDF, or full-text HTML of the
published version.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
provided that the name of the author and
the Crown Government affiliation is clearly
indicated; (2) to grant the same rights to
others to print or publish the work; and (3)
to collect royalty fees. ALL AUTHORS
must be official Crown Government
employees in order to claim this
exemption in its entirety. If at least one
author is not a Crown Government
employee, copyright must be transferred
to ASCE by that author.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
Manuscript Click here to
access/download;Manuscript;SFE_Kriging_Manuscript_Rev1_
14 This paper presents a framework for assessing the reliability of fiber-reinforced polymer
15 (FRP) strengthened reinforced concrete (RC) beams in flexure using stochastic nonlinear
16 finite element (SNFE) analysis and k-w-means clustered active learning Kriging Monte Carlo
17 Simulation (AK-MCS), where the spatial variation in the concrete and bond material
19 available non-linear finite element analysis software and automate the process for conducting
20 the SNFE-Clustered AK-MCS analysis. The k-w-means clustering is based on the U learning
22 by allowing parallel computation of the SNFE models. Parametric analysis indicated the
23 accuracy of the reliability prediction of the examined member and proved the efficiency of
24 the proposed analysis in reducing the number of calls to SNFE models compared with
26
27
29 Concrete, Spatial Variability, Stochastic Finite Element, Active Learning Kriging, Stopping
1
31 PRACTICAL APPLICATIONS
32 The quality of FRP-to-concrete bond is affected by the concrete integrity at the interface,
33 which varies across the dimensions of the strengthened member causing added uncertainty in
34 predicting the structural response, and hence, the reliability of the FRP strengthened member.
35 This research proposes a computationally efficient approach to assess the reliability of FRP
36 strengthened concrete members by considering the spatial variation in the concrete properties
37 (compressive strength, tensile strength, bulk modulus) and the quality of the FRP-to-concrete
38 bond (shear and normal bond strength) by using an adaptive machine learning technique. The
39 proposed framework can be utilized by engineers to design FRP strengthening systems for
40 concrete members experiencing variation in the concrete properties due to poor quality
42 INTRODUCTION
44 concrete mechanical properties including compressive strength, tensile strength, and modulus
45 of elasticity (MacGregor et al., 1997; Bartlett and MacGregor, 1999; Bartlett, 2007). The
47 exposure, and accuracy of the strength evaluation technique (i.e., accuracy of property testing
48 procedure). The effect of spatial variability on the structural resistance is either overlooked or
51 spatial variability of material properties through the use of stochastic (or random) nonlinear
52 finite element method (SNFE) exist (Hunter et al., 2021; Oudah and Alhashmi, 2022; Petrie
53 and Oudah, 2023), they are seldom utilized in practice due to efficiency challenges with using
54 SNFE.
2
55 SNFE utilizes random fields to represent the spatial distribution of the constitutive
57 patterns of spatial distributions. SNFE has been extensively used in geotechnical engineering
58 to model soil variability in three-dimensional (3D) analysis. It has been used for shallow
59 foundation (e.g., Soubra et al., 2019; Li et al., 2020), single and group pile reliability analysis
60 under gravity and lateral loads (e.g., Xiao-ling et al., 2021; Khorramian et al., 2023), among
61 other applications. However, the use of SNFE in the structural reliability assessment of
63 needed when the effect of spatial variability on the reliability assessment is significant such
64 as the case when the structural member is exposed to deterioration actions that vary spatially
65 (like corrosion, freeze-thaw damage, and alkali silica reactivity (ASR)), or when the member
66 capacity relies on the concrete cohesive characteristics to achieve a composite action (like
67 externally bonded fiber-reinforced polymer (FRP)). For the latter, digital image correlation
68 (DIC) processing of experimental pullout tests of FRP plates externally bonded to concrete
69 prisms indicated the significant variability in the strain distribution of FRP plates, which is an
70 indicator of the variability in the tensile capacity of the concrete and quality of workmanship
71 (Wieghaus and Atadero, 2011). To address the effect of spatial variability of mechanical
72 properties on the reliability of FRP strengthened RC members, Petrie and Oudah (2023)
73 developed a computer code to perform SNFE analysis using LS-DYNA by considering five
74 correlated random fields: concrete-FRP normal and shear bond strengths, bulk modulus,
75 concrete compressive strength, and concrete tensile strength. The framework is based on
76 building a resistance distribution using the SNFE models followed by using Monte Carlo
77 Simulation (MCS) to perform the reliability calculation for a predefined load model as per
3
79 A major challenge of the developed SNFE-MCS framework proposed by Petrie and Oudah
80 (2023) is the computational cost required to perform the analysis. Approximately 250 SNFE
81 analyses were required to build a resistance distribution per individual structural beam
83 used by practicing engineers to evaluate the reliability of FRP structural designs. Balancing
87 the structural reliability. To improve the efficiency of the reliability assessment (i.e., reduce
88 number of required SNFE models), clustered active learning Kriging (AK) reliability analysis
91 performance function, using a set of design sites, 𝑆(𝑋), and corresponding outputs, 𝑌(𝑋),
92 called together the design of experiment (DoE) (Khorramian and Oudah, 2022). The AK
93 method uses a set of correlation, regression, and learning functions to first train, and then
94 update, the predictor using the DoE through an enrichment process, where the accuracy of the
95 predictor is incrementally improved around the considered limit state (Moustapha et al.,
96 2022; Khorramian and Oudah, 2023). The enrichment process is composed of a learning
97 function and stopping criteria, where efficiency of the AK analysis is sensitive to both
98 (Khorramian and Oudah, 2024), and it can be augmented with a multi-point clustering
99 method to improve efficiency (i.e., evaluate multiple 𝑆(𝑋) and 𝑌(𝑋) per iteration). This
100 method is ideal for evaluating performance functions with high computational cost, such as
102 Current AK research trends are focused on improving the efficiency of the analysis by
103 utilizing parallel computing and multi-point enrichment processes. Wen et al. (2016)
4
104 developed a sequential Kriging reliability analysis to facilitate parallelizability. El Haj and
105 Soubra (2020) developed an efficient framework utilizing k-means for adaptive Kriging
106 assessment of SNFE models. Teixeira et al. (2020) augmented AK analysis with density
107 scanning to improve efficiency. Xing et al. (2020) developed a method to improve the
108 efficiency of parallel computing for AK analysis based on expect improvement (EI) function.
109 Wang et al. (2023) employed k-means clustering and sampling in n-ball for structural
110 assessment of engineering systems. The literature related to parallel computing and clustered
111 AK analysis, including literature that utilized AK analysis for SNFE, utilized theoretical
112 engineering examples (Kaymaz, 2005; Buckley et al., 2021; Godin-Hebert et al., 2024;
113 Echard et al., 2011; Moustapha et al., 2022). Also, the clustered AK analyses developed in
114 literature were not utilized for the reliability estimation of FRP strengthened RC beams
115 modeled using SNFE, despite the enormous studies that indicate the sensitivity of reliability
116 estimation of FRP strengthened structures to the non-uniform bond characteristic at the
117 concrete-FRP interface (e.g., Okeil et al., 2002; Wang et al., 2010; Shi et al., 2015). The AK
118 method has been applied to several frameworks to perform reliability analysis, such as AK-
119 MCS, AK-FORM, AK-IS, and AK-SS (Owen et al., 2019; Xiao et al., 2019; Yang et al.,
121 The objectives of this research are to 1) develop a framework to assess the reliability of
122 FRP strengthened RC beams using SNFE-Clustered AK-MCS; and 2) evaluate the effect of
123 stopping criteria on the number of required SNFE simulations (i.e., efficiency). The paper is
125 SNFE, review of AK, and automation), application of framework to assess a FRP
128 Overview
5
129 The framework of performing SNFE-Clustered AK-MCS to assess the structural reliability of
130 FRP strengthened RC member at ultimate limit state (ULS) is illustrated in Figure 1. The
131 limit state for flexure is presented as 𝐺(𝑋, 𝑌) = 𝑅(𝑋) – 𝑄(𝑌), where 𝐺 is the performance
132 function, R is the resistance model, Q is the load model, 𝑋 is a vector of resistance random
133 variables, and 𝑌 is a vector of load random variables. The framework consists of two main
135 overview of the components is provided herein while detailed information are included in the
137 In SNFE generation, the beam is discretized into two types of meshes: finite element mesh
138 (FEM) and stochastic element mesh (SEM), where the latter is a subset of the former (i.e.,
139 SEM density is equal to or less than FEM density). The concrete volume is discretized into
140 three dimensional (3D) elements, while the FRP is discretized into two-dimensional (2D)
141 elements of the same surface area as the individual elements of the concrete volume. Five
142 correlated random fields are considered to represent the spatial variation of the concrete
143 compressive strength, 𝑓𝑐′ , concrete tensile strength, 𝑓𝑡′ , concrete bulk modulus, K, and the
144 concrete-FRP bond normal strength, NFLS, and shear strength, SFLS, while the steel tensile
145 strength, 𝑓𝑦 , FRP tensile strength, 𝑓𝐹𝑅𝑃𝑢 , and FRP modulus of elasticity, 𝐸𝐹𝑅𝑃 , are considered
146 as random variables. The random fields realizations are generated at the centroids of the
147 stochastic elements and are assigned to all finite elements contained within each stochastic
149 analysis by Petrie and Oudah (2023) (i.e., each 3D stochastic element contains 2x2x2 = 8
150 finite elements). An automated algorithm was developed in commercially available coding
151 language to generate and run SNFE using the nonlinear software LS-DYNA, where the
152 nonlinearity in the concrete material, steel reinforcement, and FRP are considered as
153 discussed in the SNFE model development section (Livemore 2007a, MATLAB 2021b).
