See discussions, stats, and author profiles for this publication at: [Link]

net/publication/382337232

Enhanced Time-History Analysis for Seismic Response Prediction of Stone

Masonry Structures: FEM vs ML

Conference Paper · June 2024

CITATIONS READS

0 51

2 authors:

Filip Đorđević Marko Marinković

University of Belgrade - Faculty of Civil Engineering University of Belgrade
19 PUBLICATIONS 66 CITATIONS 81 PUBLICATIONS 430 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Filip Đorđević on 18 July 2024.

The user has requested enhancement of the downloaded file.

 ENHANCED TIME-HISTORY ANALYSIS FOR SEISMIC RESPONSE
PREDICTION OF STONE MASONRY STRUCTURES: FEM VS ML

F. Đorđević1, M. Marinković2

1 PhD Candidate, Faculty of Civil Engineering, Belgrade, Serbia, fdjordjevic@[Link]

2 Assistant Professor, Faculty of Civil Engineering, Belgrade, Serbia

Abstract: The seismic vulnerability of masonry structures is a critical concern due to their heterogeneous and
nonlinear behavior during earthquakes. Traditional Finite Element Method (FEM) has been currently the most
common approach for predicting structural responses to seismic events. However, it has limitations in
capturing intricate and evolving damage patterns within structures, especially when faced with strong
earthquakes. This research evaluates the performance of traditional numerical analysis conducted in SAP2000
software and compares it with a data-driven approach using the eXtreme Gradient Boosting (XGBoost)
Machine Learning (ML) algorithm. The experimental tests and numerical analyses of two adjacent interacting
unreinforced stone masonry structures served for a comparison of the two approaches. The data has been
collected within the Horizon 2020 project, The Seismology and Earthquake Engineering Research
Infrastructure Alliance for Europe – Seismic Testing of Adjacent Interacting Masonry Structures (SERA-AIMS).
A particular focus of the project was on assessing the structural behavior during the earthquake occurred in
Petrovac, Montenegro in 1979. The results indicate that the novel methodology offers significant improvements
in predicting the seismic response of stone masonry structures when compared to the conventional finite
element method, even for a small amount of the available data. The outcomes show that the XGBoost method
processes data quickly, but they also highlight how crucial it is to have a representative database in order to
train and evaluate the ML model. They also designate the need for careful data selection, and fine-tuning of
hyperparameters, but also the sensitivity of ML model to changes in the level of shaking. These findings
emphasize the potential for integrating data-driven techniques into seismic assessment practices, facilitating
more accurate and efficient forecasting of structural behavior during seismic excitations.

1 Introduction
The seismic resilience of Unreinforced Masonry Structures (URMs) is a subject of paramount importance,
particularly in regions prone to seismic activity. Earthquakes have throughout history, rendered these
structures vulnerable to damage and collapse. The devastating consequences of earthquakes on masonry
structures underscore the urgent need for effective seismic response forecasting and mitigation strategies.
Conventional techniques for structural modeling are mostly based on FEM which allows for the simulation of
the structural behavior, stress distribution, and potential failure mechanisms, but also requires the adoption of
the modeling assumptions that sometimes can lead to significant design errors. Hence, to protect and preserve
masonry structures against the forces of nature, particularly seismic events, the fusion of advanced
engineering methodologies and predictive analytics has emerged as a compelling approach.
WCEE2024 Đorđević & Marinković

There are a large number of studies that deal with seismic analysis of structures, either using FEM or ML
(Butenweg et al., 2019; Đorđević and Marinković, 2023a, 2024; Kazemi et al., 2023; Marinković and Butenweg,
2022; Yu et al., 2019). As part of the H2020 SERA-AIMS project (Tomić et al., 2023), which served as the
basis for the data extraction used in the current study, several teams performed static and dynamic analyses
of two adjacent interacting masonry aggregates using FEM, with varying degrees of success (AlShawa et al.,
2022; Đorđević, 2024; Ramaglia et al., 2022). Successful guidelines for execution of nonlinear dynamic Time-
History (TH) analyses of masonry structures are defined by (Betti et al., 2015; Penna et al., 2013; Vaculik and
Griffith, 2008; Zheng et al., 2012).

