Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Engineering Structures 301 (2024) 117295

Contents lists available at ScienceDirect

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

Machine learning predictions of code-based seismic vulnerability for


reinforced concrete and masonry buildings: Insights from a 300-building
database
Angelo Aloisio a ,∗, Yuri De Santis a , Francesco Irti b , Dag Pasquale Pasca c , Leonardo Scimia d ,
Massimo Fragiacomo a
a
Department of Civil, Construction-Architectural and Environmental Engineering, Università degli Studi dell’Aquila, L’Aquila, Italy
b
Territorial Company for Residential Buildings of the L’Aquila Province, L’Aquila, Italy
c
Norsk Treteknisk Institutt (Norwegian Institute of Wood Technology), Oslo, Norway
d
Municipality of L’Aquila, L’Aquila, Italy

ARTICLE INFO ABSTRACT

Keywords: This paper proposes a data-driven model for predicting the code-based seismic vulnerability index calibrated
Seismic vulnerability index on a dataset comprising almost 300 buildings. The vulnerability index, estimated following the Italian Seismic
Data-driven model Code, involved rigorous investigations, including geometric surveys, experimental tests, and numerical mod-
Italian Seismic Code
elling. Leveraging data from these investigations, encompassing approximately 15 categorical and numerical
Artificial neural networks
explanatory variables, the authors developed several regression and classification predictive models, such as
Binary classification
Logistic Regression and Artificial Neural Networks (ANN). The optimal models perform binary classification
to determine the categorization into two macro-classes, defined by an arbitrary vulnerability threshold. The
ANN model stands out as the best performer. When adjusting the vulnerability threshold to obtain a balanced
dataset, such a model achieves an accuracy higher than 85%. The paper also discusses the importance of
each feature by calculating the SHapley Additive exPlanations (SHAP) values. The proposed model can aid
decision-makers in allocating resources effectively to mitigate seismic risks of built environments.

1. Introduction Mechanics-based seismic vulnerability analyses are commonly used


to prioritize seismic retrofitting interventions. These analyses encom-
In seismic analyses, the vulnerability index is a measure used to pass distinct stages, beginning with knowledge acquisition and ex-
assess the potential damage or performance of structures subjected tending to building modelling, followed by structural analysis. The
outcomes of these analyses result in vulnerability indices, expressing
to earthquakes. It quantifies the level of vulnerability of a building
the ratio between structural capacity and demand. Many mechanics-
or infrastructure and helps engineers and decision-makers prioritize
based approaches with reduced computational efforts [8] have been
retrofitting and risk reduction measures. Seismic vulnerability assess-
developed for seismic vulnerability assessment of existing buildings at
ment commonly employs three main approaches [1–3]: (i) damage
a regional scale [9,10]. In the case of masonry, for instance, macro-
probability matrices, (ii) vulnerability indices, and (iii) methods based
models are less demanding from a computational standpoint, resulting
on capacity curves. These methods often facilitate the generation of
in suitable practice-oriented analyses of large structures [11,12]. The
practical fragility curves [4,5]. For large-scale assessments, hazard-
outcomes of these analyses are crucial in prioritizing retrofitting inter-
based and physical vulnerability indexes are commonly used. The
ventions and identifying critical scenarios. Nevertheless, the vulnera-
hazard-based index correlates seismic hazard levels (e.g., peak ground
bility indexes represent a semi-physical measure due to hypotheses and
acceleration, spectral acceleration) with expected damage or loss. On
approximations adopted in the analysis [13].
the other hand, the physical vulnerability index considers building
If adequately accurate, data-driven models can simplify the process
characteristics like height, age, materials, structural type, and quality
of prioritizing seismic retrofitting interventions without the need for
to quantify susceptibility to damage [6,7].
elaborate vulnerability analyses, especially considering the similarity

∗ Corresponding author.
E-mail address: angelo.aloisio1@univaq.it (A. Aloisio).

https://doi.org/10.1016/j.engstruct.2023.117295
Received 28 August 2023; Received in revised form 11 November 2023; Accepted 4 December 2023
Available online 16 December 2023
0141-0296/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Aloisio et al. Engineering Structures 301 (2024) 117295

in existing building stocks. Having reliable and robust models that metrics tailored to datasets with class imbalance. As a result, while
predict vulnerability class or index based on building characteristics accuracy rates sometimes surpass 95%, the actual performance for
and expected seismic hazard can be highly valuable when the goal is the minority class barely reaches 70%. Notably, their research focuses
to determine where to allocate resources [14–16]. on data derived from post-earthquake assessments. Similarly, Aloisio
Recently, many researchers in structural seismic vulnerability have et al. [51] developed ML models to predict the seismic risk class, given
applied diverse techniques, including machine learning (ML) methods, the recorded damage information retrieved from the AeDES database
to develop empirical vulnerability prediction models [17–19]. The data after the earthquake in L’Aquila.
feeding the training sets are generally retrieved from in-field earth- Beyond the issue of damage, empirical data may also be used
quake observations or numerical analyses. Observations of past damage for impact assessments of social [52–55] and economic loss [56,57].
have primarily been used to develop functions that relate the intensity Recently, statistical analyses of observed damage and usability data
of an event to the damage caused to a building in terms of damage have also enabled models to be calibrated to predict the usability of
probability matrices or empirical fragility curves [20–26]. Numerous a building [58,59]. Nevertheless, despite these advances, as remarked
illustrative instances exist within this research field. After the 2015 by [50], models still generally lack a focus on trying to understand
Nepal earthquake, Ghimire et al. (2022) [27] combined inspection the likely outcomes of subjective decisions on building usability. These
data with artificial intelligence techniques to construct a regional pre- studies have collectively contributed to enhancing the empirical seismic
diction model [28]. Addressing inherent uncertainties in structural
vulnerability assessment. However, their focus has revolved around ad-
damage assessment, Guo et al. (2022) [29] introduced a fuzzy global
dressing structural vulnerability through in-field observations without
seismic vulnerability analysis framework. For earthquake-prone areas,
referring to code-based vulnerability assessments.
Bernardo et al. (2022) [30] and Rezvani et al. (2022) [31] developed
In contrast, this paper deals with a relatively less-explored dimen-
three-dimensional finite element models for representative structures,
sion: the reliability of data-driven models in predicting the code-based
deriving vulnerability comparison curves through nonlinear and prob-
seismic vulnerability assessment (not empirical damage data) of ex-
abilistic modal analyses. In the aftermath of the Azores earthquake in
isting structures using readily available explanatory variables. Since
1998, Ferreira et al. (2017) [32] used global geographic information
many existing buildings do not conform to building codes, public
systems to investigate post-earthquake building conditions, leading
administrations’ resource allocation often leans towards supporting en-
to an empirical vulnerability model based on 192 buildings. Further
contributions come from Demirel et al. (2022) [33], who conducted gineers conducting simplified vulnerability analyses over intervention
an on-site seismic loss inspection following the Samos earthquake, initiatives. This study evaluates the feasibility of data-driven models
collecting seismic damage data for typical reinforced concrete struc- in estimating the vulnerability index or classification, capitalizing on
tures in the Izmir region. Shafapourtehreny et al. (2022) [34] analysed easily accessible structural information. This research uses a compre-
regional structural earthquake damage data in Istanbul, employing ML hensive database of nearly 300 buildings. The aim is to calibrate
techniques to generate vulnerability regression and prediction curves. regression and classification models through various methods, from
Sheshov et al. (2022) [35] conducted a field survey rating vulnerability traditional logistic or linear regression techniques to advanced ML
in Dulles, Albania, while Debnath et al. (2022) [36] observed unrein- algorithms.
forced masonry structures damaged in the 2017 Anbar earthquake in The database comprises reinforced concrete (RC) and masonry
India. Misseri et al. (2020) [37] validated an element model for wood- buildings, which underwent evaluation with funds from ATER, the Re-
soil reinforced structures using field inspection data from the 2010 gional Agency for Residential Construction of the province of L’Aquila
Maule earthquake in Chile. Ebrahimiyan et al. (2022) [38] presented (Italy). Following the Italian Seismic Code, engineers, supervised by
seismic loss data of RC structures under typical near- and far-field ATER, assigned seismic vulnerability indices to each building. The
earthquakes, creating damage models using soft computing techniques. assessment process followed a systematic approach involving acquiring
Pitilakis and Petridis (2022) [39] considered nonlinear dynamic fea- knowledge and subsequent structural analysis. This research has several
tures of building types, establishing seismic vulnerability comparison elements of originality. At its core, real-world data sets it apart, given
curves for existing RC structures. Gautam et al. (2022) [40] analysed the prevalent reliance on numerical studies that often employ structural
survey data from Bhutan’s earthquakes, generating vulnerability mode archetypes for broader generalizations. While data-driven methodolo-
comparison curves for group masonry structures. Italy saw significant gies increasingly gain traction in various domains, their exploration
progress in empirical seismic vulnerability research. Del Gaudio et al. remains relatively embryonal in seismic vulnerability prediction. The
(2017) [41] analysed RC buildings in the 2009 L’Aquila earthquake, crucial point is focusing on non-conforming buildings since the dataset
assessing vulnerability levels and developing statistical distributions. has vulnerability indexes below one.
Formisano and Chieffo (2022) [42] explored the vulnerability of an- The paper is structured as follows: It briefly introduces the method-
cient buildings in Morris, Italy, while Cardinali et al. (2022) [43] ology for conducting seismic vulnerability analyses on individual build-
assessed masonry structure damage in Florence, conducting empir- ings (Section 2). Section 3 presents an initial exploration of the
ical vulnerability and nonlinear static analyses. In China, frequent database. The results are shown in Section 4, divided into two sub-
earthquakes prompted many empirical vulnerability studies. Qu et al. sections: the first focusing on regression models and the second on
(2015) [44] and Sun et al. (2019) [45] investigated structural damage classification models. Finally, the last section deals with the ML model
after the 2013 Lushan earthquake, obtaining empirical fragility matri- interpretability.
ces. Gong et al. (2015) [46] studied China’s 2010 Yushu earthquake,
developing empirical fragility databases. Li et al. [7,47,48] investigated
structural damage from the 2008 Wenchuan earthquake, creating em- 2. Problem formulation
pirical vulnerability curves and matrices. Xi (2022) [49] established a
relationship between the building structure and economic vulnerability The paper aims to predict the vulnerability index for a dataset
risk parameters post-Wenchuan earthquake. comprising nearly 300 buildings. The vulnerability index has been
Recently, Tocchi et al. [50] explored the application of ML tech- determined following the Italian Seismic Code. The input data from
niques for post-earthquake building usability assessments. Their paper these investigations, encompassing 15 features, is the foundation for
underscores the challenges of predicting risk classes after a seismic developing regression and classification models. This section outlines
event, particularly with imbalanced datasets. In such datasets, the the methodologies employed to evaluate the seismic vulnerability of
majority class (no damage) vastly overshadows the damage classes. the examined building stock. Subsequently, the following subsection
However, the authors overlooked the need for appropriate performance presents the mathematical background of the predictive models.