6
154 In the AK-MCS analysis, 𝑛 realizations of the random variables 𝑋 and 𝑌 are generated. A
155 select number of realizations are chosen to formulate the initial Kriging predictor, that
156 generate SNFE models, each with a different set of random fields and variables, to evaluate
157 their nominal flexural strength, assess the performance function, and train a Kriging predictor.
158 The Kriging predictor is then used to predict surrogate performance function 𝐺̂ (𝑋, 𝑌) for all
159 𝑛 realizations where the probability of failure 𝑃𝑓0 and coefficient of variation, 𝐶𝑂𝑉𝑃𝑓0 , can be
160 calculated using Eq. (1) and Eq. (2), respectively. Following that, the clustered enrichment
161 process commences by determining a set of next candidates to train the Kriging predictor
162 using a learning function and k-w-means clustering (Lelièvre et al., 2018). The enrichment
163 process terminates when the stopping criteria is met, where multiple stopping criteria can be
164 considered to improve efficiency (see stopping criteria section). The accuracy (i.e. margin of
165 error in 𝑃𝑓𝐺̂ prediction) and efficiency (required number of training points) of the AK process
166 are sensitive to the utilized learning function and the stopping criteria (Khorramian and
167 Oudah, 2024). Several learning functions and associated stopping criteria are formulated in
168 literature including U (Echard et al., 2011), H (Lv et al., 2015), LIF (Sun et al., 2017), LF
169 (Chai et al., 2019), KO (Khorramian and Oudah, 2023), etc. While the formulations of the
170 learning functions differ based on the fundamental approach taken to assess the learning
171 process, they all seek to find the next candidate for enriching the DoE. The U learning
172 function has been utilized in several research studies for SNFE analysis, and thus, is utilized
173 in this research (Khorramian et al., 2023). Two stopping criteria are considered for this
174 framework: minimum value of U function based (Echard et al., 2011), and probability based
1− 𝑃𝑓0
177 𝐶𝑂𝑉𝑃𝑓0 = √ 𝑛 𝑃 (2)
𝑓0
7
178 where, 𝑛 is the number of trials included in the MCS, and 𝐼[𝐺̂ (𝑋, 𝑌)𝑗 ] is the 𝑗 𝑡ℎ value of the
179 indicator function denoting the sign of the Kriging estimated performance function.
180 Multi-point clustering using k-w-means is utilized to facilitate parallelization of the SNEF
181 analysis to improve efficiency. It implies that multiple SNFE analyses can be performed in
182 parallel to improve the convergence of the analysis. This is especially useful when high
183 computing power and licensing arrangements (LS-DYNA license with multiple concurrent
186 The expansion optimal linear estimation (EOLE) method was used to discretize the random
187 fields (Li and Der-Kiureghian, 1993; Sudret and Der-Kiureghian, 2000). The EOLE method
188 represents the random field as a linear function with nodal random variables and a series of
189 shape functions. The shape functions are found by minimizing the error variance.
190 Discretization was required to generate realizations of the field at the centroids of the
191 stochastic mesh element, for a given stochastic mesh density. In other words, all finite
192 elements composing a stochastic element are assigned a uniform realization based on the
193 random field discretization process, where a stochastic element is composed of one or more
196 ̂ (𝑌, 𝜃), as expressed in Eq. (3). The field realizations are
of a gaussian random field, 𝐻
197 ̂ (𝑌, 𝜃) as
obtained at the geometric centroid of all stochastic elements. Field generation of 𝐻
198 expressed in Eq. (4) requires five user-input parameters: mean, μ𝑦 , standard deviation, σ𝑦 ,
199 correlation function, correlation lengths in 3D space (a𝑥 , a𝑦 , a𝑧 ) and number of standard
201 𝐻̂ ̂
𝑙𝑛 (𝑌, 𝜃) = exp{𝐻 (𝑌, θ)} (3)
8
202 ̂ (𝑌, θ) = μ𝑙𝑛𝑦 + σ𝑙𝑛𝑦 ∑𝑟𝑖=1 ξ𝒊 (θ) 𝛙𝑇𝑖 𝑪𝑌,𝑌
𝐻 (4)
a 𝑖 √ 𝑖
203 where, 𝑪𝑌,𝑌𝑖 is the 𝑖 𝑡ℎ vector of the covariance matrix, 𝛙𝑖 is the 𝑖 𝑡ℎ greatest eigenvector of
204 the standard normal field (𝑖 𝑡ℎ out of 𝑟), 𝜉𝒊 (𝜃) is the 𝑖 𝑡ℎ randomly generated standard normal
205 variable (𝑖 𝑡ℎ out of 𝑟), a𝑖 𝑖 𝑡ℎ greatest eigenvalue of the standard normal field (𝑖 𝑡ℎ out of 𝑟).
206 The lognormal standard deviation, σ𝑙𝑛𝑦 , and mean, μ𝑙𝑛𝑦 , are obtained using Eq. (5) and (6),
207 respectively. The covariance matrix 𝑪𝒀𝒀 (Eq. (7)) is established based on a squared
208 exponential correlation function to determine the correlation between two points 𝑌 (𝑖) and 𝑌 (𝑗) ,
209 𝜌𝑖𝑗 , (Eq. (8)), and the coordinates of the geometric centroids of the SNFE (𝑦𝑥 , 𝑦𝑦 , 𝑦𝑧 ). Nataf
210 transformation is used to convert the correlation values of a lognormally distributed random
211 field to a standard normal distribution (Eq. (9)) (Nataf, 1962). The number of standard
212 normal variables r is determined such that the ratio of the eigenvalues divided by the trace of
213 the covariance matrix, Q, does not exceed 0.8, as described in Eq. (10).
σ2
214 σ𝑙𝑛𝑦 = √ln (1 + μ2𝑦 ) (5)
𝑦
1
215 μ𝑙𝑛𝑦 = ln(μ𝑦 ) − 2 σ2𝑙𝑛𝑦 (6)
ln(1+ ρij v𝑖 v𝑗 )
218 ρ′𝑖𝑗 = 𝑠ρ𝑖𝑗 = (9)
ρ𝑖𝑗 √ln(1+v2𝑖 ) ln(1+v2𝑗 )
∑𝑟𝑖=1 𝑪𝑌,𝑌
219 𝑄= 𝑖
(10)
𝑡𝑟(𝑪𝑌𝑌 )
220 where, ρ′𝑖𝑗 is the correlation for the standard normal field between two points 𝑌 (𝑖) and 𝑌 (𝑗) ,
221 and 𝑠 is a conversion factor converting the lognormal field. The terms v𝑖 and v𝑗 are the
222 coefficient of variation of the mesh points 𝑌 (𝑖) and 𝑌 (𝑗) , respectively.
9
223
225 Kriging
226 Kriging, also called the optimal estimator, is a non-linear stochastic regression method in
227 which a surrogate metamodel (referred to as the kriging predictor herein) is composed of two
228 components: a regression component, and a random process component. The kriging
229 predictor, 𝐺̂ (𝑋), can be trained to evaluate the desired limit state using the two components to
230 form it, as expressed in Eq. (11) (note that 𝑌 notation in Fig. 1 is dropped for clarity of
233 where, ℱ(𝑏, 𝑋) is the regression function of the random inputs, 𝑋, and unknown coefficient 𝑏,
235 By setting the mean of the predictor to zero, and minimizing the mean squared error (MSE)
236 or the variance of the squared error, Eq. (12) to (17) are used to estimate the mean and
237 variance of the Kriging predictor, respectively (Kaymaz 2005; Khorramian and Oudah 2022):
239 σ2𝐺̂ (𝑋) = σ2 . (1 + 𝒖𝑇 (𝑭𝑇 𝑹−1 𝑭)−1 𝒖 − 𝒓(𝑋)𝑇 𝑹−1 𝒓(𝑋)) (13)
244 Where the mean value of the Kriging predictor, μ𝐺̂ (𝑋), is considered to be the estimate of
245 the limit state, 𝐺̂ (𝑋), as shown in Eq. (12), and the standard deviation of the field, σ2𝐺̂ (𝑋),
246 also referred to as the mean square error (MSE), is given by Eq. (13). The terms 𝑭 and 𝒇(𝑿)
10
247 are the regression function, and the associated term 𝒃∗ is the regression parameter for chosen
248 regression function. The term 𝒖 is a vector that is formulated for computational ease as it
250 The correlation, 𝒓 and 𝑹, given by Eq. (18a) and Eq. (18b), respectively, where the term 𝑳
251 is the vector of autocorrelation length (called ‘correlation length’ herein). The term r(X) is
252 the correlation between the unknown site, 𝑿, with all known design sites 𝑺𝒊 . The term 𝑹(𝑿)
253 is the correlation between the design sites 𝑺𝒊 and 𝑺𝒋 . The vector 𝑳 is calculated by minimizing
254 the term in Eq. (19), in an iterative process, where 𝑚 is the number of design sites (i.e., 𝑆 =
255 [𝑆1 , … , 𝑆𝑚 ]), and 𝑘 represents the 𝑘th dimension (i.e., 𝑘 = [1,2,3] in a 3D problem).