Several authors attempt to predict the response of the system during the entire duration of the earthquake,
using ML algorithms such as Long-Short Term Memory (LSTM) (Zhang et al., 2019), WaveNet, and
Convolutional Neural Networks (CNNs) (Ning et al., 2023). Definitely based on the above, ML algorithms have
surged to the forefront of decision-making processes in engineering and scientific research (Chen and Guestrin,
2016; Xu et al., 2022; Zhang et al., 2021).

This paper presents the pioneering initiative to compare the results obtained from numerical analysis using
FEM within SAP2000 software and the predictive prowess of the XGBoost ML algorithm, to predict the TH
response of the masonry structures. The proposed algorithm has gained prominence for its ability to model
intricate relationships within data and make accurate predictions, making it an ideal candidate for estimation
of the structural seismic response. The overarching goal of this paper is to advance the forecasting of the TH
seismic behavior of masonry structures, using novel computational tool that encourage the improvement of
conservative FEM procedure.

2 Problem Definition
The goal of the SERA-AIMS project was to investigate the behavior of two adjacent masonry buildings by
testing a prototype of two buildings connected by dry joints, exposed to earthquake occurred in Petrovac,
Montenegro in 1979. The masonry walls of adjacent buildings are often connected by the so-called dry joints,
primarily due to the fact that those buildings were built at different times. Based on the experience of
earthquakes, it could be observed that such dry joints are very vulnerable places to the action of earthquake
impulses, which leads to the opening of joints and cracks between building units. From the above, it is clear
that there is a desire to improve the results of predicting the behavior of masonry structures by applying
advanced tools and improving current FEM-based modeling principles.

2.1 Structural Properties

Geometrical and material properties of the structure proposed in this work, can be found in (Tomić et al., 2023).
The FEM model was developed in SAP2000 software (see Fig. 1(a)), as a prototype without physical
connection (with a gap of 3 mm) between the units in order to simulate the real model as closely as possible.
It is important to note that walls are modelled using Shell-Layered-Nonlinear elements with released in-plane
and out-of-plane nonlinearities (Đorđević, 2024). The structure is free restrained, in order to avoid the activation
of bending moments, which would happen if the structure were fully restrained. Beams and walls are connected
with body constraints, while the finite elements within the walls are connected by edge constraints. Fig. 1(b)
presents a labels of the nodes in which the sensors are placed.

Time-History analysis allows a linear or non-linear dynamic response assessment, under a load that can vary,
depending on a given time-dependent function. Dynamic equilibrium equations can be solved using modal
methods or direct integration methods. Acceleration records as a function of time for two orthogonal directions
X and Y are defined in the next section. It should be noted that the actual sampling period from the base station
was 0.005 s, but based on the adopted parameter scaling principle, the sampling period applied in the FEM
analysis was 0.003536 s. Nevertheless, in order to show the robustness of the XGBoost ML algorithm even
with a smaller amount of input data, the original sampling period of 0.005s (200Hz frequency) was used.

2
WCEE2024 Đorđević & Marinković

(a) SAP2000 - 3D model (b) Sensor locations

Figure 1. (a) Numerical 3D model (SAP2000 software), (b) Arrangement of sensors for
measuring the response of the structure (Đorđević, 2024; Tomić et al., 2023)

2.2 Experimental Shake-Table Tests - Dataset

The database for this research was collected based on the experimental results of the shake-table test,
performed in the National Laboratory for Civil Engineering (LNEC), located in Lisbon, Portugal. Unlike the
planned earthquake execution program, in this study two intensities (25% and 50%) were considered, until
significant damage occurred in the experimental model, to demonstrate the performance of the FEM and ML
models on a relatively small dataset consisted of time series records. Fig.2 shows original ground motion
(acceleration) records of the earthquake incorporated in the shaking table test.

Figure 2. Acceleration records for 100% scaling factor of N-S and E-W directions

Table 1 contains the order of loading with the corresponding earthquake intensities. Within each intensity, the
acceleration records are given individually in the X and Y directions, but also simultaneously.

