Integrating automated machine learning and interpretability analysis in architecture, engineering
and construction industry: A case of identifying failure modes of reinforced concrete shear walls
Dong Liang 1 and Fan Xue 2∗
This is the peer-reviewed post-print version of the paper:
Liang, D. & Xue, F. (2023). Integrating automated machine learning and
interpretability analysis in architecture, engineering and construction industry: A case
of identifying failure modes of reinforced concrete shear walls. Computers in Industry,
145, 103883. Doi: 10.1016/j.compind.2023.103883
The final version of this paper is available at https://doi.org/10.1016/j.compind.2023.103883
The use of this file must follow the Creative Commons Attribution Non-Commercial No
Derivatives License, as required by Elsevier’s policy.
Abstract
Machine learning (ML) has been recognized by researchers in the architecture, engineering,
and construction (AEC) industry but undermined in practice by (i) complex processes relying
on data expertise and (ii) untrustworthy ‘black box’ models. As a result, ML results of complex
non-linear AEC problems, such as the failure mechanism of reinforced concrete (RC) shear
walls, are not comparable with empirical and mechanics-based models. This paper aims to
integrate automated ML (AutoML) and interpretability analysis to study the failure mechanism
of RC shear walls. In this study, we collected a dataset of 351 comprehensive samples for the
failure mode identification of RC shear walls. First, the AutoML model trained using the
dataset outperformed a set of conventional ML methods in terms of the F1 accuracy score. Then,
three model-agnostic interpretability analysis methods confirmed the trustworthiness of the
AutoML model. The contribution of this paper is three-fold. First, AutoML sheds light on the
automatic identification of failure modes of RC sheer walls. Second, the interpretability
analysis can validate ‘black-box’ ML models against long-established domain knowledge in
solving non-linear AEC problems. Third, for AEC industrial practitioners, the whole process
is automatic, accurate, less reliant on data expertise, and interpretable.
Highlights
•

Integration of automated Machine Learning (AutoML) and interpretable analysis for
accurate and trustworthy ML

•

A comprehensive dataset of 351 tests of failure mode identification of RC shear walls

•

A ‘black box’ AutoML model outperformed five conventional ML algorithms in terms
of F1 accuracy score

•

Complementing local and global interpretability analysis methods for explaining and
validating the AutoML model

•

Trustworthiness of the AutoML model confirmed with domain knowledge via
interpretability analysis results

1 Dong LIANG, MSc, PhD student. leodong@connect.hku.hk ,

: https://orcid.org/0000-0002-2918-7898
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong SAR, China; Email:
2 Fan XUE, PhD, Assistant Professor. xuef@hku.hk ,

: http://orcid.org/0000-0003-2217-3693
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong SAR, China; Email:
*: Corresponding author, Tel: +852 3917 4174, Fax: +852 2559 9457; Email: xuef@hku.hk

1

Keywords
Machine learning; Interpretability analysis; RC shear walls; Failure mode identification;
AutoML

5

10

15

20

25

30

35

1 Introduction
Architecture, engineering, and construction (AEC) researchers have been increasingly
interested in complex non-linear problems (Su et al. 2021). In civil engineering, for example,
traditional physics-based models often cannot achieve accurate results and efficiency at the
same time. A typical problem is the failure mode identification of reinforced concrete (RC)
shear walls, a vital lateral-force-resisting component for buildings in wind-prone or
earthquake-prone areas (Farrar et al. 1993). The failure mechanism, involving various
geometrical configurations and material parameters, is too complex to formalize empirical or
mechanics-based models. The failure mode identification of RC shear walls used to depend on
computation-intensive and time-consuming continuum-based finite element models (Dashti et
al. 2014); the high costs restricted its capability in many cases, such as building vulnerability
analysis and regional risk assessment (Mangalathu et al. 2020).
Machine learning (ML) has been used to analyze AEC problems (Singh 2021) due to
the data availability (Vadyala et al. 2022). In ML, target variables are predicted using data
science assumptions without references to rigorous physical theoretical inference (Montáns et
al. 2019). Example applications are structural behavior prediction (Luo & Paal 2022) and
structural health monitoring (Zhang et al. 2020). In fact, advanced ML techniques provide
possible alternatives to physics-based models for computationally demanding numerical
modeling, as well as dangerous and expensive experimental testing (Deka 2019).
However, ML applications in AEC have lagged behind expectations (Deka 2019) due
to two groups of possible reasons (Kuwajima et al. 2020): (i) complex data-expertise-reliant
modeling process and (ii) untrustworthy ‘black box’ models (Feng et al. 2021). First,
developing a high-quality ML model is time-consuming and requires multiple steps and expert
knowledge (Hidalgo-Mompeán et al. 2021). Furthermore, most advanced ML methods, such
as artificial neural network (ANN), support vector machines (SVM), bagging, boosting, and
deep learning, are often ‘black box’ models, as opposed to ‘white box’ models based on
mechanics or empirical regression (Feng et al. 2021). The missing or uninterpreted step-bystep details in these ‘black box’ ML models lead to untrustworthiness for industrial
practitioners and regulators who are sensitive to risks in healthcare, legislation (e.g., European
Union’s GDPR), and critical engineering (Linardatos et al. 2020).
Two new emerging technologies, i.e., automated ML (AutoML) and interpretability
analysis, have the potential to solve the two issues in complex AEC problems. AutoML is a
technique of automating the time-consuming and expertise-reliant procedures in selecting and
training ML models; it has been successfully applied in many science and engineering areas
such as medical science (Tran et al. 2021; Anwar 2021), public service (Li et al. 2019), finance
(Vakhrushev et al. 2021), materials (Bangaru et al. 2019), and the construction domain (Zhao
2

40

45

50

55

60

65

et al. 2023). Interpretability analysis refers to developing human-readable explanations to
facilitate practitioners’ comprehension of why an ML model makes certain decisions or
predictions (Molnar 2020; Guo et al. 2022).
This study aims to integrate AutoML and interpretability analysis to study failure mode
identification of RC shear walls. This research involves a dataset of 351 shear wall failure
samples collected from the literature, 15 low-level ML algorithms in AutoML, and three
interpretability analysis methods for validating the trained AutoML model. The contribution of
this paper is three-fold. First, AutoML sheds light on the automatic identification of the failure
modes of RC sheer walls. Second, interpretability analysis can validate the ‘black-box’ models
with long-established AEC domain knowledge. Third, for AEC industrial practitioners, the
whole process is automatic, accurate, less reliant on data expertise, and interpretable.
2 Overview of AutoML and interpretability analysis
2.1 AutoML
AutoML is the process of automating the time-consuming and iterative procedures involved in
building ML models (Feurer et al. 2020). Figure 1 shows the full process of developing a highquality ML model. Traditionally, appropriate ML algorithms, extensive computational
resources, time, and domain expertise are essential for training or building an ML model
(Gijsbers et al. 2019). Even professional data scientists may be confused in each phase by many
factors, such as unbalanced class, parameter noise, and selection of performance measures. As
a result, there is an increasing interest in automating and democratizing the processes in Figure
1.

Figure 1 Workflow of machine learning, where the gray box indicates the scope of AutoML
AutoML, as shown in the gray area in Figure 1, automatically completes up to four
tasks, including feature engineering, algorithm selection, parameter tuning, and evaluation
(Elshawi et al. 2019). Thus, AutoML is more efficient and productive while maintaining the
ML model quality. For example, advanced AutoML algorithms outperformed human experts
in five out of six construction management datasets (Zhao et al. 2023). In the software market,
there exist commercial AutoML platforms, such as Google Cloud AutoML and Microsoft
Azure, and open-source libraries, such as Auto-sklearn, MLBox, and TPOT (Elshawi et al.
2019). Yet, the ‘black box’ problem of ML models remains regardless of the use of AutoML;
3

70

75

80

85

90

to a certain extent, an AutoML model is even harder to explain due to the sophisticated
parameter controls in AutoML algorithms.
2.2 Interpretability analysis and applications in civil engineering
Interpretability analysis aims to explain the rationale behind the decisions and outcomes of a
computerized model so that a human can comprehend it (Gilpin et al. 2018). There are two
categories of interpretability analysis: model-specific and model-agnostic (Molnar 2020).
Model-specific interpretation tools are limited to intrinsically interpretable ML models, such
as a decision tree model of project-related waste generation behaviors (Yang et al. 2021). In
comparison, a model-agnostic interpretation can be applied to any ML model after the training
process. Furthermore, model-agnostic interpretable methods can be divided into two groups:
global methods and local methods. Global model-agnostic methods describe the average
behavior of an ML model (Friedman 2001), while local model-agnostic methods, such as local
surrogate (LIME) (Ribeiro et al. 2016), explain a subset of prediction results.
Model-agnostic interpretability analysis has been applied to ML models in the AEC
industry in recent years. Table 1 summarizes the application problems, ML task, selected
interpretability analysis methods, sample size, contribution, and limitations in seven related
studies. In short, existing studies focused on the global model-agnostic interpretability of the
regression model’s partial dependent plot (PDP) and coalitional-game-theory-based SHapley
Additive exPlanations (SHAP) are popular methods. Yet, disadvantages were observed as well.
For example, PDP was criticized for the feature interaction that can produce wrong
interpretation results (Molnar 2020). SHAP has three limitations in unintuitive feature
attributions, possibly biased and misleading interpretations, and unclear estimation thresholds.
Furthermore, local model-agnostic methods received less attention. Thus, there is a need to
gauge and compare global and local interpretability analysis methods. In this paper, two global
model-agnostic interpretation methods: PDP and ALE, and one local method, LIME, are
studied for the failure mode identification problem.
Table 1 List of recent model-agnostic interpretability work of ML models in civil engineering
Local /
Sample
Tools#
Source
Contribution
Global*
size
Ultimate load of Regression
G PD analysis Le & Phan
99 PD analysis of each
rectangular
(2020)
input variable’
CFST columns
influence on the
ultimate load
Shear Strength Regression L & G SHAP
Feng et al.
434 Coalitional gamePrediction of
(2021)
theory based SHAP to
Squat RC Walls
explain predictions
from both local and
global aspects
Targets of ML

