REVIEW

published: 16 March 2022

doi: 10.3389/fbuil.2022.811460

Applications of Machine Learning to

Wind Engineering
Teng Wu 1* and Reda Snaiki 2
1
Department of Civil, Structural and Environmental Engineering, University at Buffalo, Buffalo, NY, United States, 2Department of
Construction Engineering, École de Technologie Supérieure, University of Quebec, Montreal, QC, Canada

Advances of the analytical, numerical, experimental and field-measurement approaches in

wind engineering offers unprecedented volume of data that, together with rapidly evolving
learning algorithms and high-performance computational hardware, provide an
opportunity for the community to embrace and harness full potential of machine
learning (ML). This contribution examines the state of research and practice of ML for
its applications to wind engineering. In addition to ML applications to wind climate, terrain/
topography, aerodynamics/aeroelasticity and structural dynamics (following traditional
Alan G. Davenport Wind Loading Chain), the review also extends to cover wind damage
assessment and wind-related hazard mitigation and response (considering emerging
performance-based and resilience-based wind design methodologies). This state-of-the-
art review suggests to what extend ML has been utilized in each of these topic areas within
wind engineering and provides a comprehensive summary to improve understanding how
Edited by:
Forrest J. Masters,
learning algorithms work and when these schemes succeed or fail. Moreover, critical
University of Florida, United States challenges and prospects of ML applications in wind engineering are identified to facilitate
Reviewed by: future research efforts.
Girma Bitsuamlak,
Western University, Canada Keywords: machine learning, wind engineering, wind climate, terrain and topography, aerodynamics and
Pedro L. Fernández-Cabán, aeroelasticity, structural dynamics, wind damage assessment, hazard mitigation and response
Clarkson University, United States
Muhammad Hajj,
Stevens Institute of Technology, 1 INTRODUCTION
United States
*Correspondence: Wind engineering is an interdisciplinary field to provide rational treatment of interaction between
Teng Wu the atmospheric boundary-layer winds and human activities (Cermak 1975). There is a long and
[email protected] significant history for machine learning (ML) applications in several subfields involved in wind
engineering, such as fluid mechanics (Brunton et al., 2020), meteorology (Chen et al., 2020) and
Specialty section: mechanics of structures (Salehi and Burgueño 2018). The application of statistical learning to
This article was submitted to turbulence modeling in early 1940s (Kolmogorov 1941) and perceptron learning to structural
Wind Engineering and Science, design in late 1980s (Adeli and Yeh 1989) are representative examples. On the other hand, it seems
a section of the journal
similar passions have not been shared by researchers in the wind engineering community. Actually,
Frontiers in Built Environment
ML-based wind engineering is still in its infancy stage and the full-capacity of ML has not been
Received: 08 November 2021
leveraged yet. However, the exceptional performance of ML to extract hidden informative features
Accepted: 27 January 2022
from data shows great promise for addressing unresolved complexities and issues originated from
Published: 16 March 2022
first principles investigations in the field of wind engineering. In addition, recent advances in
Citation:
performance-/resilience-based wind engineering have placed new demands on wind
Wu T and Snaiki R (2022) Applications
of Machine Learning to
characterization, aerodynamics modeling and structural analysis that need powerful simulation
Wind Engineering. tools such as ML to overcome the emerging challenges by simultaneously achieving high
Front. Built Environ. 8:811460. computational efficiency and accuracy. It is reasonable to expect the revitalization of ML
doi: 10.3389/fbuil.2022.811460 within the wind engineering field that is fueled by 1) rapidly evolving learning algorithms and

Frontiers in Built Environment | www.frontiersin.org 1 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

high-performance computational hardware, 2) unprecedented extended from synoptic events measured by airport wind
volume of data generated with improved wind engineering observation system with traditional anemometers to non-
techniques and methodologies, and 3) urgent needs for more synoptic events measured by several field campaigns with
accurate and efficient learning and modeling of complex advanced doppler radars and Lidars (Light Detection and
phenomena in wind-related problems. Ranging) [e.g., Verification of the Origins of Rotation in
As a key subfield of artificial intelligence (AI) [that together with Tornadoes Experiment (VORTEX) and Radar Observations of
natural intelligence plays a role of the computational part of the Tornadoes and Thunderstorms Experiment (ROTATE)
ability to achieve goals in the world (McCarthy 2007)], ML campaigns for tornado events and Severe Convective OUtflow
develops learning algorithms that use inputs from a sample in Thunderstorms (SCOUT) and Wind Ports and Sea (WPS)
generator and observations from a system to generate an campaigns for thunderstorm downburst events]. Massive wind
approximation of its outputs (Cherkassky and Mulier 2007). data over complex terrain/topography are collected by continuous-
The evolution of learning algorithms started when McCulloch wave short-range WindScanner systems (e.g., Berg et al., 2013).
and Pitts (1943) invented the first mathematical model of a neural The low Reynolds-number, straight-line-wind, stationary
network. In 1952, Arthur Samuel from IBM introduced the first aerodynamics data generated in conventional boundary-layer
self-learning computer program to play the game of checkers wind tunnels are extended to 1) high-Reynolds-number
(Wiederhold et al., 1990). Then, Rosenblatt (1957) designed the aerodynamics data resulting from recently built large-scale
first neural network for computers (the perceptron) that set the facilities [e.g., windstorm simulation facility at Insurance
foundation of deep neural networks (DNNs). Kelley (1960) Institute for Business and Home Safety (IBHS), Wall of Wind
presented the method of gradients (or method of steepest (WOW) at Florida International University and Wind Engineering
descent) in his analytical development of flight performance Energy and Environment (WindEEE) at Western University], 2)
optimization, which was used to develop the basics of a vortex-flow aerodynamics data produced by tornado simulators
continuous backpropagation model for training feedforward (e.g., tornado-like vortex simulator at Iowa State University and
neural networks (Rumelhart et al., 1986). On the other hand, VorTECH at Texas Tech University), and 3) transient
Hopfield (1982) created a feedback neural network that was aerodynamics data generated in emerging actively controlled
considered as the first recurrent neural network (RNN). LeCun wind tunnels (e.g., individually-controlled multi-fan wind
et al. (1989) combined convolutional neural network (CNN) and tunnels at Tongji University, University at Buffalo and
backpropagation algorithm to recognize handwritten digits. University of Florida). Also, significant nonlinear and inelastic
Watkins (1989) introduced the concept of Q-learning based on structural dynamics data under strong winds are being created in
Markov process to significantly enhance the practicability and laboratories due to advances in performance-based wind design
feasibility of reinforcement learning. Later, Cortes and Vapnik methodology (Abdullah et al., 2020). In addition to the
(1995) designed a support-vector network considered as a new experimental and field-measurement approaches the
learning machine for two-group classification problems with high comprehensive data are further enriched by high-fidelity large-
generalization ability. Hochreiter and Schmidhuber (1997) scale simulation tools that are advanced by theoretical
introduced a long short-term memory cell to address the long- developments in wind engineering field (Blocken 2014; Kareem
term dependency issue in RNN. To overcome the learning 2020), such as computational fluid dynamics/computational
difficulty in DNNs, Hinton et al. (2006) derived a fast, greedy structural dynamics (CFS/CSI)-based hybrid modeling of
algorithm that can learn deep, directed belief networks one layer at transient structural response (Hao and Wu 2018) and statistics-
a time and hence facilitate the rapid development of deep learning. based synthesis of nonstationary wind field (Wang and Wu 2021).
Recently, Goodfellow et al. (2014) proposed a generative The Computational Modeling and Simulation Center (SimCenter)
adversarial network consisting of two models (i.e., generative of the Natural Hazards Engineering Research Infrastructure
and discriminative models) that compete with each other in a (NHERI) program provides an effective way to integrate various
zero-sum game. The sophisticated ML algorithm needs the help of simulation tools (Deierlein and Zsarnóczay 2021). Furthermore,
advanced computational hardware [e.g., graphics processing unit novel real-time aerodynamics hybrid simulation techniques are
(GPU) and tensor processing unit (TPU)] to unlock its full emerging to effectively generate nonlinear and full-scale data in
potential (Berggren et al., 2020). For example, the great success wind engineering by seamlessly stitching the numerical modeling
of AlexNet (a deep CNN on GPU) is essentially attributed to its in computer and physical testing in wind tunnel (Wu et al., 2019;
ability to leverage GPU for training (Krizhevsky et al., 2012). Wu and Song 2019). Data quality is essential to facilitate curation
Equipped with both sophisticated algorithms and advanced and reuse of the diverse and large datasets generated in the field of
computational hardware, the learning machine (LM) is driven by wind engineering. There are numerous methods and criteria
data. Both the quantity (data rich and comprehensive) and quality specified by various wind engineering research groups/centers
of the training/testing data are important to ensure good to ensure the high data quality, and the NHERI DesignSafe
performance of ML applications. Wind engineering by nature is cyberinfrastructure platform recently suggested the best
a data-rich field (e.g., high spatial and temporal resolution), and it practices for detailed data quality assessment in terms of
is rapidly becoming a data-comprehensive domain due to recent metadata quality, data content quality, data completeness and
advances of analytical, numerical, experimental and field- representation and data publications review (Rathje et al., 2017).
measurement methods (Kareem and Wu 2013; Hangan et al., The improved understanding concerning the complex nature of
2017). The data of spatiotemporally varying wind flows are wind fields (e.g., nonstationary and non-Gaussian features), the

Frontiers in Built Environment | www.frontiersin.org 2 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

associated structural aerodynamics/aeroelasticity (e.g., transient and engineering field, and concluded with critical research gaps and
nonlinear features) and the resulting load effects (e.g., nonlinear and future prospects. While ML can augment the analytical
inelastic structural response), as well as the necessary shift from a approaches [e.g., data-driven discovery of closure models
prescriptive design approach to performance-based design (Raissi et al., 2019)], numerical schemes [e.g., data-driven
methodology and further to resilience-based design philosophy turbulence modeling (Duraisamy et al., 2019)], experimental
(i.e., improving the rapidity, robustness, resourcefulness and tests [e.g., data-driven active control of transient wind
redundancy), poses new challenges in wind engineering field. simulation (Li et al., 2021a)] and field measurements [e.g.,
Hence, there is an urgent need of more accurate and efficient data-driven sparse sensor placement (Manohar et al., 2018)]
learning and modeling tools for effective solutions. The in wind engineering, the review only focuses on its role to
conventional stationary and linear analysis framework for wind- complement existing methodologies and hence potentially
structure interactions established by Robert H. Scanlan extend/transform current lines of wind engineering research
(1914–2001) and Alan G. Davenport (1932–2009) has been very and practice.
successful due to its simplicity and applicability, however, its
shortcomings have begun to surface since the underlying
complexities associated with many wind engineering problems 2 BACKGROUND OF MACHINE LEARNING
clearly show a departure from implicit assumptions of
stationarity, Gaussianity and linear features. A number of semi- Machine learning (ML) is a subclass of artificial intelligence (AI)
empirical nonlinear reduced-order models have been developed in that extracts the underlying pattern within a set of data (e.g.,
this context and improvement in their efficiency and robustness is a Murphy 2012; Goodfellow et al., 2016; Mohri et al., 2018). To
topic of cutting-edge research in the wind engineering community acquire the hidden pattern and knowledge of a problem, the
(Wu 2013). Unfortunately, these reduced-order models do not learning process involves in general five important steps, namely
always have a satisfactory representation of the full nonlinear data collection, data preparation, training, evaluation and
equations which govern the complex phenomena in wind-related parameters tuning. Once the learning machine is trained based
problems. An alternate way is to utilize the CFD techniques, on the available data (usually retrieved from analytical solutions,
however, their computational effort is too high considering the numerical simulations, experimental tests or full-scale
three-dimensional nature of winds and associated bluff-body measurements), it can predict future or unseen events. Based
aerodynamics. While CFD plays a significant role in generating on the data fed into the learning machine, ML algorithms can be
high-fidelity data of complex wind-structure interactions, its high classified into four categories, namely supervised learning,
computational cost makes it not easy to be used either in an unsupervised learning, semi-supervised learning and
informational mode to enhance wind hazard-related planning reinforcement learning (Figure 1).
and development activities (e.g., risk mitigation that needs to To train the algorithm, the supervised learning fully depends
quickly run thousands of scenarios at minimal computational on labeled data, the unsupervised learning relies purely on
expense) or in an operational mode to support emergency unlabeled data and the semi-supervised learning combines
management and response associated with a wind hazard (e.g., limited labeled data with a large amount of unlabeled data.
decision making that needs real-time prediction capability under an For reinforcement learning (RL), there is essentially no
uncertain environment). To address the emerging challenges, data- predefined data. Although RL is occasionally treated as semi-
driven machine learning offers a promising approach that is capable supervised learning considering the agent learns from its own
of processing big data in wind engineering field as well as modeling experiences in terms of infrequent and partial rewards, it is
associated complex phenomena with high computational efficiency classified here into separate category to highlight there is no
and simulation accuracy. explicit, external supervisory information provided to the
With the rapid development of ML applications in wind learning agent. It is noted the kriging and polynomial chaos
engineering due to the confluence of advanced learning expansions as two widely-used, data-driven statistical
algorithms, high-performance computational hardware and interpolation approaches are not reviewed in this study.
big data, it is believed that a systematic review on this
subject is important to suggest to what extend ML has been 2.1 Supervised Learning
utilized in each of the topic areas within wind engineering and Supervised learning models are a set of algorithms that learn the
provide a comprehensive summary to improve understanding mapping, from given labeled training data, between known inputs
how learning algorithms work and when these schemes succeed and outputs. The trainable parameters of these models are
or fail. Specifically, a total of 65 ML algorithms (Appendix A) determined based on the minimization of the loss function.
are identified for their applications in the five topic areas of wind Supervised learning models usually require a large amount of
climate, terrain/topography, aerodynamics/aeroelasticity, reliable and unbiased data for training which might not be always
structural dynamics and damage assessment, and mitigation available. These algorithms can be employed for two important
and response. This review first presents technical background of tasks, namely regression and classification.
typical ML approaches in terms of supervised learning,
unsupervised learning, semi-supervised learning and 2.1.1 Regression
reinforcement learning (RL), followed by the state of research Regression is a type of supervised learning in which the output is a
and practice of ML applications to each topic area within wind numeric variable. Among many regression models, feed-forward

Frontiers in Built Environment | www.frontiersin.org 3 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

FIGURE 1 | Machine learning categories.

FIGURE 2 | Architecture of a typical FFNN.

neural networks (FFNN) are widely utilized in wind engineering

field [Figure 2]. They are statistical models inspired by biological
learning (McCulloch and Pitts 1943) and characterized by
adaptive weights between neurons which are tuned using a
learning algorithm from observed training data. For simplicity,
the FFNN is also denoted as artificial neural network (ANN) in FIGURE 3 | Architecture of a typical random forest classifier.
this study.
Deep neural networks (DNN) are also a type of FFNN
characterized by a deep architecture equipped with multiple 2.1.2 Classification
layers, and hence allows for better generalization and accuracy Classification is another type of supervised learning in which the
(Deng and Yu 2014; Pouyanfar et al., 2018). The convolutional output is a categorical variable or a class. Support vector machines
neural networks (CNN) is another important FFNN with sparse (SVM) (Scholkopf and Smola 2018) and random forest (RF)
convolutional matrices that are usually employed for pattern (Breiman 2001) are two classical examples of classification
recognition and image classification (Krizhevsky et al., 2012; algorithms. SVM classifier identifies a hyperplane in a high-
Goodfellow et al., 2016). Recurrent neural networks (RNN) are dimensional space in which a simple linear classification can be
a class of feedback neural networks that allow previous outputs to performed. RF classifier, on the other hand, fits a number of
be used as inputs while having hidden states and are suited to decision tree classifiers on various sub-samples of the dataset,
model time-dependent regression problems (e.g., Medsker and Jain then averages the results to improve outcome accuracy [Figure 3].
1999; Mandic and Chambers 2001). Long short-term memory
(LSTM) are an advanced version of RNN to alleviate the gradient 2.2 Unsupervised Learning
vanishing and exploding issue by only keeping necessary past Unsupervised learning models draw inferences from datasets to
information in future model states (Bengio et al., 1994). describe hidden structures from unlabeled data based on

Frontiers in Built Environment | www.frontiersin.org 4 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

FIGURE 5 | Architecture of a typical GAN model.

FIGURE 4 | Architecture of a typical autoencoder model. has been attracting attention in fluid mechanics community for
efficient development of reduced-order models.

inherent characteristics (Russell and Norvig 2016). These 2.3 Semi-Supervised Learning
models usually group instances of input data using a defined Semi-supervised learning models operate based on limited
similarity index (global criterion). Clustering and labeled data with a large amount of unlabeled data. Hence,
dimensionality reduction are two standard examples of they can be regarded as combination results of supervised
unsupervised learning applications. learning and unsupervised learning algorithms. The generative
adversarial network (GAN) is a well-known semi-supervised
2.2.1 Clustering learning algorithm for estimating generative models via an
Clustering is an unsupervised learning task used for pattern adversarial process. One important feature of semi-supervised
recognition that automatically discovers natural groups or clusters learning algorithms is their labelled-data efficiency. To this end, it
in data. A cluster refers to a collection of data points aggregated may be reasonable to consider the physics-informed deep
together with similar features (Maulik and Bandyopadhyay 2002). The learning (PIDL) as a semi-supervised model that leverages
k-means clustering is one of the simplest unsupervised ML models. It physics-based equations in the augmented loss function to
is a centroid-based algorithm that partitions the data into k clusters. significantly reduce the data demand during training process.
Mean-shift clustering is another unsupervised model with a sliding-
window-based algorithm to identify dense areas of data points. Other 2.3.1 Generative Adversarial Network
clustering algorithms such as the density-based spatial clustering of The GAN model consists of two competing neural networks,
applications with noise, the expectation–maximization clustering namely the generator and the discriminator (Goodfellow et al.,
using gaussian mixture models and the agglomerative hierarchical 2014). It generates new data based on a probability distribution
clustering are also popularly used for statistical data analysis. that approximately represents the training data (true or labelled
data). Specifically, the generator produces fake samples to imitate
2.2.2 Dimensionality Reduction the distribution of a real dataset, then the discriminator tries to
Dimensionality reduction aims to find the most important distinguish (through a classification process) between the real
features within the dataset by identifying lower-dimensional samples and fake ones (from the generator). The GAN model is
representations for high-dimensional data. It minimizes the trained such that the new generated samples accurately represent
storage space, reduces the computation time and avoids the underlying mechanisms of the studied system. The
overfitting. The ML-based dimensionality reduction can be architecture of a typical GAN model is illustrated in Figure 5.
divided into linear and nonlinear algorithms. The principal
component analysis (PCA) is a commonly used linear 2.3.2 Physics-Informed Deep Learning
technique that can be regarded as a two-layer neural network The concept of PIDL models was originally proposed several
with a linear activation function. It essentially provides new decades ago (Psichogios and Ungar 1992; Dissanayake and Phan-
uncorrelated variables, also denoted as principal components, Thien 1994) in which prior knowledge (in terms of the physics-
which maximize the variance. The nonlinear autoencoder is a based governing equations) is integrated within the neural
specific type of FFNN that compresses the initial input space into networks to reduce the high-volume of required training data.
a reduced dimensional space using the encoder and then Typically, a small amount of labelled data along with a large
decompresses the obtained latent space back to the original number of unlabeled data that satisfy the underlying physics of
input space using the decoder. Accordingly, deep autoencoders the system of interest (also denoted as collocations points) are
have a “bottleneck” architecture designed for extraction of used to train these models. Hence, self-supervision plays a
representative features [Figure 4]. The autoencoder algorithm significant role in PIDL models. Recently, Raissi et al. (2017a,

Frontiers in Built Environment | www.frontiersin.org 5 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

FIGURE 6 | Architecture of a typical PIDL model.

dynamics (following traditional Alan G. Davenport Wind

Loading Chain), the review also extends to cover wind damage
assessment and wind-related hazard mitigation and response
(considering emerging performance-based and resilience-based
wind design methodologies). Considering the overwhelming
number of existing research publications, this review is by no
means exhaustive. Rather, it attempts to provide a state-of-the-art
perspective on ML applications to wind engineering-related
fields.