2
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 1. Examples of destructive and non-destructive tests to achieve adequate structural knowledge: (a) extraction of a concrete core; (b) estimation of the rebar diameter; (c)
measurement of the spacing of the stirrups; and (d) evaluation of the floor behaviour (Pictures taken by the authors).

2.1. Vulnerability index aligning closely with the Eurocodes. On the other hand, the demand is
computed through modal superposition, using the Complete Quadratic
The following equation defines the seismic vulnerability index (𝛼): Combination (CQC) method [61], as per Italian Standards. Specifically,
𝐶(𝒙, 𝜣) the seismic action’s effects are calculated for each vibration mode and
𝛼= (1)
𝐷(𝒙, 𝜣) combined. All modes with a participating mass greater than 5% are
Here, 𝐶 represents the structural capacity, 𝐷 is the structural demand, considered, reaching the total participating mass of at least 85%.
𝒙 collects the deterministic model parameters, and 𝜣 is the unknown
model parameters. Numerous methods exist for evaluating 𝛼. In Italy, 2.1.1. Input data
professional engineers must adhere to a systematic procedure provided In this phase, all building and soil characteristics data are collected.
by the seismic code. Acknowledging that the vulnerability index is This phase, often called the knowledge phase, holds paramount signif-
not an objective metric is essential. It is markedly influenced by the icance. In this context, varying levels of knowledge can be achieved,
engineers conducting the assessment. Subjectivity and conventionality categorized as limited, sufficient, and precise based on the increasing
become integral to the process from various perspectives. Subjectiv- levels of information. These levels are closely tied to corresponding
ity manifests as personal judgment regarding appropriate modelling confidence factors outlined by regulatory documents [62]. The path of
choices, substantially influencing the final estimation. In this instance, knowledge can be traced through the following steps:
predefined approaches were established in collaboration with the fund-
ing agency ATER. Conventionality manifests in adopting confidence • Identifying the construction, location concerning specific risk ar-
factors based on the acquired level of knowledge. The investigation eas, and integration within the urban context. This involves sur-
comprised two main tasks: (i) collecting all input data and (ii) con- veying the building and identifying valuable elements that could
ducting the structural analysis to compute the seismic vulnerability impact the risk level. It also entails performing a geometric survey
index. of the current state of the construction, including any cracks
The seismic vulnerability index is obtained from linear dynamic and deformation occurrences. Additionally, materials are sam-
analyses with the site-specific elastic acceleration response spectra pled to comprehensively identify the load-bearing structure while
provided by the Italian Seismic Code. The index is the smallest value considering their quality and preservation condition.
among the vulnerability indices associated with all 𝑛 structural ele- • Investigating the construction’s evolution involves studying the
ments for each building. This can be mathematically expressed as: sequence of transformation phases from its original configuration
{ } to its present state.
𝐶1 𝐶2 𝐶 • Identifying seismic-resistant components, paying close attention
𝛼(𝒙, 𝜣) = min , ,…, 𝑛 (2)
𝐷1 𝐷2 𝐷𝑛 to construction techniques, structural details, and connections
Here, 𝐶𝑖 and 𝐷𝑖 indicate the capacity and demand of the 𝑖th structural between elements.
element, respectively. The capacity for each structural element is calcu- • Analysing the materials, along with their level of deterioration
lated from the analytical expressions specified in the Italian Code [60], and mechanical properties.

3
A. Aloisio et al. Engineering Structures 301 (2024) 117295

was performed to evaluate the compressive strength of concrete in


reinforced concrete buildings. The soil properties were mainly re-
trieved from nearby buildings. This allowed evaluation of the soil class
according to the Italian Seismic Code.

2.1.2. Structural analyses


The second phase takes all input data to develop a representa-
tive structural model. Linear dynamic analyses with seismic response
spectra have been carried out on the entire dataset. Specifically, the
structural analysis comprised the following sub-phases:

• Determine the response spectrum: Compute the design response


spectrum based on the site-specific hazard parameters specified
by the Italian Seismic code.
• Model the building: Develop a structural model of the building.
In this study, commercial software was employed for building
modelling using the finite element method.
• Conduct linear dynamic analysis and assess seismic demand: Ap-
ply the response spectrum and estimate the seismic demand for
all structural elements.
• Evaluate building’s capacity: Calculate the seismic capacity of all
structural elements.
• Compute the seismic vulnerability index.

In essence, the unified approach prescribed by ATER consists of


performing linear dynamic analyses and adopting the FE method to
perform the structural analysis, given a knowledge level LC2. Regard-
ing the modelling choices for the RC structures, a 3D frame mod-
elling approach was employed. The contribution of the infill was ne-
glected, following a classical simplified approach. The well-known
Frame-by-Macro-Element (FME) approach was adopted for masonry
structures [63]. Piers and spandrels are modelled as deformable macro-
elements. In contrast, the nodes between vertical elements and hori-
zontal ones are considered rigid parts, considering that they exhibited
very little damage under earthquakes. After the whole macro-elements
structural model is set up, it is transformed into the classic Equivalent
Frame Model (EFM).

2.2. Data-driven models

The authors aim to understand if the vulnerability index, despite


being influenced by subjective choices and conventional approaches,
can be predicted using readily available numerical and categorical ex-
planatory variables. Therefore, the authors have developed regression
and classification models. The regression models can be formulated as
follows:

𝛼 = 𝑓 (𝒙, 𝜣) + 𝜖 (3)

Fig. 2. Colour maps of municipalities in the L’Aquila province, Abruzzo (Italy), where 𝑓 is the regression function, and 𝜖 represents the error term,
considered in this investigation. (a) represents the number of analysed buildings for accounting for the difference between the predicted value from the
each municipality, and (b) displays the transition from red to green, corresponding to regression function and the actual calculated value of 𝛼. The authors
the average seismic vulnerability index. (For interpretation of the references to colour will use the following state-of-the-art models: (i) Linear regression [64],
in this figure legend, the reader is referred to the web version of this article.)
(ii) Decision Tree Regression [65], (iii) Support Vector Regression
(SVR) [66], (iv) Artificial Neural Network (ANN) [67]. Specifically,
linear regression is a fundamental statistical method that assumes a lin-
• Gaining insight into the subsoil and foundation structures while ear relationship between the predictor variables and the response [68].
considering any changes or failures over time. Regression trees are a type of decision tree algorithm used for regres-
The current analyses are based on a knowledge level LC2 [60] for sion tasks. Such models work by recursively partitioning the data into
all buildings, resulting in a confidence factor for the material properties subsets based on the predictor variables. Each partition forms a tree
equal to 1.2. Fig. 1 shows examples of destructive and non-destructive leaf, and each leaf’s mean vulnerability index is used as the predicted
tests to achieve adequate structural knowledge: (a) extraction of a value [69]. Regression trees can capture non-linear relationships and
concrete core; (b) estimation of the rebar diameter; (c) measurement interactions among variables. SVR is a regression technique that uses
of the spacing of the stirrups and (d) evaluation of the floor behaviour. support vector machines to find a hyperplane that best fits the data
Certified laboratory testing facilities and companies conducted both while minimizing the prediction error [70]. SVR is particularly ef-
destructive and non-destructive tests. The assessment primarily encom- fective when dealing with high-dimensional data, which is the case,
passed methods such as Pachometer and ultrasonic testing, radiogra- and can handle both linear and non-linear relationships. ANNs are
phy, and infrared thermography. Additionally, concrete core extraction models inspired by the human brain’s structure and can learn complex