ML task

Ground motion Regression L & G SHAP
characteristics
(PGV and PGA)

Somala et
al.
(2021)

4

4,000 SHAP methods to
explain the influence
factors for estimating
PGA&PGV

Limitations
1) PD analysis ignores
feature dependence.
1) TreeSHAP (a modelagnostic subclass of
SHAP method) leading to
unintuitive attributions,
2) Possibly problematic
explanations by SHAP,
3) Unanswered thresholds
by the SHAP.
Ditto.

95

100

105

110

115

Auxetic
behavior of
cementitious
cellular
composites

Regression

G

Magnitude of
moment
capacity at
failure; and the
failure
mode/causes

Regression
and
Classificati
on

G

Structural
Regression
performance of
GFRP
elastic gridshells
Seismic
Regression
response of R/C
buildings

G

*
#

G

SHAP

Lyngdoh
et al.
(2022)

850 SHAP to analyze the 1) No local
influence of each input interpretability analysis
variable on the Pois- involved,
son’s ratios
2) Lack validation of
interpretation with expert
knowledge.
PDP,
Naser
103 An explainability case 1) Many interpretable
SHAP,
et al.
study on reinforced
methods without crosssurrogate, & (2021)
concrete (RC) beams references,
feature
strengthened with fiber- 2) Feature interaction not
interaction
reinforced polymer
considered in other
(FRP) composite
methods such as PDP,
laminates
3) No local
interpretability analysis
involved.
PDP, ALE, Kookalani
400 Three interpretable
1) No local
& SHAP
et al.
methods to analyze the interpretability analysis
(2022)
influences of variables involved.
on the stress prediction
PDP &
Demertzis 5,850 Use two methods to
1) Lacking of
SHAP
et al.
analyze the influence of comprehensive
(2023)
Housner intensity on interpretability analysis
the maximum interstory on all parameters,
drift ratio
2) No local
interpretability analysis
involved.

: L: Local; G: Global.
: PD: Partial dependence; SHAP: SHapley Additive exPlanations; PDP: PD plot; ALE: accumulated local effects.

3 The problem of failure mode identification of RC shear walls
RC shear walls are the primary lateral load-carrying structure elements in many high-rise
buildings due to superior high lateral strength and stiffness (Paulay 1975). For the desired
ductile failure modes, RC walls are designed to have a ‘strong shear and weak flexure,’ which
is commonly accomplished by restricting the axial compressive force ratio and adding
boundary components to the structure (Lefas et al. 1990). According to the aspect ratio (shear
span length divided by wall length), shear walls are typically classified as either squat (lowrise) or slender (tall, high-rise) (Birely 2012). A ductile failure mechanism characterized by
flexural yielding toward the base is more likely to occur in slender walls. Squat walls are more
prone to a nonductile shear-controlled failure mechanism, which is best defined as an abrupt
reduction of stiffness and strength under seismic loads. According to ASCE/SEI 41-17 (ASCE
2017), walls with an aspect ratio larger than 3.0 (often controlled by flexure) are regarded as
slender. In contrast, those with an aspect ratio less than 1.5 (mostly controlled by shear) are
considered short and squat. The behavior of walls having an aspect ratio between 1.5 and 3.0
is influenced by both shear and flexure. It is also suggested in FEMA 306 (ATC 1998) that
well-designed slender shear walls (with aspect ratios of more than 3.0) are more likely to
experience ductile flexural failure.
Ductile flexural failure is caused by concrete crushing or a longitudinal reinforcement
fracture at the plastic hinge zone, as shown in Figure 2(Ⅰ). Generally, flexural failure is
uncommon in squat shear walls, particularly those with aspect ratios below 1.0. However,
5

120

125

130

135

140

depending on the reinforcing characteristics, flexural failure is possible in such walls, and shear
failure may also be seen. The nonductile shear failure mechanism of squat shear walls was
divided into three categories by Paulay and Priestley (1992) sliding shear failure, diagonal
tension failure, and diagonal compression failure, as shown in Figures 2(Ⅱ) and 2(Ⅲ). The
presence of one or more corner-to-corner diagonal cracks, as seen in Figure 2(Ⅱ), is diagnostic
of diagonal tension failure (T), which may happen when a wall does not have enough horizontal
shear reinforcement. The concrete crushes under this compression develop widespread crack
patterns in diagonal compression failure mode (C), as illustrated in Figure 2(Ⅲ). Sliding shear
failure (S) may be caused by two different things: (1) several major cracks at the wall base and
(2) the buckling of rebars and the crushing of concrete in a narrow band along the base of the
wall, as illustrated in Figure 2(Ⅳ), following considerable flexural reinforcement yielding. In
RC design and retrofitting, designers and engineers are more concerned about the ductile or
non-ductile failure mode of shear walls than detailed classification. Therefore, this study
divides the classification work into two phases: the first stage, classifying ductile and nonductile failure modes, and the second stage, classifying three detailed non-ductile failure modes;
the first stage often satisfies most engineering needs.

Figure 2 Typical failure modes of RC shear walls, where dashed boxes indicate the first-stage
classification
4 Research methods
Figure 3 briefs the three-step flowchart: dataset preparation, AutoML using Auto-sklearn, and
interpretability analysis in failure mode identification of the RC shear wall in this study.

Figure 3 The proposed workflow for failure mode identification of RC shear wall
6

145

150

155

4.1 Feature settings
Since there are few studies on the failure mode identification of RC shear walls using datadriven methods, this study mainly refers to national regulations about design criteria of
structural concrete, including JGJ 3-2010 (China code) (MOHURD 2010), ACI 318-19 (U.S.
code) (ACI 2019), EC-8 (Europe code) (CEN TC-250 2004), mechanics theories (Vecchio &
Collins 1986), and an existing study by (Mangalathu et al. 2020) to set up features. Among
these existing references, a total of four types of features, material property (fy,vw, fy,hw, fy,cb, fc),
reinforcement details (ρvw, ρhw, ρvb, ρhb), geographic details (lw/tw, M/Vlw, Ab/Ag, CS), and load
(P/fcAg), were considered to influence the seismic behavior of RC shear wall designs.
Specifically, 13 features were selected for classifying the failure mode of RC shear walls, and
a description of each feature is given in Table 2. The study by Salonikios et al. (1999) described
more details on the sensors and methods for measuring these features in laboratory tests. These
13 features are also considered essential to influence the seismic behavior of other load-bearing
structures in many studies related to data-driven structure behavior prediction (Sujith
Mangalathu 2018; Hosein Naderpour 2021).
Table 2 Feature setting (excluding classification target) for the problem
Group
Material
property
Reinforcement
details
Geometric
characteristics

Load

Symbol
fy,vw
fy,hw
fy,cb
fc
ρvw
ρhw
ρvb
ρhb
lw/tw
M/Vlw
Ab/Ag
CS
P/fcAg

Unit
MPa
MPa
MPa
MPa
103
103
103
103
103
103

Feature description
Yield stresses of vertical bars in web
Yield stresses of horizontal reinforcement in web
Yield stress of confinement reinforcement in boundaryelement
Concrete compressive strength
Web vertical reinforcement ratio
Web horizontal reinforcement ratio
Boundary region vertical reinforcement ratio
Boundary region horizontal reinforcement ratio
Proportion of length and thickness
Aspect ratio
Proportion of the area of the boundary element and the gross sectional
area
Cross-section type, i.e., Rectangular (R), Barbell (B), and Flanged (F)
Axial load ratio

Figure 4 Schematic diagram of RC shear wall tests with three types of cross-sections
7