Table 1. Earthquake excitation levels

Level of shaking Substep I Substep II Substep III
25% intensity Y direction X direction Bidirection
PGA 0.156/0.219g (N-S/E-W) (Run 1.1) (Run 1.2) (Run 1.3)
50% intensity Y direction X direction Bidirection
PGA 0.313/0.438g (N-S/E-W) (Run 2.1) (Run 2.2) (Run 2.3)

3
WCEE2024 Đorđević & Marinković

3 Existing Solutions
In the available literature, there are attempts to develop a more efficient alternative procedures for surrogate
modeling of structural response based on either data-driven or physics-driven approaches. The authors
singled out three examples dealing with the prediction of nonlinear dynamic TH response, two of which used
the popular LSTM technique for data analysis, while one study deals with the application of FEM. The study
developed on the FEM model was created as part of the SERA-AIMS project (Đorđević, 2024) and as such is
used in this paper as the basis for the development of a novel alternative ML technique.

TH records presented in Fig. 3(a)-(b) are related to the implementation of data-driven LSTM paradigm, in an
attempt to assess the seismic behavior of the 6-story Pasadena hotel built in California-San Bernandino in
1970 (Zhang et al., 2019). It should be noted that the LSTM model was developed on the basis of a large
database containing 21 ground motion records from several earthquakes occurred between 1987 and 2018.
Such a large database is one of the reason for the very precise results obtained, which is confirmed by the
high agreement of experimental and predictive results. The same authors have also tried to develop LSTM
model in a physics-driven fashion, which was employed for response prediction of 3-story Moment Resisting
Frame (MRF) (Zhang et al., 2020).

The Fig. 3(c)-(d) show the efforts of (Đorđević and Marinković, 2023b) to solve the current problem also by
applying the LSTM model to a relatively small amount of available data for training and testing purposes.
Despite the lack of a representative data set, very satisfactory results were obtained.

(a) San Bernandino – LSTM full response (b) San Bernandino – LSTM detail

(c) SERA-AIMS - LSTM, Rd2 (25%) full response (d) SERA-AIMS – LSTM, Rd2 (25%) detail

Figure 3. Seismic responses obtained using LSTM models: (a) San Bernandino - Full, (b) San
Bernandino - Detail, (c) SERA-AIMS - Full (d) SERA-AIMS – Detail (Đorđević and
Marinković, 2023b; Zhang et al., 2019)

Results presented in Fig. 4(a)-(d) are developed by (Đorđević, 2024) under SERA-AIMS project, using
SAP2000 software for numerical modeling of two adjacent interacting stone masonry structures. It will be seen
in the next section that the presented FEM solution only approximately enable the assessment of the state of

4
WCEE2024 Đorđević & Marinković

the structure during the action of TH ground motion records. It is important to note that at higher earthquake
intensities, and especially with the presence of damage, the numerical model is not able to fully predict the
maximum displacements. The model developed on the FEM has also certain requirements that include:
Small-time computation steps and fine element mesh discretization, significant time consumption for model
calibration, the necessity of conducting numerous simulations within Incremental Dynamic Analysis (IDA),
initiation of modeling assumptions and parameter uncertainties that can lead to significant design errors.

(a) Rd2 – 25% (E-W) (b) Rd2 – 50% (E-W)

(c) Rd3 – 25% (N-S) (d) Rd3 – 50% (N-S)

Figure 4. Structural responses obtained by numerical analyses in SAP2000:

(a) Rd2 25%, (b) Rd2 50%, (c) Rd3 25%, (d) Rd3 50% (Đorđević, 2024)

4 Methodology
This work employed XGBoost as a robust and versatile algorithm to capture complex relationships between
seismic excitation and structural response in a data-driven manner. It is a scalable and effective distributed
gradient boosting tool for training machine learning models. XGBoost builds an ensemble of decision trees in
parallel (see Fig. 5). Each tree is constructed independently and concurrently, which allows for high parallelism.
The paradigm can handle a wide range of data types, including structured and unstructured data, and is
suitable for various machine learning tasks. As opposed to LSTM algorithm which involves sequential data
handling, XGBoost has built-in support for parallel processing, making it possible to train models on large
datasets in a reasonable amount of time. Important advantage of the proposed algorithm is that it often does
not require data normalization or standardization as a preprocessing step, which saves time during model
development. However, in such problems, the application of the XGBoost algorithm requires that the time
series dataset must be prepared by prior conversion into a supervised learning problem to start with a training
and testing phases.