4
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 3. Four samples RC buildings in (a) Celano, (b) Cerchio, (c) San Benedetto dei Marsi and (d) Rocca di Mezzo (Abruzzo, Italy).

relationships in data. They consist of interconnected nodes (neurons) where 𝑓 (𝒙) is the output or score generated by the classification model
arranged in layers. ANNs can capture non-linear relationships and for the given input features 𝒙, and ̂ is the predicted class label that
interactions between variables effectively [71]. corresponds to the highest output value among all possible classes. The
For Linear Regression, the authors used ordinary least squares to classification model is trained using a labelled dataset, associating each
estimate the linear model coefficients. The CART algorithm (Classi- building with its corresponding class label. The model learns to link
fication and Regression Trees) [65] was adopted for the regression input features with their respective class labels through training. Once
trees. The algorithm recursively partitions the data based on predictor trained, it can predict class labels for new, unseen buildings based on
variables, optimizing for reduction in mean squared error. For SVR, their features.
the authors calibrated the model using the Sequential Minimal Opti- The authors employed three models: (i) multinomial logistic re-
mization (SMO) algorithm [72]. For the ANN, the authors implemented gression [78], (ii) classification trees [65], (iii) ANN and (iv) XGBoost
a multi-layer neuron perception network with Adam optimizer [73], (Extreme Gradient Boosting) [79]. Multinomial logistic regression ex-
which adapts the learning rate dynamically based on the training tends binary logistic regression to handle multiple categories. This
progress. The chosen activation function exhibiting the best perfor- method employs input features to predict the likelihood of a building
mance was the Rectified Linear Units (ReLU). The default learning rate belonging to different vulnerability classes [78]. The predicted class
was 0.001, and the epochs were set to 100 [74–76]. The training and label is the one with the highest likelihood. This model captures re-
validation process employed the hold-out method, where the dataset lationships between input features and categorical classes. The other
two models were introduced in the preceding paragraphs. XGBoost [79]
was split into a training set and a validation set, with an 80% to
is a ML algorithm that uses a gradient boosting framework. It is
20% ratio, respectively. The authors then calculated the models’ perfor-
designed for speed and performance and is widely used for structured
mance metrics on the validation set. To assess convergence, the authors
and tabular data. The default hyperparameters provided by Scikit-
monitored the performance metrics in the training and validation sets
Learn [80] were adopted. The authors applied the same approaches for
during each training iteration, setting a threshold equal to 1% [71]. The
model calibration using the hold-out method. In the case of multinomial
authors employed the integer encoding method to manage categorical
logistic regression, calibration involves maximizing the log-likelihood,
variables. This technique converts categorical variables into progressive
following a procedure outlined in [81]. This method is highly efficient
integers, each representing a specific category within the categorical
for concave log-likelihood functions and utilizes the Newton–Raphson
variable [77].
technique.
Given the challenges in accurately assessing the exact vulnerability
index, the authors adopted a classification approach in the second step. 3. Analysis of the dataset
In this approach, vulnerability indices are grouped into distinct classes,
and the model’s objective is to predict the corresponding class for each The considered buildings, comprising nearly 300 reinforced con-
building. The goal is to predict the class label for each building based crete (RC) and masonry buildings, are located in the province of
on its features. The vulnerability index, denoted as 𝛼, is divided into L’Aquila in Abruzzo (Italy). Fig. 2 plots the geographic distribution of
multiple classes denoted as 1 , 2 , … , 𝑘 . The classification model aims the building stock. Specifically, (a) represents the number of buildings
to find the class label ̂ that best describes the building’s vulnerability for each municipality, and (b) illustrates the transition from red to
index. Given an input vector 𝒙 representing the features of a building, green, indicating the average vulnerability index, always lower than
the classification models aim to predict the building’s class label 𝑖 : 1.
Table 1 provides an overview of the explanatory variables used
̂ = arg max 𝑓 (𝒙) (4)
𝑖 in the study, indicating their data types (categorical or numerical)

5
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 5 shows selected histogram plots, the (a) structural typology,


(b) the concrete compression strength, (c) the height of the building,
(d) the construction year, (e) the soil type and (b) the structural
configuration.
The majority of the buildings (88%) in the dataset are constructed
using reinforced concrete. The construction period ranges from the
1930s to the early 2000s, with a prevalent concentration of buildings
built between the 1970s and 1980s. The quality of the concrete used
in these buildings is relatively poor, as indicated by an average com-
pressive strength of about 20 MPa, a relatively low value for structural
concrete. All the buildings are mid-rise, with an average of 2 to 3 floors.
Additionally, a significant number of buildings have pilotis floors.
Fig. 4 illustrates the municipalities in the L’Aquila province, colour-
coded according to (a) the average concrete compressive strength and
(b) the average standard deviation. Most of the buildings have a low
concrete compressive strength, relatively uniform across municipalities
with a standard deviation for each municipality generally less than 6
MPa (see Fig. 6).
Structural interventions after the 2009 L’Aquila earthquake were
meant to repair the buildings, even though they did not suffer severe
damage. In some cases, localized structural interventions, mainly joint
wrapping, were carried out. The Peak Ground Acceleration (PGA) data
do not show significant variability, with an average of 0.24 g, as
depicted in Fig. 7.
Fig. 7 presents the histogram plot of the vulnerability index. The
plot reveals a substantial number of buildings with a vulnerability level
lower than 0.2. This observation underscores the high vulnerability of
the building stock. However, it is essential to note that this measure is
conventional and lacks a direct physical significance. In other words,
the buildings are not considered to fail if the intensity measures exceed
those considered in the simulations. As mentioned in the introduction,
the authors will undertake two main predictive approaches. The first
attempt involves predicting the actual value of the intensity measures.
Subsequently, the vulnerability index will be categorized into classes
with increasing ranges to assess the improvement in accuracy as the
models become less informative in the prediction. This step aims to ex-
plore how the predictive models perform when dealing with a broader
classification of vulnerability levels, offering insights into the models’
effectiveness across different levels of vulnerability.

3.1. PCA and k-mean clustering

Data visualization is essential in this study due to the


high-dimensional nature of the data. To better understand the differ-
ences within the considered dataset, the authors conducted a prelim-
inary Principal Component Analysis (PCA) [82] followed by 𝑘-means
clustering [83]. PCA was employed to reduce the dataset’s dimen-
sionality before applying the clustering technique. K-means, a widely
used unsupervised clustering algorithm, categorized the buildings into
Fig. 4. Colour maps of municipalities in the L’Aquila Province, Abruzzo (Italy), 𝑘 clusters based on their similarity. The process begins with the random
considered in this investigation. (a) represents the average construction year for each selection of 𝑘 centroids and iteratively assigns each data point to the
municipality, and (b) the average height in meters. (For interpretation of the references nearest centroid. These centroids are updated based on the mean of
to colour in this figure legend, the reader is referred to the web version of this article.)
the assigned points. The ultimate goal is to categorize buildings with
shared attributes, facilitating an understanding of the underlying data
structure and potentially revealing patterns or trends that might not
and the corresponding categories or data statistics. The data types have been apparent in the original high-dimensional space.
are highlighted using different cell colours, with a light pink shade Fig. 8 shows a bar plot of the principal components and the first
indicating categorical data and a light green shade denoting numerical three clusters.
data. The category for concrete resistance reports the minimum value All the first principal components have high values, suggesting that
for the concrete resistance obtained from concrete samples extracted there might be significant variability in the data along multiple dimen-
from the building. This category is only valid for RC buildings. sions, and no single principal component is dominant in explaining the
Fig. 3 displays four representative RC buildings of the considered variance. This might depend on many factors. The original features of
building stock. the dataset might exhibit high variability, and PCA is capturing this
Fig. 4 illustrates the municipalities in the L’Aquila Province, colour- variability across multiple dimensions. Additionally, the features are
coded according to (a) the average year of construction and (b) the highly correlated, leading to multiple principal components capturing
average building height in meters. similar patterns of variance. Therefore, the data might inherently have

6
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 5. Histogram plot of selected explanatory variables, the (a) structural typology, (b) the compression strength of concrete, (c) the height of the building, (d) the construction
year, (e) the soil type and (b) the structural configuration.