160

165

170

175

180

185

190

4.2 Data collection
In the process of collecting data, the data needs to first meet the requirement of containing all
13 features listed in Table 2. Some extra rules are followed to reduce the impact of irregular
data for the main purpose of this study. Specifically, the cross sections of all selected shear
walls are one-story, one-bay, symmetric, and without openings. In addition, all selected walls
have continuous longitudinal reinforcement without lap splice, as well as deformed and straight
reinforcement. Finally, the aspect ratio (M/Vlw or hw/lw) is considered the dominant feature for
the seismic behavior of RC shear walls according to the ASCE/SEI 41-17 (ASCE 2017).
According to the aspect ratio, RC shear walls are conventionally classified into slender walls
(aspect ratio＞3.0), moderate walls (1.5 ≤ aspect ratio ≤ 3.0), and squat walls (aspect ratio＜
1.5). Therefore, the data collected in this study equally covers all three types of walls to ensure
the applicability of the proposed method.
A total of 351 records were collected from laboratory tests reported in the literature. For
example, the first 242 records were from the ACI 445B Shear Wall Database (Usta et al. 2017),
with the criterion that each data record covers all features in Table 2. The rest of the 109 records
were also collected from published experiments, including Chile (Massone et al. 2009), Canada
(Abdulridha & Palermo 2017; Mohamed et al. 2014; Palermo et al. 2002), China (Chiou et al.
2004; Deng et al. 2008; Hwang et al. 2004; Jiang et al. 2013; Kuang & Ho 2008; Lu et al. 2018;
Ren et al. 2018; Su & Wong 2007; Wen-yuan et al. 2018; Zhou et al. 2010), the United States
(Terzioglu et al. 2018), Australia and New Zealand (Dabbagh 2005; Gebreyohaness et al. 2014;
Tripathi et al. 2019; Yanez et al. 1991), Japan (Kabeyasawa & Matsumoto 1992; Kimura et al.
1996; Nakamura et al. 2009; Tokunaga & Nakachi 2012), Slovenia (Tomazevic et al. 1996),
and Greece (Christidis & Trezos 2017). The existing open-source library—ACI 445B Shear
Wall Database was also used by Mangalathu et al. (2020) for verifying the feasibility of
identifying failure modes of shear walls by ML methods. A total of eight algorithms—Naïve
Bayes, K-Nearest Neighbors, Decision Tree, Random Forest, AdaBoost, XGBoost, LightGBM,
CatBoost, and XGBoost—are used, and the highest accuracy was 86% from Random Forest
(Mangalathu et al. 2020). The quality and reliability of existing experiment-based data were
initially confirmed.
Table 3 lists an excerpt of the collected dataset, while the full data table is available in the
supplemental materials of this paper. Figure 5 briefs the distributions and statistical details of
the dataset, respectively. The dataset has a wide range of distribution, and common engineering
values are covered. Figure 5 (m) shows that three types of cross-sections accounted for 69.2%
(R), 18.9% (B), and 11.9% (F), respectively. The distribution of failure modes for two-stage
classification is shown in Figure 5(n)(o).

195

8

Table 3: Excerpt of the collected dataset with 351 records of RC shear walls failure tests
No.

fy,vw

fy,hw

M/Vlw

1
2
3
…
351

585
585
610
…
345

610
610
610
…
345

1.00
1.00
1.00
…
2.5

140
120

(a)

fy，vw (MPa)
Min: 258.2
Max: 806
Median: 446
Mean: 441

Count

100
80
60
40

… CS

…

…
…
…
…
…

…
…
…
…
…

(b)

R
R
R
…
R

Failure
mode (1st)
ND
ND
ND
…
ND

Failure
mode (2nd)
S
S
S
…
S

(c)

fy，cb (MPa)
Min: 0
Max: 1262
Median: 443
Mean: 456

fy，h w (MPa)
Min: 216
Max: 1262
Median: 443
Mean: 456

Source
(Salonikios et al. 1999)
(Salonikios et al. 1999)
(Salonikios et al. 1999)
…
(Zhou et al. 2010)
(d)

fc (MPa)
Min: 13.7
Max: 130.8
Median: 34
Mean: 35.7

20
0

300 400 500 600 700 800

400 600 800 1000 1200

200 400 600 800 10001200

3

ρy，vb (1×10 )
Min: 1
Max: 36.7
Median: 5.4
Mean: 6.5

(g)

ρy，h w (1×10 )
Min: 1.1
Max: 99.1
Median: 22.4
Mean: 27.6

(h)

60

5 10 15 20 25 30 35 40

40

80

Ab/Ag
Min: 0
Max: 0.38
Median: 0
Mean: 0.1

(l)

P/fcAg
Min: 0
Max: 0.5
Median: 0.03
Mean: 0.07

0.6

0.1

ρy，vw (1×10 )
Min: 1
Max: 36.7
Median: 5
Mean: 7.1

(f)

5 10 15 20 25 30 35 40

20

(e)

Count

320
280
240
200
160
120
80
40
0
0
250
200

lw/tw
Min: 4.4
Max: 45
Median: 10.8
Mean: 13

(i)

Count

150
100

3

40

80

100

Aspect ratio
Min: 0.28
Max: 5
Median: 1.3
Mean: 1.6

(j)

3

(k)

20 40 60 80 100 120 140

ρy，h b (1×103)
Min: 0
Max: 169.2
Median: 3.7
Mean: 10.2

120

160

50
0
0
(m)

10

20

30

40

50

Section

1

2

3

4

5

(n)

Failure mode - first stage

24.5%

0.2

41.3%

0.4

0.8

0.2

0.3

0.4

0.5

(o) Failure mode - second stage
10.7%
39.8%

14%

61.5%
Barbell (B)
Flanged (F)
Rectangular (R)

200

205

58.7%

49.5%
Ductile failure mode (D)
Non-ductile failure mode (ND)

Diagonal tension failure （T）
Diagonal compression failure (C)
Sliding shear failure (S)

Figure 5: The distributions of the features in the dataset. (a) – (l): Numeric features; (m) – (o)
categorical features and targets
Furthermore, Figure 6 shows the matrix of Pearson’s linear correlations between the
features. Statistically, the circles in Figure 6 indicate one close relation among fy,hw, fy,cb, and
fy,vw, and another between CS and Ab/Ag. The latter’s close relation is in line with domain
knowledge—Ab/Ag—which highly depends on the cross-section type. The imbalanced crosssection type shown in Figure 5 (n) might also lead to the first relation since fy,hw, fy,cb, and fy,vw

9

are not highly linearly correlated in theory. It should be noted that the circled relations might
break the independence assumptions of some interpretability analysis methods, such as PDP.

210

215

Figure 6 Pearson’s correlation matrix of the 13 features
4.3 Classification using Auto-sklearn
This study applies Auto-sklearn, an open-source AutoML library (Feurer et al. 2020), to the
problem of RC shear wall failure mode identification. Auto-sklearn has been tested
successfully in a large study based on 209 metadata sets and 39 datasets (Feurer et al. 2020).
As shown in Figure 7, Auto-sklearn comprises 15 low-level ML algorithms, 14 preprocessing
methods, and four data preprocessing techniques. Auto-sklearn employs Bayesian optimization
to automate the processes of feature engineering, algorithm selection, parameter tuning, and
the entire modeling process shown in Figure 7. Therefore, AutoML can save human experts
from heavy reliance on data expertise (Elshawi et al. 2019). More details of Auto-sklearn can
be found in Feurer et al. (2015b).

10

220

Figure 7 Structural diagram and process details in Auto-sklearn
The training target is set to F1 accuracy, which is an unbiased performance indicator:
TP
Precision =
,
(1)
TP + FP
TP
Recall =
,
(2)
TP + FN
n

225

Macro=
F1 (∑ 2 ×
i =1

230

235

240

Recall(i ) × Precision(i )
)/n,
Recall(i ) + Precision(i )

(3)

where TP, TN, FP, and FN indicate true positives, true negatives, false positives, and false
negatives for the binary shear wall failure mode classification problem, respectively. It should
be noted that F1 is able to measure performance objectively when the class balance is skewed
(Li et al. 2017) and is widely adopted in the latest ML-related studies (Humphrey et al. 2022;
Chen & Xue 2023). After the AutoML model is trained, five manual ML algorithms—Naive
Bayes, Random Forest, SVM, Artificial Neural Network, and gradient boosting—that are
widely adopted in the literature are compared with the AutoML model according to the F1
value. In addition, 10-fold cross-validation is used to validate ML models for a limited sample
of data. Such a method is superior to split sample validation (Ellis & Mookim 2013; Wong &
Yeh 2020; Yuan et al. 2020) because using only a single split of full samples is sensitive to the
particular split sample created.
4.4 Interpretability analysis
Three interpretability analysis methods are implemented to explain the trained AutoML from
quantitative and qualitative aspects in this study. Two global model-agnostic interpretability
analysis methods are PDP and ALE, and a local model-agnostic interpretability analysis
method is LIME. Overall, the collective results of the three independent methods can jointly
verify and confirm the explainability of the trained AutoML model.
4.4.1 Partial dependent plot (PDP)

11

245

PDP illustrates the marginal impact of one or two features on an ML model’s predicted output
(Friedman 2001). PDP demonstrates whether a result and a feature have a linear, monotonic,
or complicated relationship. The partial dependence function is defined as (Molnar 2020):
∧
∧

=
f S ( xS ) E=
X C  f ( xS , X C ) 



250

255

260

265

270

275

∧

∫ f ( x , X )d ( X ) .
S

C

C

(4)

The xS are features related to the partial dependency function, whereas the XC are other features
utilized in the ML model f^. The features are handled as random variables. Partial dependency
operates by marginalizing the output of the ML model across the distribution of features in set
C, such that the function reveals the link between set S and the predicted result. By
marginalizing the features, one can get a dependency function on S features, including
interactions with other features. The PDP method is intuitive, clear, and easy to implement.
However, the PDP presumes the marginal feature set is uncorrelated with the feature set S.
If the assumption does not hold, the results of the PDP may be unreliable (Molnar 2020).
Figure 8 illustrates the estimation of PDPs. With the feature ‘aspect ratio’ as an example,
the estimating process is as follows:
1) Choose a feature, such as ‘Aspect ratio.’
2) Create a grid.
3) For each grid value:
a. Replace the feature with the grid value, such as setting all aspect ratios to 1.5 in
Figure 8(b).
b. Calculate the individual conditional expectation (light blue curves) for each
single data piece).
c. Average the predictions of an individual conditional expectation.
4) Complete a curve, shown as the black curve in Figure 8(a).