3.1 Wind Climate

FIGURE 7 | Architecture of a typical RL. The review of ML applications to wind climate is organized by
classifying it into classical boundary-layer winds, tropical
cyclones and non-synoptic events. By leveraging the
b) advanced the PIDL models by leveraging the automatic increasingly available datasets (e.g., satellite data), ML has
differentiation technique to solve partial differential equations. become a supporting tool or even a reliable competitor of
The architecture of a typical PIDL model is presented in Figure 6. classical approaches for wind climate modeling (e.g., CFD).
Most reviewed articles employed ML algorithms as a
2.4 Reinforcement Learning regression (e.g., long-term prediction of surface wind speed) or
RL algorithm is usually formulated based on Markov decision a classification (e.g., downburst occurrence prediction) tool. The
process (Sutton and Barto, 2018). The core part of RL is its agent selected metrics to evaluate the performance of ML algorithms
that interacts with its environment. Accordingly, the agent learns a included the root mean square (RMS), coefficient of correlation,
policy that maps the states to the actions by maximizing the expected mean squared error (MSE), mean absolute error (MAE), mean
cumulative reward using an automated trial-and-error process (e.g., absolute percentage error (MAPE), coefficient of determination
Mnih et al., 2015; Silver et al., 2017). Typical reinforcement learning (R2), among others.
models include value-based models (e.g., Q-learning or deep
Q-learning) (Watkins and Dayan 1992), policy-based models 3.1.1 Classical Boundary-Layer Winds
(e.g., deep deterministic policy gradient) (Lillicrap et al., 2015) Air movement in the planetary boundary layer plays a
and hybrid models (e.g., actor-critic) (Williams 1992). Recently, fundamental role in current wind design of structures and
the deep RL (with DNN-based policy) has been gaining attention in infrastructure. Although a detailed universal description of
wind engineering community as an efficient way for dynamic flow characteristics in the boundary-layer region has not been
control and shape optimization (Li et al., 2021a; 2021b). The possible, the classical boundary-layer winds in gales from large
architecture of a typical deep RL is depicted in Figure 7. depressions or in monsoons can be well represented by a number
of empirical or semi-empirical models [e.g., power-law profile for
distribution of mean wind speed (Davenport 1960) and power
3 APPLICATIONS OF MACHINE LEARNING spectrum for turbulent fluctuations (Panofsky and McCormick
TO WIND ENGINEERING 1960)]. The major research efforts have been focused on the
accurate estimate of design wind speed in a statistical analysis
This section provides a comprehensive review of the state of framework (Simiu and Scanlan 1978). Specifically, long-term
research and practice of ML for its applications to wind wind data from meteorological observations are analyzed
engineering. In addition to ML applications to wind climate, based on extreme value theory to obtain the design wind
terrain/topography, aerodynamics/aeroelasticity and structural speed at each location. However, the accurate forecast of

Frontiers in Built Environment | www.frontiersin.org 6 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

classical boundary-layer winds is very challenging since it 3.1.2.1 Tropical Cyclone Genesis
involves a large range of various temporal and spatial scales TC genesis requires several necessary environmental conditions
(e.g., from fractions of a meter to several thousand kilometers for (e.g., existence of low-pressure area and sea surface temperature
spatial scale and from fractions of seconds to several years for of at least 26°C), however, the exact mechanisms of TC formation
time scales). Usually, the temporal and spatial resolutions from are still not well understood (Gray 1968, 1979; Emanuel 2003;
the state-of-the-art weather forecast models [e.g., global forecast Holton and Hakim 2013). To predict the TC genesis, both
system from National Oceanic and Atmospheric Administration numerical and statistical models were developed. The
(NOAA)] are not sufficient for wind engineering purpose. On the numerical models (e.g., global forecast system) are essentially
other hand, the unprecedented volumes of data from field based on the physical principles and their performance heavily
measurements (e.g., weather station and satellite) provide a depends on improved understanding of TC genesis mechanism.
solid foundation to advance ML applications for classical The statistical models (e.g., Michael 2017; Chen and Duan 2018;
boundary-layer winds. Cui and Caracoglia 2019) linearly relate the TC genesis to a few
Table 1 presents the reviewed applications of ML for classical selected environmental factors, and hence show poor
boundary-layer winds, where the ML model, training scheme, interpolation and limited predictability. The lack of a deep
input data, output data, data source and performance metric are understanding of underlying mechanisms stimulated data-
summarized for each application. The training/testing data were driven techniques for TC genesis simulations. As a result,
essentially retrieved from field measurements. From Table 1, it increasing ML applications are available to accurately predict
can be concluded that most applications used ML as a regression TC genesis. Table 2i presents the reviewed applications of ML for
model for prediction of mean wind speed (averaging time TC genesis, where the ML model, training scheme, input data,
ranged from minutes to months), while the short-term output data, data source and performance metric are summarized
prediction of turbulent fluctuations that are very important for each application. The training/testing data were essentially
to structural dynamics is very limited. In many applications, retrieved from satellite measurements along with reanalysis
the selection of ML models is simply based on gut feeling or past results. It is expected the improved spatial resolution of
experience. Although several researchers conducted comparison currently available datasets will further enhance simulation
studies to select good ML models for their specific applications, results of ML models. From Table 2i, it can be concluded that
it might be very challenging to generalize the obtained results to most applications used ML as a classification model for either
other applications due to a lack of a systematic comparison short-term or long-term forecasting of TC genesis. Although
framework. more dynamic and thermodynamic environmental factors can be
retrieved using advanced remote sensing technologies in recent
3.1.2 Tropical Cyclones years, the identification of the most appropriate set of inputs to
Tropical cyclones (TCs), also commonly known as hurricanes ML models (predictors) is still very challenging.
in North Atlantic, typhoons in western North Pacific and
cyclones in Australia, are low-pressure storms that form over 3.1.2.2 Tropical Cyclone Translation
a warm ocean surface (Holton and Hakim 2013). With an Numerical forecast models have been successfully applied in
average of 90 events reported annually (Zhao et al., 2012), forecasting normal TC trajectories, but they are
TCs and their cascading hazards (e.g., wind, rain, storm surge computationally expensive. Although several statistical models
and wave) pose a serious threat to public safety, livelihoods were also developed based on a large amount of historical TC path
and local economies in many coastal regions around the records (e.g., Vickery et al., 2000,2009; Emanuel et al., 2006; Hall
globe. Hence, significant efforts have been made in and Jewson 2007; Chen and Duan 2018; Snaiki and Wu 2020a;
modeling and predicting TCs and relatively well- Snaiki and Wu 2020b), their linear nature makes them incapable
established mesoscale numerical weather prediction of capturing the inherent nonlinearities in such a complex
frameworks [e.g., Weather Research and Forecasting dynamic system (Zhang and Nishijima 2012). Both numerical
(WRF) model] are available for high-fidelity simulations. and statistical models or their combinations (statistical-dynamics
However, the high-fidelity computationally expensive models) show poor performance in forecasting sudden speed
models might not be always appropriate for planning change, recurvature and stagnation in TC movement (Chen et al.,
activities in an uncertain environment where Monte Carlo 2020). To satisfy both simulation accuracy and efficiency,
simulations are needed or emergency managements where increasing ML applications emerged for TC path prediction.
real-time or near-real-time predictions are required. The Table 2ii presents the reviewed applications of ML for TC
high demand for a rapid and reliable technique used to translation, where the ML model, training scheme, input data,
assist decision-makers and planers results in many ML output data, data source and performance metric are summarized
models for efficient simulations of key stages in the life for each application. The training/testing data were essentially
cycle of a TC. These ML applications to TCs are fueled by retrieved from meteorological databases (e.g., satellite data) and
increasingly available remotely-sensed and high-fidelity reanalysis results. Typically, the TC track information is available
numerical data. The review in this section is organized only at each 6-h interval. From Table 2ii, it can be concluded that
following the four important components of full track of a most applications used ML as a regression model for TC path
TC, namely genesis, translation, intensity and wind field. prediction. Since the forecast of TC track can be regarded as a

Frontiers in Built Environment | www.frontiersin.org 7 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 1 | Summary of ML applications for classical boundary-layer winds.

Application ML model Training scheme Input data Output data Data source Performance Remarks
metric

Forecasting LNN-ANN-NLN- Gradient descent, Past mean hourly Next mean hourly Field RMS NLN with logic
mean hourly RBF-ANFIS-ERNN Levenberg Marquardt data (six past wind speed measurements at rule outperformed
wind speed measurements) the Odigitria of the all other models
time series Greek island of
Sfetsos (2000) Crete on March
1996 (total of
744 h)
Forecasting ANN-JRNN Back-propagation, Past daily-, Next daily, weekly Field Coefficient of Best performance
daily, weekly cascade correlation weekly- and and monthly measurements correlation by RNN trained
and monthly monthly averaged averaged mean from 1989 to 2000 with the cascade
mean wind mean wind speed wind speeds in two locations in correlation
speeds More Mumbai, India
and Deo
(2003)
Prediction of ANN-SVM Levenberg-Marquardt mean daily wind Next daily mean 12 years of mean MSE SVM model
the next daily speed of previous wind speed daily wind speed in outperformed the
mean wind days (ranging Medina city, Saudi ANN model
speed between 1 and 11) Arabia
Mohandes
et al. (2004)
Long-term IIRANN, DRNN, Global recursive 3-days forecast of Hourly mean wind Atmospheric MAE-RMS Similar
wind speed LAFMN prediction error wind speed and speed and power modeling system performance
and power direction provided for up to 72-h SKIRON and wind results for the
forecasting by meteorological turbines data from three models
Barbounis models at four April 1st, 2000 until
et al. (2006) nearby sites 31 December
2000 in Rokas’
wind park on the
Greek island of
Crete
Short-term ANFIS least-squares estimator 4 to 6 past mean Next mean wind 21-month time Mean absolute ANFIS model
mean wind and the gradient descent wind speeds and speed and series of 2.5 min percentage error outperformed a
speed direction with a direction at mean wind from locally developed
forecasting 2.5 min time step 2.5 min Hydro Tasmania at persistence
Potter and Tasmania, model
Negnevitsky Australia
(2006)
Prediction of ANN, RBF, ALEN Levenberg–Marquardt Past hourly mean Next hourly mean Anemometers Mean absolute ANN
the next hourly wind speed wind speed data for 1 year error and RMSE outperformed
mean wind observations (up (2002) in two sites other models
speed Li and to 8 observations) in North Dakota
Shi (2010)
Prediction of SVM (radial basis - Past hourly mean Next hourly mean Sidi Daoud wind RMSE and MAE Satisfactory
the hourly kernel) wind speed and wind speed and farm in Tunisia results
mean wind direction in the site direction for a from 2010 to 2011
speed and (up to 10 past lead time up
direction samples) to 10 h
Lahouar and
Slama (2014)
Short-term ANN, SVM (RBF Gradient descent Wind direction at Next 1–2 min 34 days data from Mean absolute SVM model
wind direction kernel) past minutes wind direction the 34th America’s error and mean outperformed the
forecasting Cup in 2013, San effectiveness ANN model
Tagliaferri et al. Francisco index
(2015)
Prediction of ANN, GRNN Missing 10 variables (e.g., Monthly averaged Data retrieved from MSE and RMSE GRNN
the monthly latitude, longitude, mean wind speed NASA outperformed the
averaged earth temperature, corresponding to ANN model
mean wind atmospheric various cities in
speed Kumar pressure) India
and Malik
(2016)
Predict of Hybrid model - Past values of the Mean wind speed Data from several MAE and RMSE WPD-DBSCAN-
short-term (wavelet packet wind speeds for up to 1 day sites in the Sichuan ENN
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 8 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 1 | (Continued) Summary of ML applications for classical boundary-layer winds.

Application ML model Training scheme Input data Output data Data source Performance Remarks
metric

mean wind decomposition determined based with a 10-min Province, China outperformed all
speed Yu et al. [WPD] + density- on the gradient time step over 16 days with other models
(2018) based spatial boosted an average wind
clustering of regression trees speed of 10 min
applications with
noise [DBSCAN] +
ENN), WPD-
ENN, ENN
Time-series ANN, SVR, FIS, Bayesian Regularization, Past values of Mean wind speed Osorio wind farm RMSE, MSE SVR, GMDH and
prediction of ANFIS, GMDH Scaled Conjugate wind speed for approximately in the south of ANFIS models
mean wind Gradient, BFGS Quasi- (number not 361-time steps Brazil preformed the
speed Newton, Levenberg mentioned) ahead with best. The
Khosravi et al. Marquardt and Resilient several time prediction
(2018a) backpropagation intervals (e.g., 5- accuracy of
min and 30-min) ANFIS was
increased when
coupled with
particle swarm
optimization
(PSO) and genetic
algorithm (GA).
The Levenberg
Marquardt
performed the
best
Prediction of ANN, SVR (with Levenberg Marquardt, Pressure, local Mean wind Wind farm in RMSE Levenberg
the mean wind radial basis Conjugate Gradient and time, temperature speed, direction Bushehr, Iran Marquardt and
speed, function), ANFIS Bayesian Regularization and relative and power (in 5- Bayesian
direction and humidity min, 10-min, 30- Regularization
power min and 1-h algorithms gave
Khosravi et al. intervals) for up the best
(2018b) to 24 h performance
for ANN.
SVR was the best
to simulate the
wind speed. Low
prediction results
were obtained by
the 3 models for
the wind direction
Short-term ESN, ANFIS, - Past values of Mean wind speed Several Nevada RMSE The best
prediction of NESN-P with wind speed and and direction at weather prediction results
wind speed polynomial direction at time 10 min intervals information given by the
and direction functions and interval of 10 min for up to 1 day stations in Reno, NESN-MP
Chitsazan et al. NESN-MP with (the exact number and 6 days, Nevada
(2019) multivariable was not specified) respectively
polynomials
Probabilistic Ensemble of 3 Adaptive momentum Past values of Wind gusts for up Sutong Cable- RMSE, MAE and The ensemble
prediction of machines models estimation wind speed (the to 72 h Stayed Bridge in the mean model achieves
the wind gusts (RF, LSTM and number was Jiangsu province absolute percent the highest
Wang et al. GPR), RF, LSTM determined based of China (sampling error (MAPE) accuracy
(2020) and GBRT on the partial frequency of 1 Hz
autocorrelation with a total of a
function) total of 720 h)
Prediction of MFQL, SVR, KNN - 7 intrinsic mode 1-min ahead National Institute of Mean Absolute MFQL
mean wind functions obtained mean wind speed Wind Energy and Percentage Error outperformed the
speed Sharma from past wind Wind Resource (MAPE) other models
et al. (2020) speed values Assessment data
using empirical portal (in ten Indian
model cities)
decomposition
technique
-
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 9 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 1 | (Continued) Summary of ML applications for classical boundary-layer winds.