Table 1
List of the explanatory variable, indicating the data type (categorical or continuous) and the list of categories and data statistics
for categorical and continuous variables, respectively.
Variable Data type Categories/Statistics (mean, std. dev., max, min)
Structural typology Categorical {Concrete, Masonry}
Concrete compressive strength [MPa] Numerical {21.06,6.93,42.5,6.630}
Total surface of the building [m2 ] Numerical {298.63,210.14,2104.60,30}
Height of the building [m] Numerical {9.40,3.00,21.50,2.47}
Volume of the building [m3 ] Numerical {2541.88,1425.10,10420,500}
No. of residential units Numerical {6.18,3.01,24,1}
Construction year Numerical {1977,12.09,2006,1938}
Structural configuration Categorical {Clustered, Isolated}
No. of elevated floors Numerical {2.76,0.83,7,1}
Basement floors Numerical {0.21,0.41,1,0}
Pilotis floor Categorical {Yes, No}
Structural interventions Categorical {Yes, No}
Damage Categorical {No, Low, Moderate, High}
Peak ground acceleration [g] Numerical {0.24,0.02,0.34,0.16}
Soil type Categorical {A,B,C,D}

Table 2
Average and standard deviation for the data (predictors and response variable) for the first three clusters. For
categorical variables, the authors reported as average the prevalent category.
Explanatory variables Cluster No.1 Cluster No.2 Cluster No.3
Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Structural typology Concrete Concrete Concrete
Concrete compressive strength [MPa] 20.87 6.77 25.97 4.95 19.39 6.36
Total surface of the building [m2 ] 283.42 141.06 666.33 544.26 243.81 131.25
Height of the building [m] 9.24 2.77 12.14 5.35 9.04 2.22
Volume of the building [m3 ] 2355.37 896.84 5798.68 1938.25 2106.26 1027.81
No. of residential units 5.88 2.15 12.13 4.13 5.69 2.90
Construction year 1978 10.87 1981 9.81 1979 8.79
Structural configuration Clustered Clustered Isolated
No. of elevated floors 3 0.66 4 1.03 2 0.64
Basement floors 0 0.42 0.00 0.00 0 0.50
Pilotis floor No No No
Structural interventions No No No
Damage No No Light
Peak ground acceleration [g] 0.25 0.02 0.23 0.02 0.24 0.02
Soil type A B C
Vulnerability index 0.23 0.13 0.25 0.14 0.14 0.13

7
A. Aloisio et al. Engineering Structures 301 (2024) 117295

• Building size and volume: Cluster No. 2 stands out as it consists


of larger buildings in terms of surface area and volume.
• Residential units: The number of residential units in Cluster No. 2
is approximately double compared to the other two clusters, reaf-
firming that these might be larger multi-residential complexes.
• Construction year: The buildings in all three clusters were built
close in time, between 1978 and 1981. This indicates that the
construction year alone is not a significant distinguishing factor
among these clusters.
• Structural configuration: An interesting distinction is that build-
ings in Cluster No. 3 are primarily isolated, in contrast to the
clustered configuration in Clusters No. 1 and No. 2.
• Damage: Cluster No. 3 has buildings with light damage, while
the other two clusters predominantly consist of buildings with no
damage

Still, the similarity of the above three clusters in terms of vulnerability


indexes confirms the dataset’s significant correlation, complexity and
high dimensionality.

4. Results

The results section is organized into two main parts, dedicated to


regression and classification models, respectively.

4.1. Regression models

This section discusses the feasibility of regression models in predict-


ing the vulnerability index based on the explanatory variables listed in
Table 1. The regression analysis is based on four distinct techniques: (i)
Linear regression [64], (ii) Decision Tree Regression [65], (iii) Support
Vector Regression (SVR) [66] and (iv) Artificial Neural Network [67].
As also highlighted in Section 2, the hold-out method was employed,
partitioning the data into 80% for training and 20% for validation
sets. This choice was strategic given the relatively small database,
comprising nearly 300 buildings. Resorting to cross-validation with
𝑘-folds would have introduced challenges due to the low number of
samples in each fold. For Linear Regression, the model was imple-
mented without any regularization. For Decision Tree Regression, the
model was calibrated with a minimum of 3 leaves with the maximum
split set to 100. The Gaussian kernel was employed for the SVR, with
the box constraint set to 1.0 and the kernel scale to 0.5. The optimal
neural network architecture consisted of three layers of fully connected
ten neurons. The optimization of the neural network architecture aimed
to mitigate the challenges of underfitting and overfitting. The ReLU
(Rectified Linear Unit) was adopted as an activation function after a
preliminary evaluation, while the learning rate was set to 0.001. The
batch size was 32, and the number of epochs 100.
Fig. 6. Colour maps of municipalities in the L’Aquila Province, Abruzzo (Italy), The chosen evaluation metrics for the regression models included
considered in this investigation. (a) represents each municipality’s average concrete the Mean Squared Error (MSE), the Root Mean Squared Error (RMSE),
compressive strength, and (b) the related standard deviation (Std. Dev.). (For interpre-
tation of the references to colour in this figure legend, the reader is referred to the
the Mean Absolute Error (MAE), the Maximum Absolute Error (MAXAE)
web version of this article.) and the R-squared (R2 ).
Fig. 9 displays predicted vs. true values of the seismic vulnerability
index for the considered regression methods: (a) linear regression, (b)
complex patterns that a single dominant principal component cannot decision tree, (c) support vector machine, (d) artificial neural network.
capture. This is also related to the high dimensionality, resulting in Each plot shows both the training and the validation set.
several principal components with high variance. Table 3 reports the error metrics corresponding to the test and
Table 2 presents an overview of the average and standard deviation calibration sets. The performance of these models is relatively poor.
of various data variables for the first three clusters. Transitioning from the simplest linear model to the complexity of the
neural network yields no discernible enhancement in model perfor-
• Concrete strength and structural typology: The three clusters pre- mance. This observation highlights that the primary challenge lies not
dominantly comprise concrete buildings. However, Cluster No. 2 in capturing intricate relationships among explanatory variables but in
shows a higher concrete resistance (25.97 MPa) compared to Clus- their insufficiency to provide robust information.
ter No. 1 (20.87 MPa) and Cluster No. 3 (19.39 MPa). This sug- As the neural network architecture grows in complexity, there is a
gests that buildings in Cluster No. 2 are built with better-quality notable boost in performance on the calibration set, even achieving an
materials. R2 as high as 0.8. However, this gain in calibration accuracy is offset

8
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 7. Histogram plots of (a) the Peak Ground Acceleration (PGA) and (b) the vulnerability index.

Fig. 8. Bar plot of the principal components and the first three clusters.

Table 3
Error metrics of the considered models for both the training and validation sets. The acronyms are Mean
Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Maximum Absolute
Error (MAXAE) and R-squared (R2 ).
Model Test set Calibration set
MSE RMSE MAE MAXAE R2 MSE RMSE MAE MAXAE R2
Linear regression 0.02 0.13 0.10 0.34 0.31 0.01 0.12 0.09 0.37 0.43
Regression tree 0.02 0.16 0.12 0.38 −0.25 0.01 0.10 0.08 0.43 0.48
Support vector machine 0.02 0.14 0.10 0.38 0.26 0.01 0.12 0.09 0.39 0.40
Neural network 0.02 0.14 0.10 0.38 0.23 0.01 0.07 0.05 0.29 0.79

by a deterioration in performance on the validation set, particularly Various factors can contribute to high 𝑝-values. In this context, two
evident when compared to simpler architectures. The Regression Tree plausible explanations are (i) sample variation and (ii) low statistical
demonstrates relatively comparable MSE and RMSE metrics among the significance. (i) The data collected might exhibit significant variability,
intermediate models. Yet, a conspicuous drop in the R2 value on the making it challenging to discern a notable effect, even if one is present.
test set indicates overfitting. The SVR exhibits more consistent MSE, (ii) Additionally, insufficient sample sizes exacerbate this issue. Conse-
RMSE, and MAE values and a moderate R2 score, revealing its relatively quently, the analysis may lack the statistical significance necessary to
lower susceptibility to overfitting. As already remarked, the neural identify a substantial effect, resulting in elevated 𝑝-values.
network demonstrates competitive error metrics, achieving a high R2 An alternative is a classification approach. The vulnerability indexes
on the calibration set. However, the discrepancy between the training care are categorized into classes. By doing so, the authors have explored
set’s R2 and the validation set’s R2 reveals overfitting. Therefore, the the potential enhancement in the performance metrics.
considered explanatory variables, though readily extractable, do not
deliver reasonable accuracy in predicting the vulnerability index, even 4.2. Classification models
in preliminary investigations.
The authors further emphasized this aspect by reporting the 𝑝-values This subsection discusses the reliability of a classification strategy
corresponding to the coefficients of the linear regression model, see for predicting vulnerability index classes. Initially, the authors grouped
Fig. 10. All the calculated 𝑝-values are significantly higher than 0.1, the seismic vulnerability indexes into four classes, which later evolved
whereas a recommended threshold is 0.05 [84]. The most statistically into three, each defined by variable thresholds. However, despite using
significant features are identified by the lowest 𝑝-values, particularly complex neural network architectures, the attained accuracy metrics
those related to building typology (masonry or concrete) and structural remained relatively modest, plateauing between 0.6 and 0.65. Such
configuration (clustered or isolated). levels of accuracy cannot be considered satisfactory for a predictive

9
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 9. Predicted vs. true values of the vulnerability index for the considered regression methods: (a) linear regression, (b) decision tree, (c) support vector machine, (d) artificial
neural network. Each plot shows both the training and the validation set.