5
WCEE2024 Đorđević & Marinković

Figure 5. Example of the Ensemble XGBoost trees architecture with a max_depth=3

4.1 Walk Forward Validation

Walk-Forward Validation (WFV) is a technique commonly used in time series forecasting and machine learning
for evaluating the performance of predictive models, particularly in scenarios where the data has a sequential
or temporal nature. This validation method extends the Rolling Cross Validation (RCV) technique, and is
designed to simulate real-world decision-making and prediction processes by incrementally updating the
model with new data as it becomes available.

In contrast, RCV is a method that allows the model to be tested and validated on multiple sub-samples of the
data, each containing a fixed number of observations or a fixed time window, with some degree of overlap
between adjacent sub-samples. The training and test sets are created by sliding a fixed-size window over the
time series data as presented in Fig. 6(b).

WFV naturally incorporates new data as it becomes available, allowing the model to adapt to shifting seismic
conditions (see Fig. 6(a)). In contrast, RCV does not inherently account for temporal dependencies and may
not capture the evolving nature of seismic responses. WFV has the benefits to use a larger amount of data for
calibration as compared to the RCV, but it assumes that the dimension of the train set varies with time. RCV
trains each time in a small dataset, that increasing the risk of overfitting at every round. Based on the above
states, WFV has better performances than traditional K-Fold Cross Validation splitting strategy for time series
forecasting, since it is not directly applicable to time series data due to its temporal nature. In this work WFV
is proposed for XGBoost hyperparameters tuning.

(a) Walk Forward Validation strategy (b) Rolling Cross Validation strategy

Figure 6. Validation strategies for time series records: (a) Walk Forward Validation,
(b) Rolling Cross Validation

4.2 Performance indicators

In data analysis, the evaluation and interpretation of results are critical processes that help researchers and
practitioners make informed decisions, validate models, and quantify the quality of predictions. Performance
indicators are essential tools for assessing the accuracy and reliability of models and analytical techniques. In
this paper, for the purpose of assessing the quality of the analysis, the Coefficient of Determination (R2), Mean
Squared Error (MSE), and Root Mean Squared Error (RMSE) were used as indicators of model performance.

6
WCEE2024 Đorđević & Marinković

n⋅ ∑ni=1(yi ⋅ y̅) n
i - ∑i=1 yi ⋅ ∑ni=1 y̅i
R2 =
√[n ⋅ (∑ni=1 yi 2 ) - (∑ni=1 yi )2 ] ⋅ [n ⋅ (∑ni=1 y̅i 2 ) - (∑ni=1 y̅i )2 ]
( )

1 n
2
MSE = ⋅∑ (yi - y̅)
i
n i=1

1 n
RMSE =√ ⋅∑ (yi - y̅) 2
i
n i=1

During the process of adjusting hyperparameters using the WFV technique, a sensitivity analysis was
performed to assess the influence of the most important hyperparameters of the XGBoost model on its
performance in predicting the TH seismic response of the stone masonry structures.

Fig. 7 illustrates the performance of the model in relation to the values of the main hyperparameters, i.e.,
n_estimators and max_depth. Hyperparameter n_estimators determines the number of boosting rounds or
decision trees to be built in the XGBoost ensemble, while max_depth specifies the maximum depth of each
decision tree within ensemble. It controls the complexity of the individual trees by limiting how many levels of
nodes the trees can have.

Based on a detailed analysis of the hyperparameters, the final model with 100 estimators and a maximum
depth of 6 was adopted. The criterion for selecting the most suitable model, utilizes a compromise between
the optimal time spent on performing the analysis with satisfactory accuracy and low complexity of the model
(see Fig. 7). Additionally, it is important to mention the value of the learning rate=0.3, as well as the
regularization parameters lambda and alpha, both with values of 0.01.

(a) Train set (b) Test set

Figure 7. Influence of the most important XGBoost hyperparameters on the results:

(a) Train, (b) Test

5 Dynamic Response Forecasting

This section contains the predicted results of nonlinear dynamic TH analysis of adjacent interacting stone
masonry structures, obtained by XGBoost algorithm. A comparative analysis of the XGBoost performance with
the traditional numerical model based on FEM was also performed. Table 2 provides information on the quality
of the results generated by the ML technique. It can be seen that authors adopted a 60/40% distribution of the
data into the training and test set. This is important due to the fact that the presence of a large test set while
maintaining a high accuracy of the model speaks of its ability to make an accurate forecast on new data, which
is significant for practical reasons.
However, it should be noted that the model achieved better results on the test set than on the training set, as
a consequence of the selected regularization factors lambda and omega, which caused the error function to