257 𝑹 = 𝑹(𝑳, 𝑺𝒊 , 𝑺𝒋 ) = ∏𝑛𝑘=1 𝑅𝑘 (𝑙𝑘 , 𝑎𝑖𝑘 , 𝑏𝑗𝑘 ) ; 𝑺𝒊 = [𝑎𝑖1 … 𝑎𝑖𝑛 ]; 𝑺𝒋 = [𝑏𝑗1 … 𝑏𝑗𝑛 ]; (18b)
1
258 𝐿∗ = min {|𝑹| 𝜎𝑘2 } ; 𝑘 = 1, … , 𝑛
𝑚 (19)
𝐿
259 In this work, a linear regression function and a spline correlation function have been chosen
260 for analysis based on a preliminary analysis. Table 1 shows the equations for both functions.
261 Note that for a linear regression function the minimum initial points in the DoE are
263 problem (Khorramian and Oudah 2022). In the present work, an initial DoE size of 𝑡 + 6 is
264 chosen. DACE MATLAB Toolbox is used to form the Kriging predictor (Lophaven et al.
265 2002a and 2002b). The formulation of the learning function and stopping criteria are
268 A learning function is required to determine the next design sites for enrichment of the
269 Kriging predictor as described in the Overview section. The use of a learning function allows
270 the user to grow the size of the DoE in a stepwise manner to find the most efficient solution
271 while maintaining a sufficient level of accuracy (Buckley et al 2020). The commonly used U-
11
272 learning function has been selected in this work to demonstrate the clustered-SNFE
273 framework, as it allows clustering of the U-learning data to generate multiple enrichment
274 sites per iteration. Shown in Eq. (20), the learning function works by calculating the
275 reliability index of the prediction, 𝑈(𝑿) , through definition of a linear equation that
276 determines the number of standard deviations, 𝜎𝐺̂ (𝑿) , between the estimated Kriging
277 prediction value, 𝐺̂ (𝑿), and the limit state (𝐺̂ (𝑿) = 0), for all design sites considered in the
μ ̂ (𝑋)
279 𝑈(𝑋) = σ𝐺 (𝑋) (20)
̂
𝐺
280 Once the Kriging predictor has calculated the μ𝐺̂ (𝑋) and σ2𝐺̂ (𝑋) vectors of the MCS trials, Eq.
283 As discussed in the Overview section, a stopping criterion is required for the learning
284 function to determine when to cease the loop for the enrichment of the Kriging predictor.
285 Two stopping criteria are considered in this analysis. The first is the criteria proposed in
286 Echard et al. (2011) where the U-value should be equal or greater than 2.0 for all realizations
287 predicted using the Kriging surrogate model. The limit of 2.0 on the U-value implies that the
288 probability of making a mistake on the sign of 𝜙(−2) corresponds to 0.023 (i.e., desired
289 reliability level of avoiding a mistake). The limit ensures all points found by the kriging
290 predictor have low MSE of the kriging predictor. The use of the U-value stopping criteria has
291 been shown by El Haj and Soubra (2020) to be conservative in the number of added points
292 needed for enrichment of the Kriging predictor. An alternative stopping criterion to the U-
293 value stopping criteria has been implemented based on the work of Schöbi et al. (2017). The
294 Schöbi stopping criteria uses the probability of failure as shown in Eq. (21), and calculates
12
𝑘 𝑘
𝑃𝑓+ − 𝑃𝑓−
297 ε𝑃𝑓 = ≤ ε𝑠𝑡𝑜𝑝 (21)
𝑃𝑓0
299 𝑘
𝑃𝑓+ = 𝑃 [(𝐺̂ (𝑋) + 𝑘. 𝜎𝐺̂ (𝑥)) < 0] (23)
300 𝑘
𝑃𝑓− = 𝑃 [(𝐺̂ (𝑋) − 𝑘. 𝜎𝐺̂ (𝑥)) < 0] (24)
301 where 𝜀𝑃𝑓 is the calculated estimator error of the predator and ε𝑠𝑡𝑜𝑝 is the stopping criteria for
302 ending enrichment of the Kriging predictor using active learning. The term 𝑃𝑓0 is the
303 probability of failure calculated using the Kriging predictor to evaluate the limit state. The
𝑘 𝑘
304 terms 𝑃𝑓+ and 𝑃𝑓− are also the probability of failure, calculated by taking the mean Kriging
305 predictor value, 𝐺̂ (𝑋), and adding or subtracting, respectively, the MSE, σ𝐺̂ (𝑥), multiplied
306 by a constant, 𝑘, to represent the number of standard deviations away from the mean. The
307 choice of 𝑘 will determine how quickly the Schöbi stopping criteria converges to the selected
310 A clustering method has been implemented to improve the efficiency of the AK method to
311 meet the stopping criteria for learning, called k-w-means clustering (Lelièvre et al. 2018).
312 When single-point enrichment is conducted, the trial with the lowest U-value identifies the
313 next candidate to be added as a design site. By choosing the trial with the lowest U-value
314 ensures that the design site with the most uncertainty can be evaluated by SNFE. k-w-means
315 clustering instead takes the set of trials with the lowest estimated U-value and breaks them to
316 form several clusters. The number of clusters chosen will determine the number of added
317 points per iteration of AK enrichment. The method makes use of the information gained by
318 evaluating the U Leaning Function by calculating geometric means of each cluster using the
319 U-values of each trial as weights, to propose a set of enrichment points. The chosen
13
320 enrichment points may not be the lowest U-value trials but are within the lowest and should
321 be sufficient to yield convergence. By iterating through this procedure, an optimal set of
322 design sites can be obtained. A summary of the steps in the selection of the next cluster of
324 Step 1. Let the term 𝐾𝑚𝑒𝑎𝑛 be the number of clusters to be considered, and the term 𝑁𝑐𝑙𝑢𝑠𝑡𝑒𝑟
326 Step 2. Calculate the U-values of the learning function based on the current Kriging predictor
327 using Eq. (20). Sort the trials in a vector, 𝑼𝑴𝒊𝒏 , in descending order of U-value.
328 Step 3. Select a total of 𝑃 (𝑃 = 𝐾𝑚𝑒𝑎𝑛 𝑥 𝑁𝑐𝑙𝑢𝑠𝑡𝑒𝑟 ) lowest ranked U-values from 𝑼𝑴𝒊𝒏 and
329 randomly select 𝐾𝑚𝑒𝑎𝑛 of the 𝑃 selected trials to be the initial cluster centroids
331 Step 4. Form a Voronoi diagram of the 𝑃 number of trials selected by assigning all trials not
(1)
332 designated as a centroid in Step 3 to belong to the 𝑖 𝑡ℎ cluster centroid, 𝑐𝑖 , that
335 (25), which applies a weight based on the U-value of the given value of 𝑋.
𝑁𝑘 ( 1
∑𝑘=1 ) 𝑋𝑘
(𝑗+1) 𝑈𝑘
336 𝑐𝑖 = 𝑁𝑘 1
; 𝑖 = [1, 2, … , 𝐾𝑀𝑒𝑎𝑛 ] (25)
∑𝑘=1( )
𝑈𝑘
337 where the terms 𝑗 and 𝑁𝑘 represent the iteration number and number of points within
338 the 𝑖 𝑡ℎ cluster, respectively. The terms 𝑈𝑘 represent the U-value of the 𝑘 𝑡ℎ trial, and
(𝑗)
340 Step 6. Calculate the error between the previous and current set of cluster centroids, 𝑐𝑖 and
(𝑗+1)
341 𝑐𝑖 , respectively, using Eq. (26):
14
343 Step 7. Compare Ω𝑐𝑙𝑢𝑠𝑡𝑒𝑟 found in Step 6 to the stopping criteria, Ω𝑠𝑡𝑜𝑝 . If the stopping
344 criteria is not met (Ω𝑐𝑙𝑢𝑠𝑡𝑒𝑟 > Ω𝑠𝑡𝑜𝑝 ), repeat Steps 2 to 6 until criteria is met. If the
346 Step 8. Calculate the points that form the smallest Euclidean distance with each established
347 centroid and select it for updating the Kriging predictor DoE. This is needed because
348 the algorithm will produce centroids that are not sample points in the design space.
349 In this work, a k-w-means clustering algorithm considering 4 clusters with 5 points (i.e.,
350 𝐾𝑚𝑒𝑎𝑛 = 4, and 𝑁𝑐𝑙𝑢𝑠𝑡𝑒𝑟 = 5) is used. A stopping criterion is set as 2% for this study (Ωstop =
351 0.02).