Figure 8 Example PDP analysis of the ‘aspect ratio’ feature. (a) PDP series plot; (b) each
point in (a) indicating the individual expectations with a condition of aspect ratio = 1.5
4.4.2 Accumulated local effects (ALE)
ALE plots average the changes in predictions (Apley & Zhu 2020), which could block the
effect of other features (Grömping 2020). Therefore, ALE plot is a less biased substitute for
PDP when the features are correlated. In ALE, the features are divided into intervals.
Essentially, ALE plots take the average of the changes in the predictions and aggregate them
throughout the intervals to evaluate local effects:
12

∧

∂ f ( xS , xC )
f S , ALE ( xS )= ∫ ( ∫
d (=
X C | X S zS )d )dzS − constant
z0,S
xC
∂xS

∧

280

285

290

295

300

xS

(5)

where zs denotes the interval value of the concerned feature of ALE, and other symbols are the
same as those in Eqn. (1).
Figure 9 shows ALE estimation of two correlated features X1 and X2. First, the feature
X1 is divided into intervals, indicated as vertical dash lines. The predicted data points are
replaced by the difference in each interval. The differences are accumulated and centered,
resulting in the final ALE curve. Overall, the ALE method is more complex and less intuitive
than PDP; therefore, ALE is often used together with the PDP method.

Figure 9 ALE results of two correlated features (X1 and X2)
4.4.3 Local interpretable model-agnostic explanations (LIME)
LIME (Ribeiro et al. 2016) trains local surrogate models, as opposed to global surrogate models
in PDP and ALE, to approximate the predictions of the underlying black box model. LIME
aims to fully understand the rationale behind the predictions of an ML model. First, LIME
creates a new dataset of samples’ perturbations and their predictions, then develops an
interpretable model of a weighted new dataset according to the closeness of the samples to the
instance of interest. Commonly interpretable ML models for LIME are linear regression,
logistic regression, and decision tree. The predictions from the ML model should be reasonably
approximated by the learned model at the local level.
Figure 10 shows the LIME algorithm’s sampling and local model training. The whole
process is based on a classification example using tabular data:
a) Prepare the ML prediction results based on features X1 and X2, such as the predicted
classes 1 (green) and 0 (orange) in Figure 10(a).
b) Select an instance of interest (big dot with cross sign) and small dots sampled.
c) Weigh the points that are located closer to the instance of interest.
d) Learn decision boundary Locally from the weighted samples, shown as the dashed line
in Figure 10(d).

13

305

310

315

320

325

Figure 10 LIME algorithm for local interpretability. (a) ML prediction results; (b) sampling;
(c) local model training; (d) local interpretability results
In summary, PDP and ALE are appropriate and complementary methods for explaining
the general mechanisms in the trained AutoML model. PDP is intuitive but better for less
correlated features, while ALE can work with correlated features. LIME is also an intuitive
local method and can explain individual predictions locally.
5 Experimental results
5.1 Experimental settings
All software was operated in single threading mode on a desktop computer with Intel i7-10700
2.9 GHz CPU, 32 GB memory, Python 3.7, and Ubuntu 18.04. The Auto-sklearn toolkit (ver.
0.14.7) was used, while five baseline ML algorithms were from OrangeML (ver. 3.32) (Demsar
et al. 2013). Three open-source tools—PDPbox (ver. 0.2), Alibi (ver. 0.8), and LIME library
(ver. 0.2.0)—were selected as efficient implementations of PDP, ALE, and LIME,
respectively.
5.2 Classification results of AutoML
Figure 11 compares the confusion matrices of AutoML against the baseline ML. In the
classification of ‘ductile (D)’ and ‘non-ductile (ND)’ modes, the Auto-sklearn model was the
best with an F1 = 90.6%, while the Auto-sklearn model was also the best (F1 = 84.5%) for
classifying the three types of ‘non-ductile (ND)’ modes. It should be noted that the next best
model in Figure 11a and the third best model in Figure 11b are Random Forest, which echoed
the findings in Mangalathu et al. (2020). The comparison results confirmed that AutoML is
automatic (saving manual interventions that relied on data expertise) and effective (less
erroneous) for the failure mode identification of RC shear walls.

14

330

335

340

345

Figure 11: Comparison of the confusion matrices of four manual ML and AutoML models
through 10-fold cross-validation of training: (a) first-stage classification; (b) second-stage
classification
The details in the Auto-sklearn training revealed more insights. In the model, an
ensemble model was built for each fold of prediction, in which different algorithms with
different hyperparameters are combined to achieve a better prediction performance. Figure 12
shows one ensemble model. In Figure 12, the ‘radom_forest’ classifier with a setting
‘max_features = 3, N_estimators = 512’ was one of the best low-level ML algorithms; the
‘gradient_boosing’ low-level classifier with a setting ‘early_stopping = true, L2_regularization
= 8.06, learning_rate = 0.092, max_iter = 128’ was another of the best low-level algorithms.
Human experts can, therefore, replicate the AutoML model manually using the tracked training
record.
5.3 Results of interpretability analyses of AutoML
5.3.1 Partial dependent plot (PDP)
Figure 13 shows the PDP curves of the trained AutoML model regarding all 11 assumingly
independent features. The x-axes in Figure 13 denote the features, while the quantiles of the
feature are the dash lines. The y-axes indicate the probabilities of non-ductile (ND) failure
modes, whereas a low value (e.g., y = 0.3) represents the ductile (D) failure mode. The
individual conditional expectation is illustrated with a light blue curve, while the PDP curve is
drawn as a thick black curve.

15

350

Figure 12: Details of the ensemble models built in a fold of AutoML training

Figure 13 PDP curves of 11 independent features regarding the trained AutoML model

16

355

360

365

370

375

380

Figure 13 (f) (i-k) shows that predictions of the AutoML model are significantly
influenced by four features, namely the boundary region vertical reinforcement ratio (ρvb), lw/tw,
aspect ratio, and axial load ratio (P/fcAg). First, the probability of non-ductile failure mode
increases with the increase in ρvb. The evidence of this phenomenon is that a larger ρvb leads to
higher flexural strength, while shear strength ensures stability. The shear walls are more likely
to fail in a non-ductile (shear-damage-based) mode with larger flexural strength (Gulec 2009).
In addition, Figure 13 (i) shows that the probability of non-ductile failure mode
obviously increases when lw/tw increases, especially when exceeding 13.3. lw/tw increases the
flexural strength due to the increase in vertical reinforcement moment with a stable crosssection of shear walls, but the shear strength does not change significantly as the area and
reinforcement amount of shear walls do not change (Lefas et al. 1990). When the threshold of
flexural strength exceeds shear strength, the probability of non-ductile failure mode may
increase obviously.
Figure 13 (j) illustrates that non-ductile failures are more likely to occur when the aspect
ratio is < 1.5; the probability of ductile failure is high when the aspect ratio exceeds 2. This
observation is generally consistent with the theory in (ASCE 2017), which stated that the aspect
ratio 1.5 (< 1.5 mostly controlled by shear damage (non-ductile) and 3.0 (>3.0 often controlled
by flexure damage [ductile]) are regarded as the thresholds of shear and flexure failure modes.
Finally, Figure 13(k) shows that the higher the P/fcAg , the larger probability of flexure
(ductile) failure mode predicted by the AutoML model. The trend can be explained by common
knowledge about the RC structure component design theory. That is, a proper axial load could
be helpful for the improvement of shear strength, and the flexural failure mode is more likely
to happen (Moehle 2015).
5.3.2 Accumulated Local Effects (ALE)
Figure 14 shows the ALE curves of the trained AutoML model, with background PDP curves
that were adjusted to start at zero. Overall, the trends of the PDP and ALE curves are similar
for most features. The interpretations of specified features by PDP curves could be explained
by the ALE. The remainder of this section focuses on two more features—Ab/Ag and CS—than
those verified by domain knowledge in Sect. 5.3.1.