7
WCEE2024 Đorđević & Marinković

be inflated by adding weighting factors. Another possible reason may lie in the selected technique of model
validation, which was improved by successively passing through the rounds until the last fold.
It should be concluded that it is a highly efficient model from the perspective of time consumption, bearing in
mind that the calculation takes many times less than the FEM approach. In addition, XGBoost performs
analyzes about 20% faster than the LSTM algorithm, making it an ideal candidate for forecasting the seismic
response of masonry structures.
Table 2. Performance scores of the XGBoost algorithm (100 estimators / 6 depth)
Model Train / Test split [%] Dataset R2 RMSE (⋅10-3m) MSE (⋅10-5m2)
Train 0.944 9.956 9.911
XGBoost 60 / 40 Test 0.956 10.440 10.900

All 0.954 8.330 6.939

Fig. 8 illustrates previously mentioned splitting train/test strategy, with special reference to comparative
analysis with experimental observations. The presented results show that the ML model gives an almost
perfect match with the experiment in the displacement range [-0.2÷0.2], while there are certain minor deviations
outside of that. One of the potential reasons lies in the fact that the structure at an intensity of 50% has become
more significantly damaged, which leads to noticeable physical and mechanical changes, and thus affects the
predicted results. Other potential consequences can be the lack of sufficient amount of data but also a
noticeable range of positive and negative experimental results with the presence of noise.

Figure 8. Comparison of the experimental and XGBoost responses with proposed Train/Test
(60/40%) data splitting strategy

(a) Complete TH responses, XGB / Exp (b) Detail of TH response, XGB / Exp

Figure 9. TH responses on all data: (a) Complete records XGB/Exp, (b) Detail XGB/Exp

Nevertheless, it can be concluded that the proposed model has sufficiently overcome all obstacles and
provided superior predictions. Fig. 9 compares the XGBoost results with the experimental ones, derived for
each ground motion record (see Table 1) extracted from sensors.

8
WCEE2024 Đorđević & Marinković

Apart from the experimental results, the performance of the developed ML model was also compared with the
results of the numerical analysis performed in SAP2000. Fig. 10 shows an example of reading the experimental
results from the Rd2 sensor, where you can see an incomparably greater match achieved by XGBoost than is
the case with SAP2000.

(a) Rd2 – XGBoost – Full (b) Rd2 – SAP2000 - Full

(c) Rd2 – XGBoost - Detail (d) Rd2 – SAP2000 - Detail

Figure 10. Comparative results of the responses obtained by the XGBoost and SAP2000 models against
the experimental results for sensor Rd2 (25%): (a) XGBoost / Experimental - Full, (b)
SAP2000 / Experimental - Full, (c) XGBoost / Experimental - Detail, (d) SAP2000 /
Experimental - Detail

Similar outcomes were recorded for the readings of other sensors. Finally, XGBoost, as one of the state-of-
the-art machine learning algorithm, offers a multitude of advantages that render it exceptionally well-suited for
seismic time history response forecasting of masonry structures. It exhibits remarkable predictive capabilities,
empowering it to discern intricate patterns within seismic data and their impact on masonry structures. Last
but not least, its efficiency and scalability make it an invaluable asset for handling large datasets and
conducting extensive simulations.

6 Conclusions
This paper presents a novel XGBoost paradigm for the prediction of seismic structural response in terms of
TH records. The proposed model is specified to contain up to 100 estimators and has a maximum depth of 6.
The adopted settings aim to create an ensemble of decision trees that collectively provide accurate predictions
while constraining the individual tree complexity to mitigate overfitting. XGBoost has proven to have low time
consumption, and very accurate predictive performances even for modeling highly nonlinear tasks that contain
heterogeneous data. After the necessary restructuring of time series data in order to develop a supervised
learning algorithm, the model has shown its robustness even on a limited set of data, overcoming noises and