Application ML model Training scheme Input data Output data Data source Performance Remarks
metric

Prediction of LSSVM optimized Past values of 1-h mean wind The training data RMSE, MAE, LSSVM optimized
mean wind with four mean (hourly) wind speed (next were sampled mean absolute with the
speed algorithms speed (50 values) 1-48 h) every 1-h from a percentile error backtracking
Zhongda Tian Backtracking wind farm in (MAPE), search
et al. (2020) search, genetic Jinzhou, China R-square and optimization
algorithm, particle reliability algorithm
swarm, and outperformed all
improved feature other models
selection

time series prediction problem, the feedback neural networks Wu 2017a; Snaiki and Wu 2017b; Snaiki and Wu 2018; Snaiki and
such as RNNs and LSTMs are preferred and lead to good Wu 2020c; Fang et al., 2018; He et al., 2019) to simulate the
performance. However, their performance within each 6-h boundary-layer wind field. However, none of these models can
interval is unknown due to the sampling limitation in the simultaneously achieve simulation accuracy and efficiency. To
training data. address this issue, increasing ML applications emerged for TC
boundary-layer wind field simulation. Table 2iv presents the
3.1.2.3 Tropical Cyclone Intensity reviewed applications of ML for TC wind field, where the ML
The TC intensity (over ocean or land) can be measured in terms model, training scheme, input data, output data, data source and
of central pressure or maximum sustained wind speed. It is performance metric are summarized for each application. The
impacted by several complicated physical phenomena (e.g., training/testing data were essentially retrieved from meteorological
atmosphere-ocean interaction and vertical wind shear), and hence databases (e.g., satellite data) and high-fidelity simulations. It is
remains one of the most challenging issues in TC forecasting especially expected the improved spatial resolution of currently available
for rapid intensification prediction. To avoid the high computational datasets will further enhance simulation results of ML models.
cost of numerical forecast models, both statistics-based (e.g., Vickey From Table 2iv, it can be concluded that most applications use ML
et al., 2000; DeMaria et al., 2005; Hall and Jewson 2007; Vickey et al., as a regression model for prediction of surface wind speed. Since
2009) and physics-based (e.g., Snaiki and Wu 2020a) tools were these ML models were often trained and fine-tuned to predict the
developed for fast prediction of TC intensity. However, neither TC wind field at a specific region, it might be very challenging to
statistical nor physical models guarantee prediction accuracy of TC generalize the obtained results to other locations. It is noted that
intensity due essentially to the over-simplification of such a only wind field at a certain altitude is available in most ML
complicated dynamic system. To improve simulation accuracy applications due essentially to training data sparsity issue in
while keeping a high efficiency, increasing ML applications are vertical dimension. The widely-used logarithmic or power-law
available for TC intensity prediction. Table 2iii presents the profiles are typically employed to obtain the TC boundary-layer
reviewed applications of ML for TC intensity, where the ML winds. Accordingly, the supergradient winds that may have
model, training scheme, input data, output data, data source and significant implications to the wind design of tall buildings is
performance metric are summarized for each application. The not captured (Snaiki and Wu 2020c).
training/testing data were essentially retrieved from meteorological
databased (e.g., satellite data) and reanalysis results. From Table 2iii, it 3.1.3 Non-synoptic Winds
can be concluded that most applications used ML as a regression (or a Unlike synoptic winds that are associated with large-scale
classification) model for estimation of intensity time series (or levels). meteorological systems characterized by horizontal scales of
Although encouraging simulation results indicate a good performance thousands of kilometers and time scales of days, the non-
of ML models in predicting TC intensity for their specific applications, synoptic wind systems are local phenomena (e.g., a horizontal
the selection of the most appropriate set of inputs (including the scale of several hundreds of meters) and short lived (e.g., a time
number of predictors and previous time steps) is still very challenging. scale of a few minutes) (Chowdhury and Wu 2021). Furthermore,
In addition, it is not easy to conduct a systematic comparison among the transient nature of non-synoptic winds makes them exhibit
reviewed ML models since the used performance metrics differ time-varying mean wind speeds and nonstationary/non-
substantially from one application to another. Gaussian fluctuations. Accordingly, the detection,
measurement, and modeling of non-synoptic wind systems lag
3.1.2.4 Tropical Cyclone Wind Field behind those of synoptic winds. However, numerous studies have
TC wind hazard is of great significance since it (directly) induces demonstrated the importance of the non-synoptic wind events on
significant damage to life and property and (indirectly) triggers the structural design (e.g., Holmes 1999; Letchford et al., 2002;
other TC-induced hazards (e.g., storm surge and waves). Hao and Wu 2017). For example, the design wind speeds with
Substantial research efforts have been made for development of relatively high return periods are usually dominated by the
numerical models (e.g., WRF) or analytical models (e.g., Snaiki and thunderstorm downbursts (Twisdale and Vickery 1992; Solari

Frontiers in Built Environment | www.frontiersin.org 10 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

et al., 2015) and the ASCE 7–22 includes the first-ever criteria for Wu 2018; Oreskovic et al., 2018; Oreskovic and Savory 2018; Iida
tornado-resistant design (ASCE, 2021). Recently, there is a rapid and Uematsu 2019) or experimentally using wind tunnels (e.g.,
development of field-measurement networks (e.g., THUNDERR Jesson et al., 2015; Jubayer et al., 2016; Hoshino et al., 2018;
project at University of Genova) and laboratory facilities (e.g., Aboutabikh et al., 2019; Asano et al., 2019; Junayed et al., 2019;
WindEEE at Western University) for improved understanding of Romanic et al., 2019). Both numerical and experimental
non-synoptic wind systems. These advances offer an approaches to obtain wind fields associated with downbursts
unprecedented volume of data, and hence provide an are very time consuming (either computational expensive or
opportunity to facilitate ML applications to non-synoptic labor intensive). This shortcoming motivated increasing use of
winds. Although the non-synoptic wind systems can be ML tools for efficient and accurate simulations of downbursts.
originated from various mechanisms (e.g., convective storm, Table 3ii presents the reviewed applications of ML for
gravity wave or negative buoyancy) (Bluestein 2021), the downbursts, where the ML model, training scheme, input data,
review only focuses on those associated with convective output data, data source and performance metric are summarized
storms. Specifically, ML applications to thunderstorms for each application. The training/testing data were essentially
(subsynoptic-scale weather system) are first presented, retrieved from field measurement. From Table 3ii, it can be
followed by detailed reviews of its applications to two concluded that most applications used ML as a classification
important types of non-synoptic wind events associated with model for prediction of the occurrence of downburst or
thunderstorms, namely downbursts and tornadoes. probability of damaging wind. There are a very limited
number of ML applications for modeling and forecasting the
3.1.3.1 Thunderstorms downburst wind field, hence more research efforts are needed in
A thunderstorm is short-lived atmospheric weather system this aspect. It is noted that the reviewed ML applications usually
accompanied by lightning and thunder, gusty winds, heavy rain, involved a high number of predictors. The employment of
and sometimes hail (Solari 2020). The life cycle of a thunderstorm relatively high number of input variables may be necessary
usually consists of cumulus stage, mature stage and dissipative stage, due to the complexity of downburst prediction. However, it
and it typically lasts around 30 min. Both mesoscale and microscale makes the ML models not easy to use since these input
numerical models have been developed for simulation of variables might not be always available.
thunderstorms (Hawbecker 2021). Mesoscale modeling covers a
large-scale computational domain (and hence fully considers physics 3.1.3.3 Tornadoes
involved), however, it is limited to a low spatiotemporal resolution. Tornadoes are characterized by a rotating column of air
Microscale modeling utilizes a high spatiotemporal resolution (and descending from supercell thunderstorms lasting from several
hence obtains important small-scale features in the simulation of minutes to few hours. They are the most intense of all non-
winds), however, it is limited to a relatively small-scale synoptic wind events, and hence result in significant damage and
computational domain resulting in insufficiently reliable collapse of structures (Hao and Wu 2016, 2020). Several
boundary conditions. To avoid shortcomings of currently analytical and empirical models have been developed to
available numerical models, ML models may provide a promising simulate the vertical and radial wind profiles of tornado-like
approach for efficient and accurate simulation of key stages in the life vortices (e.g., Wen and Chu 1973; Baker and Sterling 2017). These
cycle of a thunderstorm. Table 3i presents the reviewed applications models are clearly over-simplified. The tornado wind fields are
of ML thunderstorms, where the ML model, training scheme, input also modeled using CFD simulations (e.g., Kuai et al., 2008;
data, output data, data source and performance metric are Ishihara et al., 2011; Liu and Ishihara 2015; Eguchi et al.,
summarized for each application. The training/testing data were 2018; Gairola and Bitsuamlak 2019; Kawaguchi et al., 2019;
essentially retrieved from meteorological databases and reanalysis Huo et al., 2020; Liu et al., 2021) or laboratory tests (e.g.,
results. From Table 3i, it can be concluded that most applications Sarkar et al., 2006; Refan and Hangan 2016; Razavi and Sarkar
used ML as either a classification or a regression model for prediction 2018; Tang et al., 2018; Ashton et al., 2019; Gillmeier et al., 2019;
of thunderstorm occurrence. Obviously, there is still room for more Hou and Sarkar 2020; Razavi and Sarkar 2021). However, CFD
comprehensive applications of ML in terms of modeling and simulations of tornadoes are computational expensive while the
forecasting each aspect of the thunderstorm from formation to laboratory tests are labor intensive. These shortcomings
dissipation. In addition, most ML applications to thunderstorm motivated increasing use of ML tools for efficient and accurate
were limited to simple models with standard algorithms (e.g., ANN modeling of tornadoes. Table 3iii presents the reviewed
with backpropagation). applications of ML for tornadoes, where the ML model,
training scheme, input data, output data, data source and
3.1.3.2 Downbursts performance metric are summarized for each application. The
Downbursts are one of the most spectacular and dangerous training/testing data were essentially retrieved from
events resulting from thunderstorms (Solari 2020). Their radial meteorological datasets (e.g., Radio-based data). From
outflows and ring vortices after touchdown produce strong wind Table 3iii, it can be concluded that most applications use ML
gusts very close to the ground and therefore lead to substantial as a classification or a regression model for prediction of tornado
structural damages (e.g., Yang et al., 2018). Downbursts are occurrence. Obviously, there is still room for more
typically simulated numerically using CFD (e.g., Mason et al., comprehensive applications of ML in terms of simulation of
2009; Aboshosha et al., 2015; Haines and Taylor 2018; Hao and the full track of a tornado (including its intensity and associated

Frontiers in Built Environment | www.frontiersin.org 11 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 2 | Summary of ML applications for tropical cyclones.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

i) Genesis Prediction of the SVR with radial - Nine predictors (e.g., El Number of TCs in Australian RMSE, MAE, R2 SVR outperformed
number of TCs in basis function Niño Southern the northwest of Government, Bureau the MLR model – The
the northwest of coupled with Oscillation) Australia of Meteorology prediction accuracy
Australia Richman sequential website was further improved
and Leslie (2012) minimal by coupling the SVR
optimization model with Quasi-
algorithm, MLR Biennial Oscillation
Prediction of TC SVM with - El Niño—Southern TC genesis Bureau of MAE SVM outperformed
genesis in the polynomial Oscillation indices, (number of TCs) in Meteorology’s LDA model. Overall
South Pacific kernel, LDA Multivariate ENSO the South Pacific National Climate prediction
Ocean and Index, El Niño Modoki Ocean and Center - Australia performance for both
Australian region Index, Dipole Mode Australian region models is low
Wijnands et al. Index and the Southern
(2014) Oscillation Index
Prediction of TCs DT (C4.5 - Sea surface TCs genesis in the Navy Operational Prediction Satisfactory results
genesis in the algorithm) temperature, rainfall western North Global Atmospheric accuracy = were obtained based
western North intensity, divergence Pacific region Prediction System (correctly classified on the C4.5 algorithm
Pacific region averaged between and the Tropical samples/number
Zhang et al. (2015) 1000- and 500-hPa Rainfall Measuring of samples in the
levels, maximum 800- Mission (TRMM) whole dataset)
hPa relative vorticity Microwave Imager
and the 300-hPa air (TMI) from 2004 to
temperature anomaly 2013
Variable selection LR and Peter- - Selected variables: TCs genesis in IBTrACS, tropical p-value and area Top ranked variables
and prediction of Clark algorithm relative vorticity region between cloud cluster (TCC) under the receiver include the relative
TC genesis (925 hPa), potential 30°N and 30°S and ERA-Interim operating vorticity (925 hPa),
Wijnands et al. vorticity (600 hPa) and (1979–2014) characteristic potential vorticity
(2016) vertical wind shear (ROC) curve (600 hPa) and vertical
(200–700 hPa) wind shear
(200–700 hPa)
Development a TC DT (C5.0 - 8 WindSat-derived TC genesis WindSat satellite data Prediction Good simulation
genesis detection algorithm) indices tested and 2 (wind and rainfall) accuracy = results were obtained
model over the were selected as the were used to extract (correctly classified
western North most dominant the training/testing samples/number
Pacific Park et al. predictors: circulation data from 2005 to of samples in the
(2016) symmetry and intensity 2009 over the whole dataset)
western North Pacific
Prediction of the SVR (with 2 - SST and El Niño 3.4 Number of Hurricane database in RMSE The SVR model gave
number of seasonal kernels: were the best seasonal TCs in the the North Atlantic enhanced prediction
TCs in the North polynomial and attributes North Atlantic basin and Hadley compared to an
Atlantic region radial basis region Centre Sea Ice and operational statistical
Richman et al. function) Sea Surface model that was
(2017) Temperature dataset developed by
Colorado State
University. The
polynomial kernel
gave a slightly
improved simulation
results compared to
the RBF kernel
Prediction of TC LR, NB, DT, - Several Genesis prediction Mesoscale convective F1-score accuracy AdaBoost algorithm
formation from KNN, ANN, QDA, thermodynamic and at different lead system (MCS) was the best
mesoscale SVM (with a radial dynamic predictors times (e.g., 6 h) dataset, IBTrACS, classifier. Both the
convective system basis function were employed in this and ERA-Interim genesis potential
Zhang et al. (2019) kernel), study (e.g., genesis (1985–2008) index and the low-
AdaBoostRF. potential index, 850- level vorticity were the
hPa vorticity and most dominant
vertical wind shear) predictors for the
tropical cyclone
genesis
Detection of TC DT, RF, SVM - 8 dynamic and Genesis detection WindSat satellite F1-score accuracy Best performance
genesis over the (with three hydrological predictors for a lead time up measurements from and PSS score from the SVM model
western North different kernels: (e.g., rain rate, circular to 30 h 2005 to 2009 over the with a radial basis
Pacific Kim M et al. linear, variance of wind western North Pacific function kernel
(2019) polynomial, and speed) basin
radial basis
functions), LDA

ii) Prediction of ANN Pseudo invert 12 h of past track 24 h of cyclone Joint Typhoon MAE Acceptable accuracy
Translation cyclone track over learning observations (in terms track over the Warning Center
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 12 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 2 | (Continued) Summary of ML applications for tropical cyclones.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

the Indian Ocean Ali of latitude and Indian Ocean at 6 h (JTWC) from 1971 to
et al. (2007) longitude) intervals 2002
Prediction of TCs ANN Levenberg 2 previous 6-h 24 h of cyclone 20 years of historical Correlation Good simulation
track of over the Marquardt positions and the track over the tack data from the coefficient results
western North current one (in terms of western North JTWC
Pacific basin Wang latitude and longitude) Pacific basin at 6
et al. (2011) hourly intervals
Trajectory RNN Genetic algorithm Past hurricane track hurricane track for National Oceanic and MAE Acceptable accuracy
Prediction of locations which are up to 12 h in Atmospheric
Atlantic Hurricanes selected by the RNN advance Administration
Moradi model (6-hourly (NOAA) from 1900 to
Kordmahalleh et al. hurricane center’s 2013
(2016) latitude and longitude)
Cyclone track MNN, RNN, Backpropagation Past hurricane 1-step of 6-h ahead JTWC between 1985 RMSE MNN-based model
prediction over the LSTM, GRU trajectories TC trajectory (in and 2013 in the South outperformed the
South Indian ocean -automatically selected terms of latitude Indian ocean three recurrent neural
Zhang et al. (2018) by the algorithm- and longitude) networks
Prediction of Grid-based RNN, Backpropagation Past hurricane Hurricane tracks NOAA database MSE, RMSE The grid-based
hurricane sparse RNN locations (6-hourly over the Atlantic algorithm
trajectories over the distributed) basin up to 120 h outperformed the
Atlantic basin sparse RNN
Alemany et al.
(2019)
Prediction of a GAN Backpropagation Satellite images Typhoon tracks in Korean Average absolute Acceptable accuracy
typhoon track in the the Korean Meteorological error
Korean Peninsula Peninsula at 6 h Administration and
Rüttgers et al. lead time the ERA-interim
(2019) databases with a total
of 76 typhoons that hit
the Korean peninsula
from 1993 till 2017
Prediction of the ConvLSTM AdaGrad Last 5 consecutive Spatial-temporal Community RMSE The error increased
spatial-temporal hurricane density- hurricane trajectory Atmospheric Model with the increasing
hurricane trajectory maps (up to 15-h) with a v5 from 1995 to 2015 leading time
Kim S et al. (2019) 3-h time steps
Tropical cyclone CNN Adam Atmospheric fields TC trajectory (in TCs data in both RMSE, MAE The proposed model
track forecasting (image-like data) terms of latitude hemispheres from outperformed the
Giffard-Roisin et al. corresponding to the and longitude) for NOAA, IBTrACS and statistical CLP5
(2020) current and past data up to 24-h leading ERA-Interim since model
(with a 6-h time step) time 1979 (more than
including the latitude, 3,000 storms with 6-h
longitude and time steps)
geospatial height fields
at three pressure
levels: 700, 500, and
225 hPa (e.g., wind
speed components)

iii) Intensity Prediction of ANN, MLR Backpropagation 11 predictors (e.g., Typhoon intensity National Centers for Average error The ANN-based
typhoon intensity initial storm intensity, changes in the Environmental model outperformed
changes in the initial storm latitude, western North Prediction/National THE MLR model
western North vertical wind shear and Pacific basin from Center for
Pacific basin Baik 850-mb horizontal 12-h and up to 72- Atmospheric
and Paek (2000) moisture flux) h (1 output) Research (NCEP/
NCAR) reanalysis
from 1983 to 1996
Prediction of the ANN, RBF, Backpropagation 5 predictors: sea Cyclone intensity Indian Meteorological RMSE, MAE ANN model provided
cyclone intensity MLR, OLR surface temperature, over the Arabian Department from the best prediction
over the Arabian central pressure, Sea and Bay of 2005 to 2010 results
Sea and Bay of pressure drop, Bengal for
Bengal Chaudhuri maximum sustained approximately 72 h
et al. (2013) surface wind speed lead time (1 output)
and total ozone column
Prediction of the ANN, MLR, SVM Backpropagation Multispectral Imagery Cyclone intensity Tropical cyclone Kappa coefficient The three models
cyclone intensity level (class labels) Nalgae data from 04/ and overall provided comparable
levels Chen et al. 08/2017 till 06/08/ accuracy (%) classification results
(2018) 2017 retrieved from
No. 4 meteorological
satellite (FY-4) of
China
Prediction of time RNN Backpropagation 3 previous time steps Time series Western North Pacific Average forecast Performance
series of typhoon along with the current prediction of typhoon database error comparable to the
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 13 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 2 | (Continued) Summary of ML applications for tropical cyclones.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

intensity Pan et al. time of typhoon intensity up to 48 h from the Chinese Japanese
(2019) location and intensity with a 6 h time step Meteorological Meteorological
Administration and Agency-Global
the Shanghai Spectral model
Typhoon Institute
from 1949 till 2016
Cyclone intensify Hybrid CNN- Gradient descent 3-D atmospheric Intensity (24-h lead International Best MAE Good simulation
forecasting over the LSTM model (2D- and Adam variables (wind time) with a 6 h Track Archive for results comparable to
Western Pacific, CNN, 3D-CNN components, time step Climate Stewardship other operational
Eastern Pacific and and LSTM) temperature, relative (IBTrACS) and ERA- forecast models (e.g.,
North Atlantic humidity and Interim reanalysis Hurricane Weather
basins Chen et al. geopotential height) and Research
(2019) and 2-D sea surface Forecasting Model)
variables (sea surface
temperature)
TC intensity CNN Adam Satellite images of TCs TC intensity in near Satellite outputs from RMSE Good simulation
prediction over the in real time real time 2003 till 2016 from the results
Pacific Northwest Meteorological
and Atlantic Ocean Satellite Research
Wei Tian et al. Cooperation Institute
(2020) and JWTC
Hurricane intensity CNN Adam Satellite images of TCs TC intensity in near U.S. Naval research RMSE Acceptable
prediction Maskey in real time real time laboratory and the simulation results
et al. (2020) NOAA Geostationary
Operational
Environmental
Satellite from 2000
through 2019

iv) Wind Estimation of ANN (a total of 3 Levenberg- ANN 1: SeaWinds ANN 1: wind speed QuikSCAT mission MAE Good simulation
field hazard surface wind field were used) Marquardt scatterometer from 0 to 20 m/s and H*Wind between results for the surface
based on satellite measurements 1999 and 2009 for all wind speed were
data Stiles et al. ANN 2: Outputs of ANN 2: corrected basins (globally) obtained
(2014) ANN 1 wind speed over
20 m/s (retrieved
from H*Wind)
ANN 3: 6 predictors ANN 3: final
(outputs of the first two optimized wind
ANNs, QuikSCAT speed with a
radiometer rain rate 12.5 km resolution
and rain impact
quantity, maximum
likelihood estimation
direction interval wind
speed and cross-track
distance)
Forecasting surface SVM with 4 - 13 features are Surface wind Central Weather RMSE -Pearson VII SVR
wind speeds during kernels: linear, considered (e.g., speed (1-h Bureau of Taiwan model is the most
tropical cyclones polynomial, radial central pressure, average) for up to from 2000 till 2012 accurate technique
Wei (2015) basis function latitude, longitude, sea 6 h over two (84 typhoon events) among all other
and Pearson VII surface pressure) offshore islands tested kernel-based
based on stepwise near Taiwan SVM models
regression method -Resolution not
discussed
Estimation of TCs LSSVM, RBFNN, - TC age, center latitude Critical wind radii of National Satellite MAE LSSVM
inner-core surface linear regression and maximum surface 34- and 50-kt Meteorological Centre outperformed all
wind structure wind speed winds in real time of China and the other models
based on infrared Shanghai Typhoon
satellite images Institute from 2005 to
Zhang et al. (2017) 2008
Simulation of TC KEDL L-BFGS-B Storm parameters Hurricane H*Wind snapshots RMSE Good simulation
boundary-layer (e.g., spatial boundary-layer results were obtained
winds Snaiki and coordinates, storm size winds
Wu (2019) and intensity)
Surface wind DNN Back- 16 inputs for Taipei and Hourly surface wind Central Weather RMSE Good consistency
simulation in near propagation 14 for Keelung field with 1-degree Bureau of Taiwan and between the
real time Wei (2019) algorithm corresponding to the by 1-degree Weather Research simulated and WRF
typhoon resolution in 2 and Forecasting (47 results
characteristics (e.g., locations in Taiwan typhoons from 2000
central pressure) and (Taipei and till 2017)
surface meteorological Keelung)
data (e.g., relative
humidity)