17

ND

0.3

0.615

0.245

0.140
ALE

Probability

0.2

0.1

0.0

-0.1

D

385

390

395

400

(l)
Section_R

Section_B

Section

Section_F

Figure 14 Comparison of PDP and ALE curves of all features
Figure 14(l) shows that barbell and flanged walls are more prone to fail in a non-ductile
mode according to the trained AutoML model. Specifically, compared to RC walls with
rectangular cross-sections, walls containing boundary components—barbells or flanges—are
more vulnerable to diagonal compression failure, which is a non-ductile failure mode. This is
because flanged and barbell walls may accommodate a larger amount of reinforcement at the
ends of the wall, which, in turn, improve the wall’s flexural strength and raise the shear
demands in the wall web (Gulec 2009).
Surprisingly, Figure 14(m) shows that the probability trend does not change much when
Ab/Ag increases from 0 to 0.38, which is not anticipated. According to civil engineering
knowledge, larger Ab/Ag leads to a higher probability of non-ductile failure since non-ductilefailure-controlled flanged and barbell walls correspond to the larger Ab/Ag than rectangular
walls in practice.
5.3.3 Local interpretable model-agnostic explanations (LIME)
Figure 15 shows the interpretation of the trained AutoML model by LIME for two typical
instances (Nos. 275 and 219). The No. 219 instance is a ductile failure mode, while No. 275 is
a non-ductile failure mode. In contrast with the two global methods, LIME focuses on
individual prediction results. Figure 15 lists the overall contributions of all features of the
instance according to their importance, while the partial effects are illustrated. The tornado
charts in Figure 15 illustrate the trained Auto-sklearn model’s key decision features’ partial
effects and thresholds.

18

Feature

Value

Non-ductile instance

(a)

(b)

Ductile

ρvb>10

0.08

Feature

Non-ductile

Section=F

0.23
Aspect ratio<=1.00
0.23
lw/tw>15.71
0.21

Ductile instance
410

415

420

0.14
0.86

fy,cb>500
0.05
fy,vw>504
0.05

Value

Ductile Non-ductile

Section=R

0.35

405

Prediction
probabilities

P/fcAg>0.10

0.23

lw/tw<=8.00
0.12
10<ρvb<=20
0.12
1.33<Aspect ratio<=2.00
0.07
34.00<fc<=43.25
0.04

Prediction
probabilities
0.99
0.01

Figure 15 LIME explanations for two typical instances in the dataset. (a) A non-ductile
instance; (b) a ductile instance.
In Figure 15(a), the final classification was a non-ductile failure mode due to the high
prediction probability—0.86 > 0.50—corresponding to the non-ductile failure mode. Among
the top six features, the section, aspect ratio, lw/tw, fy,cb, and fy,vw contributed to the non-ductile
failure mode with scores at 0.23, 0.23, 0.21, 0.05, and 0.05, respectively. In contrast, ρvb
contributed to the ductile failure mode with a score of 0.08. The flanges section contributes
most to the non-ductile failure mode classification result. The evidence is that diagonal
compression failure, one type of non-ductile failure mode, is more likely to occur in walls with
boundary elements (such as barbells or flanges) (Gulec 2009). Meanwhile, the aspect ratio <=
1.00 contributes to the non-ductile failure mode, and it corresponds to the theory that the shear
walls with an aspect ratio < 1.5 are mostly controlled by shear damage (non-ductile) (ASCE
2017). The evidence of other features, including lw/tw and ρvb, can be found in the explanation
of Figure 13(i) and (f), respectively, in Sect. 5.3.1.
In Figure 15(b), all top six features (i.e., section, P/fcAg, lw/tw, ρvb, aspect ratio, and fc)
contribute to the ductile failure mode, of which the probability is at 0.99, and the corresponding
scores are at 0.35, 0.23, 0.12, 0.12, 0.07, and 0.04, respectively. The related explanation is also
consistent with existing knowledge. For instance, the rectangular section contributes most to
the ductile failure mode since it has smaller flexural strength and is more prone to flexural
failure mode (ductile mode).

19

425

430

435

440

445

450

455

In summary, PDP interpretability analysis indicates four uncorrelated features
influencing predictions: boundary region vertical reinforcement ratio (ρvb), lw/tw, aspect ratio,
and axial load ratio (P/fcAg). In detail, the probability of non-ductile failure mode increases
with the increase of ρvb. Non-ductile failures are more likely to occur when the aspect ratio is
< 1.5, but the probability of ductile failure is high when the aspect ratio exceeds 2. The higher
P/fcAg leads to a larger probability of flexure (ductile) failure mode of shear walls. In addition,
ALE curves exhibit similar trends as that in PDP analysis for most uncorrelated features. For
uncorrelated features, ALE shows that barbell and flanged walls are more prone to fail in a
non-ductile mode according to the trained AutoML model. Finally, the LIME provides a
detailed explanation for individual prediction, including importance contributions, partial
effects, and corresponding thresholds of all features.
6 Discussion
6.1 Application
As the major lateral load-carrying components in structure systems, shear walls have received
much attention due to the higher structural lateral resistance ability requirement for high-rise
buildings. However, the existing design process or retrofitting evaluation relies on computingintensive simulation and expensive experimental testing. The fast and reliable failure mode
identification of the shear walls and related explanations for prediction results can help improve
the existing practices in designing or retrofitting RC shear walls. As shown in Figure 16,
designers and engineers can first use AutoML tools to examine the failure mode of shear walls
after acquiring a preliminary design plan based on architectural requirements and structural
standards. Generally, the ductile failure mode is preferred to achieve better seismic
performance. Therefore, if they get non-ductile failure results in the first step, they can use the
interpretation to acquire direct information about how to reverse the failure mode by input
factors modifications. Especially if the tendency, contribution ratio, and threshold of each
feature are detailed and illustrated in LIME, designers could modify the feature according to
the interpretation result. This modification could be an iterative process until satisfactory
results are obtained, meaning the ductile failure mode of the shear wall. A similar iterative
modification process based on LIME interpretation analysis results is applied in the automotive
industry (Pana & Stark 2022).
The iterative RC wall design modification process in Figure 16 can thus take advantage
of the trained AutoML model and interpretability analysis. Compared to the conventional
computing-intensive simulation and expensive experiment, the novel iterative process can
improve traditional simulation-based and test-based design processes in terms of efficiency and
costs of money, manpower, and simulation and experiment resources (Chiarello et al. 2021).

20

460

465

470

475

480

485

Figure 16 An iterative RC wall design process based on the AutoML model and
interpretability analysis
6.2 Implications
By resolving the complex data-expertise-reliant modeling process and the untrustworthy ‘black
box’ of ML application in failure mode identification, the proposed integration of automated
ML and interpretability analysis has the potential to solve more complex AEC problems using
‘black box’ ML techniques. Example problems are the prediction of the behaviors of structural
elements and systems (Huang et al. 2022), structure design optimization (Fang et al. 2022),
structure health monitoring (Yang et al. 2022), off-site modular building production (Li et al.
2021b), tower crane optimization (Huang et al. 2021), site work packaging (Li et al. 2021a),
project-related solid waste generation behaviors (Yang et al. 2021), the semantic enrichment
of buildings and the city (Xue et al. 2021; Li et al. 2022), and the prediction of the mechanical
properties of the materials used in AEC (Nguyen et al. 2020).
6.3 Pros and Cons of this Study
This study has three advantages:
1. AutoML automates and improves the ML model’s establishment, including algorithm
selection, parameter tuning, and model evaluation. Auto-sklearn, the AutoML selected
in this study, outperformed other 5 manual algorithms in the prediction of the ML of
walls’ failure mode according to Sect. 5.2. In addition, Auto-sklearn is open source for
academia and industry.
2. The global interpretation methods tested in Sects. 5.3.1 and 5.3.2 visualized the input–
output correlations and sensitivities of the trained AutoML model. Civil engineers can
validate (or reject) the AutoML model against the features with domain knowledge. For
example, the interpretation of AutoML indicates that aspect ratios 1.5 (< 1.5 mostly
controlled by shear damage (non-ductile)) and 3.0 (> 3.0 often controlled by flexure
damage (ductile)) are regarded as the threshold of shear and flexure failure mode, which
is consistent with existing knowledge such as code (ASCE 2017). The local
interpretation could help users gain a deeper understanding of individual classification
results.
21

490

495

500

505

510

515

520

525

3. Last but not least, the interpretation presented in this study triangulates the learned ML
models with domain knowledge in civil engineering. Then, the mechanisms of the case
AEC problem can be interpreted in more comprehensive (against all features), visual,
and quantitative ways. For example, the quantified threshold of ρvb of RC shear walls
between the ductile and non-ductile failure modes was detected in PDP in Sect. 5.3.1,
whereas the only trend that greater ρvb leads to a greater probability of non-ductile
failure was known in past civil engineering knowledge.
However, there also exist four limitations in this study:
1. Compared with general AEC knowledge, the interpretation of the ML model focuses
on the marginal effect (i.e., the learned discrimination “areas” in the feature space). The
interpretability in this study may thus be unable to explain the whole feature space.
2. Although some interpretability methods, such as ALE, could be used to interpret the
marginal effect of correlated features, the correlation among different features may
influence the accuracy of the interpretability result. For example, the significance of
Ab/Ag was not captured in Sect. 5.3.2. One possible reason may be that too much weight
has been given to the section, which shares a similar meaning with the Ab/Ag.
3. Due to the small data samples, the time cost of different ML models is limited and,
therefore, is not taken into consideration. Time cost could be a good performance
indicator when meeting more complex AEC problems and larger datasets.
4. Several assumptions, such as balanced classes or continuous and average feature
distribution, were required for the interpretability analyses in this study. Features of
interest unable to meet the assumptions may result in unexpected curves or trends in
PDP or ALE plots. For example, in the ALE analysis of the feature Ab/Ag, the result was
not in line with the domain knowledge, where a possible reason was that half of the
instances in the dataset had low values (0) for feature Ab/Ag.
7 Conclusion
Although ML showed success in terms of error metrics in civil engineering, two fatal
problems—the complex data-expertise-relied modeling process and the untrustworthy ‘black
box’ nature—have undermined ML adoptions and applications due to critiques from civil
engineers and legislators. This paper presents an integration of the AutoML algorithm and
interpretability analysis to solve the two fatal problems with a complex AEC problem—the
identification of failure modes of RC shear walls. A training dataset of 351 cyclic tests of RC
shear walls was collected in this study. Based on the training dataset, 10-fold cross-validation
was utilized for training and validating the AutoML model. As a result, the trained AutoML
model outperformed the conventional ML algorithms. Regarding interpretability, two global
model-agnostic methods—PDP and ALE—and a local model-agnostic method, LIME, were
adopted to explain the trained AutoML model per feature and per instance, respectively. The
interpretability results confirmed that most observed trends and correlations in the trained
AutoML model were in line with domain knowledge. Overall, the proposed integration of the