353 An automated computer code has been developed to conduct the SNFE-Clustered AK-MCS
354 reliability analysis. The code establishes an interface between LS-DYNA and the developed
355 random field, and AK-MCS algorithms to perform the analysis. The user is required to build
356 what is called the “base FE” model using LS-DYNA. Developing the base model includes,
357 building the geometry, discretizing it into FEM, defining material models, apply boundary
358 conditions, and defining loads. The user then submits the base model to the user developed
359 SNFE-Clustered AK-MCS algorithm. The code functions as follows. The algorithm
360 generates realizations of the random fields and variables based on a predefined statistical
361 parameters and calls to read the LS-DYNA base FE model, discretize the FEM into SEM,
362 and assign the generated random variables and fields to the LS-DYNA model to form SNFE
363 (multiple models can be generated as required). Once all required SNFE models are
364 generated, the algorithm calls LS-DYNA to evaluate the ultimate strength of the model in a
365 displacement or load-controlled mode, using an LS-DYNA SMP double precision implicit
366 solver. Once SNFE simulation is complete, the algorithm is called to extract the load-
367 deflection response of the simulations and determine the nominal flexural strength required to
15
368 evaluate the performance function and perform clustered AK analysis using the algorithms
372 The SNFE-Clustered AK-MCS analysis procedure was applied to assess the reliability of an
373 experimentally tested FRP strengthened RC beam subjected to four-point bending at ULS
374 (Zhang et al. 2006). The beam geometry and reinforcement details are shown in Fig. 2. The
375 beam length, depth and width are 2500 mm, 250 mm, and 120 mm, respectively. The flexural
376 reinforcement consists of 2-12 mm rebars at bottom, and 2-10 mm rebars at top, while 8 mm
377 stirrups equally spaced at 100 mm are used to resist the shear force. Dual layer carbon FRP
378 laminates of 0.22 mm in thickness and 120 mm in width are externally bonded to the concrete
379 beam for flexural strengthening. The beam span length is 2500 mm while the CFRP length is
380 1800 mm centered at mid-span. The steel yield strength, 𝑓𝑦 , is 335 MPa, while the concrete
′
381 compressive strength at 28 days, 𝑓𝑐 , is 25 MPa. The specified CFRP tensile strength, 𝑓𝑓𝑟𝑝𝑢 ,
382 and CFRP modulus of elasticity, 𝐸𝑓𝑟𝑝 , are 3350 MPa and 235 GPa, respectively. The beam
383 factored moment of resistance is 21.5 kNm and the failure mode is tension-controlled per
386 The performance function corresponding to ULS, resistance model, 𝑅(𝑋), and load model,
391 𝑅(𝑋) is the product of the nominal flexural strength of SNFE, 𝑅𝑆𝐹𝐸 (𝑋), and a model error
392 𝜖𝑆𝐹𝐸 . Five correlated random fields, including 𝑓𝑐′ , 𝑓𝑡′ , 𝐾, 𝑁𝐹𝐿𝑆, and 𝑆𝐹𝐿𝑆, and three random
16
393 variables including 𝑓𝑦 , 𝑓𝑓𝑟𝑝𝑢 , and 𝐸𝑓𝑟𝑝 were considered in developing the SNFE simulations,
394 and hence, in 𝑅𝑆𝐹𝐸 (𝑋). The five random fields are fully correlated to 𝑓𝑐′ random field, which
395 implies that the number of standard normal variables in EOLE generation and training the
397 𝑄(𝑌) includes three random variables, moment due to dead load, 𝑀𝐷, moment due to live
398 load, 𝑀𝐿, and transformation of live load to load effects, 𝑀𝑇. The total number of random
399 variables is 11 (4 for random field generation + 3 for material random variables + 1 for 𝜖𝑆𝐹𝐸
400 + 3 for 𝑄(𝑌)). The statistical parameters (bias, distribution type, mean, and coefficient of
401 variation (COV)) of 𝑅(𝑋) and 𝑄(𝑌) are included in Table 2, where values for 𝑅(𝑋) are based
402 on relevant literature and values for 𝑄(𝑌) are based on values used in the calibration of the
403 National Building Code of Canada (NBCC 2020). The mean values for 𝑅(𝑋) are based on the
404 experimental values reported by Zhang et al (2006), while the mean values for 𝑄(𝑌)
405 correspond to load effects that produce a fully utilized section (demand-to-capacity ratio of
406 unity). This procedure was adapted to examine the reliability of the beam when it is moment
407 of resistance is fully utilized per NBCC 2020 load combination (1.25MD + 1.5ML) for
408 MD/ML ratio of 1. Similar approaches for reliability assessment have been used in literature
411 The Karagozian & Case (K&C) concrete model in LS DYNA (Livermore, 2007b) was used
412 to model the concrete response (*MAT 072R3 – Concrete Damage REL3). The Poisson’s
413 ratio was set to 0.2, while the concrete density was set to 2350 kg/m3. The FEM and SEM
414 were generated as per the framework described in Figure 1. 𝑓𝑐′ random field was generated
415 using EOLE method as described in the Discretization of the Random Fields using EOLE
416 section. 𝑓𝑡′ was generated using Eq. (30) as per CEB-FIP Model Code 1990, while K was
417 calculated using the equation of state (i.e., the piecewise description of the relationship
17
418 between pressure and volume strain). Both 𝑓𝑡′ and K were considered as correlated random
2 1/3
𝑓𝑐′
420 𝑓𝑡′ = 1.58 (145) (30)
421 A plastic kinematic material model was used to model the reinforcing steel with a
422 Poisson’s ratio, tangent modulus, and density of 0.2, 463 MPa, and 7700 kg/m3, respectively
423 (*MAT 003 – Plastic Kinematic). The CFRP plate was modeled using a damage-based
424 material model in LS-DYNA (*MAT 054 – Enhanced Composite Damage), with a density of
426 The concrete was modeled using 4 node elements with full integration, while the
427 reinforcing steel was modeled using beam elements with translational degrees of freedom
428 (dof) constrained by the concrete element dof (i.e., perfect bond). The FRP sheets were
429 modeled using shell elements with full integration. An automatic tie-break contact was used
430 to model the FRP-concrete interface bond, where the failure criteria of the bond is expressed
431 in Eq. (31). 𝜎𝑛 describes the normal surface stress, 𝜎𝑠 is the shear surface stress, 𝑁𝐹𝐿𝑆 is the
432 normal failure limit stress and 𝑆𝐹𝐿𝑆 is the shear failure limit stress. The NFLS and SFLS
433 random fields were considered as fully correlated to 𝑓𝑐′ random field realizations at the beam
434 soffit and were defined using the expression derived by Neale et al. (2005), as expressed in
435 Eq. (32) to (34). The expressions were validated in literature using extensive FE analysis
436 (Almusallam et al. 2015; Elsanadedy et al. 2013). The term β𝑤 is a shape factor that depends
437 on the ratio of 𝑏𝑓 / 𝑏𝑐 . Where, 𝑏𝑓 is the width of the CFRP sheet, and 𝑏𝑐 is the width of the
|σ | 2
𝑛 𝑠 |σ | 2
439 (𝑁𝐹𝐿𝑆 ) + (𝑆𝐹𝐿𝑆 ) ≥ 1.0 (31)
18
441 𝑆𝐹𝐿𝑆 = 1.5𝛽𝑤 𝑁𝐹𝐿𝑆 (33)
𝑏𝑓
2.25−
𝑏𝑐
442 β𝑤 = √ 𝑏𝑓 (34)
1.25+
𝑏𝑐
443 A base NFE model (i.e., does not include random fields) was first developed and
444 calibrated using experimental test data reported in Zhang et al. (2005) (experimentally tested
445 Beam A20). The load displacement curve for the model was validated by Petrie and Oudah
446 (2023). The loads at yield and ultimate of the model are 1.03 and 1.06 times the loads of the
447 experimental test, respectively, while the displacement at yield and ultimate of the model are
448 0.70 and 0.93 times the displacement of the experimental test, respectively. For the SNFE
449 simulations, the SEM-to-FEM ratio was set to 2.0 and the correlation length for EOLE
450 random field generation was set to 600 as recommended in Petrie and Oudah (2023) for FRP
453 Automated SNFE-Clustered AK-MCS analysis was performed using the developed computer
454 code. A total of 5x105 MCS trials were generated, where 17 DoE were used to form the initial
455 Kriging predictor and four new points (design site and response) are added to the DoE in each
456 iteration of the enrichment process by using k-w-means clustering. Fig. 3 illustrates the
457 capacity evaluation of select beams developed in the first iteration of the enrichment process,
458 where visualization of 𝑓𝑐′ random field, plots of the maximum principal strains at nominal
459 flexural strength, 𝑀𝑛 , and plots of the extracted moment versus mid-span deflection are
460 shown. Strain plots indicate spatial variability in the material response at ultimate capacity as
461 evident in the non-uniform strain in FRP and non-symmetric crack patterns along the beam
463 determine 𝑀𝑛 , where the latter is utilized to update the Kriging predictor for improved
19
465 A sensitivity analysis on the U-value and Schöbi stopping criteria with varying scale
466 factors, 𝑘, was conducted. A summary of the considered cases for the sensitivity analysis of
467 the stopping criteria and analysis results are shown in Table 3. Results include the number of
468 added points during the clustered AK enrichment process, total number of calls of SNFE
469 simulation, reliability index, β, of the strengthened beam calculated once the stopping criteria
470 is reached, and duration of analysis in days. The calculated reliability index ranges from 3.68
471 to 3.72, which falls within the typical range of values expected for a code compliance design
472 of FRP reinforced members (Hassan and Oudah, 2022). All runs were processed on a
473 windows-based server with dual 2.20GHz processors (48 cores), a standard graphics card,
474 and 128 Gb of available RAM. The SNFE-Clustered AK-MCS analysis was run until all
475 stopping criteria were met. Figure 4 shows a plot of the three probabilities of failures, 𝑃𝑓0 ,
𝑘 𝑘
476 𝑃𝑓+ , 𝑃𝑓− for the various values of 𝑘 considered. The solid line represents the Kriging
𝑘 𝑘
477 predictor probability of failure, 𝑃𝑓0 and the dashed lines each represent a value of 𝑃𝑓+ , 𝑃𝑓− .