Frontiers in Built Environment | www.frontiersin.org 14 March 2022 | Volume 8 | Article 811460

Frontiers in Built Environment | www.frontiersin.org

Wu and Snaiki
TABLE 3 | Summary of ML applications for non-synoptic winds.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

i) Prediction of severe ANN Backpropagation Lifted index and surface Value between 0 and 1 Centralized Storm critical success Acceptable results
Thunderstorm thunderstorms McCann moisture convergence representing the likelihood Information System of the index
(1992) of the thunderstorm National Severe Storms
occurrence for a 3-7 h lead Forecast Center (NSSFC)
time from April to August 1990
over the eastern two-thirds
of the United States
Prediction of the surface ANFIS, ANN, Gradient descent and Lift index, Convective Surface peak gust wind Radiosonde and RMSE, MAE ANFIS model outperformed the
peak gust wind speed RBFNN, MLR the least squares Inhibition Energy, speed in Kolkata, India rawinsonde from the other machine learning models
during thunderstorm estimate Convective Available with a lead time up to 12 h Department of
events Chaudhuri and Potential Energy and bulk Atmospheric Sciences,
Middey (2011) Richardson number University of Wyoming for
the location of Kolkata,
India from 1997 till 2009
Prediction of ANN Levenberg Marquardt, Wind speed, humidity and Hourly temperature during Indian meteorological RMSE, MAE, Best results with
thunderstorms Momentum, Conjugate mean sea level pressure thunderstorm, proxy for department from 2007 to correlation Levenberg–Marquardt learning
occurrences Litta et al. Gradient, Delta Bar thunderstorm occurrence, 2009 (hourly data) coefficient algorithm
(2012) Delta, Quick Propagation over the northeastern
and Step region of India
Prediction of severe ANN, KNN Gradient descent 2 predictors at 5 Likelihood of occurrence of Indian Meteorological Correlation KNN model was the best classifier
thunderstorms geopotential heights: dry severe thunderstorms with Department from 1969 to coefficient
occurrences Chakrabarty adiabatic lapse rate and a lead time between 10 2008
et al. (2013) moisture difference (a total and 14 h over the
of 10 inputs) northeastern region of
India
15

Prediction of thunderstorm ANN, Bayes Network, Artificial Bee Colony 31 thermodynamic and Thunderstorm occurrence METeorological Aerodrome Accuracy, AUC, and ANN model optimized with ABC
occurrence Yasen et al. C4.5 decision (ABC), gradient descent dynamic predictors Reports and Surface F-measure algorithm outperformed the other
(2017) Tree, KNN Synoptic observation from classifiers in detecting
December 2015 to thunderstorms
November 2016 at lake
Charles airport in Louisiana
Prediction of thunderstorm ANN Scaled conjugate 15 inputs (e.g., most Thunderstorm occurrence ERA-Interim database from Heidke skill score Acceptable results
occurrence Ukkonen et al. gradient unstable lifted index and in the next 6-h period 2002 to 2015 over Finland
(2017) relative humidity near
700 hPa) identified based
on skill scores
Forecasting SD-AE Stochastic gradient 38 features (e.g., total Thunderstorm occurrence North American Mesoscale Peirce skill score The SD-AE model outperformed
thunderstorms descent predictable water and through cloud-to-ground Forecast System and the an ANN model developed by
occurrence Kamangir et al. convective precipitation) lightning parameter for a National Lightning Data Collins and Tissot (2015, 2016) for
(2020) maximum lead time of 15 h Network from the the same region and with similar
and within 400 km2 of a 2004–2012 lead time
March 2022 | Volume 8 | Article 811460

selected site in South

ML Applications to Wind Engineering

Texas
Forecasting the CNN Backpropagation 91-min time series of wind Event type (thunderstorm Automated Surface F1 score and Reliable classifier for
occurrence of speed and direction or non-thunderstorm Observing System (ASOS) average success thunderstorms occurrences
thunderstorms events event) (1-min averaged data) from rate
(Chen and Lombardo 2000 to 2018 with a total of
2020) 76,480 time series of
91 min of wind speed and
direction
(Continued on following page)
Frontiers in Built Environment | www.frontiersin.org

Wu and Snaiki
TABLE 3 | (Continued) Summary of ML applications for non-synoptic winds.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

ii) Downburst Prediction of damaging ANN Conjugate gradient 23 radar-derived predictors Probability of damaging National Severe Storms Fraction Correct Acceptable results
wind from tornadic and characterizing the wind (with a damaging Laboratory (NSSL) and Heidke’s Skill
straight-line events circulations (e.g., depth of wind excessing 25 m/s) Mesocyclone Detection Score
(including downbursts) circulation, maximum with a lead time of 20-min Algorithm (MDA)
(Marzban and Stumpf rotational velocity and low
1998) altitude shear)
Classification of damaging LDA - 26 reflectivity and radial Severity of downburst WSR-88D radars (in several median Heidke skill Acceptable results for the
downburst winds (Smith velocity-based attributes winds (severe or non- locations within the U.S.) prediction of severe downburst
et al. (2004) (e.g., cell volume, max severe events) with a from the National Climatic events
reflectivity and height of the maximum lead time of Data Center’s Storm. It
max reflectivity) 15 min contains 91 events that
produced severe
downbursts and 1247
events that did not produce
severe downbursts
Prediction of the LR, LR with an elastic Gradient descent 431 predictors. They can Probability of occurrence Near-surface wind AUC - The simulation results indicated
probability of occurrence network, ANN, RF, GBTE be divided into 4 main of damaging winds with a observations (from the that storm motion and sounding
of damaging straight-line categories, namely radar lead time up to 90 min Meteorological Assimilation indices are the dominant
winds (including statistics, storm shape Data Ingest System, the predictors
downbursts) from storm parameters, storm motion Oklahoma Mesonet, and - Both random forest and
cells Lagerquist et al. and sounding indices the National Weather gradient-boosted tree ensembles
(2017) Service), radar scans (from gave the best simulation results
the Multiyear Reanalysis of
Remotely Sensed Storms)
and soundings (from the
16

Rapid Update Cycle and

the North American
Regional Reanalysis) [from
2001 to 2011]
Downburst wind speed LSSVM (coupled with - Time series of downburst Time series of downburst Time series of downburst MAE, RMSE, 2- The combined Morlet wavelet and
forecasting Li and Li (2018) variational mode wind (up to 1600 s) wind from 1600 s through wind from two norms relative error radial basis kernel functions (RBF)
decomposition and 1800 s measurements data and Pearson gave the best simulation results
particle swarm) – with consisting of 450 sample correlation
several kernels (linear, points with a sampling coefficient
polynomial, Mexican Hat, frequency of 0.25 Hz for a
radial basis function, and total of 1800 s (the data
Morlet wavelet) source was not mentioned)
Identification of the RF - 8 dual-polarization radar Downburst related events Weather observation Mean Decrease Although the model provided
downburst occurrence signatures (e.g., maximum or null events around the towers around the Cape Accuracy (MDA) and good simulation results, strong
Medina et al. (2019) vertically integrated liquid Cape Canaveral Air Force Canaveral Air Force Station Mean Decrease events were better classified
and temperature colder Station and Kennedy and Kennedy Space Center Gini (MDG) compared to weaker ones
March 2022 | Volume 8 | Article 811460

than 0°C) Space Center from 2015 to 2016

ML Applications to Wind Engineering

iii) Tornado Prediction of the tornado’s ANN Conjugate Gradient 23 input variables (e.g., occurrence/non- National Severe Storms Critical Success The ANN model outperformed
occurrences Marzban and maximum shear, low- and occurrence of tornados for Laboratory’s (NSSL) Index other statistical models such as
Stumpf (1996); Marzban mid-altitude convergence) a given mesoscale Mesocyclone Detection the discriminant analysis bur still
et al. (1997); Marzban circulation in the next Algorithm (MDA) with a total the performance is low
(2000) 20 min of 3258 circulation events
Detection of the tornado’s ANN Resilient 13 features (e.g., rotational Tornado occurrence from National Severe Storms Heidke Skill Score Simulation results acceptable
occurrences Lakshmanan backpropagation velocity) given circulations in the Laboratory based on the
et al. (2005) next 20 min Mesocyclone Detection
Algorithm (MDA) and the
near-storm environment
(NSE) with 110 storm days
(Continued on following page)
Frontiers in Built Environment | www.frontiersin.org

Wu and Snaiki
TABLE 3 | (Continued) Summary of ML applications for non-synoptic winds.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

Prediction of the tornado SVM, LDA, BNN Backpropagation 34 input features (e.g., Tornado occurrence in the Weather Surveillance Radar Heidke Skill Score - linear programming support
occurrence Santosa meso core depth and meso next 20 min 1998 Doppler vector machine was used for
(2007) low-level shear) feature selection
- BNN model gave the best
performance in detecting
tornados from given circulations
Prediction of the tornado SVM with 3 kernels (linear, Backpropagation 53 input features (e.g., Tornado occurrence in the Radar measurements from Heidke Skill Score -The best classifier was the SVM
occurrence Adrianto et al. polynomial and RBF), azimuthal shear low level next 30 min the National Climatic Data model with the RBF kernel
(2009) ANN, LDA average, gradient direction center with a total of 33 -SVM model outperformed the
maximum and reflectivity storm days sampled at other algorithms
aloft average) 30 min
Prediction of the tornado SVM (radial basis function - 22 attributes (e.g., wind Tornado occurrence from MDA and NSE databases Heidke Skill Score -Feature selection was performed
occurrence Trafalis et al. kernel), LR, RF, rotation shear and humidity) mesocyclones events (no with 111 storm days using the SVM-Recursive Feature
(2014) forest leading time indicated) Elimination algorithm with a radial
basis function kernel
- SVM with threshold adjustment
outperformed all other classifiers
17

Prediction of the CNN Adam Storm-centered radar Probability of occurrence Multiyear Reanalysis of Area under the Excellent simulation results
probability of occurrence image and a proximity of a tornado in the next- Remotely Sensed Storms receiver-operating-
of a tornado Lagerquist sounding hour (MRRSS) in the [period from characteristic curve
et al. (2018), (2020) 2000 to 2011] and Gridded (AUC) score
NEXRAD WSR-88D Radar
(GNWR) [period from 2011
to 2018]
Predicting property ANN (2) AdaGrad Storm, land cover, ANN1: occurrence or non- NOAA’s tornado database, AUC, MSE, R2 -Only the initial tornado
damage from tornadoes socioeconomic and occurrence of damage due the National Land Cover coordinates are accounted for
Diaz and Joseph (2019) demographic features to a tornado event database and the American rather than the tornado path
ANN2: level of damage Community Survey -Acceptable results
when it occurs
Prediction of the RF, CNN Stochastic gradient 222 input features at Tornadic and non-tornadic Rapid Update Cycle Overall accuracy -The input feature selection was
occurrence of tornadic descent various geopotential events sounding data from the score (in %) carried out using RF which
events Coffer et al. (2020) heights were initially National Climatic Data indicated that the pressure terms
selected (e.g., temperature, Center from 2003 to 2017 are not as important as the other
March 2022 | Volume 8 | Article 811460

pressure) –Exact final environmental parameters (e.g.,

ML Applications to Wind Engineering

parameters not mentioned- v-wind component)
- RF outperformed CNN
Wu and Snaiki ML Applications to Wind Engineering

TABLE 4 | Summary of ML applications for terrain and topography.

Application ML model Training scheme Input data Output data Data source Performance Remarks
metric

Modeling the effects of ANN Cascade correlation 6 inputs including Fractional speed- CFD simulations R2 coefficient -Comparison
topography on the simple geometric up ratio corresponding to with experimental
wind profile properties different data from wind
Bitsuamlak (2004); (i.e., “Windward topographic tunnel
Bitsuamlak et al. slope of the hill”, configurations: -Good
(2002), (2006), (2007) “Distance between single and multiple performance
hills”, “Height from hills and
the crest of the hill” escarpments
and “Longitudinal
location”),
roughness element
and hill count
Wind field simulation ANN Backpropagation Terrain roughness, Wind velocity time The time series of MSE -The conditional
considering terrain combined mean wind profile series (3 min of wind speed were simulation
effects with and spectral density time series with a generated using technique
Martínez-Vázquez and conditional time step of 0.1 s) the procedure of significantly
Rodríguez-Cuevas simulation at different points Simiu and Scanlan decreased the
(2007) technique (1978) at two number of
heights (i.e., 10 required layers in
and 200 m) with 11 the ANN
local velocities -Good simulation
(from 0.5 to 100 m/ results
s) and surface
roughness
between 0.001
and 0.050 m
Estimation of the effect ANN Bayesian regularization 4 inputs: 10-min Annual average Meteo-Galicia RMS -Wind direction is
of wind direction on mean wind speed wind speed at a during 2003 at the important to be
wind speed prediction from 3 stations given site with Galicia region in the considered to
in complex terrain nearby and wind complex terrain northwest Spain improve the
Lopez et al. (2008) direction from configuration corresponding to 5 simulation results
another nearby stations and for a site with
station representing complex terrain
various terrain
conditions (e.g.,
inland and offshore
conditions) and
elevations
Prediction of typhoon ANN Backpropagation Upstream wind Wind speed and Reynolds- MAE Good simulation
wind speed and profile speed and direction direction at height averaged Navier- results for both
over complex terrain at height z z on a bridge site Stokes simulations wind speed and
Huang and Xu (2013) which provides the direction
wind profiles at the
bridge site given an
inlet upstream
wind field (which
does not account
for topographic
effects)
Prediction of the wind DNN Proximal adagrad 3 cartesian Mean wind speed 4 CFD simulations MAE Acceptable
flow over complex coordinates (x,y,z) over a given site of the wind field in a simulation results
topographies Mayo of the selected point with complex given coastal dune
et al. (2018) and the incoming topography system with
uniform mean wind complex terrain
speed
Selection of the ANN (2) cascade correlation ANN1: height from ANN1: mean RWDI USA LLC No error scores - Visual
experimental floor, the bottom- longitudinal wind wind tunnel in were provided inspection of the
hardware within a wind spire width, the velocity and Miramar, Florida predicted wind
tunnel Abdi et al. surface roughness turbulence profile and
(2009) and the top spire intensity turbulence
width intensity of the
first neural
network
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 18 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 4 | (Continued) Summary of ML applications for terrain and topography.