9
WCEE2024 Đorđević & Marinković

showing high adaptability. Traditional FEM procedures have specific modeling requirements, which can
occasionally compromise the reliability and accuracy of such a model, especially for complex tasks involving
dynamic structural analyses. It is important to note that the application of such ML methods in future research
may lead to mitigation of structural vulnerabilities to earthquake excitations. It is necessary to improve such
paradigms into physics-driven methods, which will be able to recognize the complex physics of problems
including geometric and material nonlinearities of structures. In order to make them accessible to the
engineering community, it is important to consider the development of ML software solutions, as well as the
introduction of suitable user-friendly graphical user interfaces to receive input parameters related to the
structural properties. Additional improvements can be made regarding the speed of performing analyses, either
by reducing the required number of iterations or by improving the existing computing paradigm (Babović, Bajat,
Barac, et al., 2023; Babović, Bajat, Đokić, et al., 2023), using the DataFlow paradigm as presented in
(Milutinovic et al., 2015; Milutinovic, Kotlar, et al., 2017; Milutinovic, Salom, et al., 2017). In related fields, there
has been an expansion of modern ML techniques that deal with time series data such as Autoregressive
Integrated Moving Average (ARIMA) (Siami-Namini et al., 2019) and Transformer Neural Network (Devlin et
al., 2018), which opens up the possibility for their potential application in structural engineering as well.

7 References
AlShawa, O., Liberatore, D. and Sorrentino, L. (2022), “Blind Prediction of Shake Table Tests of Adjacent
Interacting Masonry Structures: Combined Finite-Discrete Element Model”, 3rd European Conference on
Earthquake Engineering & Seismology.
Babović, Z., Bajat, B., Barac, D., Bengin, V., Đokić, V., Đorđević, F., Drašković, D., et al. (2023), “Teaching
computing for complex problems in civil engineering and geosciences using big data and machine
learning: synergizing four different computing paradigms and four different management domains”,
Journal of Big Data, Vol. 10 No. 1, p. 89.
Babović, Z., Bajat, B., Đokić, V., Đorđević, F., Drašković, D., Filipović, N., Furht, B., et al. (2023), “Research
in computing-intensive simulations for nature-oriented civil-engineering and related scientific fields, using
machine learning and big data: an overview of open problems”, Journal of Big Data, Vol. 10 No. 1, p. 73.
Betti, M., Galano, L. and Vignoli, A. (2015), “Time-history seismic analysis of masonry buildings: A comparison
between two non-linear modelling approaches”, Buildings, Vol. 5 No. 2, pp. 597–621.
Butenweg, C., Marinković, M. and Salatić, R. (2019), “Experimental results of reinforced concrete frames with
masonry infills under combined quasi-static in-plane and out-of-plane seismic loading”, Bulletin of
Earthquake Engineering, Vol. 17 No. 6, pp. 3397–3422.
Chen, T. and Guestrin, C. (2016), “XGBoost: A scalable tree boosting system”, Proceedings of the ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, Vol. 13-17-Augu, pp. 785–
794.
Devlin, J., Chang, M.-W., Lee, K. and Toutanova, K. (2018), “{BERT:} Pre-training of Deep Bidirectional
Transformers for Language Understanding”, CoRR, Vol. abs/1810.0, available at:
[Link]
Đorđević, F. (2024), “Numerical Modeling of Two Adjacent Interacting URM Structures”, IPSI Transactions on
Internet Research.
Đorđević, F. and Marinković, M. (2023a), “Implementation of Hybrid ANN-GWO Algorithm for Estimation of the
Fundamental Period of RC-Frame Structures”, Second Serbian International Conference on Applied
Artificial Intelligence, Kragujevac, Serbia, Kragujevac, Serbia.
Đorđević, F. and Marinković, M. (2023b), “A Comparative Study of ML and FEM Models for the Prediction of
Seismic Structural Behavior”, Natural Resources and Environmental Risks: Towards a Sustainable
Future, Branch of the Serbian Academy of Sciences and Arts in Novi Sad, Serbia.
Đorđević, F. and Marinković, M. (2024), “Advanced ANN Regularization-Based Algorithm for Prediction of The
Fundamental Period of Masonry-Infilled RC Frames”, Journal of Big Data, Springer, Submitted for review.
Kazemi, F., Asgarkhani, N. and Jankowski, R. (2023), “Machine learning-based seismic response and
performance assessment of reinforced concrete buildings”, Archives of Civil and Mechanical Engineering,
Springer London, Vol. 23 No. 2, pp. 1–21.
Marinković, M. and Butenweg, C. (2022), “Experimental testing of decoupled masonry infills with steel anchors
for out-of-plane support under combined in-plane and out-of-plane seismic loading”, Construction and
Building Materials, Vol. 318, p. 126041.
Milutinovic, V., Kotlar, M., Stojanovic, M., Dundic, I., Trifunovic, N. and Babovic, Z. (2017), “Implementing