Following the binary classification, the threshold was initially set at


0.1, representing a very low threshold, deemed critical according to the
funding agency ATER. Four methods have been compared: logistic re-
gression [78], classification tree [65], ANN and XGBoost. As explained
in Section 2, the authors implemented a feed-forward fully-connected
multi-layer perceptron architecture [74–76], where each layer had 15
neurons optimized following [85]. The same hyperparameters adopted
in the regression models for classification trees and ANN are used
here. The comparison was limited to these four models since other
options, such as SVM, random forest or naive Bayesian classifiers,
performed similarly to the first two. As a result, the four were selected
Fig. 10. Bar plot illustrating the p-values corresponding to the coefficients of the linear for comparison.
regression model. The ascending numbering denotes the explanatory variable to which Fig. 11 illustrates the average confusion matrix for calibration and
the 𝑝-value pertains based on Table 1.
validation sets. These matrices present the model’s predictions against
each category’s actual true class labels. The diagonal elements, from
the top left to the bottom right, represent accurate predictions for each
model, given the considerable uncertainty and error intrinsic to the class. Conversely, the off-diagonal elements report the misclassifica-
assessment process. tions, where predictions did not align with the true class labels. In these
In contrast, the final classification approach demonstrated more confusion matrices, additional metrics are reported: recall, precision,
promising outcomes. By dividing the data into two groups, separated and accuracy. The right column, shaded in grey, reports the recall
by a specified vulnerability threshold (𝛼𝑡 ), buildings with seismic vul-
or true positive rate, indicating the model’s performance in correctly
nerability indexes below 𝛼𝑡 were classified as ‘Severe’. In contrast, those
identifying positive instances within each class. The bottom row, also
with higher indexes were labelled as ‘Moderate’. Binary classifications
shaded in grey, presents precision, representing the proportion of true
offer several advantages. They simplify the complexity of the problem
positive predictions among all predicted positive instances for each
and often yield clearer distinctions between classes, making the models
more interpretable and potentially more robust. Following this simpli- class. Moreover, the bottom-right cell in the matrix shows the model’s
fied methodology, the authors observed a substantial enhancement in overall accuracy, the ratio of accurately classified instances to the
accuracy, often surpassing 0.85 in several instances. Such an accuracy total predictions. Accuracy provides a holistic view of the model’s
level can be deemed acceptable and reasonable for predictive models. comprehensive performance.
Consequently, this subsection presents the results of the classification Nonetheless, relying solely on accuracy as a performance metric can
models based on a binary framework, distinguishing between ‘Severe’ be misleading in datasets with class imbalance, such as the one under
and ‘Moderate’ vulnerability classes. consideration with threshold 0.1. The ‘Severe’ class, being the minority,

10
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 11. Average confusion matrix for (a), (c), (e) calibration and (b), (d), (f) validation sets using the (a)–(b) logistic regression, (c)–(d) classification tree and (e)–(f) artificial
neural network model assuming a threshold 0.1 for separating the ‘Severe’ and ‘Moderate’ vulnerability classes.

has only 67 instances, while the ‘Moderate’ class, the majority, has 209 recall for the severe class. Consequently, the models achieve a satis-
elements. High-accuracy models would be of limited practical use if factory accuracy, higher than 70%, by merely predicting the majority
they fail to classify the minority class correctly. Therefore, Table 4 class. The only models presenting more balanced performance between
reports relevant performance metrics like precision, recall, and the the majority and minority classes are the ANN and XGBoost models.
F1-score. Nonetheless, the most reliable remains the ANN, which has equal recall
for both. Therefore, the ANN is the best model for such an imbalanced
The ‘‘Severe’’ (minority) class has considerably lower precision
dataset among the four in this context. This model also has the highest
values across all models when compared to the moderate (majority). accuracy, nearly equal to 90%.
In such imbalanced datasets, the models tend to be biased towards In a second step, the authors assessed the dependence of the per-
the majority class. As a result, the models do not learn the distin- formance metrics (accuracy and recall) on the vulnerability threshold
guishing features of the minority class as effectively. This imbalance used to distinguish between the ‘Moderate’ and ‘Severe’ classes. As ex-
results in many false negatives for the minority class, leading to lower pected, the performance metrics of all employed models, ranging from

11
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Fig. 12. Performance metrics of the four models given random selections of the validation set, as a function of the vulnerability threshold used to separate the two vulnerability
classes, ‘Moderate’ vs. ‘Severe’.

Table 4
Performance metrics comparison for logistic regression, tree model, and artificial neural network on the ‘Severe’ and ‘Moderate’ classes.
Metric Logistic regression Tree model Artificial neural network XGBoost
Severe (%) Moderate (%) Severe (%) Moderate (%) Severe (%) Moderate (%) Severe (%) Moderate (%)
Precision 29.41 90.38 20.00 87.10 60.00 96.80 73.44 86.24
Recall 50.00 79.66 33.33 77.10 85.71 88.20 51.60 93.53
F1-Score 37.33 84.62 25.00 81.86 70.59 92.31 59.10 89.64
Accuracy 75.36 70.73 87.80 83.54

binary logistic regression to ANN, displayed a significant dependence vulnerability threshold and maintains good performance even in im-
on the selected vulnerability threshold. Fig. 12 displays the dependence balanced datasets. For a more cautious approach, it is preferable to
of the performance metrics estimations (accuracy and recall) for the classify a building as ‘Severe’ rather than ‘Moderate’. Hence, even if
validation sets on the vulnerability threshold. The validation sets have the funding agency might suggest a lower vulnerability threshold to
been randomly selected as 20% of the database following a Montecarlo identify significantly non-conforming buildings, it is recommended to
approach until convergence of the recall, corresponding to a relative
set the minority class as ‘Moderate’. Nonetheless, ANN models can be
variation of the coefficient of variation (CoV) less than 1%.
customized to meet the specific needs of public administrations, who
Fig. 12 shows how the threshold influences the average performance
must prioritize interventions based on vulnerability thresholds.
metrics. For logistic, classification tree and XGBoost models, the re-
duction in recall for minority classes is concerning, dropping below
0.2. Even with minor class imbalances, recall hovers around 0.2 in the
logistic regression for the minority class. While the recall reduction for 5. Model interpretation
the classification tree is more gradual, it remains significant. However,
for the ANN, the drop in recall is relatively modest, always staying
above 0.6. The accuracy remains satisfactory, with a minimum of 86% The authors estimated the feature importance using the optimized
for balanced datasets, which is achieved with a threshold of 0.2. ANN model with a threshold of 0.2 to have a balanced dataset. The
This study demonstrates that the ANN is the most reliable model SHapley Additive exPlanations (SHAP) values were calculated to inter-
among those tested. This model retains flexibility in choosing the pret the model’s predictions [69].

12
A. Aloisio et al. Engineering Structures 301 (2024) 117295

the influence of each feature. It likens each feature of an instance to a


‘player’ in a coalition game, where the prediction is analogous to the
game’s payoff, and SHAP values provide a fair distribution of the payoff
among the players [87]. Assuming a model prediction 𝑓 (𝑥) based on a
single input variable 𝑥, applying an additive feature attribution method,
the prediction 𝑓 (𝑥) can be expressed as [86,88,89]:

𝑀
𝑓 (𝑥) = 𝑔(𝑥′ ) = 𝜙0 + 𝜙𝑖 𝑥′𝑖 (5)
𝑖=1

where 𝑥′ is the simplified input that maps to the original inputs through
a mapping function 𝑥 = ℎ(𝑥′ ); 𝜙0 = 𝑓 (ℎ(𝑥0 )) is a constant when all
inputs are null; 𝜙𝑖 are the attribution values for feature 𝑖; and 𝑀 is
the total number of the features. The SHAP values can be computed
by solving Eq. (5) with satisfying three desirable properties (i.e., local
accuracy, missingness, and consistency). This is particularly useful for
models like neural networks, where it is unclear how each feature in-
fluences the output because of complex interactions and non-linearities
within the model.