Application ML model Training scheme Input data Output data Data source Performance Remarks
metric

indicated good
simulation results
ANN2: target mean ANN2: difference - Results from
longitudinal wind between top and ANN2 were not
speed, target bottom spire satisfactory
turbulence intensity width and the
and height from floor surface roughness
Prediction of wind ANN (2) Lavenberg–Marquardt ANN1: number of ANN1: mean wind Boundary-layer No error scores - Visual
properties in urban roughness speed, turbulence wind tunnel tests of were provided inspection of the
environments based elements, number of intensity and the National Wind predicted results
on wind tunnel tests barriers, height from length scale factor Tunnel Facility in indicated
Varshney and Poddar floor and slot width Kanpur, India (18 satisfactory
(2012) ANN2: number of ANN2: configurations) simulations
roughness instantaneous
elements, number of velocity
barriers and slot
width
Designing laboratory ANN (2) RPROP Riedmiller and ANN1: basis barrier ANN1: mean wind Boundary-layer R2 -ANN1: except
wind simulations Braun (1993) height, barrier speed, turbulent wind tunnel at the the turbulent
Križan et al. (2015) castellation height, intensities (in the Technische length scale in the
surface roughness three directions), Universität x-direction (not
elements’ spacing length scales (in München with a that accurate) all
density, surface the three total of 23 other results were
roughness directions) and configurations of good
elements’ height turbulent hardware setups
and height of Reynolds stress
measurement
points
ANN2: basis barrier ANN2: power - ANN2: good
height, surface spectral densities simulation results
roughness of the velocity were obtained
elements’ spacing fluctuations in the
density, surface three directions
roughness
elements’ height,
frequency and
height of
measurement
points

wind field). Just like ML applications to downbursts, a high single hill) through correction factors. To examine the effects of
number of input variables (predictors) were utilized for the complex terrain condition on wind fields, wind tunnel tests are
reviewed ML models. The identification of the most usually employed with a very small geometric scale (e.g., 1:500).
appropriate set of predictors is still very challenging, and a Alternatively, numerical schemes such as the mass-conservation
trail-and error approach was typically employed. In addition, or momentum-conservation model can be used to capture the
it is not easy to conduct a systematic comparison among reviewed terrain effects on oncoming wind fields. Although the
ML models since the used performance metrics differ topographic effects can be well simulated based on
substantially from one application to another. momentum-conservation models (e.g., using Reynolds-
averaged Navier-Stokes equations), the needed computational
3.2 Terrain and Topography time makes it impractical for use as a real-time decision support
Wind characteristics including mean wind speeds and turbulent tool. The mass-conservation model computes wind fields over
fluctuations are much affected by the surrounding terrain and complex terrain in seconds to a few minutes (Forthofer et al.,
topography. As a consequence, careful consideration of local 2014a; 2014b), but the accuracy of simulation may be poor
terrain roughness and topographic features as well as because nonlinear momentum effects are not considered
surrounding obstacles is vital to the accurate determination of (Jackson and Hunt 1975). Considering the complex terrain-
wind pressures on structures and pedestrian level winds. Wind wind data from high-fidelity CFD simulations, wind tunnel
codes and standards consider the terrain effects corresponding to tests and field measurements are increasingly available, ML
limited (and simplified) terrain geometries (e.g., escarpment and tools can be utilized (as computationally efficient reduced-

Frontiers in Built Environment | www.frontiersin.org 19 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

order models that possess high simulation accuracy of complex and accuracy. It is noted that there are numerous ML applications
nonlinear systems) to provide rapid estimation of wind flows over to aerodynamics and aeroelasticity of both bluff bodies (e.g.,
various terrain conditions. However, ML development for terrain circular cylinder) and streamlined bodies (e.g., airfoil) in fluid
and topographic considerations is still at an early stage with a mechanics community (e.g., Kutz 2017; Brunton et al., 2020),
limited number of studies reported in the literature. Table 4 however, they are not discussed here. The review in this section
presents the reviewed applications of ML for terrain and only covers wind-sensitive structures in civil engineering. The ML
topography, where the ML model, training scheme, input data, applications for bridge aerodynamics and aeroelasticity are first
output data, data source and performance metric are summarized reviewed in Table 5i and then followed by buildings and other
for each application. The training/testing data were essentially structures in Table 5ii, where the ML model, training scheme,
retrieved from either CFD simulations or wind tunnel tests. From input data, output data, data source and performance metric are
Table 4, it can be concluded that most applications used ML as a summarized for each application. The training/testing data were
regression model for prediction of wind fields over various terrain essentially retrieved from either CFD simulations or wind tunnel
conditions and topographic configurations. There are a few tests. From Table 5, it can be concluded that most applications
studies that applied ML techniques to assist in efficient search used ML as a regression model for prediction of steady-state force
for a correct layout of passive flow altering devices (e.g., spires and coefficients, flutter derivatives and vortex-induced vibrations
roughness elements) in the boundary-layer wind tunnel. It is (VIV) of various bridges and for modeling of wind pressure
noted that the current ML applications to consider topographic coefficients of various buildings (as well as estimation of the
effects on wind fields are usually limited to terrain configurations interference factors for adjacent buildings). The different
that can be characterized by several parameters, hence, the aerodynamic representations in bridges (mainly using global
employed ML models and training schemes are simple and quantities such as force coefficients) and buildings (mainly
standard (e.g., ANN with backpropagation). However, several using local quantities such as pressure coefficients) are
advanced ML models such as autoencoder (e.g., Fukami et al., partially due to available data types from wind tunnel tests.
2019) and GAN (Kim and Lee 2020) have been utilized to assist in Although satisfactory ML simulation results have been
the generation of turbulent inflow (as a realistic inlet boundary obtained (in terms of interpolations), most reviewed
condition of CFD simulations). applications do not necessarily have good performance in
terms of extrapolations outside the training datasets. It is
3.3 Aerodynamics and Aeroelasticity noted that the currently available ML models of aerodynamics
The bluff-body aerodynamics and aeroelasticity play a critical and aeroelasticity are developed for the main purpose of being
role in the safe and cost-effective design of wind-sensitive used as preliminary design tools to avoid the high-cost wind
structures, and their considerations rely heavily on boundary- tunnel tests in the early design stage. There is a lack of systematic
layer wind tunnels. In addition to the Reynolds number effects comparison among various ML models, hence, their selection for
(due to very small model scales), wind tunnel tests are very time specific applications is rather rudimentary.
consuming and labor intensive. To this end, CFD techniques have
been rapidly developed for simulations of structural
aerodynamics (gust-induced effects) and aeroelasticity 3.4 Structural Dynamics and Damage
(motion-induced effects). The purpose is to make CFD Assessment
simulations serve as a complementary or even alternative Due to the computational complexity of numerical techniques
approach to wind tunnel tests. Despite significant advances of (e.g., finite element method) for solving wind-induced nonlinear
hardware and algorithms, the reliable CFD simulations of wind- structural response, reduced-order models (e.g., ANN) have been
structure interactions are still computationally very expensive due developed to alleviate the computational cost of the high-fidelity
to three-dimensional nature of wakes and intensive flow models. The ML models have been used for structural dynamics
separations from structures. Hence, a number of reduced- and damage assessment for several decades mainly in the field of
order models have been developed to efficiently model earthquake engineering (e.g., Wu et al., 1992; Masri et al., 1993;
structural aerodynamics and aeroelasticity (Wu and Kareem Jiang and Adeli 2005; Pei et al., 2005; Gholizadeh et al., 2009;
2013). Unfortunately, these reduced-order models do not Facchini et al., 2014; Derkevorkian et al., 2015; Liang 2019; Wu
always have a satisfactory representation of the full nonlinear and Jahanshahi 2019; Yu et al., 2020). However, similar
equations that govern the wind-structure interactions. applications have not emerged in wind engineering
Specifically, modern bridge decks and super tall buildings with community until recently due essentially to the linear
unusually geometries all exhibit nonlinear unsteady consideration of the wind-induced structural response [ASCE
aerodynamics and aeroelasticity that limit the applicability of 7-16 (ASCE, 2017)]. Recent advances of performance-based wind
the state-of-the-art reduced-order modeling methodologies. On design methodology have placed increasing importance on
the other hand, the Kolmogorov Neural Network existence effective simulations of nonlinear, inelastic structural dynamics
theorem offers mathematical foundation for applying response under strong winds. The numerical estimation of wind-
multilayer neural networks to approximate arbitrary nonlinear induced nonlinear structural response using a high-fidelity finite
systems with any precision (Huang and Lippmann 1988; Hornik, element model is computationally very expensive due to its small
1991). With high-fidelity data and advanced algorithms, ML time-step size and long simulation duration. Accordingly, several
models can simultaneously achieve great simulation efficiency ML applications to wind-induced structural dynamics have been

Frontiers in Built Environment | www.frontiersin.org 20 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 5 | Summary of ML applications for aerodynamics and aeroelasticity.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

i) Bridges Estimation of ANN Resilient 100 inputs:90 inputs Flutter derivatives (6) Wind tunnel test MSE Acceptable
aeroelastic parameters backpropagation representing the of a rectangular (total of 17 performance
of bridge decks Jung Riedmiller and Braun section geometry and section experiments)
et al. (2004) (1993) 10 inputs for the
nondimensional
velocity
Prediction of flutter ANN (total Gradient descent Width-to-depth ratio 8 flutter derivatives Experimental data No error metrics From the graphical
derivatives of a of 8) and a set of reduced (each given by 1 from wind tunnel results, the
rectangular section frequency ANN separately) of tests simulation results
model Chen et al. rectangular section were in good
(2008) model agreement with the
experimental ones
Prediction of flutter SVM (RBF - Non-dimensional 8 flutter derivatives Wind tunnel tests MSE Good simulation
derivatives of a cable kernel) velocity and width to of a cable stayed were retrieved form results
stayed bridge Lute depth ratio of bridge bridge Matsumoto et al.
et al. (2009) deck (1996)
Estimation of flutter ANN (total Backpropagation Width-to-depth ratio, 8 flutter derivatives CFD simulations and No error metrics Good performance
derivatives of a of 8) reduced frequency (each given by 1 forced-vibration test
rectangular section and reduced ANN separately) of in a wind tunnel
Chung et al. (2012) velocities rectangular section
model
Modeling vortex- DT, SVR - DT: incoming wind DT: VIV modes (a Field measurements RMSE, accuracy Good simulation
induced vibration of a (with speed and direction total of 6) of a full-scale (%), squared results
long-span suspension Gaussian at three locations on suspension bridge correlation
bridge Li et al. (2018) radial basis the bridge deck over a period of 6- coefficient
kernel) SVR: same inputs as SVR: VIV amplitudes years (2010–2015)
DT model at the located in the
current step along eastern ocean of
with the response of China
the previous step
Prediction of nonlinear LSTM Back-pass algorithm Bridge deck motions Motion-induced CFD simulations No error metrics Excellent
unsteady bridge aerodynamic forces (total of agreement (through
aerodynamics Li et al. 14,880 input-output visual inspection)
(2020) data corresponding between the LSTM
to a 2-D bridge deck model and CFD
cross-section) was obtained
Prediction of ANN Levenberg-Marquardt 18 inputs Normalized lift force 2 dimensional CFD MSE Good simulation
aeroelastic response of corresponding to the and torsional simulations for the results were
bridge decks Abbas response for heave moment coefficients two bridge cross- obtained
et al. (2020) and pitch (in terms of at current time step sections
displacement,
velocity and
acceleration) at
previous time steps
with three lag terms
Prediction of the flutter ANN Levenberg-Marquardt 1st ANN category: Critical flutter Wind tunnel R2 While the
velocity of suspension (different deck chord, deck velocity of experiments along performance of the
bridges Rizzo and topologies) weight or the ratio suspension bridge with finite element- ANN models varied
Caracoglia (2020) between the 1st with closed box based simulation according to the
torsional and the 1st deck sections corresponding to topology, their
vertical circular various geometrical performance was
frequencies of the and mechanical good
bridge, structural parameters of the
damping, air density bridge deck cross-
and the flutter section
derivatives
2nd ANN category:
deck chord, the ratio
between the first
torsional and the first
vertical circular
frequencies of the
bridge, and the flutter
derivatives
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 21 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 5 | (Continued) Summary of ML applications for aerodynamics and aeroelasticity.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

ii) Building Prediction of wind load ANN Gradient descent Hemispherical Steady-state wind Wind tunnel tests R2 Acceptable results
& other distribution for air- membrane: internal pressure coefficient
structures supported structures pressure ratio and for air-supported
Turkkan and two spatial structures (e.g.,
Srivastava (1995) orientations cylindrical and
hemispherical
membranes)
Cylindrical
membrane: Similar
inputs as the first
case + membrane
aspect ratio
Modelling wind- ANN Generalized delta rule Spacing between two Mean and dynamic Wind-tunnel tests No error metrics -
induced interference adjacent buildings in along- and across- from two references
effects on high-rise the along- and wind directions Saunders and
buildings Khanduri across-wind interference factors Melbourne (1980);
et al. (1997) directions Taniike and Inaoka
(1988)
Modelling wind- ANN Backpropagation Building aspect ratio, Interference index Wind tunnel tests No error metrics -
induced interference normalized from several sources
effects on high-rise separation distance e.g., Zambrano and
buildings English and and power law index Peterka (1978);
Fricke (1999) Blessmann and
Riera (1985)
Interpolation of wind- ANN Levenberg–Marquardt 4 adjacent Wind pressure Wind tunnel tests of R2 Good simulation
induced pressure time experimental coefficient at the a 1:50 scale model results
series on a scaled pressure taps at the next time step
model Chen et al. next time step (t+1)
(2002) and values of the
pressure taps at
current & two
previous time steps in
the target tap
(central one)
Prediction of pressure ANN (2 Levenberg–Marquardt Roof height, wind ANN1: mean Wind tunnels MSE Good simulation
coefficients on roofs of models) direction and two pressure coefficients experimental data results
low buildings Chen normalized roof on a gable roof of
et al. (2003) coordinates (for the low-rise building
two models) ANN2: root-mean-
square pressure
coefficients on a
gable roof of low-rise
building
Prediction of building ANN, Backpropagation Ground roughness, Inference factor Experimental data MSE RBF outperformed
interference effects RBFNN (with relative orientation of from literature e.g., the ANN model
Zhang and Zhang Gaussian two buildings Bailey and Kwok
(2004) kernel) (1985)
Prediction of wind FNN (2 Backpropagation FNN1: wind direction FNN1: Mean Boundary-layer wind MSE - Acceptable results
loads on a large flat models) and the positions of pressure coefficients tunnels tests for the 1st FNN
roof Fu et al., (2006), the available pressure on a large flat roof model
(2007) taps
FNN2: wind direction FNN2: Power - No error metrics
and the frequency for spectral density (at were reported for
the few selected tap given input the 2nd FNN model
locations frequencies) at few
locations in the roof
corners and leading
edge
Wind load evaluation ANN (4 Quickprop algorithm ‘2 geometric Statistics of wind Experimental wind Predefined error - Acceptable results
for the design of roof models) Fahlman (1988) parameters of the pressure coefficient tunnel tests index (normalized
cladding of spherical dome’, ‘2 on the roof of a by the standard
domes Uematsu and coordinates spherical dome: deviation of the
Tsuruishi (2008) parameters x and y’, mean, standard target data)
‘turbulence intensity deviation, skewness
of the incoming wind and kurtosis
at the mean roof
height’
Estimation of the wind ANN, Backpropagation Aspect & side ratio Along-wind mean Wind tunnel tests RMSE RBFNN
force coefficients on a RBFNN, and ground coefficient of base outperformed all
rectangular building GRNN roughness shear of a other models
Wang et al. (2013), rectangular building
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 22 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 5 | (Continued) Summary of ML applications for aerodynamics and aeroelasticity.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

Wang and Cheng

(2015), (2017)
Wind load prediction of GRNN - rise-span ratio, Statistics of the Wind tunnel tests R2 While the mean
large-span dry coal depth-span ratio, pressure pressure coefficient
sheds Sun et al. (2017) wind angle, and local coefficients: mean, was predicted
coordinates RMS. Skewness, accurately, the
kurtosis of pressure kurtosis of the
coefficients, three pressure coefficient
auto-correlation was poorly
coefficients and predicted
coherence exponent
Prediction of wind ANN (2 Levenberg–Marquardt ANN1: coordinates ANN1: mean or Wind tunnel tests RMSE - No error metric
loads on high-rise models) (x, y, z) of the root-mean-square was reported for
building Huang et al. pressure taps pressure coefficients the 1st ANN model
(2017) on a high-rise
building
ANN2: coordinates ANN2: time series of - Good simulation
(x, y, z) of the wind-induced results for the 2nd
pressure taps and pressures on a high- ANN model based
time rise building on RMSE.
Prediction of wind ANN (3 Levenberg–Marquardt Wind direction and Mean pressure Tokyo Polytechnique MSE, R2 Good simulation
pressure coefficients models for building coefficients over few University accuracies
on building surfaces flat-, gable, characteristics (1 locations on the experimental
Bre et al. (2018) and hip- parameter for the flat- roofs and walls (5 database
roofed low- roofed building, and 2 outputs for the flat-
rise parameters for the roofed, 6 for the
buildings) gable roofed and hip- gable-roofed and 8
roofed buildings) for the hip-roofed)
Prediction of roof ANN Levenberg–Marquardt Turbulence intensity Mean, root-mean- Wind tunnel tests RMSE, MAE, R2 The accuracy of
pressures on a low-rise (at eave height) and 2 square, and peak simulation results
structure normalized roof pressure coefficients depends on the
Fernández-Cabán coordinates on the roof (at 152 pressure taps
et al. (2018) roof taps) of 3 scaled location
low-rise buildings (1:
50, 1:30, and 1:20)
Modeling for unsteady CNN-AE + Adam Temporal variation of Temporal variation Direct numerical MSE -The use of CNN-
flows around bluff LSTM the flow field around (next time step) of simulation (DNS): AE allows the
bodies of various different 2-D cross- the flow field around 100 different bluff- mapping between
shapes Hasegawa sections shapes: 2 different 2-D cross- bodies shapes (in 2- the high-
et al. (2019), (2020) velocity components sections shapes: 2 D space) with 500 dimensional space
and pressure velocity components instantaneous time- and a low-
and pressure series flow fields dimensional latent
each space which
facilitates the
training of the LSTM
model
- Excellent
performance
Prediction of wind DT, RF, - GAN: wind direction GAN: mean and Aerodynamic R2 The GANs-based
pressures on a tall XGBoost, and location of the fluctuating pressure database of Tokyo model
building under GAN interfering building coefficients over all Polytechnic outperformed the
interference effects Hu faces of the building University other three machine
et al. (2020) DT, RF, XGBoost: DT, RF, XGBoost: learning algorithms
wind direction, the mean and and provided
coordinates of the fluctuating pressure accurate mean and
pressure tap and the coefficient at one fluctuating pressure
location of the point on the building coefficients on the
interfering building surface principle building
Prediction of low-rise DNN Levenberg–Marquardt Prediction’s location Mean and peak wind Wind tunnel tests R2 Excellent
gable roof building (x, y, z) and the pressure coefficients performance
pressures Jianqiao incoming wind on the surface of a results
Tian et al. (2020) direction scale model
corresponding to a
low-rise, gable roof
building
Predicting wind RF, DT, Gradient descent Turbulence intensity, Mean and From published R2 GBRT
pressures around GBRT incoming wind, fluctuating wind papers e.g., Cheng outperformed all
circular cylinders Hu Reynolds number pressures around a et al. (2016), Gao other models
and Kwok (2020) and circumferential circular cylinder for et al. (2017)
angle of the cylinder high Reynolds
numbers