10
 WCEE2024 Đorđević & Marinković

Neural Networks by Using the DataFlow Paradigm”, DataFlow Supercomputing Essentials: Algorithms,
Applications and Implementations, Springer International Publishing, Cham, pp. 3–44.
Milutinovic, V., Salom, J., Trifunovic, N. and Giorgi, R. (2015), Guide to DataFlow Supercomputing,
SpringerISBN: 9783319162287, available at:[Link]
Milutinovic, V., Salom, J., Veljovic, D., Korolija, N., Markovic, D. and Petrovic, L. (2017), DataFlow
Supercomputing Essentials: Research, Development and Education, Springer, available
at:[Link]
Ning, C., Xie, Y. and Sun, L. (2023), “LSTM, WaveNet, and 2D CNN for nonlinear time history prediction of
seismic responses”, Engineering Structures, Elsevier Ltd, Vol. 286 No. October 2022, p. 116083.
Penna, A., Rota, M., Mouyiannou, A. and Magenes, G. (2013), “Issues on the use of time-history analysis for
the design and assessment of masonry structures”, ECCOMAS Thematic Conference - COMPDYN
2013: 4th International Conference on Computational Methods in Structural Dynamics and Earthquake
Engineering, Proceedings - An IACM Special Interest Conference, No. January, pp. 669–686.
Ramaglia, G., Lignola, G.P. and Prota, A. (2022), “Simplified evaluation for the SERA AIMS blind prediction:
seismic behaviour of scaled masonry adjacent buildings”, 3rd European Conference on Earthquake
Engineering & Seismology.
Siami-Namini, S., Tavakoli, N. and Siami Namin, A. (2019), “A Comparison of ARIMA and LSTM in Forecasting
Time Series”, Proceedings - 17th IEEE International Conference on Machine Learning and Applications,
ICMLA 2018, No. April 2019, pp. 1394–1401.
Tomić, I., Penna, A., DeJong, M., Butenweg, C., Correia, A.A., Candeias, P.X., Senaldi, I., et al. (2023),
“Shake-table testing of a stone masonry building aggregate: overview of blind prediction study”, Bulletin
of Earthquake Engineering, available at:[Link]
Vaculik, J. and Griffith, M.C. (2008), “Time-history analysis for unreinforced masonry walls in two-way bending”,
The 14 World Conference on Earthquake Engineering, No. November.
Xu, J., He, J., Gu, J., Wu, H., Wang, L., Zhu, Y., Wang, T., et al. (2022), “Financial Time Series Prediction
Based on XGBoost and Generative Adversarial Networks”, International Journal of Circuits, Systems and
Signal Processing, Vol. 16, pp. 637–645.
Yu, Y., Wang, C., Gu, X. and Li, J. (2019), “A novel deep learning-based method for damage identification of
smart building structures”, Structural Health Monitoring, Vol. 18 No. 1, pp. 143–163.
Zhang, L., Bian, W., Qu, W., Tuo, L. and Wang, Y. (2021), “Time series forecast of sales volume based on
XGBoost”, Journal of Physics: Conference Series, Vol. 1873 No. 1, available
at:[Link]
Zhang, R., Chen, Z., Chen, S., Zheng, J., Büyüköztürk, O. and Sun, H. (2019), “Deep long short-term memory
networks for nonlinear structural seismic response prediction”, Computers and Structures, Elsevier Ltd,
Vol. 220, pp. 55–68.
Zhang, R., Liu, Y. and Sun, H. (2020), “Physics-informed multi-LSTM networks for metamodeling of nonlinear
structures”, Computer Methods in Applied Mechanics and Engineering, Elsevier B.V., Vol. 369, p. 113226.
Zheng, N., Zhou, J., Yin, Y., Han, J. and Ji, S. (2012), “Non-linear Time History Response Analysis of Low
Masonry Structure with tie-columns”, 15th World Conference on Earthquake Engineering, pp. 1–8.

11

View publication stats