5.2. Model interpretation

SHAP values highlight which features are most influential in the


decision-making process of the ML model. Features with a higher abso-
lute value (positive or negative) are more relevant in the predictions.
These values are calculated by considering all possible combinations of
features and observing the change in prediction with and without the
feature in question.
Fig. 13(a) provides the feature importance index of each input
feature. Indirectly, the feature importance underscores the role of a
variable when missing in the predictive model. If the feature impor-
tance is relatively minor, the omission of that particular variable may
not significantly impact the overall prediction. The features are listed
along the 𝑌 -axis, with their corresponding SHAP values on the 𝑋-axis.
A longer bar indicates a higher absolute impact on the output, whether
positive or negative. The key determinants affecting the outcome are
the concrete compressive strength, the geometry of the building (vol-
ume, height, surface and number of floors) and the construction year.
Then follows the structural typology, RC or masonry, and the presence
of the pilots’ floors. Other variables like the preexistence of damage,
the structural interventions, and the number of units or the basement
floors are less influential. The fact that the compressive strength of
concrete stands out as the key indicator gives value to the building’s
knowledge, revealing that not only the geometry of the buildings but
also the material properties are fundamental in the assessment. The
geometric properties are also relevant because they relate to the build-
ing dynamics and, consequently, the structural demand. The volume,
surface and height consider multiple aspects not directly expressed,
such as eccentricity in plan and how masses are distributed. The PGA
is important as it should be in vulnerability analysis, but it is still not
at the top, possibly due to its reduced variability. The third position
is the construction year. The construction year has one of the highest
feature importance scores. This is related to the evolution of building
codes, materials, and architectural designs. These findings indicate that
Fig. 13. Global interpretations of the ANN model by SHAP values: (a) SHAP feature the primary discriminator is the quality of the material rather than the
importance and (b) SHAP summary plot. (For interpretation of the references to colour type of construction material used, which is eighth in this ranking.
in this figure legend, the reader is referred to the web version of this article.) Fig. 13(b) gives the summary plot of SHAP values for each fea-
ture, with Table 5 displaying the integer encoding of the categorical
variables. Each dot represents the SHAP value of a feature for a
single prediction. The spread of the dots indicates variability in the
importance of the feature across different predictions. For example, if
5.1. SHapley Additive exPlanations (SHAP)
the dots for a feature are widely spread along the 𝑋-axis, this feature
has a varying impact on the model’s predictions. The colour coding,
SHAP, or SHapley Additive exPlanations, is a method used to un- from pink to blue, indicates the feature value. In the case of categorical
derstand the decision-making of ML algorithms by assigning SHAP variables, the integer encoding in Table 5 has been adopted. Namely,
values derived from cooperative game theory principles [86]. The key positive SHAP values indicate that the feature value tends to influence
point of SHAP is deconstructing a model’s prediction by quantifying the vulnerability (i.e., reduction) positively.

13
A. Aloisio et al. Engineering Structures 301 (2024) 117295

Table 5
Integer encoding of the categorical variables.
Integer encoder Structural typology Configuration Pilotis Structural intervention Damage Soil type
0 Masonry Isolated No No No A
1 Reinforced concrete Clustered Yes Yes Light B
2 Moderate C
3 Severe

Fig. 14. Global interpretations of the ANN model by SHAP values in terms of feature dependence scatter plots.

The detailed dependence of the nine features on the SHAP values is 6. Conclusions
shown in Fig. 14.
Each subplot represents Fig. 13(b) in the plane feature vs. SHAP This study examines the possible use of Machine Learning (ML)
values. A colour coding from blue to pink is used to consider the predictive models for code-based seismic vulnerability assessments.
dependence of an additional feature in the plot. The concrete resistance The authors used a database of nearly 300 buildings, encompassing
and the construction year positively correlate with the SHAP values. both reinforced concrete and masonry structures, from the regional
It means that a higher compressive strength and construction year agency of residential construction (ATER) in the province of L’Aquila,
leads to better seismic performance, i.e. an increase in the vulnerability Abruzzo, Italy. These buildings underwent a comprehensive seismic
index since the buildings are more likely to belong to the ‘Moderate’ vulnerability assessment following all required stages, ranging from
knowledge acquisition to structural analysis. The chosen metric for
vulnerability class. A similar trend is observed for the volume. Larger-
prediction was the vulnerability index, the ratio between capacity and
sized buildings are less vulnerable than smaller ones. Regarding the
demand.
building height, surface, PGA and the number of floors, the trend is
The findings indicated that regression and multiclass classification
more complex, with no manifest correlation with the SHAP values.
models yielded unsatisfactory results and cannot be deemed suffi-
The structural typology and the presence of the pilotis floors, the only
ciently accurate for reliable predictions. On the other hand, binary
categorical variables in Fig. 14 exhibit an evident trend. In the label
classification tasks emerged as promising. Despite their simplicity,
encoding, masonry structures corresponding to 0 have a higher vulner-
binary classification models, predicting whether a building falls below
ability than the RC ones. Then, as expected, the absence of the pilotis or above a certain vulnerability threshold, greatly benefit public ad-
floors (0) has marked beneficial effects on the seismic vulnerability ministrations. Such models allow administrators to define acceptable
performance. vulnerability levels for their managed building stock.
In conclusion, it should be remarked that the analysed building The most effective classification model was the ANN. The authors
stock represents the typical residential construction in Southern Europe. highlighted the essential role of the vulnerability threshold in demar-
Consequently, while not easily generalizable, these findings can serve cating the ‘Severe’ and ‘Moderate’ vulnerability classes and their influ-
as a reference for interpreting the ML models obtained from other ence on performance metrics. Should the chosen vulnerability threshold
datasets. introduce considerable class imbalance, the logistic regression and

14
A. Aloisio et al. Engineering Structures 301 (2024) 117295

classification tree models do not accurately predict the minority class. Acknowledgements
Notably, recall for these models can drop to 20%, even with marginal
imbalances. In contrast, the ANN effectively predicts the minority class, The authors acknowledge the regional residential construction
maintaining a recall of no less than 60% even with significant class agency in the province of L’Aquila (ATER), Abruzzo, Italy, for providing
imbalances. The ANN achieves an accuracy of approximately 85% for the data used in this investigation.
a balanced dataset, corresponding to a vulnerability threshold of 0.2 to
separate the ‘Moderate’ and ‘Severe’ classes. References
The feature importance based on the SHapley Additive exPlana-
tions (SHAP) values reveals that the concrete compressive strength, [1] Barbat AH, Carreño ML, Pujades LG, Lantada N, Cardona OD, Marulanda MC.
Seismic vulnerability and risk evaluation methods for urban areas. A review with
the building geometry (volume, height and average floor surface) and
application to a pilot area. Struct Infrastruct Eng 2010;6(1–2):17–38.
construction year emerge as the features with the highest significance. [2] Formisano A. Theoretical and numerical seismic analysis of masonry building
In essence, the material quality is more important in this context aggregates: case studies in San Pio Delle Camere (L’Aquila, Italy). J Earthq Eng
than the building typology, whether RC or masonry. The building 2017;21(2):227–45.
[3] Brando G, De Matteis G, Spacone E. Predictive model for the seismic vulnerability
geometry indirectly reflects the building dynamics’ role in estimating
assessment of small historic centres: application to the inner Abruzzi Region in
structural performance. Additionally, the construction year highlights Italy. Eng Struct 2017;153:81–96.
the shifts in building standards, materials, and architectural evolution [4] Singhal A, Kiremidjian AS. Method for probabilistic evaluation of seismic
over time. Parallelly, other variables, such as structural configuration, structural damage. J Struct Eng 1996;122(12):1459–67.
PGA, and soil type, hold less importance. Given the relevance of soil [5] Benedetti D, Petrini V. On seismic vulnerability of masonry buildings: proposal
of an evaluation procedure. Ind Costr 1984;18(2):66–74.
type and seismic-related features in structural assessments, this may
[6] Kourehpaz P, Molina Hutt C. Machine learning for enhanced regional seismic
seem counterintuitive. However, in the analysed dataset, these factors risk assessments. J Struct Eng 2022;148(9):04022126.
do not showcase significant variability, thus not substantially impacting [7] Li S-Q, Gardoni P. Empirical seismic vulnerability models for building clusters
model predictions. considering hybrid intensity measures. J Build Eng 2023;68:106130.
[8] Brando G, Rapone D, Spacone E, Masciotta MG. MUDis: A low computational
Developing data-driven models for code-based seismic vulnerability
effort multi-unit discretization procedure for modelling masonry walls with
assessment carries notable economic advantages. When agile mod- periodic arrangement. In: Structures, Vol. 43. Elsevier; 2022, p. 1380–406.
els for predicting interventions are in place, redirecting funds from [9] Formisano A, Florio G, Landolfo R, Mazzolani FM. Numerical calibration of an
comprehensive vulnerability analyses to seismic retrofitting becomes easy method for seismic behaviour assessment on large scale of masonry building
aggregates. Adv Eng Softw 2015;80:116–38.
feasible. Additionally, these models address a crucial concern for public
[10] Vettore M, Donà M, Carpanese P, Follador V, da Porto F, Valluzzi MR. A
administrations, focusing on code-based seismic vulnerability, which multilevel procedure at urban scale to assess the vulnerability and the exposure
holds legal significance in certain countries like Italy despite lacking of residential masonry buildings: The case study of Pordenone, Northeast Italy.
direct physical interpretation. Heritage 2020;3(4):1433–68.
Future research endeavours will focus on enhancing classification [11] Kita A, Cavalagli N, Masciotta MG, Lourenço PB, Ubertini F. Rapid post-
earthquake damage localization and quantification in masonry structures through
accuracy by exploring larger datasets and more explanatory variables. multidimensional non-linear seismic IDA. Eng Struct 2020;219:110841.
This will involve comparing the predictability of in-field observations [12] Pellegrini D, Girardi M, Lourenço PB, Masciotta MG, Mendes N, Padovani C,
with code-based vulnerability assessments. Ramos LF. Modal analysis of historical masonry structures: Linear perturbation
and software benchmarking. Constr Build Mater 2018;189:1232–50.
[13] XuanRui Y. Developing an artificial neural network model to predict the
CRediT authorship contribution statement durability of the RC beam by machine learning approaches. Case Stud Constr
Mater 2022;17:e01382.
Angelo Aloisio: Conceptualization, Data curation, Formal analy- [14] Wen W, Zhang C, Zhai C. Rapid seismic response prediction of RC frames
based on deep learning and limited building information. Eng Struct 2022;267.
sis, Funding acquisition, Investigation, Methodology, Project admin-
http://dx.doi.org/10.1016/j.engstruct.2022.114638, cited By 4.
istration, Resources, Software, Supervision, Validation, Visualization, [15] Stojadinović Z, Kovačević M, Marinković D, Stojadinović B. Rapid earthquake
Writing – original draft, Writing – review & editing. Yuri De San- loss assessment based on machine learning and representative sampling. Earthq
tis: Conceptualization, Data curation, Formal analysis, Investigation, Spectra 2022;38(1):152–77. http://dx.doi.org/10.1177/87552930211042393.
[16] Xu J-G, Feng D-C, Mangalathu S, Jeon J-S. Data-driven rapid damage evaluation
Methodology, Software, Validation, Visualization, Writing – original
for life-cycle seismic assessment of regional reinforced concrete bridges. Earthq
draft, Writing – review & editing. Francesco Irti: Data curation, Fund- Eng Struct Dyn 2022;51(11):2730–51.
ing acquisition, Investigation, Project administration, Resources. Dag [17] Iervolino I. Asymptotic behavior of seismic hazard curves. Struct Saf 2022;99.
Pasquale Pasca: Data curation, Investigation, Methodology, Software, http://dx.doi.org/10.1016/j.strusafe.2022.102264.
Validation. Leonardo Scimia: Conceptualization, Methodology, Project [18] Lagomarsino S, Giovinazzi S. Macroseismic and mechanical models for the
vulnerability and damage assessment of current buildings. Bull Earthq Eng
administration. Massimo Fragiacomo: Supervision. 2006;4(4):415–43. http://dx.doi.org/10.1007/s10518-006-9024-z.
[19] Lagomarsino S, Cattari S, Angiolilli M, Bracchi S, Rota M, Penna A. Modelling
Declaration of competing interest and seismic response analysis of existing URM structures. Part 2: Archetypes of
Italian historical buildings. J Earthq Eng 2023;27(7):1849–74. http://dx.doi.org/
10.1080/13632469.2022.2087800.
All authors have participated in (a) conception and design, or anal- [20] Braga F, Dolce M, Liberatore D. A statistical study on damaged buildings and
ysis and interpretation of the data; (b) drafting the article or revising an ensuing review of the MSK-76 scale. In: Proceedings of the seventh European
it critically for important intellectual content; and (c) approval of the conference on earthquake engineering. 1982, p. 431–50, cited By 165.
[21] Spence R, Coburn A, Pomonis A, Sakai S. Correlation of ground motion with
final version.
building damage: The definition of a new damage-based seismic intensity scale.
This manuscript has not been submitted to, nor is under review at, In: Proceedings of the 10th world conference on earthquake engineering, Vol. 1.
another journal or other publishing venue. 1992, p. 551–6, cited By 82.
The authors have no affiliation with any organization with a direct [22] Yamaguchi N, Yamazaki F. Fragility curves for buildings in Japan based on
damage surveys after the 1995 Kobe earthquake. In: Proc. 12th world conference
or indirect financial interest in the subject matter discussed in the
on earthquake engineering. 2000, cited By 26.
manuscript [23] Rossetto T. Vulnerability curves for the seismic assessment of reinforced concrete
building populations. In: Vulnerability curves for the seismic assessment of
reinforced concrete buildings populations. 2004, cited By 20.
Data availability
[24] Rota M, Penna A, Strobbia C. Processing Italian damage data to derive ty-
pological fragility curves. Soil Dyn Earthq Eng 2008;28(10–11):933–47. http:
The data that has been used is confidential. //dx.doi.org/10.1016/j.soildyn.2007.10.010, cited By 224.