Frontiers in Built Environment | www.frontiersin.org 23 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 6 | Summary of ML applications for structural dynamics and damage assessment.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

i) Structural Modeling ANN Gradient descent 12 inputs: mean Vertical (torsional) Tongji-1 wind No error metrics - Cellular
dynamics hysteretic wind velocity in acceleration of tunnel at State Key automata-
nonlinear the current and the bridge deck Lab in Tongji based system
behavior of bridge next time steps, section in the next University was employed
aerodynamics fluctuating time step to optimize the
Wu and Kareem components in ANN
(2011) the longitudinal configuration
and vertical - The visual
direction in the inspection of the
current and next results indicated
time steps, and the good
the vertical and agreement
torsional between the
displacement simulated and
with their first measured
and second - ANN model
derivatives in the showed good
current time step promise in
simulating the
hysteretic
nonlinear
behavior of the
bridge deck
which interacts
with the
incoming
fluctuating wind
Analysis of tall ANN Backpropagation Building shape Shear force and Data generated No error metrics From visual
building for (height, breadth bending from numerical inspection, it
across wind and depth), the moments of tall examples can be
response terrain category buildings concluded that
Vyavahare et al. and incoming a good
(2012) wind speed agreement
between the
simulated and
numerical
results has been
obtained
Identification of ERBFN Genetic Wind speed and Column stress of Wind tunnel tests RMSE, Good simulation
the dynamic algorithm direction a tall building maximum error results were
properties high- subjected to wind between the obtained
rise buildings loads measured and
subjected to wind estimated values
Oh et al. (2017)
Identification of ANN Backpropagation Building Dynamic Dataset were RMSE Good simulation
the dynamic geometry response in the generated based results were
properties high- (height, breadth along-wind and on the Indian Wind obtained
rise buildings and depth), across-wind in Code (IWC) for
subjected to wind incoming wind terms of base various building
Nikose and velocity and shear and base configurations
Sonparote terrain category bending moment
(2019a); (2019b),
(2020)
Wind-induced CNN Backpropagation Top-level (top Maximum and Wind tunnel tests RMSE Good simulation
response floor of a tall minimum strains results were
estimation for tall building) wind of the building obtained
buildings Oh et al. induced columns
(2019) displacement in
both time and
frequency
domain and
measured wind
speed in the
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 24 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 6 | (Continued) Summary of ML applications for structural dynamics and damage assessment.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

frequency
domain
Wind-induced KE-LSTM AdaMax Wavelet Normalized Numerically for the MAE - The governing
nonlinear coefficients of structural case of SDOF and equation of
structural the normalized displacement at MDOF motion was
dynamic analysis wind excitation different nodes embedded
Wang and Wu (external wind within the loss
(2020) force) function
- Excellent
simulation
results
Prediction of AWN Backpropagation Wind load and Maximum Dataset generated RMSE - AWN
structural high- absolute numerically parameters
response of wind- performance acceleration of corresponding to were updated
excited tall control systems the structure a 39-stoery steel- sequentially
buildings Micheli (HPCS) frame system each time data
et al. (2020) characteristics building subjected arrives (online
to wind load and training)
equipped with - Good
several equipment simulation
(e.g., damping results
devices, sensors
and global
controller)
ii) Damage Constructing and Ensemble - 10 predictors Classification: The data contains Accuracy (%) - The results of
assessment validating models were identified Structures that the damage states and customized this study
geographically composed (e.g., number of were correctly or and corresponds error metric suggest that
refined HAZUS- of 50 floors, terrain not well predicted to approximately HAZUS-MH4
MH4 hurricane bagged DT roughness, wind by HAZUS-MH4 700,000 fragility curves
wind risk models speed and (in terms of residences in the for certain home
Subramanian direction) hurricane induced Harris County types, need to
et al. (2013) wind damage) in following hurricane be refined to
1-km square Ike (2008) improve the
blocks prediction
results
Probabilistic ANN Backpropagation 8 predictors: Mean damage Boundary-layer Accuracy (%) Good
damage wind speed, ratio of an asphalt wind tunnel tests performance
estimation for angle of attack, shingle roof from the University
asphalt shingle shingle of Western Ontario
roofing Huang resistance,
et al. (2015) building length,
building width,
building height,
roof slope and
surface
roughness
Estimatin of the SVR (with - Gross vehicle Daily equivalent Traffic data from a RMSE, MAE, Good simulation
fatigue damage of Gaussian weight; 10-min fatigue damage cable-stayed MAPE results were
coastal bridges kernel) wind speed; accumulation bridge located in obtained
under coupled significant wave southern China
loads Zhu and height; and peak coastal regions &
Zhang (2018) wave period the wind/wave
data from
Meteorological
Observatory near
the bridge location
from 1980 to 2012
Performance ANN Levenberg- Maximum mean Fragility values Numerically Absolute Various
assessment of a Marquadt tangential associated with generated in differences architectures
vertical structure velocity of the each intensity which the Monte were tested and
subjected to non- tornado and its measures Carlo simulation the best ANN
stationary, radial length combination was employed model has one
tornadic wind scale
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 25 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 6 | (Continued) Summary of ML applications for structural dynamics and damage assessment.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

loads Le and hidden layer

Caracoglia (2020) with 4 neurons
Object detection Series of Backpropagation Digital images Bounding boxes -The models were Mean average Acceptable
in aerial imagery CNN trained and videos of the ground pretrained on the precision results
for disaster using objects of interest common objects
response and transfer (i.e., flooded area, in context/visual
recovery after the learning building roofs object classes
occurrence of damage, debris, (COCO/VOC)
hurricanes Pi et al. vegetation and databases
(2020) cars) and their Everingham et al.
corresponding (2010); Lin et al.
class labels (e.g., (2017)
damaged or - Then they were
undamaged) retrained on new
aerial video
dataset Volan
2018
(corresponding to
hurricanes that
occurred in
2017–2018)
obtained using
web mining
algorithms

developed in recent years for simultaneously achieving high appropriate set of inputs to ML models for damage assessment
simulation accuracy and efficiency. The performance-based (predictors or features) is still very challenging.
(and further resilience-ba sed) wind design philosophies also
require accurate damage assessment of structures and 3.5 Mitigation and Response
infrastructure under extreme storms. The structural damages Both long-term and short-term strategies are needed to enhance
under winds depend on numerous factors including wind resilience of individual structures or communities to withstand
features (e.g., wind speed/direction and topography) and built wind-related hazards. One important long-term consideration is
environment characteristics (e.g., building opening and roof to mitigate structural response/vibration subjected to winds
slope), hence its assessment and quantification are extremely through structural optimization and/or control. For structural
challenging. On the other hand, increasingly available field- optimization under winds, the shape optimization is probably the
measurement data characterizing structural damages under most effective approach to reduce aerodynamic loading. For
strong wind events [e.g., resulting from post-disaster wind-induced vibration control, both aerodynamic and
reconnaissance activities such NHERI Natural Hazards mechanical measures are well recognized in wind engineering
Reconnaissance (RAPID) Facility and NSF Structural Extreme community. Although the structural performance evaluation
Events Reconnaissance (StEER) Network] provide a great under winds is typically a very complicated task, the
opportunity to learn from data by using various ML models. corresponding simulations during optimization or (active)
The ML applications for structural dynamics are first reviewed control process is required to be efficient and accurate because
in Table 6i and then followed by damage assessment in Table 6ii, they need to be conducted either repeatedly for numerous
where the ML model, training scheme, input data, output data, data scenarios or in a (near) real-time sense. As noted earlier, the
source and performance metric are summarized for each ML models are very promising to simultaneously achieve the high
application. The training/testing data were essentially retrieved simulation efficiency and accuracy goal. In addition, the RL
from numerical simulations, wind tunnel tests and field models that have gained increasing popularity in recent years
measurements. From Table 6, it can be concluded that most can be used as very effective optimization or control algorithms
applications used ML as a regression model for modeling compared to conventional approaches (Silver et al., 2017). In the
structural dynamics and as a regression or a classification model consideration of short-term actions, efficient management
for structural damage assessment. While many applications strategies are critically important. Although the ML models
employed simple ML models and standard training schemes used in the disaster (including wind-related hazard)
(e.g., ANN with backpropagation), some advanced schemes management framework (i.e., covering preparedness, response
such as knowledge-enhanced LSTM have been successfully and recovery) have recently been systematically reviewed (e.g.,
applied to predict time series of wind-induced nonlinear Sun et al., 2020), its applications to social media-informed
structural response. It is noted that the selection of the most response are still discussed here since the unprecedentedly

Frontiers in Built Environment | www.frontiersin.org 26 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 7 | Summary of ML applications for mitigation and response.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

i) Structural Vibration control of ANN (2 Backpropagation ANN1&2: 20 inputs- ANN1: 4-time steps Numerically generated RMS and defined The coupled ANN
optimization & wind-induced models) absolute wind- ahead the absolute using a SIMULINK dimensionless models were able to
control under response of tall induced acceleration of three model with a total of performance reduce substantially
winds buildings with active acceleration of 3 floors (i.e., 50th, 60th 50 s of data and a indexes the peak
tuned mass damper selected floors at and 70th) sampling time of 0.001s displacement and the
Bani-Hani (2007) the current and ANN2: future control for a tall building with absolute acceleration
previous 4-time force at the next time 76-stoery (data response of the
steps, and the active step of the active generated with and building storeys
tuned mass damper tuned mass damper without random white
control forces at the noise control force of up
current and to 5 Hz frequency)
previous 4-time
steps
Aerodynamic shape ANN with a - Geometric variables Objective function = LES simulations of a R2 - Good simulation
optimization of tall genetic of the cross section the mean drag two-dimensional flow results
buildings Elshaer algorithm and the wind angle coefficient or the corresponding to - Significant
et al. (2016), (2017) of attacks standard deviation of different geometric optimization of the
the lift coefficient properties of the cross mean drag coefficient
section and standard
deviation of the lift
coefficient
Aerodynamic shape KE-DRL Gradient descent State: external Action: design RANS and LES - - Both specific direct-
optimization of tall shape of the adjustment of the simulation of a 2-D domain and cross-
buildings Li et al. structure cross section to cross section example domain knowledge
(2021a) maximize the are leveraged through
aerodynamic transfer-learning and
mitigation (by meta-learning
minimizing the drag - The deep
of a high-rise building) deterministic policy
gradient algorithm
(DDPG) was used for
the RL algorithm
- RL-based shape
optimizer
outperformed the
basic gradient
descent, particle
swarm optimization
(PSO) and typical RL
without knowledge
Bluff body active flow DRL Adam States: drag and lift Action: ratio of the Entropy-viscosity- - -The Twin Delayed
control in coefficients rotation rate for each based large eddy Deep Deterministic
experiments and rotating cylinder and simulation (LES) (for the policy gradient
simulations Fan et al. the maximum numerical simulation) algorithm was
(2020) rotation rate to and an experimental selected as the RL-
minimize the drag in setup algorithm to update
both simulations and the agent
experiments - The RL-agent was
capable to efficiently
learn a control
strategy, for both
experiment and
simulation, that will
allow the
reattachment of flow
behind the cylinder
and reduce the drag
coefficient
ii) Disaster Information NB (2 - NB1: tweets NB1: classification of 206,764 tweets F1 score Acceptable
response classification from classifiers) tweets as personal, collected during the performance
informed by disaster-related direct informative, Joplin tornado of 2011
social media messages in twitter indirect informative in Joplin, Missouri
Imran et al. (2013) direct-indirect
informative and other
following the tornado
event in Joplin,
Missouri (2011)
(Continued on following page)

Frontiers in Built Environment | www.frontiersin.org 27 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

TABLE 7 | (Continued) Summary of ML applications for mitigation and response.

Application ML model Training Input Output Data Performance Remarks

scheme data data source metric

NB2: informative NB2: classification of

tweets informative tweets as
caution, donation,
advice, or information
source
Classification of SLDA, LR, - Tweets Classification of 17 million tweets AUC The LR was the best
tweets to inform KNN, NB, DT tweets to identify collected during 12 classifier
disaster response those that reported different natural
Ashktorab et al. human casualties or disasters in the U.S
(2014) structural damage since 2006 (e.g.,
(requiring tornado and hurricane)
intervention)
Information RF, SVM, NB - Tweets 9 classes (e.g., 52 million tweets for Area under ROC - Good results were
classification from injured or dead events related to 19 curve obtained for all
disaster-related people, infrastructure natural hazards and classes (for the three
messages in twitter and utilities damage, crisis (e.g., typhoon, classifiers) except for
Imran et al. (2016) displaced people and floods and earthquake) the “missing trapped
evacuations, caution occurring between or found people” -
and advice) 2013 and 2015 in poor classification-
different parts of the
world was used
Information SVM, KNN, - Images Image classes in The images were Average SVM-based model
classification from GNB, MNB, terms of human roles: collected from August precision gave the best
disaster-related BNB, rescuees or rescuers 17th to 3 September prediction accuracy
events O’Neal et al. DT, SGD 2017 based on private
(2018) social media platforms
(e.g., twitter) during
Hurricane Harvey
(2017)
Real-time disaster VGG-16 Adam Tweeter-based VGG-16 CNN: A total of 17,483 Accuracy Acceptable
communication CNN, ANN images informative features images were extracted simulation results
Robertson et al. (pre-storm, landfall from Twitter between
(2019) and the period after 17th August and 17
landfall) September 2017 from
ANN: urgency level Hurricane Harvey
(highly urgent, (2017)
moderately urgent,
somewhat urgent,
not urgent, and
unrelated to the
hurricane event)
Information ANN, SVM, - Tweets 2 classes: crisis- 6 crisis related datasets Accuracy ANN classifier
classification from NB, LR related tweets and were used (e.g., outperformed all
disaster-related non-crisis-related hurricane Harvey 2017 other classifiers
events Manna and tweets and the 2011 Joplin
Nakai (2019) Tornado) with
approximately 10,000
tweets for each event
Real-time information CNN, RMSprop Tweets 5 classes: 3 manually labeled Accuracy CNN outperformed
classification from SVM, LR Information Sources, datasets were used other classifiers
hurricane-related Caution and Advice, corresponding to
events Yu et al., Infrastructure and hurricane Sandy (2012),
(2019) Resources, Irma (2017) and Harvey
Casualties and (2017), respectively
Damage, and with approximatively
Donation and Aid 2000–3000 tweets per
each event
Identification of social DT, SVM, - Tweets Tweets from people 2,072,715 tweets F1-score CNN, SVM and ANN
media-based ANN, LR, requiring or not related to Hurricane achieve the best
requests for urgent NB, urgent rescue by first Harvey (2017) event simulation results
help during AdaBoost, responders
hurricanes Devaraj RR
et al., (2020)

abundant data from various powerful communication tools (e.g., model, training scheme, input data, output data, data source and
Twitter) greatly facilitate the rapid ML model developments in performance metric are summarized for each application. The
this field. Table 7i,ii respectively present the reviewed training/testing data were essentially retrieved from CFD
applications of ML for mitigation and response, where the ML simulations and experimental tests for structural mitigation or

Frontiers in Built Environment | www.frontiersin.org 28 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

FIGURE 8 | Overview of reviewed ML applications in wind engineering (following Alan G. Davenport Wind Loading Chain).

from social media platforms for disaster response. From Table 7i, tunnel tests and CFD simulations. In general, the supervised learning
it can be concluded that the structural performance evaluations in dominates the ML applications in wind engineering with the podium
mitigation applications usually used ML as a regression model position attributed to simple models with standard algorithms (e.g.,
while RL was typically utilized as an effective optimization or ANN with backpropagation). Actually, the selection of various ML
control algorithm. It is noted that relatively few ML applications models is rather rudimentary since there is a lack of systematic
for structural optimization and control under winds have been comparison among them (e.g., in terms of model complexity and
generated compared to those in earthquake engineering performance). It is noted that the great potential of semi-supervised
community (e.g., Ghaboussi and Joghataie 1995; Adam and learning and unsupervised learning (as well as RL) with little or no
Smith 2008; Jiang and Adeli 2008; Yakut and Alli 2011; labelled data is not leveraged yet. Accordingly, the current ML
Subasri et al., 2014; Khodabandehlou et al., 2018; developments in wind engineering heavily rely on available
Khalatbarisoltani et al., 2019; Hayashi and Ohsaki 2020). From labelled data. For example, the ML applications to non-synoptic
Table 7ii, it can be concluded that most social media-informed winds are much less than those of synoptic winds due essentially to
response applications used ML as a classification model for the difficulty in obtaining the data of local and short-lived storms.
disaster rescue and relief information dissemination. Although On the other hand, the recent emergence of numerous ML
these ML applications present promising results in terms of applications to social media-informed disaster response is due
effectively supporting timely decision-making, there is a mainly to the unprecedentedly abundant data from various
concern of using information from social media platforms due powerful communication tools. For the reviewed ML
to a lack of data quality control. applications, the training/testing data are retrieved from several
major sources (e.g., field measurements, wind tunnel tests,
3.6 Summary numerical simulations and social media platforms). In the
The ML applications in each topical area of wind engineering are determination of ML model inputs and outputs, a good
summarized in Figure 8. As shown in the figure, ML models are understanding of underlying physics of each application is critical
unevenly distributed among these areas. The wind climate area to effectively select an appropriate set of predictors (ML inputs)
has the most ML applications followed by the aerodynamics and while the output types heavily depend on the needs of traditional
aeroelasticity area, and they are respectively contributed by wind analysis procedure in each application (e.g., local wind pressures for
engineering-related fields of meteorology and fluid mechanics. building design and global wind forces for bridge design).
On the other hand, the wind engineering-exclusive field of
terrain and topography has the least applications of ML.
Although ML models have been instrumental in modern 4 CHALLENGES AND PROSPECTS
structural design for winds, their developments are in a very
preliminary stage and there is still a long way to go before they The rapidly increasing ML applications to wind engineering have
can complement or even replace existing approaches of wind generated a large volume of datasets associated with a large set of