478 A comparison of the Schöbi and U Learning Function stopping criteria for the AK
479 analysis shows a difference in the number of added points to meet convergence for the beam
480 considered. The U-value criteria took the most added points to converge while the Schöbi
481 stopping criteria varied based on the constant 𝑘 used in Eq. (21) (𝑘 is the number of standard
482 deviations away from 𝑃𝑓 that the upper and lower error bounds are calculated at). As
483 expected, the lower value of 𝑘 in Eq. (23) and Eq. (24) allowed the stopping criteria to be
484 met for the Schöbi equation in fewer iterations of enrichment of the DoE. The number of
485 added points slightly changed the calculated reliability of the strengthened beam, in the range
486 of 3.6 to 3.8. Results show an agreeance between the 𝑈, 𝑃𝑓1 , and 𝑃𝑓2 stopping criteria, while
487 the 𝑃𝑓0.5 showed a higher reliability index value. The probability-based stopping criteria of 𝑃𝑓1 ,
488 and 𝑃𝑓2 yielded 70% and 40% less simulation time as compared with U-value stopping
489 criteria.
20
490 The k-w-means clustering technique allowed the parallelization of the enrichment of the
491 DoE through the selection of multiple points for each iteration of training the Kriging
492 predictor. The added cluster of points in most cases did not coincide with the lowest U-value
493 trial but still offered improved enrichment. This is due to having multiple points, and the fact
494 that the cluster evaluates a set of trials that are still within the lowest ranked U-value of the
496 Clustering improves the efficiency of the AK method most during the initial iterations and
497 was shown to reduce in effectiveness in later iterations. This reduction in improvement in
498 learning is due to the next candidates selected being of high certainty (high U-value of
499 candidate), meaning the improvement in the Kriging predictor decreases after this point is
500 reached. As the number of added points reaches above 160 points, fewer than 5 trials have a
501 U-value lower than 2 (stopping criteria for U). This increase in average U-value means that as
502 the enrichment iterations increase to meet the U stopping criteria, the points taken into the
503 clustering are of high U-value (𝑈 >> 2) and may not yield an improvement in the Kriging
504 predictor. It is recommended that a switch from multi-point clustered enrichment to single-
505 point enrichment of the lowest U-value trial is made once a certain threshold of average U-
506 value is reached to ensure efficient convergence. While the SNFE-Clustered AK-MCS
507 analysis showed promising results in terms of simulation efficiency, the accuracy of AK
508 simulation should be interpreted considering the sensitivity of the following four parameters:
509 - AK configuration. It refers to the choice of the correlation, regression, and learning
510 functions (Buckley et al., 2021). Extensive parametric analysis indicates the variability in
511 the analysis solution based on the input AK configuration. Although recommendations of
512 suitable AK configurations along with specialized metrics for a diverse class of reliability
513 problems have been made in literature and used in this research, these configurations are
21
515 - Initial training set. The Kriging predictor is generally sensitive to the initial set of
516 training design sites (17 in this study). Performing the analysis for at least two different
518 - Reliability estimation technique. MCS was utilized in literature, while other estimation
519 techniques like importance sampling (IS) and first order reliability method (FORM) may
520 yield marginally different results and may perform the best with a different set of AK
521 configurations.
522 - Number of random variables. The number of standard normal variables, r, required to
523 generate the random field using EOLE (Eq. (4)) is a function of the desired accuracy and
524 the correlation length (i.e., the larger the correlation length, the higher the number of r for
525 a given desired accuracy of the random field). The use of AK is challenged when a high
526 number of r is needed to represent the random field since the number of calls to SNFE
527 analysis and the required computational memory will be high. A balance between
528 including enough standard normal variables for representation quality of the random field
529 and keeping the number of variables in the DoE low must be achieved.
531 This study is the first to establish a framework for reliability analysis using clustered active
532 learning Kriging (AK) for fiber-reinforced polymer (FRP) strengthened RC members, where
533 the spatial variation in the bond strength (normal and shear) and mechanical properties of the
534 concrete (compressive strength, tensile strength, and bulk modulus) are considered as random
535 fields. The random fields were generated using expansion optimal linear estimation (EOLE)
536 and utilized to discretize stochastic nonlinear finite element (SNFE) models. The framework
537 utilizes k-w-means clustered learning with U learning function to improve the efficiency of
538 the enrichment process of the design of experiment (DoE) used for training the Kriging
539 predictor. Sensitivity analysis concerning the AK stopping criteria was conducted to assess
22
540 the efficiency of the reliability method, where efficiency was gauged by the number of
541 simulations required for convergence. The analysis considered several stopping criteria (U-
542 value and three variants of the Schöbi criteria). The following conclusions were derived:
543 The SNFE-Clustered AK-MCS framework was validated, and it yielded reduced number
545 The implemented k-w-means clustering allowed efficient parallelization of the SNFE
547 The choice of stopping criteria directly affected the convergence of solution. It was found
548 that the Schöbi stopping criteria with a constant, k = 1.0, is sufficient as a stopping
549 criterion to allow efficient convergence while maintaining accuracy relative to a more
551 Research is needed into the choice of AK configuration for the assessment of FRP
552 strengthened beams, and into calibration of the inputs for generating a representative
554 The practical use of SNFE based reliability analysis to assess the safety of reinforced
555 concrete members strengthened externally in flexure with FRP, where the nominal flexural
556 strength is sensitive to the spatial variation of the mechanical properties, is hindered due to
557 the significant computational cost of performing the analysis. Optimization of the analysis
558 will aid in making them more viable to be run by practicing engineers and code developers
559 when performing reliability analysis. By facilitating the incorporation of real-world data,
560 these frameworks enhance the efficiency of reliability analyses and aim to provide valuable
23
562 DATA AVAILABILITY STATEMENT
563 Some or all data, models, or code generated or used during the study are proprietary or
564 confidential in nature and may only be provided with restrictions. The clustering-based
565 Kriging computer code and the random field discretization computer code are restricted.
566 ACKNOWLEDGMENTS
567 The authors wish to acknowledge the financial contributions of Dalhousie University, the
568 Mathematics of Information Technology and Complex Systems (MITACS) program, and
570 REFERENCES
571 ACI (American Concrete Institute). (2023). “Guide for the design and construction of
572 externally bonded FRP systems for strengthening concrete structures”. ACI PRC-
574 Almusallam, T. H., Elsanadedy, H. M., and Al-Salloum, Y. A. 2015. “Effect of longitudinal
575 steel ratio on behavior of RC beams strengthened with FRP composites: Experimental
579 Bartlett, F.M. 2007. “Canadian standards association standard A23.3-04 resistance factor for
581 Bartlett, F. M., and MacGregor, J. G. 1999. “Variation of in-place concrete strength in
583 Bartlett, F. M., Hong, H. P., and Zhou, W., 2003. “Load factor calibration for the proposed
584 2005 edition of the National Building Code of Canada: Statistics of loads and load
586 Buckley, E., Khorramian, K., and Oudah, F. 2021. “Application of adaptive Kriging method
24
587 in bridge reliability analysis.” Canadian Society of Civil Engineering Annual
589 Castaldo, P., Gino, D., and Mancini, G. 2019. “Safety formats for non-linear finite element
592 Chai, X., Sun, Z., Wang, J., Zhang, Y., and Z. Yu. 2019. “A new Kriging-based learning
593 function for reliability analysis and its application to fatigue crack reliability.” IEEE
595 CEB-FIP Model Code 1990: Design Code. 1993. London :T. Telford.
596 Canadian Standrads Assocciation (CSA). 2019. “Canadian Highway Bridge Design Code”.
598 Canadian Standrads Assocciation (CSA). Re-approved 2017. “Design, and construction of
600 ON.
601 Echard, B., Gayton, N., and Lemaire, M. 2011. “AK-MCS: An active learning reliability
602 method combining Kriging and Monte Carlo Simulation.” Structural Safety. 33(2):
603 145–154.
604 El Haj, A. K., and Soubra, A. H. 2020. “Efficient estimation of the failure probability of a
607 Elsanadedy, H. M., Almusallam, T. H., Alsayed, S. H., and Al-Salloum, Y. A. 2013.
610 Godin-Hebert, E., Khorramian, K., and Oudah, F. (2024). “Recommendations for active-
611 learning Kriging reliability analysis of bridge structures.” ASCE Journal of Bridge
25
612 Engineering. (under review)
613 Hassan, A., and Oudah, F. (2022). “Reliability of compression-controlled FRP RC flexural
614 members designed using North American codes and standards: Comparison and FRP
617 Huang, X., Sui, L., Xing, F., Zhou, Y., and Wu, Y. 2019. “Reliability assessment for flexural
620 Hunter, M.D., Ferche, A.C., and Vecchio, F.J. 2021. “Stochastic finite element analysis of
624 Khorramian, K., Alhashmi, A.E, and Oudah, F. 2023. “Optimized active learning Kriging
625 reliability based assessment of laterally loaded pile groups modeled using random
627 Khorramian, K. and Oudah, F., 2022. “Active Learning Kriging-Based Reliability for
628 Assessing the Safety of Structures: Theory and Application.” In: Leveraging Artificial
629 intelligence into Engineering, Management, and Safety of Infrastructure. s.l.:Taylor &
631 Khorramian, K. and Oudah, F. 2023. “New learning functions for active learning Kriging
634 Khorramian, K. and Oudah, F. 2024. “Metric systems for performance evaluation of active
635 learning Kriging configurations for reliability analysis.” ASCE-ASME Journal of Risk
26
637 Lelièvre, N., Beaurepaire, P., Mattrand, C., and Gayton, N. 2018. “AK-MCSi: A Kriging-
638 based method to deal with small failure probabilities and time-consuming models.”