22

530

535

540

545

550

555

560

AutoML algorithm and interpretability analysis is confirmed to address the two fatal issues for
the case problem.
The following conclusions can be drawn from this paper:
1) As a high-performance, easy-to-use, and open-source AutoML software, Auto-sklearn
was confirmed to be effective and fast in the failure identification of RC shear walls.
Evidence is that Auto-sklearn’s output outperformed five manual ML algorithms.
2) PDP analysis based on the partial effects confirmed a list of significant features,
including boundary region vertical reinforcement ratio (ρvb), lw/tw, aspect ratio, and axial
load ratio (P/fcAg). The features’ significance was consistent with long-established
domain knowledge. For example, low values (<1.5) of the aspect ratio led to a high
probability of non-ductile failure mode, and high values (>2) contributed to a high
probability of ductile failure mode.
3) ALE found similar partial effects with PDP for uncorrelated features, of which two
have different estimation mechanisms. For correlated features, such as Ab/Ag and
section, interpretation results show that flanged and barbell RC shear walls are more
prone to fail in a non-ductile mode.
4) LIME provides a local interpretation for an individual prediction. The partial effect of
each feature and related detailed evidence is given, which could provide detailed
design/retrofitting suggestions for a single building case.
The contribution to knowledge from this study lies mainly in structural engineering, but
also in applied computing. For the former, a theory-based, comprehensive data sampling is
provided. More importantly, integration of an AutoML method and interpretability analysis
provides an automatic and interpretable problem solution, which resolves the two fatal
problems—meaning the complex data-expertise-relied modeling process and the untrustworthy
‘black box’ nature—of the application of ML in AEC industry. For the latter, empirical
evidence is provided to the real-world performance and applicability of AutoML and three
interpretability methods for a complex AEC problem. In the future, the proposed integrated
workflow can be applied to extracting engineering information and knowledge from complex
AEC problems. In addition, big data and other advanced interpretability methods can be
adopted for interpreting complex AEC informatics with sophisticated ML methods.
Acknowledgment
This study was supported by Hong Kong Research Grants Council (RGC) (Nos. T22-504/21R, C7080–22GF)) and in part by Shenzhen Science Technology and Innovation Commission
(SZSTI) (No. SGDX20201103093600002). We appreciate the open data sources offered in the
literature listed in section 4.2.

565

References
Abdulridha, A. & Palermo, D. (2017). Behaviour and modelling of hybrid SMA-steel
reinforced concrete slender shear wall. Engineering Structures, 147(15), 77-89.
doi:10.1016/j.engstruct.2017.04.058
23

570

575

580

585

590

595

600

605

610

615

ACI. (2019). Building Code Requirements for Structural Concrete (ACI 318-19) and
Commentary. Farmington Hills, MI, USA: American Concrete Institute.
doi:10.14359/51716937
Anwar, T. (2021). COVID19 diagnosis using AutoML from 3D CT scans. 2021 IEEE/CVF
International Conference on Computer Vision Workshops (ICCVW) (pp. 203-507).
ICCV. doi:10.1109/ICCVW54120.2021.00061
Apley, D. W. & Zhu, J. (2020). Visualizing the effects of predictor variables in black box
supervised learning models. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 82(4), 1059-1086. doi:10.1111/rssb.12377
ASCE. (2017). Seismic Evaluation and Retrofit of Existing Buildings (ASCE/SEI 41-17).
Reston, Virginia, USA: ASCE. doi:10.1061/9780784414859
ATC. (1998). Evaluation of Earthquake Damaged Concrete and Masonry Buildings: Basic
Procedures Manual (FEMA 306). Washington, DC: Federal Emergency Management
Agency.
Bangaru, S. S., Wang, C., Hassan, M., Jeon, H. W. & Ayiluri, T. (2019). Estimation of the
degree of hydration of concrete through automated machine learning based
microstructure analysis – A study on effect of image magnification. Advanced
Engineering Informatics, 42, 100975. doi:10.1016/j.aei.2019.100975
Birely, A. C. (2012). Seismic Performance of Slender Reinforced Concrete Structural Walls.
Ann Arbor: University of Washington (Doctoral dissertation).
CEN TC-250. (2004). Eurocode-8: Design of Structures for Earthquake Resistance—Part 1:
General Rules, Seismic Actions and Rules for Buildings. Brussels: CEN.
Chen, S.-H. & Xue, F. (2023). Automatic BIM detailing using deep features of 3D views.
Automation in Construction, 148, 104780. doi:10.1016/j.autcon.2023.104780
Chiarello, F., Belingheri, P. & Fantoni, G. (2021). Data science for engineering design: State
of the art and future directions. Computers in Industry, 129, 103447.
doi:10.1016/j.compind.2021.103447
Chiou, Y. J., Mo, Y. L., Hsiao, F. P., Liou, Y. W. & Sheu, M. S. (2004). Behavior of high
seismic performance walls. 13th World Conference on Earthquake Engineering, (p. No.
3180). Vancouver.
Christidis, K. I. & Trezos, K. G. (2017). Experimental investigation of existing non-conforming
RC
shear
walls.
Engineering
Structures,
140(1),
26-38.
doi:10.1016/j.engstruct.2017.02.063
Dabbagh, H. (2005). Strength and Ductility of High-strength Concrete Shear Walls Under
Reversed Cyclic Loading. Sydney, Australia: University of New South Wales, Doctoral
dissertation.
Dashti, F., Dhakal, R. & Pampanin, S. (2014). Numerical simulation of shear wall failure
mechanisms. The 2014 New Zealand Society for Earthquake Engineering Conference.
Auckland,
New
Zealand.
Retrieved
from
https://www.nzsee.org.nz/db/2014/oral/22_Dashti.pdf
Deka, P. C. (2019). A Primer on Machine Learning Applications in Civil Engineering. New
York: CRC Press. doi:10.1201/9780429451423
Demertzis, K., Kostinakis, K., Morfidis, K. & Iliadis, L. (2023). An interpretable machine
learning method for the prediction of R/C buildings' seismic response. Journal of
Building Engineering, 63(Part A), 105493. doi:10.1016/j.jobe.2022.105493
Demsar, J., Curk, T., Erjavec, A., Gorup, C., Hocevar, T., Milutinovic, M., Mozina, M.,
Polajnar, M., Toplak, M., Staric, A., Stajdohar, M., Umek, L., Zagar, L., Zbontar, J.,
24