Frontiers in Built Environment | www.frontiersin.org 29 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

domain-specific algorithms. It is strongly believed that the create extra points in currently data-scarce dimensions. Wind
platforms encouraging open sharing of these datasets and engineering data could be short in time span of their collection.
algorithms would greatly benefit the ML research progress in For example, the climate changing impacts are not easy to be
wind engineering. The openly available wind engineering datasets considered based on the currently available wind data since their
will greatly reduce efforts for their creation/collection and pre- record period is much shorter than the time scale of climate
processing, and open-source ML algorithms will save significant changing. Also, few structural performance data under winds are
time for their re-implementation. The reduced need of time and long enough to take the life-span deterioration behaviors into
effort to use the state-of-the-art or latest developed ML tools account. Essentially, the learning machine based on current wind-
under such a culture of openness would spur interests among structure interaction data cannot be used for accurately
researchers in wind engineering, and hence result in more related predicting future long-term behaviors of the same wind-
ML applications. Moreover, the developed cyberinfrastructure to structure system. Wind engineering data could be highly
store and share data usually has a systematic curation procedure heterogeneous for collaborative or large-scale ML applications.
to ensure the high quality of its standardized benchmark datasets. Many complex tasks (e.g., life-cycle performance evaluation of
Also, the open-source software allows the hidden bugs/tricks of structures under winds) and/or real-world problems (e.g.,
ML algorithms to be easily uncovered and accordingly makes hurricane resilience assessment of coastal communities) in
them more robust. In addition to availability, the reproducibility wind engineering need collaborative efforts and/or large-scale
and testability of wind engineering data and domain-specific implementations. The datasets generated from these activities
algorithms due to a culture of openness would also facilitate may result from various CFD simulation tools or field
the adoption of the obtained transparent and trustworthy ML measurement devices, and they are typically interpreted by
tools in real-world problems. Although the wind engineering different entities before sent to a central processing platform.
community has started to embrace the prevalent openness of ML Accordingly, significant processing efforts (e.g., data cleaning,
community (e.g., NHERI DesignSafe platform), the culture of data aggregation, dimension reduction and data standardization)
openness is still in its early stage. It is expected that more are needed for these heterogeneous datasets with high variability
incentives based on the existing reward system (e.g., a digital of data types and formats (e.g., mixtures of structured, semi-
object identifier for each dataset or algorithm published by the structured and unstructured data). In addition, advanced
platform) are needed to motivate the ML wind engineering powerful learning machines are necessary to generate new
community towards open science. Given a potential open- knowledge from large, heterogeneous sets of wind
science environment with openly available datasets and open- engineering data.
source algorithms (supported by open-access scientific
publications), some remaining challenges and future prospects 4.1.2 Machine Learning Algorithm Challenges
are discussed in terms of data in wind engineering and algorithms ML algorithms commonly-used in wind engineering are standard
in ML. It is noted that both challenge and prospect lists are not ones designed for solving problems in other fields (e.g., handwriting
exhaustive. recognition or computer vision). While these classical algorithms
(e.g., ANN with backpropagation) achieved great success for simple
4.1 Challenges and Research Gaps wind engineering applications, they are not necessarily concise and
The reviewed various ML models for a wide range of topics in efficient. More importantly, the immediate applications of these
wind engineering suggests that their cross field has recently popular algorithms to modern wind engineering (involving
attracted much interest. However, there are still numerous nonstationary and non-Gaussian wind flow, transient and
challenges to advance ML applications to wind engineering nonlinear aerodynamics, nonlinear and inelastic structural
from conception and research into practice. These remaining dynamics, or time-variant wind-structure system under a
challenges of data in wind engineering and algorithms in ML are changing climate) may be very challenging. On the other hand,
discussed in this sub-section. the newly developed ML algorithms (e.g., advanced LSTM and
GAN) need to be carefully scrutinized for their applicability to these
4.1.1 Wind Engineering Data Challenges complex problems. ML algorithms commonly-used in wind
Wind engineering data could be rich in some dimensions but may engineering are supervised ones that need a significant amount of
be poor in others. For example, a large volume of flow data or labelled data. Although the cost of obtaining/collecting the data from
pressure data could be obtained by one wind tunnel test (using various sources (e.g., numerical simulations, wind tunnel tests, or
advanced measurement systems with high resolution in space and field measurements) is greatly reduced and accordingly
high sampling rate in time), however, all these data would be unprecedented volume of data are increasingly available, these
located at a point in the Reynolds number dimension. For datasets may be limited to unlabeled due to a lack of sufficient
structural response under winds, most of the data are located human resources (with expert knowledge) for data labeling. ML
in the linear elastic domain, while very limited nonlinear inelastic algorithms commonly-used in wind engineering are purely data-
data needed to advance implementation of performance-based driven ones that are usually consider as black boxes. Furthermore,
wind design are available. Another example is that the currently available ML models usually present a conflict between
anemometric monitoring network typically generates abundant their advances (and hence performance) and explainability. One
data in time dimension but sparse data in space. More important feature of human intelligence is the ability to explain the
importantly, it is usually very challenging or expensive to rationale behind its decisions to others, hence, the explainability of

Frontiers in Built Environment | www.frontiersin.org 30 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

learning machines is often an essential prerequisite for establishing a wind fields (resulting from tropical cyclones, extratropical
trust relationship between human intelligence and artificial cyclones or local non-synoptic storms) need to be generated
intelligence. The highly non-transparent nature of ML algorithms by global climate models coupled with accurate and efficient
may be acceptable for some applications in wind engineering (e.g., a downscaling exercises under projected climate conditions [e.g.,
CNN mapping the oncoming winds to pressure fields on or velocity various RCP (Representative Concentration Pathway) scenarios].
fields around various bridge decks), however, it may be a clear To effectively learn from heterogeneous data that need to be first
drawback for many high-stake applications (e.g., evacuation unified, they can be efficiently processed by advanced big data
planning or transportation infrastructure management under a analytics. For example, unsupervised or semi-supervised
landfalling hurricane) since any error in prediction may have clustering techniques could be used for data cleaning, data
catastrophic consequences. It is noted that the high-stake fusion techniques of Kalman filters could be used for data
applications also place a high demand for quantification of aggregation, and linear principal component analysis or
uncertainties involved in ML algorithm selection, training and nonlinear self-organizing map could be used for dimensional
performance evaluation (along with data collection), whereas the reduction.
formalization of uncertainty quantification for purely data-
driven approaches is very challenging and not well 4.2.2 ML Algorithm Prospects
established yet. ML algorithms commonly-used in wind To facilitate ML applications to complex wind engineering
engineering are typically selected based on past experience problems, the state-of-the-art or latest algorithms emerging in
(or simply by “gut feeling”) and the associated model ML community could be leveraged. For example, the GAN could
hyperparameters (e.g., layer and neuron numbers, be used for effectively generating nonstationary and non-
activation function and learning rate) are usually obtained Gaussian wind flow through its two competing sub-networks,
by extensive trial and error. While the selected ML the CNN could be employed for efficiently mapping oncoming
algorithms present good performance for the particular winds to pressure fields (characterizing transient and nonlinear
applications of interest, they are not necessarily an aerodynamics) on structures with an arbitrary shape because it is
optimal choice. A systematic approach to identify the particularly good at handling input-output data with a known
most appropriate ML model and associated best grid-like topology, the LSTM could be utilized for accurately
hyperparameters essentially needs a global optimization simulating nonlinear and inelastic structural dynamics since its
within a high dimensional space, and is currently very forget gates ensure a reliable consideration of long-term
challenging for wind engineering applications. dependencies (where the structural response at the current
time depends on not only the current wind load but also the
4.2 Prospects and Future Directions load history), and the lifelong learning networks should be
The remaining challenges, while not trivial, provide new research explored for adaptively modeling time-variant wind-structure
opportunities for the development of more effective ML tools. system assuming their underlying parameters can be
The identified prospects of data in wind engineering and continuously modified to accommodate new data inputs. The
algorithms in ML are discussed in this sub-section. direct or immediate applications of the advanced ML algorithms
to complex wind engineering problems may not necessarily result
4.2.1 Wind Engineering Data Prospects in parsimonious models that may need specialized customization
To generate/collect wind engineering data that are scarce in for each application. To reduce the demand for labelled data in
certain dimensions, advanced full-scale/laboratory/numerical ML applications to wind engineering, both unsupervised learning
tools and technologies need to be utilized or developed. In and semi-supervised learning (including physics-informed
addition to large-scale facilities (e.g., WindEEE), various high- machine learning) are promising alternatives to popularly used
fidelity and efficient modern CFD techniques (e.g., hybrid large supervised learning. In addition, advanced ML algorithms have
eddy simulation/Reynolds-averaged Navier-Stokes schemes) been emerging (e.g., reservoir computing) for processing
should be exploited to generate data of high-Reynolds number information generated by complicated dynamical systems
scenarios. The rational loading protocols for extreme wind using very small training datasets. To open the ML black box,
performance cyclic testing of deformation-controlled MWFRS model explainability and interpretability in wind engineering
(Main Wind Force Resisting System) members need to be applications needs to be enhanced. Various general techniques
designed to generate the wind-induced nonlinear inelastic have been developed to improve understanding of the ML model
structural response data. Also, data reconstructions using predictions, such as sensitivity analysis and layer-wise relevance
linear/nonlinear dimensionality reduction techniques (e.g., propagation. On the other hand, the definitions of explainability
singular value decomposition/autoencoder) should be and interpretability are typically domain dependent, hence, the
employed to enhance spatial resolution of full-scale domain knowledge in wind engineering should be leveraged for
measurements. To generate/collect wind engineering data that enhanced explainability/interpretability of each ML application.
cover sufficiently-long time span of structural behaviors, more It is expected that the explainability/interpretability analysis
reliable long-term structural health monitoring systems should be (along with uncertain quantification) will likely become a
established in addition to high-fidelity modeling of aging and fundamental building block for bounding the overall
deterioration of wind-sensitive structures. For the consideration confidence in ML applications in wind engineering (parallel to
of wind engineering data under a changing climate, synthesized verification and validation in CFD simulations). To enable an

Frontiers in Built Environment | www.frontiersin.org 31 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

automatic search of ML model hyperparameters in wind engineering formulas, probabilistic relations, logic rules, simulation results,
applications, increasingly available optimization schemes with field observations, human feedback, and others) in various
improved efficiency and accuracy (e.g., grid search, random subfields of wind engineering and then to effectively integrate
search, Bayesian optimization and population-based training) can them into each module of machine learning pipeline (data
be utilized to find the best configuration for each task. On the other preparation, model selection, model training, and others).
hand, it is believed that a practical guide to selection of ML models in While domain knowledge could be employed to enhance
wind engineering applications will greatly facilitate their appropriate purely data-driven ML tools, it is expected that learning
use. The best practices for model selection in each application are machines could be utilized for harnessing data to discover new
essentially consistent with the principle of Ockham’s razo by first knowledge in wind engineering (e.g., governing laws
testing simple linear ML models (due to their easy to implement and characterizing transport of turbulence quantities or
high model explainability), and then followed by more complex optimization of wind-structure system).
nonlinear models (without data overfitting). Among ML models
with similar complexity, a predetermined performance metric is
typically used for further model selection. Since iteration is generally 5 CONCLUDING REMARKS
needed in a purely performance-driven ML model selection, the
domain knowledge is suggested to be utilized for a more effective A total of 65 machine learning (ML) algorithms were reviewed in
search process. terms of their applications to each topical area of wind
engineering, namely wind climate, terrain/topography,
aerodynamics/aeroelasticity, structural dynamics, wind damage
4.3 Knowledge-Enhanced Machine assessment and wind-related hazard mitigation and response.
Learning The most ML applications were found in wind climate area, while
As discussed in preceding sections, domain knowledge could be the terrain/topography area had the least applications of ML.
leveraged for improved selection of ML model and its inputs and Although the ML-based wind engineering is fueled by the
outputs in wind engineering applications. Hence, a good unprecedented volume of analytical, numerical,
understanding of fundamental physics and other types of experimental and field-measurement data together with
domain knowledge underlying each subfield of wind rapidly evolving learning algorithms and high-
engineering would enable more effective use of ML tools. It is performance computational hardware, it is still at an early
noted that the fundamental physics in terms of governing stage of development. Most of wind engineering applications
equations is a special type of domain knowledge, and recent employed supervised learning with standard ML models
studies have demonstrated that the required labelled datasets designed for solving problems in other fields, and the
could be significantly reduced by incorporating the underlying promising unsupervised and semi-supervised learning
physics into training process (and hence enhancing the tools were rarely used to reduce the high demand of
regularization mechanism) (e.g., Raissi et al., 2017a; 2017b). labelled data. For the selection of ML models and
Other equation-based domain knowledge such as empirical/ associated hyperparameters in wind engineering
semi-empirical formulas were also employed as part of the applications, it was typically based on expertise and
loss function in deep learning to provide machine-readable extensive trial and error. In this review, the culture of
prior knowledge that facilitates the effective regularization openness, explainability/interpretability and uncertainty
of the neural networks for simulations of tropical cyclone quantification were identified as important research gaps
winds (Snaiki and Wu 2019) and nonlinear structural that need to be addressed in ML-based wind engineering
dynamics (Wang and Wu 2020). In addition, the equation-free community. Furthermore, the knowledge-enhanced machine
domain knowledge has been integrated into a deep RL-based learning was considered as a very promising scheme to
aerodynamic shape optimizer (via the transfer-learning and enhance ML applications to wind engineering.
meta-learning techniques) to remarkably enhance the training
efficiency for wind engineering applications (Li et al., 2021a).
These emerging successful applications indicate that this novel AUTHOR CONTRIBUTIONS
scheme of knowledge-enhanced machine learning (KEML) could
significantly enhance ML applications to wind engineering. To All authors contributed to the study conception and design, data
fully embrace the promising potential of KEML, systematic collection, analysis and interpretation of results, drafted
research efforts are needed to efficiently identify knowledge manuscript preparation, reviewed the results, and approved
representations (invariances, physics equations, empirical the final version of the manuscript.

Abdi, D., Levine, S., and Bitsuamlak, G. “Application of an Artificial Neural

REFERENCES Network Model for Boundary Layer Wind Tunnel Profile Development,” in
Proceedingsof the 11th Americas Conference on Wind Engineering, San Juan,
Abbas, T., Kavrakov, I., Morgenthal, G., and Lahmer, T. (2020). Prediction of Puerto Rico, June 2009.
Aeroelastic Response of Bridge Decks Using Artificial Neural Networks. Abdullah, S. A., Aswegan, K., Jaberansari, S., Klemencic, R., and Wallace, J. W.
Comput. Structures 231, 106198. doi:10.1016/j.compstruc.2020.106198 (2020). Performance of Reinforced Concrete Coupling Beams Subjected to