15
A. Aloisio et al. Engineering Structures 301 (2024) 117295

[25] Del Gaudio C, De Martino G, Di Ludovico M, Manfredi G, Prota A, Ricci P, [48] Li S-Q. Analysis of an empirical seismic fragility prediction model of wooden
Verderame GM. Empirical fragility curves for masonry buildings after the 2009 roof truss buildings. Case Stud Constr Mater 2022;17. http://dx.doi.org/10.1016/
L’Aquila, Italy, earthquake. Bull Earthq Eng 2019;17:6301–30. j.cscm.2022.e01420.
[26] Rosti A, Rota M, Penna A. Empirical fragility curves for Italian URM buildings. [49] Xi J. Relationship between the organizational structure in implementing Post-
Bull Earthq Eng 2021;19(8):3057–76. http://dx.doi.org/10.1007/s10518-020- Disaster Housing reconstruction and outcome characteristics: A study on urban
00845-9, cited By 80. dujiangyan after the Wenchuan Earthquake. Int J Disaster Risk Reduct 2022;80.
[27] Ghimire S, Guéguen P, Giffard-Roisin S, Schorlemmer D. Testing machine http://dx.doi.org/10.1016/j.ijdrr.2022.103221, cited By 1.
learning models for seismic damage prediction at a regional scale using building- [50] Tocchi G, Misra S, Padgett JE, Polese M, Di Ludovico M. The use of
damage dataset compiled after the 2015 Gorkha Nepal earthquake. Earthq machine-learning methods for post-earthquake building usability assessment: A
Spectra 2022;38(4):2970–93. http://dx.doi.org/10.1177/87552930221106495. predictive model for seismic-risk impact analyses. Int J Disaster Risk Reduct
[28] Chen W, Zhang L. An automated machine learning approach for earthquake 2023;97:104033.
casualty rate and economic loss prediction. Reliab Eng Syst Saf 2022;225. [51] Aloisio A, Rosso MM, De Leo AM, Fragiacomo M, Basi M. Damage classification
http://dx.doi.org/10.1016/j.ress.2022.108645. after the 2009 L’Aquila earthquake using multinomial logistic regression and
[29] Guo M, Huang H, Xue C, Huang M. Assessment of fuzzy global seismic neural networks. Int J Disaster Risk Reduct 2023;96:103959.
vulnerability for RC structures. J Build Eng 2022;57. http://dx.doi.org/10.1016/ [52] Coburn A, Spence R, Pomonis A. Factors determining human casualty levels in
j.jobe.2022.104952. earthquakes: Mortality prediction in building collapse. In: Proceedings of the
[30] Bernardo V, Campos Costa A, Candeias P, Costa A. Seismic vulnerability assess- 10th world conference on earthquake engineering. 1992, p. 5989–94, cited By
ment and fragility analysis of pre-code masonry buildings in Portugal. Bull Earthq 152.
Eng 2022;20(11):6229–65. http://dx.doi.org/10.1007/s10518-022-01434-8. [53] Chang S, Pasion C, Tatebe K, Ahmad R. Linking lifeline infrastructure per-
[31] Rezvani R, Soroushian S, Zaghi A, Maragakis M. Numerical seismic fragility formance and community disaster resilience: Models and multi-stakeholder
analysis for suspended ceilings with various geometries. J Build Eng 2022;54. processes. In: Linking lifeline infrastructure performance and community disaster
http://dx.doi.org/10.1016/j.jobe.2022.104627. resilience: Models and multi-stakeholder processes. 2008, cited By 22.
[32] Ferreira T, Maio R, Vicente R. Seismic vulnerability assessment of the old city [54] Zuccaro G, Cacace F. Revisione dell’inventario a scala nazionale delle classi
centre of Horta, Azores: calibration and application of a seismic vulnerability tipologiche di vulnerabilità ed aggiornamento delle mappe nazionali di rischio
index method. Bull Earthq Eng 2017;15(7):2879–99. http://dx.doi.org/10.1007/ sismico. In: Revisione dell’inventario a scala nazionale delle classi tipologiche di
s10518-016-0071-9. vulnerabilità ed aggiornamento delle mappe nazionali di rischio sismico. Atti del
[33] Demirel I, Yakut A, Binici B. Seismic performance of mid-rise reinforced concrete XIII convegno ANIDIS. 2009, cited By 9.
buildings in Izmir Bayrakli after the 2020 Samos earthquake. Eng Fail Anal [55] Spence R, So E, Scawthorn C. A global earthquake building damage and
2022;137. http://dx.doi.org/10.1016/j.engfailanal.2022.106277. casualty database. In: Human casualties in earthquakes: Progress in modelling
[34] Shafapourtehrany M, Yariyan P, Özener H, Pradhan B, Shabani F. Evaluating
and mitigation. 2011, cited By 43.
the application of K-mean clustering in Earthquake vulnerability mapping of
[56] De Martino G, Di Ludovico M, Prota A, Moroni C, Manfredi G, Dolce M.
Istanbul, Turkey. Int J Disaster Risk Reduct 2022;79. http://dx.doi.org/10.1016/
Estimation of repair costs for RC and masonry residential buildings based on
j.ijdrr.2022.103154.
damage data collected by post-earthquake visual inspection. Bull Earthq Eng
[35] Sheshov V, Apostolska R, Bozinovski Z, Vitanova M, Stojanoski B, Edip K,
2017;15(4):1681–706. http://dx.doi.org/10.1007/s10518-016-0039-9, cited By
Bogdanovic A, Salic R, Jekic G, Zafirov T, Zlateski A, Chapragoski G, Tomic D,
54.
Zurovski A, Trajchevski J, Markovski I. Reconnaissance analysis on buildings
[57] Di Ludovico M, De Martino G, Prota A, Manfredi G, Dolce M. Relationships
damaged during Durres earthquake Mw6.4, 26 November 2019, Albania: effects
between empirical damage and direct/indirect costs for the assessment of seismic
to non-structural elements. Bull Earthq Eng 2022;20(2):795–817. http://dx.doi.
loss scenarios. Bull Earthq Eng 2022;20(1):229–54. http://dx.doi.org/10.1007/
org/10.1007/s10518-021-01271-1.
s10518-021-01235-5, cited By 11.
[36] Debnath P, Halder L, Chandra Dutta S. Damage survey and seismic vulnera-
[58] Bertelli S, Rossetto T, Ioannou I. Derivation of empirical fragility functions
bility assessment of unreinforced masonry structures in low-intensity Ambasa
from the 2009 Aquila earthquake. In: 16th European conference on earthquake
earthquake of northeast India. Structures 2022;44:372–88. http://dx.doi.org/10.
engineering, Thessaloniki. 2018, p. 1–12, cited By 6.
1016/j.istruc.2022.08.005.
[59] Zucconi M, Romano F, Ferracuti B. Typological fragility curves for RC buildings:
[37] Misseri G, Palazzi C, Rovero L. Seismic vulnerability of timber-reinforced
influence of damage index and building sample selection. Eng Struct 2022;266.
earthen structures through standard and non–standard limit analysis. Eng Struct
http://dx.doi.org/10.1016/j.engstruct.2022.114627, cited By 7.