640 Li, C., and Der-Kiureghian, A. 1993. “Optimal discretization of random fields.” ASCE-ASME
642 Li, Y., Liu, K., Zhang, B., and Xu, N. 2020. “Reliability of shape factors for bearing capacity
645 Livermore Software Technology Corporation (LSTC). 2007. “LS-DYNA Keyword User
646 Manual – Theory Manual), V3. Version 971”. LSTC 2007a, Livermore, California.
647 Livermore Software Technology Corporation (LSTC). 2007. “LS-DYNA Keyword User
650 Lophaven, S. N., Nielsen, H. B. and Søndergaard, J. 2002. “A Matlab Kriging toolbox, s.l.”
651 Technical University of Denmark, Kongens Lyngby, Technical Report No. IMM-TR-
652 2002-12.
653 Lophaven, S. N., Nielsen, H. B. and Søndergaard, J. 2002. “DACE: a Matlab Kriging toolbox.
654 Vol. 2.” IMM Informatics and Mathematical Modelling. The Technical University of
656 Lv, Z., Lu, Z., and Wang, P. 2015. “A new learning function for Kriging and its applications
659 MacGregor, J. G., Kennedy, D.J.L., Bartlett, M., Chernenko, D., Maes, M.A., Dunaszegi, L.
660 1997. “Design criteria and load resistance factors for the confederation bridge.”
27
662 M. Moustapha, S. Marelli and B. Sudret. 2022. “Active learning for structural reliability:
664 Nataf, A. 1962. “Determination des distribution dont les marges sont donnees.” Comptes
666 National Building Code of Canada (NBCC). 2020. “National building code of Canada”.
668 Neale, K. W., Ebead, U., Abdel Baky, H., Elsayed, W., and Godat, A. 2005. “Modelling of
670 of the International Symposium on Bond Behaviour of FRP in Structures (BBFS 2005).
671 Nowak, A. S., and Szerszen, M. M. 2003. “Calibration of design code for buildings (ACI
672 318): Part 1 - Statistical models for resistance.” ACI Structural Journal. 100(3):377–
673 382.
674 Okeil, A. M., El-Tawil, S., and Shahawy, M. 2002. “Flexural reliability of reinforced
675 concrete bridge girders strengthened with carbon fiber-reinforced polymer laminates.”
677 Oudah, F. 2023. “Time-dependent reliability-based charts to evaluate the structural safety of
678 RC wharf decks exposed to corrosion and freeze-thaw effect.” Engineering Structures.
680 Oudah, F., and Alhashmi, A.E. 2022. “Time-dependent reliability analysis of degrading
681 structural elements using stochastic FE and LSTM learning.” Canadian Society of
683 Petrie, C. 2022. “Reliability analysis of externally bonded FRP strengthened beams
686 Canada.
28
687 Petire, C. and Oudah, F. 2023. “Stochastic finite element approach to assess the reliability of
688 fiber reinforced polymer strengthened concrete beams.” ACI Structural Journal.
690 The MathWorks, Inc. (2022). MATLAB version: 9.13.0 (R2021b). MATLAB 2021b.
692 Schöbi, R., Sudret, B., and Marelli, S. 2017. “Rare event estimation using polynomial-chaos
693 Kriging.” ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part
695 Shi, J., Wu, Z., Wang, X., and Noori, M. 2015. “Reliability analysis of intermediate crack-
698 Soubra, A.-H., Al-Bittar, T., Thajeel, J., and Ahmed, A. 2019. “Probabilistic analysis of strip
699 footings resting on spattially varying soils using Kriging metamodeling and
701 Sudret, B., and Der-Kiureghian, A. 2000. Stochastic finite elements and reliability - A state-
703 Berkeley.
704 Sun, Z., Wang, J., Li, R., and Tong, C. 2017. “LIF: A new Kriging based learning function
705 and its application to structural reliability analysis.” Reliability Engineering & System
707 Teixeira, R., Nogal, M., O’Connor, Al., Martinez-Pastor, B. 2020. “Reliability assessment
708 with density scanned adaptive Kriging.” Reliability Engineering and System Safety,
710 Wen, Z., Pei, H., Liu, H., Yue, Z. 2016. “A sequential Kriging reliability analysis method
29
712 Engineering and System Safety, 153: 170-179.
713 Wang, J., Cao, Z., Xu, G., Yang, J., Kareem, A. 2023. “An adaptive Kriging method based on
714 K-means clustering and sampling in n-ball for structural reliability analysis.”
716 Wang, N., Ellingwood, B. R., and Zureick, A. 2010. “Reliability-based evaluation of flexural
719 Wieghaus, K. T., and Atadero, R. A. 2011. “Effect of existing structure and FRP uncertainties
722 Xiao-ling, Z., Bo-han, J., Yan, H., Shong-loong, C., and Xiu-yu, L. 2021. “Random field
723 model of soil parameters and the application in reliability analysis of laterally loaded
725 Xing, J., Luo, Y., Gao, Z. 2020. “A global optimization strategy based on the Kriging
728 Zhang, A. H., Jin, W. L., and Li, G. B. 2006. “Behavior of preloaded RC beams strengthened
729 with CFRP laminates.” Journal of Zhejiang University: Science. 7(3): 436–444.
730
30
731 Table 1. Active-learning Kriging: Regression and correlation functions.
733
739
31
740 Notation
741
742
ax , ay , az correlation lengths in 3D space (x, y, z);
β reliability index
βw bond shape factor;
εstop stopping criteria for AK active learning;
εPf estimator error of the predator calculated probability of failure;
ϵSFE model error of stochastic FE model;
K concrete bulk modulus;
μy mean;
μlny lognormal mean;
μĜ mean of predictor performance function vector;
vi , vj coefficient of variation of the mesh points;
ξ𝐢 () 𝑖 𝑡ℎ randomly generated standard normal variable;
′
ρij correlation for the standard normal field between two points 𝑌 (𝑖) and 𝑌 (𝑗) ;
σy standard deviation;
σlny lognormal standard deviation;
σs shear stress calculated by non-linear finite element analysis;
σn normal stress calculated by non-linear finite element analysis;
σĜ standard deviation of predictor performance function vector;
yx , yy , yz coordinates denoting the centroid of a SEM;
𝛙i 𝑖 𝑡ℎ greatest eigenvector of the standard normal field;
Ωcluster error between the previous and current set of cluster centroids;
Ωstop stopping criteria for clustering algorithm;
𝑏𝑐 width of concrete beam;
𝑏𝑓 width of the externally bonded CFRP sheet;
𝒃∗ regression parameter for chosen regression function;
(𝑛)
𝑐𝑘 kth cluster centroid within the nth cluster;
𝑪𝑌𝑌 covariance matrix;
𝐶𝑂𝑉𝑃𝑓0 , coefficient of variation of the probability of failure based on kriging
predictor;
DoE design of experiment (kriging);
dof degrees of freedom;
𝐸𝐹𝑅𝑃 FRP modulus of elasticity;
FEM finite element mesh;
𝑓𝑐′ ultimate concrete compressive strength;
𝑓𝑡′ ultimate concrete tensile strength;
𝑓𝑦 steel tensile strength;
𝑓𝐹𝑅𝑃𝑢 , FRP tensile strength;
ℱ(𝑏, 𝑋) regression function of the random inputs, 𝑋, and unknown coefficient 𝑏;
𝑭 regression function;
𝒇(𝑿) regression function;
𝐺 performance function;
𝐺̂ kriging predictor;
𝐻̂ vector of lognormal random field realizations;
32
𝐻̂𝑙𝑛 vector of gaussian random field realizations;
𝐼[ ] Indicator function given 0 or 1 depending on the sign of 𝐺 or 𝐺̂ ;
𝑘 𝑘th dimension;
𝐾 number of standard deviations away from 𝑃𝑓 (upper/lower error bounds);
𝐾𝑚𝑒𝑎𝑛 mean number of points included in each cluster;
𝑳 vector of autocorrelation length;
𝑀𝑛 nominal flexural strength;
𝑀𝐷 moment due to dead load;
𝑀𝐿 moment due to live load;
𝑀𝑇 transformation of live load to load effect factor;
𝑁𝑐𝑙𝑢𝑠𝑡𝑒𝑟 number of clusters;
NFLS normal failure limit stress;
n number of MCS trials;
𝑃 total number of points included in clustering;
𝑃𝑓0 probability of failure;
𝑃𝑓𝐺̂ probability of failure based on kriging predictor;
𝑃𝑓0 probability of failure using mean Kriging predictor to evaluate the limit
state;
𝑘 probability of failure using Kriging predictor value adding 𝜎𝐺̂ (𝑥)
𝑃𝑓+
multiplied by a constant 𝑘;
𝑘 probability of failure using Kriging predictor value subtracting 𝜎𝐺̂ (𝑥)
𝑃𝑓−
multiplied by a constant 𝑘;
𝑄 load model;
r number of standard normal variables included in truncation;
𝑅 resistance model;
𝒓 correlation vector between unknown point and design sites;
𝑹 correlation matrix between design sites;
𝑅𝑆𝐹𝐸 resistance model of stochastic FE model;
𝑆 vector of design sites;
𝑠 conversion factor converting the lognormal field;
𝑆𝐸𝑀 stochastic element mesh;
SFLS shear failure limit stress;
𝑡 number of random variables in the problem;
𝒖 vector formulated for computational ease;
𝑼 vector of U-learning values calculated using Eq. (20);
𝑼𝑴𝒊𝒏 vector of 𝑃 U-learning values, sorted in descending order;
𝑋 vector of resistance random variables;
𝑌 vector of load random variables, vector of response sites;
𝑌 (𝑖) , 𝑌 (𝑗) ith and jth point of 3D SEM denoted by 𝑦𝑥 , 𝑦𝑦 , 𝑦𝑧 ; and
𝑧(𝑋) stochastic process of the random inputs, 𝑋.