620

625

630

635

640

645

650

655

660

Zitnik, M. & Zupan, B. (2013). Orange: Data Mining Toolbox in Python. Journal of
Machine Learning Research, 14(1), 2349-2353.
Deng, M., Liang, X. & Yang, K. (2008). Experimental study on seismic behavior of high
performance concrete shear wall with new strategy of transverse confining stirrups.
14th World Conference on Earthquake Engineering. Beijing.
Ellis, R. P. & Mookim, P. G. (2013). K-Fold Cross-Validation is Superior to Split Sample
Validation for Risk Adjustment Models. Boston University-Department of Economics.
Elshawi, R., Maher, M. & Sakr, S. (2019). Automated machine learning: State-of-the-art and
open challenges. arXiv preprint, arXiv:1906.02287. https://arxiv.org/abs/1906.02287
Fang, C., Ping, Y., Gao, Y., Zheng, Y. & Chen, Y. (2022). Machine learning-aided multiobjective optimization of structures with hybrid braces–Framework and case study.
Engineering Structures, 269, 114808. doi:10.1016/j.engstruct.2022.114808
Farrar, C. R., Reed, J. W. & Salmon, M. W. (1993). Failure modes of low-rise shear walls.
Journal of Energy Engineering, 119(2), 119-138. doi:10.1061/(ASCE)07339402(1993)119:2(119)
Feng, D.-C., Wang, W.-J., Mangalathu, S. & Taciroglu, E. (2021). Interpretable XGBoostSHAP machine-learning model for shear strength prediction of squat RC walls. Journal
of
Structural
Engineering,
147(11),
04021173.
doi:10.1061/(ASCE)ST.1943541X.0003115
Feurer, M., Eggensperger, K., Falkner, S., Lindauer, M. & Hutter, F. (2020). Auto-sklearn 2.0:
The
next
generation.
arXiv
preprint,
arXiv:2007.04074,
24.
https://arxiv.org/abs/2007.04074, 24
Feurer, M., Springenberg, J. T. & Hutter, F. (2015b). Initializing Bayesian hyperparameter
optimization via metalearning. Proceedings of the AAAI Conference on Artificial
Intelligence. 29, pp. 1128-1135. Palo Alto, California USA: AAAI Press.
doi:10.1609/aaai.v29i1.9354
Friedman, J. (2001). Greedy function approximation: A gradient boosting machine. Annals of
Statistics, 29(5), 1189-1232. doi:10.1214/aos/1013203451
Gebreyohaness, A., Clifton, C., Butterworth, J. & Ingham, J. (2014). Experimental assessment
of inadequately detailed reinforced concrete wall components. ACI Structural Journal,
111(2), 279-290. doi:10.14359/51686520
Gijsbers, P., LeDell, E., Thomas, J., Poirier, S., Bischl, B. & Vanschoren, J. (2019). An open
source
AutoML
benchmark.
arXiv
preprint,
arXiv:1907.00909.
https://arxiv.org/abs/1907.00909
Gilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., Specter, M. & Kagal, L. (2018). Explaining
explanations: An overview of interpretability of machine learning. IEEE 5th
International Conference on data science and advanced analytics (DSAA) (pp. 80-89).
IEEE. doi:10.1109/DSAA.2018.00018
Grömping, U. (2020). Model-Agnostic effects plots for interpreting machine learning models.
Berlin: Beuth University of Applied Sciences Berlin.
Gulec, C. K. (2009). Performance-based Assessment and Design of Squat Reinforced Concrete
Shear Walls. Buffalo, NY, USA: State University of New York at Buffalo.
Guo, H., Zhang, Y. & Zhu, K. (2022). Interpretable deep learning approach for tool wear
monitoring in high-speed milling. Computers in Industry, 138, 103638.
doi:10.1016/j.compind.2022.103638
Hidalgo-Mompeán, F., Fernández, J. F., Cerruela-García, G. & Márquez, A. C. (2021).
Dimensionality analysis in machine learning failure detection models. A case study
25

665

670

675

680

685

690

695

700

705

with
LNG
compressors.
Computers
in
Industry,
128,
103434.
doi:10.1016/j.compind.2021.103434
Hosein Naderpour, M. M. (2021). Failure mode prediction of reinforced concrete columns
using machine learning methods. Engineering Structures, 248, 113263.
doi:10.1016/j.engstruct.2021.113263
Huang, C., Li, W., Lu, W., Xue, F., Liu, M. & Liu, Z. (2021). Optimization of multiple-crane
service schedules in overlapping areas through consideration of transportation
efficiency and operational safety. Automation in Construction, 127, 103716.
doi:10.1016/j.autcon.2021.103716
Huang, C., Li, Y., Gu, Q. & Liu, a. J. (2022). Machine learning–based hysteretic lateral forcedisplacement models of reinforced concrete columns. Journal of Structural
Engineering, 148(3), 04021291. doi:10.1061/(ASCE)ST.1943-541X.0003257
Humphrey, A., Kuberski, W., Bialek, J., Perrakis, N., Cools, W., Nuyttens, N., Elakhrass, H.
& Cunha, P. A. (2022). Machine-learning classification of astronomical sources:
estimating F1-score in the absence of ground truth. Monthly Notices of the Royal
Astronomical Society: Letters, 517(1), L116–L120. doi:10.1093/mnrasl/slac120
Hwang, S.-J., Tu, Y.-S., Yeh, Y.-H. & Chiou, T.-C. (2004). Reinforced concrete partition walls
retrofitted with carbon fiber reinforced polymer. ANCER Annual Meeting: Networking
of Young Earthquake Engineering Researchers and Professionals.
Jiang, H., Wang, B. & Lu, X. (2013). Experimental study on damage behavior of reinforced
concrete shear walls subjected to cyclic loads. Journal of Earthquake Engineering,
17(7), 958-971. doi:10.1080/13632469.2013.791895
Kabeyasawa, T. & Matsumoto, K. (1992). Test and analyses of ultra-high strength reinforced
concrete shear walls. Earthquake Engineering, 10th World Conference. Balkema.
Kimura, H., Sugano, S. & Nagashima, T. (1996). Seismic behavior of reinforced concrete
columns using ultra-high-strength concrete under high axial load. 4th Int. Sym. on the
Utilization of HS/HPC, (p. 184). Paris.
Kookalani, S., Cheng, B. & Torres, J. L. (2022). Structural performance assessment of GFRP
elastic gridshells by machine learning interpretability methods. Frontiers of Structural
and Civil Engineering, 1-18. doi:10.1007/s11709-022-0858-5
Kuang, J. & Ho, Y. (2008). Seismic behavior and ductility of squat reinforced concrete shear
walls with nonseismic detailing. ACI Structural Journal, 105(2), 225-231.
doi:10.14359/19738
Kuwajima, H., Yasuoka, H. & Nakae, T. (2020). Engineering problems in machine learning
systems. Machine Learning, 109(5), 1103-1126. doi:10.1007/s10994-020-05872-w
Le, T.-T. & Phan, H. C. (2020). Prediction of ultimate load of rectangular CFST columns using
interpretable machine learning method. Advances in Civil Engineering, 8855069.
doi:10.1155/2020/8855069
Lefas, I. D., Kotsovos, M. D. & Ambraseys, N. N. (1990). Behavior of reinforced concrete
structural walls: Strength, deformation characteristics, and failure mechanism.
Structural Journal, 87(1), 23-31. doi:10.14359/2911
Li, D.-C., Hu, S. C., Lin, L.-S. & Yeh, C.-W. (2017). Detecting representative data and
generating synthetic samples to improve learning accuracy with imbalanced data sets.
Plos One, 12(8), e0181853. doi:10.1371/journal.pone.0181853
Li, M., Xue, F., Wu, Y. & Yeh, A. G. (2022). A room with a view: Automatic assessment of
window views for high-rise high-density areas using City Information Models and deep
26

710

715

720

725

730

735

740

745

750

755

transfer
learning.
Landscape
and
Urban
Planning,
226,
104505.
doi:10.1016/j.landurbplan.2022.104505
Li, X., Chi, H. L., Lu, W., Xue, F., Zeng, J. & Li, C. Z. (2021a). Federated transfer learning
enabled smart work packaging for preserving personal image information of
construction
worker.
Automation
in
Construction,
128,
103738.
doi:10.1016/j.autcon.2021.103738
Li, X., Wu, L., Zhao, R., Lu, W. & Xue, F. (2021b). Two-layer Adaptive Blockchain-based
Supervision model for off-site modular housing production. Computers in Industry, 128,
103437. doi:10.1016/j.compind.2021.103437
Li, Z., Guo, H., Wang, W. M., Guan, Y., Barenji, A. V., Huang, G. Q., McFall, K. S. & Chen,
X. (2019). A Blockchain and AutoML Approach for Open and Automated Customer
Service. IEEE Transactions on Industrial Informatics, 15(6), 3642-3651.
doi:10.1109/TII.2019.2900987
Linardatos, P., Papastefanopoulos, V. & Kotsiantis, S. (2020). Explainable ai: A review of
machine learning interpretability methods. Entropy, 13(1), 18. doi:10.3390/e23010018
Lu, X., Yang, B. & Zhao, B. (2018). Shake-table testing of a self-centering precast reinforced
concrete frame with shear walls. Earthquake Engineering and Engineering Vibration,
17(2), 221-233. doi:10.1007/s11803-018-0436-y
Luo, H. & Paal, S. G. (2022). Data-driven seismic response prediction of structural components.
Earthquake Spectra, 38(2), 1382-1416. doi:10.1177/87552930211053345
Lyngdoh, G. A., Kelter, N.-K., Doner, S., Krishnan, N. A. & Das, S. (2022). Elucidating the
auxetic behavior of cementitious cellular composites using finite element analysis and
interpretable machine learning. Materials & Design, 213, 110341.
doi:10.1016/j.matdes.2021.110341
Mangalathu, S., Jang, H., Hwang, S.-H. & Jeon, J.-S. (2020). Data-driven machine-learningbased seismic failure mode identification of reinforced concrete shear walls.
Engineering Structures, 208, 110331. doi:10.1016/j.engstruct.2020.110331
Massone, L. M., Orakcal, K. & Wallace, J. W. (2009). Modelling of Squat Structural Walls
Controlled by Shear. ACI Structural Journal, 106(5), 646-655. doi:10.14359/51663105
Moehle, J. (2015). Seismic Design of Reinforced Concrete Buildings. New York: McGraw-Hill
Education.
Mohamed, N., Farghaly, A. S., Benmokrane, B. & Neale, K. W. (2014). Experimental
investigation of concrete shear walls reinforced with glass fiber–reinforced bars under
lateral cyclic loading. Journal of Composites for Construction, 18(3), A4014001.
doi:10.1061/(asce)cc.1943-5614.0000393
MOHURD. (2010). Technical Specification for Concrete Structures of Tall Building (JGJ 32010) (in Chinese). Beijing: Ministry of Housing and Vrban-Rural Development, PRC.
Molnar, C. (2020). Interpretable Machine Learning. A Guide for Making Black Box Models
Explainable. Victoria, Canada: Leanpub.
Montáns, F., Chinesta, F., Gomez-Bombarelli, R. & Kutz, J. N. (2019). Data-driven modeling
and learning in science and engineering. Comptes Rendus Mécanique, 347(11), 845855. doi:10.1016/j.crme.2019.11.009
Nakamura, N., Tsunashima, N., Nakano, T. & Tachibana, E. (2009). Analytical study on
energy consumption and damage to cylindrical and I-shaped reinforced concrete shear
walls subjected to cyclic loading. Engineering Structures, 31(4), 999-1009.
doi:10.1016/j.engstruct.2008.12.013
27