2020;215. http://dx.doi.org/10.1016/j.engstruct.2020.110663.
[60] delle Infrastrutture e dei Trasporti M. Norme tecniche per le costruzioni. Decreto
[38] Ebrahimiyan F, Hadianfard M, Naderpour H, Jankowski R. Fragility analysis
Minist 2008.
of structural pounding between adjacent structures arranged in series with
[61] Humar J. Dynamics of structures. CRC Press; 2012.
various alignment configurations under near-field earthquakes. Bull Earthq Eng
[62] e dei Trasporti MdI. NTC2008–Norme tecniche per le costruzioni. Tech. rep., DM
2022;20(13):7215–40. http://dx.doi.org/10.1007/s10518-022-01471-3.
14/01/2008; 2008, [In Italian].
[39] Pitilakis D, Petridis C. Fragility curves for existing reinforced concrete buildings,
[63] Lagomarsino S, Penna A, Galasco A, Cattari S. TREMURI program: an equivalent
including soil–structure interaction and site amplification effects. Eng Struct
frame model for the nonlinear seismic analysis of masonry buildings. Eng Struct
2022;269. http://dx.doi.org/10.1016/j.engstruct.2022.114733.
[40] Gautam D, Chettri N, Tempa K, Rodrigues H, Rupakhety R. Seismic vulnerability 2013;56:1787–99.
of bhutanese vernacular stone masonry buildings: From damage observation to [64] Neter J, Wasserman W, Kutner MH. Applied linear regression models. Richard
fragility analysis. Soil Dyn Earthq Eng 2022;160. http://dx.doi.org/10.1016/j. D. Irwin; 1983.
soildyn.2022.107351. [65] Breiman L, Friedman J, Olshen R, Stone C. Classification and regression trees.
[41] Del Gaudio C, De Martino G, Di Ludovico M, Manfredi G, Prota A, Ricci P, Boca Raton, Florida: crc press; 1984.
Verderame G. Empirical fragility curves from damage data on RC buildings [66] Vapnik V, Vapnik V. Statistical learning theory, Vol. 1. New York: Wiley; 1998,
after the 2009 L’Aquila earthquake. Bull Earthq Eng 2017;15(4):1425–50. http: p. 2, (624).
//dx.doi.org/10.1007/s10518-016-0026-1. [67] Haykin S. Neural networks: A comprehensive foundation. Prentice Hall PTR;
[42] Chieffo N, Formisano A, Landolfo R, Milani G. A vulnerability index based- 1998.
approach for the historical centre of the city of Latronico (Potenza, Southern [68] James G, Witten D, Hastie T, Tibshirani R, et al. An introduction to statistical
Italy). Eng Fail Anal 2022;136. http://dx.doi.org/10.1016/j.engfailanal.2022. learning, Vol. 112. Springer; 2013.
106207. [69] Hastie T, Tibshirani R, Friedman JH, Friedman JH. The elements of statistical
[43] Cardinali V, Tanganelli M, Bento R. Seismic assessment of the XX century learning: data mining, inference, and prediction, Vol. 2. Springer; 2009.
masonry buildings in Florence: Vulnerability insights based on urban data [70] Bishop CM, Nasrabadi NM. Pattern recognition and machine learning, Vol. 4.
acquisition and nonlinear static analysis. J Build Eng 2022;57. http://dx.doi. Springer; 2006.
org/10.1016/j.jobe.2022.104801. [71] Goodfellow I, Bengio Y, Courville A. Deep learning. MIT Press; 2016.
[44] Qu Z, Dutu A, Zhong J, Sun J. Seismic damage to masonry-infilled tim- [72] Platt JC. 12 Fast training of support vector machines using sequential minimal
ber houses in the 2013 M7.0 Lushan, China, earthquake. Earthq Spectra optimization. In: Advances in kernel methods. Cambridge: MIT press; 1999, p.
2015;31(3):1859–74. http://dx.doi.org/10.1193/012914EQS023T. 185–208.
[45] Sun X, He M, Li Z, Lam F. Seismic performance of energy-dissipating [73] Kingma DP, Ba J. Adam: A method for stochastic optimization. 2014, arXiv
post-tensioned CLT shear wall structures II: Dynamic analysis and dissipater preprint arXiv:1412.6980.
comparison. Soil Dyn Earthq Eng 2020;130:105980. [74] Géron A. Hands-on machine learning with scikit-learn, keras, and tensorflow. "
[46] Gong M, Lin S, Sun J, Li S, Dai J, Xie L. Seismic intensity map and typical O’Reilly Media, Inc."; 2022.
structural damage of 2010 Ms 7.1 Yushu earthquake in China. Nat Hazards [75] Raschka S. Python machine learning. Packt publishing ltd; 2015.
2015;77(2):847–66. http://dx.doi.org/10.1007/s11069-015-1631-z. [76] Aggarwal CC, et al. Neural networks and deep learning, Vol. 10. Springer; 2018,
[47] Li S-Q, Chen Y-S. Analysis of the probability matrix model for the seismic p. 3, (978).
damage vulnerability of empirical structures. Nat Hazards 2020;104(1):705–30. [77] Muller AC, Guido S. Introduction to machine learning with Python. O’Reilly;
http://dx.doi.org/10.1007/s11069-020-04187-2. 2017.

16
A. Aloisio et al. Engineering Structures 301 (2024) 117295

[78] Hosmer Jr DW, Lemeshow S, Sturdivant RX. Applied logistic regression, Vol. [84] Casella G, Berger RL. Statistical inference. Cengage Learning; 2021.
398. John Wiley & Sons; 2013. [85] Aloisio A, Contento A, Alaggio R, Quaranta G. Physics-based models, surrogate
[79] Chen T, Guestrin C. Xgboost: A scalable tree boosting system. In: Proceedings of models and experimental assessment of the vehicle–bridge interaction in braking
the 22nd acm sigkdd international conference on knowledge discovery and data conditions. Mech Syst Signal Process 2023;194:110276.
mining. 2016, p. 785–94. [86] Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Adv
[80] Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Neural Inf Process Syst 2017;30.
Prettenhofer P, Weiss R, Dubourg V, et al. Scikit-learn: Machine learning in [87] Molnar C. Interpretable machine learning. Lulu. com; 2020.
Python. J Mach Learn Res 2011;12:2825–30. [88] Feng D-C, Wang W-J, Mangalathu S, Taciroglu E. Interpretable XGBoost-SHAP
[81] Harrell FE. Regression modeling strategies. Berlin, Heidelberg: Springer-Verlag; machine-learning model for shear strength prediction of squat RC walls. J Struct
2006. Eng 2021;147(11):04021173.
[82] Jolliffe IT. Principal component analysis for special types of data. Springer; 2002. [89] Mangalathu S, Karthikeyan K, Feng D-C, Jeon J-S. Machine-learning interpretabil-
[83] MacQueen J, et al. Some methods for classification and analysis of multivariate ity techniques for seismic performance assessment of infrastructure systems. Eng
observations. In: Proceedings of the fifth Berkeley symposium on mathematical Struct 2022;250:112883.
statistics and probability, Vol. 1. Oakland, CA, USA; 1967, p. 281–97.

17

You might also like