743
744
33
745 Figure Caption
746 Figure 1. Schematic illustration of the SNFE-Clustered AK-MCS framework (numbers are
34
Figure 1 Click here to access/download;Figure;Fig_1.tiff
Figure 2 Click here to access/download;Figure;Fig_2.tiff
Figure 3 Click here to access/download;Figure;Fig_3.tiff
Figure 4 Click here to access/download;Figure;Fig_4.tiff
Response to Editors/Reviewers Comments Click here to access/download;Response to Editors/Reviewers
Comments;SFE_Kriging_Response-to-
Editor’s Comments:
The manuscript should be edited to avoid Removal of the names of python and matlab have
commercial names in its main body (e.g., L18). been completed and replaced with alternative text
Commercial names (materials, supplies, as suggested. The use of LS DYNA is considered
equipment, software, etc.) may be given in the needed, as the solution is uniquely developed for
references (and in the acknowledgements if that specific non-linear finite element software.
applicable), preferably including links to online As the formulations of the materials, elements,
data sheets for materials and user manuals for and solving algorithms may change the outcome
software. For example, in L18, it is suggested to of experiment. The authors kindly request an
edit the text to the effect of “Using commercially exemption for referring to LS-DYNA in the text.
available codes, an algorithm was developed to
automate the SNFE-Clustered AK-MCS analysis”
(or something similar) instead of “LSDYNA-
Python-MATLAB code was developed to
automate the SNFE-Clustered AK-MCS analysis”.
L45, etc: when a list of references is presented, the Addressed, thank you.
order should be chronological. The rest of
the manuscript should be checked accordingly.
L441: please use official ACI or CSA Addressed, thank you. “Nominal Flexural
nomenclature for Mn instead of “ultimate nominal Strength” has been adopted throughout the text.
moment”. For example, “nominal moment
capacity” or “nominal flexural strength”. Same for
the other parameters used in the manuscript; please
check the other parameters accordingly and edit
if/as needed.
Page 1 of 11
L542: please follow ASCE’s guide for authors for The code reference is updated to ACI PRC-440.2-
the references format. Here, “ACI (American 23 in the revised manuscript.
Concrete Institute). (2017). “Guide for the design
and construction of externally bonded FRP systems
for strengthening concrete structures. ACI PRC-
440.2-17, Farmington Hills, MI.” Also, if
applicable, please consider updating the reference
since this guide has just been superseded by ACI
PRC-440.2-23.
Given the number of parameters, it is A notation table has been added at the end of the
recommended that a Notation section be added text along with the tables. Greek letters have been
following the instructions in ASCE’s guide for unitalicized.
authors. Throughout the manuscript, italics should
be used for notation except for numbers and Greek
letters, as detailed in ASCE’s guide for authors.
This format should be used consistently in the text,
tables, visuals and Notation section.
“fiber-reinforced polymer…” instead of “fiber Addressed, thank you for the remarks.
reinforced polymer…”, consistent with the format
used in ACI, AASHTO, CSA, etc., and preferred in
this journal. Please revise the manuscript
accordingly, now both versions occur in
the manuscript.
Fig. 2: make all text labels readable, perhaps by Thank you for the comment. The figures text has
increasing the font size. Please add concrete cover been made larger. The cover is not given in the
dimensions. experimental work by Zhang et al. (2006).
Table 2: use a reasonable number of digits (and the Table 2 has been updated. All other instances have
same in each column). Also, use ‘kN-m’ or ‘kNm’ also been updated.
instead of ‘kN.m’.
Reviewer #1:
Response
Comments Authors’ Response No.
Page 2 of 11
The variable P_fG^ in Eq. (2) Updated, thank you. 1
should be P_f0. Please check and
modify.
Why discretize random fields? In The following statement was added to the revised
addition, the basic principle of manuscript to elaborate on the fundamentals of the
EOLE should be given before the EOLE method and provide reasoning regarding the
realizations of lognormal random need for discretization.
field and gaussian random field.
Line 187:
“The EOLE method represents the random field as a
linear function with nodal random variables and a
series of shape functions. The shape functions are
found by minimizing the error variance. Discretization
was required to generate realizations of the field at the
centroids of the stochastic mesh element, for a given
stochastic mesh density. In other words, all finite
elements composing a stochastic element are assigned
a uniform realization based on the random field
discretization process, where a stochastic element is
composed of one or more finite elements.”
Lines 239-240, the two points A sentence has been added and the section updated to 2
considered are the design point Si clarify this point. The correlation matrix R (Eq. 18b) is
and all other points Xj, why does the correlation matrix between the design sites
the R(X) denote the correlation themselves. The equation for the correlation between
between the design points Si and an unknown point X and the design sites, r, has been
Sj? added (Eq.18a).
Line 250:
“ The correlation, r and R, given by Eq. (18a) and Eq.
(18b), respectively, where the term L is the vector of
autocorrelation length (called ‘correlation length’
herein). The term r(X) is the correlation between the
unknown site, X, with all known design sites S_i. The
term R(X) is the correlation between the design sites
S_i and S_j. The vector L is calculated by minimizing
the term in Eq. (19), in an iterative process, where m
is the number of design sites (i.e., S=[S_1,…,S_m]),
and k represents the kth dimension (i.e., k = [1,2,3] in
a 3D problem).”
Line 259, the reason to propose the The reviewer is kindly referred to lines 160 to 175 of 3
U-learning function is totally the “Overview” section for a detailed explanation of
wrong. The logic in Lines 247-260 how the active learning process is performed. The
is bad. What is the active learning? Overview section provides the sequential events of the
What kind of learning functions active learning process and provides a review of the
are widely used in reliability learning functions available in literature, and their role
analysis? This information is very in the enrichment process.
important when introducing the
Kriging surrogate model with an The authors purposely avoided providing a detailed
active learning strategy. explanation in the referenced Line 259 to avoid
repetition of the information presented in the
Page 3 of 11
“Overview” Section. The purpose of the “U Learning
Function” section of the paper (where Line 259
belongs) is to only present the formulation of the
function and not the reasoning behind choosing it since
the latter was described in the “Overview” section as
mentioned above.
There is no logic between the As presented in the authors’ response No. 3, the roles 4
sections "U Learning Function" of the learning function and stopping criteria are
and "Stopping Criteria". Why described in detail in the “Overview” section of the
should introduce the stopping manuscript. The “Stopping Criteria” section of the
criteria? manuscript includes the formulation of the stopping
criteria described in the “Overview” section. This
explains why the discussion about the role of stopping
criteria in the “Stopping Criteria” section was limited
in the first submission.
Line 284:
“As discussed in the Overview section, a stopping
criterion is required for the learning function to
determine when to cease the loop for the enrichment of
the Kriging predictor.”
Line 269, what is the meaning of The following was added to the revised submission: 5
the U-value of 2.0 for the stopping
Page 4 of 11
criteria, from the probability point Line 286:
of view? “Two stopping criteria are considered in this analysis.
The first is the criteria proposed in Echard et al.
(2011) where the U-value should be equal or greater
than 2.0 for all realizations predicted using the
Kriging surrogate model. The limit of 2.0 on the U-
value implies that the probability of making a mistake
on the sign of ϕ(-2) corresponds to 0.023 (i.e., desired
reliability level of avoiding a mistake). The limit
ensures all points found by the kriging predictor have
low MSE of the kriging predictor.”
Line 201, what is the value of the The following was added to the revised manuscript to 6
prescribed variance limit used in address the reviewer’s comment:
this manuscript?
Line 211:
“The number of standard normal variables r is
determined such that the ratio of the eigenvalues
divided by the trace of the covariance matrix, Q, does
not exceed 0.8, as described in Eq. (10).”
∑𝑟𝑖=1 𝑪𝑌,𝑌
𝑄= 𝑖
(10)
𝑡𝑟(𝑪𝑌𝑌 )
In general, the model error is The authors appreciate the reviewer’s concern about 7
defined as the ratio of the the model error used in the paper. The reviewer’s
experimental value to the proposition is valid if the model error is derived based
predicted value. Since the on a reasonable number of tests of the same beam
experimental values are referred to configuration, but this is not the case. Zhang et al.,
the reference Zhang et al 2006, the 2006 only tested one specimen of the beam
real model error and the configuration considered in the proposed manuscript.
corresponding random feature can Therefore, the authors believe that relying on other
be obtained through the predicted established sources for FE model error is justified for
values. It is better than the value the purpose of our analysis.
given by another reference
Castaldo et al 2019. Besides, for Table 2 was updated to include the mean and not the
the random features of model bias as per the reviewer’s request.
error, the mean value and cov
value should be given, not the Bias
value.