760

765

770

775

780

785

790

795

800

Naser, M. (2021). An engineer's guide to eXplainable Artificial Intelligence and Interpretable
Machine Learning: Navigating causality, forced goodness, and the false perception of
inference.
Automation
in
Construction,
129,
103821.
doi:10.1016/j.autcon.2021.103821
Nguyen, T.-D., Tran, T.-H. & Hoang, N.-D. (2020). Prediction of interface yield stress and
plastic viscosity of fresh concrete using a hybrid machine learning approach. Advanced
Engineering Informatics, 44, 101057. doi:10.1016/j.aei.2020.101057
Palermo, D., Vecchio, F. J. & Solanki, H. (2002). Behavior of three-dimensional reinforced
concrete shear walls. ACI Structural Journal, 99(1), 81-89. doi:10.14359/11038
Pana, Y. & Stark, R. (2022). An interpretable machine learning approach for engineering
change management decision support in automotive industry. Computers in Industry,
138, 103633. doi:10.1016/j.compind.2022.103633
Paulay, T. & Priestley, M. J. (1992). Seismic Design of Reinforced Concrete and Masonry
Buildings (Vol. 768). New York: Wiley.
Paulay, T. (1975). Design aspects of shear walls for seismic areas. Canadian Journal of Civil
Engineering, 2(3), 321-344. doi:10.1139/l75-030
Ren, F., Chen, J., Chen, G., Guo, Y. & Jiang, T. (2018). Seismic behavior of composite shear
walls incorporating concrete-filled steel and FRP tubes as boundary elements.
Engineering Structures, 168(1), 405-419. doi:10.1016/j.engstruct.2018.04.032
Ribeiro, M. T., Singh, S. & Guestrin, C. (2016). "Why should I trust you?": Explaining the
predictions of any classifier. The 22nd ACM SIGKDD international conference on
knowledge
discovery
and
data
mining
(pp.
1135-1144).
ACM.
doi:10.1145/2939672.2939778
Ribeiro, M. T., Singh, S. & Guestrin, C. (2016). Model-agnostic interpretability of machine
learning. arXiv:1606.05386. https://arxiv.org/abs/1606.05386
Salonikios, T. N., Kappos, A. J., Tegos, L. A. & Penelis, G. G. (1999). Cyclic load behavior of
low-slenderness reinforced concrete walls: Design basis and test results. ACI Structural
Journal, 96(4), 649-660. doi:10.14359/703
Singh, P. (2021). Data Mining techniques and its application in civil engineering—A review.
In A. K. Verma & P. K. Kapur, Asset Analytics (pp. 175-183). Singapore: Springer.
doi:10.1007/978-981-16-0037-1_15
Somala, S. N., Chanda, S., Karthikeyan, K. & Mangalathu, S. (2021). Explainable Machine
learning on New Zealand strong motion for PGV and PGA. Structures, 34, 4977-4985.
doi:10.1016/j.istruc.2021.10.085
Su, M., Peng, H. & Li, S. (2021). A visualized bibliometric analysis of mapping research trends
of machine learning in engineering (MLE). Expert Systems with Applications, 186,
115728. doi:10.1016/j.eswa.2021.115728
Su, R. & Wong, S. (2007). Seismic behaviour of slender reinforced concrete shear walls under
high
axial
load
ratio.
Engineering
Structures,
29(8),
1957-1965.
doi:10.1016/j.engstruct.2006.10.020
Sujith Mangalathu, J.-S. J. (2018). Classification of failure mode and prediction of shear
strength for reinforced concrete beam-column joints using machine learning techniques.
Engineering Structures, 160, 85-94. doi:10.1016/j.engstruct.2018.01.008
Terzioglu, T., Orakcal, K. & M.Massone, L. (2018). Cyclic lateral load behavior of squat
reinforced
concrete walls. Engineering Structures, 160(1), 147-160.
doi:10.1016/j.engstruct.2018.01.024
28

805

810

815

820

825

830

835

840

845

Tokunaga, R. & Nakachi, T. (2012). Experimental study on edge confinement of reinforced
concrete core walls. 15th World Conference on Earthquake Engineering. Lisbon.
Tomazevic, M., Lutman, M., Capuder, F. & Petkovic, L. (1996). Seismic behaviour of RC
shear-walls: an experimental study. 11th World conference on Earthquake Engineering,
(p. No.381). Acapulco, Mexico.
Tran, N. K., Howard, T., Walsh, R., Pepper, J., Loegering, J., Phinney, B., Salemi, M. R. &
Rashidi, H. H. (2021). Novel application of automated machine learning with MALDITOF-MS for rapid high-throughput screening of COVID-19: A proof of concept.
Scientific Reports, 11, 8219. doi:10.1038/s41598-021-87463-w
Tripathi, M., Dhakal, R. P. & Dashti., F. (2019). Bar buckling in ductile RC walls with different
boundary zone detailing: Experimental investigation. Engineering Structures, 198,
109544. doi:10.1016/j.engstruct.2019.109544
Usta, M., Pujol, S., 445B, A. S., Puranam, A., Song, C. & Wang, Y. (2017). ACI 445B Shear
Wall Database. Purdue University Research Reposito. doi:https://doi.
org/10.4231/R7HH6H39
Vadyala, S. R., Betgeri, S. N., C.Matthews, J. & Matthews, E. (2022). A review of physicsbased machine learning in civil engineering. Results in Engineering, 13, 100316.
doi:10.1016/j.rineng.2021.100316
Vakhrushev, A., Ryzhkov, A., Savchenko, M., Simakov, D., Damdinov, R. & Tuzhilin, A.
(2021). LightAutoML: AutoML solution for a large financial services ecosystem.
arXiv:2109.01528v2. https://arxiv.org/abs/2109.01528v2
Vecchio, F. J. & Collins, M. P. (1986). The modified compression-field theory for reinforced
concrete elements subjected to shear. ACI Journal, 83(2), 219-231. doi:10.14359/10416
Wen-yuan, Z., Ke, W., Qiang, W., Yong, C., Yu, Z. & Yu-kun, D. (2018). Aseismic behavior
of composite shear wall with stiffened double steel plates and infilled concrete.
Engineering Mechanics, 35(11), 125-133. doi:10.6052/j.issn.1000-4750.2017.06.0482
Wong, T.-T. & Yeh, P.-Y. (2020). Reliable Accuracy Estimates from k-Fold Cross Validation.
IEEE Transactions on Knowledge and Data Engineering, 32(8), 1586-1594.
doi:10.1109/TKDE.2019.2912815
Xue, F., Wu, L. & Lu, W. (2021). Semantic enrichment of building and city information models:
A ten-year review. Advanced Engineering Informatics, 47, 101245.
doi:10.1016/j.aei.2020.101245
Yanez, F., Park, R. & Paulay, T. (1991). Seismic behavior of reinforced concrete structural
walls with irregular openings. Pacific Conference of Earthquake Engineering, (pp. 6778). New Zealand.
Yang, X., Gao, Y., Fang, C., Zheng, Y. & Wang, W. (2022). Deep learning‐based bolt
loosening detection for wind turbine towers. Structural Control and Health Monitoring,
29(6), e2943. doi:10.1002/stc.2943
Yang, Z., Xue, F. & Lu, W. (2021). Handling missing data for construction waste management:
machine learning based on aggregated waste generation behaviors. Resources,
Conservation and Recycling, 175, 105809. doi:10.1016/j.resconrec.2021.105809
Yuan, L., Guo, J. & Wang, Q. (2020). Automatic classification of common building materials
from 3D terrestrial laser scan data. Automation in Construction, 110, 103017.
doi:10.1016/j.autcon.2019.103017
Zhang, M., Lia, M., Zhang, J., Liu, L. & Li, H. (2020). Onset detection of ultrasonic signals
for the testing of concrete foundation piles by coupled continuous wavelet transform
29

850

855

and machine learning algorithms. Advanced Engineering Informatics, 43, 101034.
doi:10.1016/j.aei.2020.101034
Zhao, R., Yang, Z. & Xue, F. (2023). Automated Machine Learning for the construction
industry: Outperformance in classification and regression benchmarking tests with a
practical guideline. Building Research & Information, (under review).
Zhou, Y., Lu, X. & Dong, Y. (2010). Seismic behaviour of composite shear walls with multiembedded steel sections. Part I: experiment. The Structural Design of Tall and Special
Buildings, 19(6), 618-636. doi:10.1002/tal.597

30