In Table 3, the duration of the The CPU configuration was included in the revised 8
computation is given. So the manuscript.
computer configuration (e.g., CPU
and some other index) should be Line 474:
given. “All runs were processed on a windows-based server
with dual 2.20GHz processors (48 cores), a standard
graphics card, and 128 Gb of available RAM”
It is difficult to check the accuracy The authors appreciate the reviewer’s comment 9
of the reliability index obtained by regarding the challenge with evaluating the accuracy
different stopping criteria, since of the reliability index. However, the reliability index
results included in Table 3 range from 3.68 to 3.72,
Page 5 of 11
there is no reference reliability which is generally in conformity with the typical range
index, i.e., the MCS result. of reliability index used in code calibration in USA and
Canada. The authors included the following statement
in the revised manuscript to enrich the discussion of
the analysis results and to address the comment by the
reviewer.
Line 472:
“The calculated reliability index ranges from 3.68 to
3.72, which falls within the typical range of values
expected for a code compliance design of FRP
reinforced members (Hassan and Oudah, 2022).”
The meaning of the Bias should A footnote to Table 2 was added to the revised 10
be given. manuscript to define the bias.
Line 741:
“ratio of actual mean-to-specified mean”
The conclusions should be The conclusion has been restructured and almost 11
reorganized in a more clear form. completely rewritten. Please see revised “Conclusion”
section of the revised manuscript.
When introducing AK in the The following was added to the revised manuscript to 12
introduction part, some methods address the comment. Four references were added.
such as AK-MCS, AK-IS, and
AK-SS should be given with Line 117:
references. “The AK method has been applied to several
frameworks to perform reliability analysis, such as
AK-MCS, AK-FORM, AK-IS, and AK-SS (Owen et al.,
2019; Xiao et al., 2019; Yang et al., 2020; and
Khorramian and Oudah, 2022).”
Reviewer #2:
Response
Comments Authors’ Response No.
…, it is unclear how the developed The authors believe that the proposed research 13
framework links the research gap directly aligns with the “practical”-oriented
with the focus on the mechanics research focus of the ASCE Journal of Composite
and manufacturing science of for Construction. The manuscript addresses a real-
composite materials and their life research gap experienced by the authors, who
applications. In addition, the study have diverse experiences working on real-life
failed to provide clarification on consulting jobs of assessing and retrofitting concrete
why such methodology is important structures. One of the main challenges with using
in the field of FRP strengthening, FRP for real-life strengthening of existing structures
specifically. Therefore, the study is the inability to reliably assess the performance of
could be a better fit with other external strengthening when the concrete structure
journals. The experimental experiences a considerable spatial variation in
specimens used for the validation of concrete properties due to active deterioration
the SNFE models is very limited
Page 6 of 11
reducing the applicability of the (spatial variation implies variation of concrete
developed methods. integrity in the 3D space of the evaluated structure).
In this research, we aim to address this challenge and
the uncertainty it causes in predicting the reliability
of the strengthened system, by combing the concept
of random fields and adaptive learning. Therefore,
the research need stems from practical experience
and speaks to structural designers working in the
area of assessment and retrofit of existing structures.
Line 32:
“The quality of FRP-to-concrete bond is affected by
the concrete integrity at the interface, which varies
across the dimensions of the strengthened member
causing added uncertainty in predicting the
structural response, and hence, the reliability of the
FRP strengthened member. This research proposes
a computationally efficient approach to assess the
reliability of FRP strengthened concrete members
by considering the spatial variation in the concrete
properties (compressive strength, tensile strength,
bulk modulus) and the quality of the FRP-to-
concrete bond (shear and normal bond strength) by
using an adaptive machine learning technique. The
proposed framework can be utilized by engineers to
design FRP strengthening systems for concrete
members experiencing variation in the concrete
properties due to poor quality control or active
deterioration.”
The scope of the study needs to be As stated in the above response to Comment No. 13, 14
stated clearly after addressing the the “Practical Application” section was completely
Page 7 of 11
research gap with regard to FRP re-written to highlight the practicality of the
strengthening of structures. proposed research.
Line 554:
“The practical use of SNFE based reliability
analysis to assess the safety of reinforced concrete
members strengthened externally in flexure with
FRP, where the nominal flexural strength is
sensitive to the spatial variation of the mechanical
properties, is hindered due to the significant
computational cost of performing the analysis.
Optimization of the analysis will aid in making them
more viable to be run by practicing engineers and
code developers when performing reliability
analysis. By facilitating the incorporation of real-
world data, these frameworks enhance the efficiency
of reliability analyses and aim to provide valuable
tools, advancing the state-of-the-art in concrete
structural safety assessment.”
Please clearly explain the The following paragraph was added to explain the 15
importance of learning functions. importance of the learning function in the process of
Also, justify selecting the U learning active learning.
function and briefly discuss the
difference between this function and Line 269:
others like H or LIF functions. The “A learning function is required to determine the
authors are encouraged to provide as next design sites for enrichment of the Kriging
much details as possible since not all predictor as described in the Overview section. The
researchers and practitioners are use of a learning function allows the user to grow
familiar with the approach used. the size of the DoE in a stepwise manner to find the
most efficient solution while maintaining a sufficient
level of accuracy (Buckley et al 2020). The
commonly used U-learning function has been
selected in this work to demonstrate the clustered-
SNFE framework, as it allows clustering of the U-
learning data to generate multiple enrichment sites
per iteration.”
Page 8 of 11
than the one used in the framework, which is the U-
Learning function. Please refer to the following
added sentence.
Line 169:
“While the formulations of the learning functions
differ based on the fundamental approach taken to
assess the learning process, they all seek to find the
next candidate for enriching the DoE. The U
learning function has been utilized in several
research studies for SNFE analysis, and thus, is
utilized in this research (Khorramian et al., 2023).”
Please provide the epoxy properties. The bond strength varies spatially along the beam 17
It is proved that behavior of FRP- dimension of the proposed framework of analysis.
strengthened RC members depends This represents the core utility of the proposed
heavily on the bond strength as the framework of analysis since it accounts for the
FRP-concrete interface. uncertainties in the bond performance along the
spatial variation of the concrete properties.
The bond stress limits for the FRP (epoxy) are based
on the work of Neal et al., 2005 as indicated in the
“SNFE Model Development” section (Eqs 31 to 34).
They are empirical formulas to estimate the normal
and shear stress limits for debonding failure and are
based on the concrete strength and the geometry of
the bond surface. Since f’c in Eq. (32) is discretized
as a 3D spatially varying random field, the resulting
NFLS and NSLF (normal and shear bond strength,
respectively) are represented by random field fully
correlated to f’c.
Line 399: What is the equation of The following was added to the revised manuscript 18
state? to address the comment.
Line 418:
“…while K was calculated using the equation of
state (i.e., the piecewise description of the
relationship between pressure and volume strain).”
Please include what exactly NFLS The normal failure stress limit (NFLS) and shear 19
and SFLS stand for. failure stress limit (SFLS) are the normal and shear
stress limits that denote bond failure in the SNFE
Page 9 of 11
analysis (Eq. (31)). The stress limits are calculated
based on the calculation of the NFLS and SFLS
based on the ultimate compressive strength of the
concrete, the ratio of the width of the concrete beam
and width of the FRP strip used for flexural
strengthening (Neale et al., 2005). Also, please refer
to Response No. 17 for further description about the
role of NFLS and SFLS in capturing the spatial
variation of the bond strength.
Line 433:
“σ_n describes the normal surface stress, σ_s is the
shear surface stress, NFLS is the normal failure
limit stress and SFLS is the shear failure limit
stress”
Reviewer #3:
Response
Comments Authors’ Response No.
The practical application does not The described method would benefit structural 22
seem to serve the intended engineers who design FRP strengthening of existing
purpose, it look like a description concrete structures with considerable spatial
of method rather than how this variability (i.e., variability of the mechanical
research could benefit the properties) along the FRP-to-concrete interface (i.e.,
practicing engineers. bond), and with the concrete volume (i.e, concrete
mechanical properties). The higher the variation in
Page 10 of 11
these properties, the higher is the uncertainty of FRP
design, and hence, its effectiveness.
Line 32:
“The quality of FRP-to-concrete bond is affected by
the concrete integrity at the interface, which varies
across the dimensions of the strengthened member
causing added uncertainty in predicting the structural
response, and hence, the reliability of the FRP
strengthened member. This research proposes a
computationally efficient approach to assess the
reliability of FRP strengthened concrete members by
considering the spatial variation in the concrete
properties (compressive strength, tensile strength,
bulk modulus) and the quality of the FRP-to-concrete
bond (shear and normal bond strength) by using an
adaptive machine learning technique. The proposed
framework can be utilized by engineers to design FRP
strengthening systems for concrete members
experiencing variation in the concrete properties due
to poor quality control or active deterioration.”
The conclusion part may be better The “Conclusion” section was completely 23
presented in bullet points to make restructured, and bullet points were added. Please refer
easier for the reader to understand to the revised manuscript.
the major findings of this research.
Page 11 of 11
Track Changes Version