Frontiers in Built Environment | www.frontiersin.org 32 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Simulated Wind Loading. ACI Struct. J. 117 (3), 283–295. doi:10.14359/ Bitsuamlak, G. T., Bédard, C., and Stathopoulos, T. (2007). Modeling the Effect of
51724555 Topography on Wind Flow Using a Combined Numerical-Neural Network
Aboshosha, H., Bitsuamlak, G., and El Damatty, A. (2015). Turbulence Approach. J. Comput. Civ. Eng. 21 (6), 384–392. doi:10.1061/(asce)0887-
Characterization of Downbursts Using LES. J. Wind Eng. Ind. Aerodynamics 3801(2007)21:6(384)
136, 44–61. doi:10.1016/j.jweia.2014.10.020 Bitsuamlak, G. T. (2004). “Evaluating the Effect of Topographic Elements on Wind
Aboutabikh, M., Ghazal, T., Chen, J., Elgamal, S., and Aboshosha, H. (2019). Flow: a Combined Numerical Simulation-Neutral Network Approach,”
Designing a Blade-System to Generate Downburst Outflows at Boundary Layer (Montreal, Quebec, Canada: Concordia University). Doctoral dissertation.
Wind Tunnel. J. Wind Eng. Ind. Aerodynamics 186, 169–191. doi:10.1016/j. Bitsuamlak, G. T., Stathopoulos, T., and Bédard, C. “Neural Network Predictions of
jweia.2019.01.005 Wind Flow over Complex Terrain,” in 4th Structural Specialty Conf. of the
Adam, B., and Smith, I. F. (2008). Reinforcement Learning for Structural Control. Canadian Society for Civil Engineering, Whistler, BC Canada, May 2002.
J. Comput. Civ. Eng. 22 (2), 133–139. doi:10.1061/(asce)0887-3801(2008)22: Blessmann, J., and Riera, J. D. (1985). Wind Excitation of Neighbouring Tall
2(133) Buildings. J. wind Eng. Ind. aerodynamics 18 (1), 91–103. doi:10.1016/0167-
Adeli, H., and Yeh, C. (1989). Perceptron Learning in Engineering Design. 6105(85)90076-5
Computer-Aided Civil Infrastructure Eng. 4 (4), 247–256. Blocken, B. (2014). 50 Years of Computational Wind Engineering: Past, Present
Adrianto, I., Trafalis, T. B., and Lakshmanan, V. (2009). Support Vector Machines and Future. J. Wind Eng. Ind. Aerodynamics 129, 69–102. doi:10.1016/j.jweia.
for Spatiotemporal Tornado Prediction. Int. J. Gen. Syst. 38, 759–776. doi:10. 2014.03.008
1080/03081070601068629 Bluestein, H. B. (2021). “The Types of Non-synoptic Wind Systems,” in The Oxford
Alemany, S., Beltran, J., Perez, A., and Ganzfried, S. (2019). Predicting Hurricane Handbook of Non-synoptic Wind Storms. Editors H. Hangan and A. Kareem
Trajectories Using a Recurrent Neural Network. Aaai 33 (01), 468–475. doi:10. (Oxford, United Kingdom: Oxford University Press). doi:10.1093/oxfordhb/
1609/aaai.v33i01.3301468 9780190670252.013.1
Ali, M. M., Kishtawal, C. M., and Jain, S. (2007). Predicting Cyclone Tracks in the Bre, F., Gimenez, J. M., and Fachinotti, V. D. (2018). Prediction of Wind Pressure
north Indian Ocean: An Artificial Neural Network Approach. Geophys. Res. Coefficients on Building Surfaces Using Artificial Neural Networks. Energy and
Lett. 34 (4), L04603. doi:10.1029/2006gl028353 Buildings 158, 1429–1441. doi:10.1016/j.enbuild.2017.11.045
American Society of Civil Engineers (2017). Minimum Design Loads and Breiman, L. (2001). Random Forests. Mach Learn. 45, 5–32. doi:10.1023/a:
Associated Criteria for Buildings and Other Structures. Reston, VA: 1010933404324
Structural Engineering Institute of American Society of Civil Engineers. Brunton, S. L., Noack, B. R., and Koumoutsakos, P. (2020). Machine Learning for
American Society of Civil Engineers (2021). Minimum Design Loads for Buildings Fluid Mechanics. Annu. Rev. Fluid Mech. 52, 477–508. doi:10.1146/annurev-
and Other Structures. Reston, VA: Structural Engineering Institute of American fluid-010719-060214
Society of Civil Engineers. Cermak, J. E. (1975). Applications of Fluid Mechanics to Wind Engineering—A
Asano, K., Iida, Y., and Uematsu, Y. (2019). Laboratory Study of Wind Loads on a Freeman Scholar Lecture. J. Fluids Eng. 97 (1), 9–38. doi:10.1115/1.3447225
Low-Rise Building in a Downburst Using a Moving Pulsed Jet Simulator and Chakrabarty, H., Murthy, C. A., and Gupta, A. D. (2013). Application of Pattern
Their Comparison with Other Types of Simulators. J. Wind Eng. Ind. Recognition Techniques to Predict Severe Thunderstorms. Ijcte 5 (6), 850–855.
Aerodynamics 184, 313–320. doi:10.1016/j.jweia.2018.11.034 doi:10.7763/ijcte.2013.v5.810
Ashktorab, Z., Brown, C., Nandi, M., and Culotta, A. “Tweedr: Mining Twitter to Chaudhuri, S., Dutta, D., Goswami, S., and Middey, A. (2013). Intensity Forecast of
Inform Disaster Response,” in Proceedings of the 11th International ISCRAM Tropical Cyclones over North Indian Ocean Using Multilayer Perceptron
Conference, Pennsylvania, USA, May 2014, 269–272. Model: Skill and Performance Verification. Nat. Hazards 65 (1), 97–113.
Ashton, R., Refan, M., Iungo, G. V., and Hangan, H. (2019). Wandering doi:10.1007/s11069-012-0346-7
Corrections from PIV Measurements of Tornado-like Vortices. J. Wind Eng. Chaudhuri, S., and Middey, A. (2011). Adaptive Neuro-Fuzzy Inference System to
Ind. Aerodynamics 189, 163–172. doi:10.1016/j.jweia.2019.02.010 Forecast Peak Gust Speed during Thunderstorms. Meteorology Atmos. Phys.
Baik, J.-J., and Paek, J.-S. (2000). A Neural Network Model for Predicting 114 (3-4), 139. doi:10.1007/s00703-011-0158-4
Typhoon Intensity. J. Meteorol. Soc. Jpn. 78 (6), 857–869. doi:10.2151/ Chen, C. H., Wu, J. C., and Chen, J. H. (2008). Prediction of Flutter Derivatives by
jmsj1965.78.6_857 Artificial Neural Networks. J. wind Eng. Ind. aerodynamics 96 (10-11),
Bailey, P. A., and Kwok, K. C. S. (1985). Interference Excitation of Twin Tall 1925–1937. doi:10.1016/j.jweia.2008.02.044
Buildings. J. Wind Eng. Ind. Aerodynamics 21 (3), 323–338. doi:10.1016/0167- Chen, G., and Lombardo, F. T. (2020). An Automated Classification Method of
6105(85)90043-1 Thunderstorm and Non-thunderstorm Wind Data Based on a Convolutional
Baker, C. J., and Sterling, M. (2017). Modelling Wind fields and Debris Flight in Neural Network. J. Wind Eng. Ind. Aerodynamics 207, 104407. doi:10.1016/j.
Tornadoes. J. Wind Eng. Ind. Aerodynamics 168, 312–321. doi:10.1016/j.jweia. jweia.2020.104407
2017.06.017 Chen, R., Wang, X., Zhang, W., Zhu, X., Li, A., and Yang, C. (2019). A Hybrid
Bani-Hani, K. A. (2007). Vibration Control of Wind-induced Response of Tall CNN-LSTM Model for Typhoon Formation Forecasting. Geoinformatica 23
Buildings with an Active Tuned Mass Damper Using Neural Networks. (3), 375–396. doi:10.1007/s10707-019-00355-0
Struct. Control. Health Monit. 14 (1), 83–108. doi:10.1002/stc.85 Chen, R., Zhang, W., and Wang, X. (2020). Machine Learning in Tropical Cyclone
Barbounis, T. G., Theocharis, J. B., Alexiadis, M. C., and Dokopoulos, P. S. (2006). Forecast Modeling: A Review. Atmosphere 11 (7), 676. doi:10.3390/
Long-term Wind Speed and Power Forecasting Using Local Recurrent Neural atmos11070676
Network Models. IEEE Trans. Energ. Convers. 21 (1), 273–284. doi:10.1109/tec. Chen, Y., and Duan, Z. (2018). A Statistical Dynamics Track Model of Tropical
2005.847954 Cyclones for Assessing Typhoon Wind hazard in the Coast of Southeast China.
Bengio, Y., Simard, P., and Frasconi, P. (1994). Learning Long-Term Dependencies J. Wind Eng. Ind. Aerodynamics 172, 325–340. doi:10.1016/j.jweia.2017.11.014
with Gradient Descent Is Difficult. IEEE Trans. Neural Netw. 5 (2), 157–166. Chen, Y., Kopp, G. A., and Surry, D. (2002). Interpolation of Wind-Induced
doi:10.1109/72.279181 Pressure Time Series with an Artificial Neural Network. J. Wind Eng. Ind.
Berg, J., Mann, J., and Patton, E. G. (2013). Lidar-observed Stress Vectors and Veer Aerodynamics 90 (6), 589–615. doi:10.1016/s0167-6105(02)00155-1
in the Atmospheric Boundary Layer. J. Atmos. oceanic Technol. 30 (9), Chen, Y., Kopp, G. A., and Surry, D. (2003). Prediction of Pressure Coefficients on
1961–1969. doi:10.1175/jtech-d-12-00266.1 Roofs of Low Buildings Using Artificial Neural Networks. J. wind Eng. Ind.
Berggren, K., Xia, Q., Likharev, K. K., Strukov, D. B., Jiang, H., Mikolajick, T., et al. aerodynamics 91 (3), 423–441. doi:10.1016/s0167-6105(02)00381-1
(2020). Roadmap on Emerging Hardware and Technology for Machine Chen, Z., Yu, X., Chen, G., and Zhou, J. “Cyclone Intensity Estimation Using
Learning. Nanotechnology 32 (1), 012002. doi:10.1088/1361-6528/aba70f Multispectral Imagery from the FY-4 Satellite,” in Proceedings of the 2018
Bitsuamlak, G., Stathopoulos, T., and Bedard, C. (2006). Effects of Upstream Two- International Conference on Audio, Language and Image Processing (ICALIP),
Dimensional hills on Design Wind Loads: a Computational Approach. Wind Shanghai, China, July 2018 (Piscataway, New Jersey, United States: IEEE),
and Structures 9 (1), 37–58. doi:10.12989/was.2006.9.1.037 46–51.

Frontiers in Built Environment | www.frontiersin.org 33 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Cheng, X. X., Zhao, L., and Ge, Y.-J. (2016). Field Measurements on Flow Past a in Proceedings of the 8th International Colloquium on Bluff Body
Circular cylinder in Transcritical Reynolds Number Regime. Acta Phys. Sin. 65 Aerodynamics and Applications, June 2016Northeastern University. Boston,
(21), 214701. doi:10.7498/aps.65.214701 Massachusetts.
Cherkassky, V., and Mulier, F. M. (2007). Learning from Data: Concepts, Theory, Emanuel, K., Ravela, S., Vivant, E., and Risi, C. (2006). A Statistical Deterministic
and Methods. Hoboken, New Jersey, United States: John Wiley & Sons. Approach to hurricane Risk Assessment. Bull. Amer. Meteorol. Soc. 87 (3),
Chitsazan, M. A., Sami Fadali, M., and Trzynadlowski, A. M. (2019). Wind Speed 299–314. doi:10.1175/bams-87-3-299
and Wind Direction Forecasting Using echo State Network with Nonlinear Emanuel, K. (2003). Tropical Cyclones. Annu. Rev. Earth Planet. Sci. 31 (1),
Functions. Renew. Energ. 131, 879–889. doi:10.1016/j.renene.2018.07.060 75–104. doi:10.1146/annurev.earth.31.100901.141259
Chowdhury, J., and Wu, T. (2021). “Aerodynamic Loading Due to Non-synoptic English, E. C., and Fricke, F. R. (1999). The Interference index and its
Wind Systems,” in The Oxford Handbook of Non-synoptic Wind Storms. Editors Prediction Using a Neural Network Analysis of Wind-Tunnel Data.
H. Hangan and A. Kareem (Oxford, United Kingdom: Oxford University J. Wind Eng. Ind. Aerodynamics 83 (1-3), 567–575. doi:10.1016/s0167-
Press), 337. 6105(99)00102-6
Chung, J., Lee, S. W., Chang, S., and Kim, Y. S. “Estimation of Flutter Derivatives of Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J., and Zisserman, A.
Various Sections Using Numerical Simulation and Neural Network,” in The (2010). The Pascal Visual Object Classes (Voc) challenge. Int. J. Comput. Vis. 88
2012 World Congress on Advances in Civil, Environmental, and Materials (2), 303–338. doi:10.1007/s11263-009-0275-4
Research (ACEM’ 12), Seoul, Korea, August 26-30, 2012. Facchini, L., Betti, M., and Biagini, P. (2014). Neural Network Based Modal
Coffer, B., Kubacki, M., Wen, Y., Zhang, T., Barajas, C. A., and Gobbert, M. K. Identification of Structural Systems through Output-Only
(2020). “Using Machine Learning Techniques for Supercell Tornado Prediction Measurement. Comput. Structures 138, 183–194. doi:10.1016/j.
with Environmental Sounding Data,” in Tech. Rep. HPCF–2020–18, UMBC compstruc.2014.01.013
High Performance Computing Facility (Baltimore County: University of Fahlman, S. E. (1988). “Faster-learning Variations of Back-Propagation: An
Maryland). Empirical Study,” in Proc. 1988 Connectionist Models Summer School.
Collins, W. G., and Tissot, P. (2016). Thunderstorm Predictions Using Artificial Editors D. Touretzky, G. Hinton, and T. Sejnowski (San Mateo, CA:
Neural networksArtificial Neural Networks-Models and Applications. London: Morgan Kaufmann), 38–51.
IntechOpen. Fang, G., Zhao, L., Cao, S., Ge, Y., and Pang, W. (2018). A Novel Analytical Model
Collins, W., and Tissot, P. (2015). An Artificial Neural Network Model to Predict for Wind Field Simulation Under Typhoon Boundary Layer Considering Multi-
Thunderstorms within 400 Km2 South Texas Domains. Met. Apps 22 (3), Field Correlation and Height-Dependency. J. Wind. Eng. Ind. Aerodyn. 175,
650–665. doi:10.1002/met.1499 77–89.
Cortes, C., and Vapnik, V. (1995). Support-vector Networks. Mach Learn. 20 (3), Fan, D., Yang, L., Wang, Z., Triantafyllou, M. S., and Karniadakis, G. E. (2020).
273–297. doi:10.1007/bf00994018 Reinforcement Learning for bluff Body Active Flow Control in Experiments
Cui, W., and Caracoglia, L. (2019). A New Stochastic Formulation for Synthetic and Simulations. Proc. Natl. Acad. Sci. USA 117 (42), 26091–26098. doi:10.
hurricane Simulation over the north Atlantic Ocean. Eng. Structures 199, 1073/pnas.2004939117
109597. doi:10.1016/j.engstruct.2019.109597 Fernández-Cabán, P. L., Masters, F. J., and Phillips, B. M. (2018). Predicting Roof
Davenport, A. G. (1960). Rationale for Determining Design Wind Velocities. Pressures on a Low-Rise Structure from Freestream Turbulence Using
J. Struct. Div. 86 (5), 39–68. doi:10.1061/jsdeag.0000521 Artificial Neural Networks. Front. Built Environ. 4, 68. doi:10.3389/fbuil.
Deierlein, G. G., and Zsarnóczay, A. (2021). State of the Art in Computational 2018.00068
Simulation for Natural Hazards Engineering. Second Edition. Plano, Texas: Forthofer, J. M., Butler, B. W., McHugh, C. W., Finney, M. A., Bradshaw, L. S.,
Center Comput. Modeling Simulation. SimCenter. Stratton, R. D., et al. (2014b). A Comparison of Three Approaches for
DeMaria, M., Mainelli, M., Shay, L. K., Knaff, J. A., and Kaplan, J. (2005). Further Simulating fine-scale Surface Winds in Support of Wildland Fire
Improvements to the Statistical hurricane Intensity Prediction Scheme (SHIPS). Management. Part II. An Exploratory Study of the Effect of Simulated
Weather Forecast. 20 (4), 531–543. doi:10.1175/waf862.1 Winds on Fire Growth Simulations. Int. J. Wildland Fire 23 (7), 982–994.
Deng, L., and Yu, D. (2014). Deep Learning: Methods and Applications. Foundations doi:10.1071/wf12090
Trends. Signal. Processing 7 (3–4), 197–387. doi:10.1561/2000000039 Forthofer, J. M., Butler, B. W., and Wagenbrenner, N. S. (2014a). A Comparison of
Derkevorkian, A., Hernandez-Garcia, M., Yun, H.-B., Masri, S. F., and Li, P. (2015). Three Approaches for Simulating fine-scale Surface Winds in Support of
Nonlinear Data-Driven Computational Models for Response Prediction and Wildland Fire Management. Part I. Model Formulation and Comparison
Change Detection. Struct. Control. Health Monit. 22 (2), 273–288. doi:10.1002/ against Measurements. Int. J. Wildland Fire 23 (7), 969–981. doi:10.1071/
stc.1673 wf12089
Devaraj, A., Murthy, D., and Dontula, A. (2020). Machine-learning Methods Fu, J. Y., Li, Q. S., and Xie, Z. N. (2006). Prediction of Wind Loads on a Large Flat
for Identifying Social media-based Requests for Urgent Help during Roof Using Fuzzy Neural Networks. Eng. Structures 28 (1), 153–161. doi:10.
Hurricanes. Int. J. Disaster Risk Reduction 51, 101757. doi:10.1016/j. 1016/j.engstruct.2005.08.006
ijdrr.2020.101757 Fu, J. Y., Liang, S. G., and Li, Q. S. (2007). Prediction of Wind-Induced Pressures on
Diaz, J., and Joseph, M. B. (2019). Predicting Property Damage from Tornadoes a Large Gymnasium Roof Using Artificial Neural Networks. Comput. structures
with Zero-Inflated Neural Networks. Weather Clim. Extremes 25, 100216. 85 (3-4), 179–192. doi:10.1016/j.compstruc.2006.08.070
doi:10.1016/j.wace.2019.100216 Fukami, K., Nabae, Y., Kawai, K., and Fukagata, K. (2019). Synthetic Turbulent
Dissanayake, M. W. M. G., and Phan-Thien, N. (1994). Neural-network-based Inflow Generator Using Machine Learning. Phys. Rev. Fluids 4 (6), 064603.
Approximations for Solving Partial Differential Equations. Commun. Numer. doi:10.1103/physrevfluids.4.064603
Meth. Engng. 10 (3), 195–201. doi:10.1002/cnm.1640100303 Gairola, A., and Bitsuamlak, G. (2019). Numerical Tornado Modeling for Common
Duraisamy, K., Iaccarino, G., and Xiao, H. (2019). Turbulence Modeling in the Age Interpretation of Experimental Simulators. J. Wind Eng. Ind. Aerodynamics
of Data. Annu. Rev. Fluid Mech. 51, 357–377. doi:10.1146/annurev-fluid- 186, 32–48. doi:10.1016/j.jweia.2018.12.013
010518-040547 Gao, D.-L., Chen, W.-L., Li, H., and Hu, H. (2017). Flow Around a Circular
Eguchi, Y., Hattori, Y., Nakao, K., James, D., and Zuo, D. (2018). Numerical cylinder with Slit. Exp. Therm. Fluid Sci. 82, 287–301. doi:10.1016/j.
Pressure Retrieval from Velocity Measurement of a Turbulent Tornado-like expthermflusci.2016.11.025
Vortex. J. Wind Eng. Ind. Aerodynamics 174, 61–68. doi:10.1016/j.jweia.2017. Ghaboussi, J., and Joghataie, A. (1995). Active Control of Structures Using Neural
12.021 Networks. J. Eng. Mech. 121 (4), 555–567. doi:10.1061/(asce)0733-9399(1995)
Elshaer, A., Bitsuamlak, G., and El Damatty, A. (2017). Enhancing Wind 121:4(555)
Performance of Tall Buildings Using Corner Aerodynamic Optimization. Gholizadeh, S., Salajegheh, J., and Salajegheh, E. (2009). An Intelligent
Eng. Structures 136, 133–148. doi:10.1016/j.engstruct.2017.01.019 Neural System for Predicting Structural Response Subject to
Elshaer, A., Bitsuamlak, G., and El Damatty, A. “June. Aerodynamic Shape Earthquakes. Adv. Eng. Softw. 40 (8), 630–639. doi:10.1016/j.
Optimization of Tall Buildings Using Twisting and Corner Modifications,” advengsoft.2008.11.008

Frontiers in Built Environment | www.frontiersin.org 34 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Giffard-Roisin, S., Yang, M., Charpiat, G., Kumler Bonfanti, C., Kégl, B., and Engineering, Copenhagen, Denmark, 21-24 June 1999. Editor A. Larsen
Monteleoni, C. (2020). Tropical Cyclone Track Forecasting Using Fused Deep (Boca Raton, Florida, United States: CRC Press), 1409–1415.
Learning from Aligned Reanalysis Data. Front. Big Data 3, 1–13. doi:10.3389/ Holton, J. R., and Hakim, G. J. (2013). An Introduction to Dynamic Meteorology.
fdata.2020.00001 Fifth edition. Amsterdam: Academic Press.
Gillmeier, S., Sterling, M., and Hemida, H. (2019). Simulating Tornado-like Flows: Hopfield, J. J. (1982). Neural Networks and Physical Systems with Emergent
the Effect of the Simulator’s Geometry. Meccanica 54 (15), 2385–2398. doi:10. Collective Computational Abilities. Proc. Natl. Acad. Sci. 79 (8), 2554–2558.
1007/s11012-019-01082-4 doi:10.1073/pnas.79.8.2554
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning, 1. Cambridge: Hornik, K. (1991). Approximation Capabilities of Multilayer Feedforward
MIT press, 2. Networks. Neural networks 4 (2), 251–257. doi:10.1016/0893-6080(91)
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., 90009-t
et al. (2014). “Generative Adversarial Nets,” in Advances in Neural Information Hoshino, N., Iida, Y., and Uematsu, Y. (2018). Effects of Non-stationarity of
Processing Systems. Editors M. I. Jordan, Y. LeCun, and S. A. Solla (Cambridge, Downburst on the Wind Loading of Buildings. J. Wind Eng. 43 (1), 1–13. doi:10.
Massachusetts, United States: MIT Press), 2672–2680. 5359/jwe.43.1
Gray, W. M. (1968). Global View of the Origin of Tropical Disturbances and Hou, F., and Sarkar, P. P. (2020). Aeroelastic Model Tests to Study Tall Building
Storms. Mon. Wea. Rev. 96 (10), 669–700. doi:10.1175/1520-0493(1968) Vibration in Boundary-Layer and Tornado Winds. Eng. Structures 207, 110259.
096<0669:gvotoo>2.0.co;2 doi:10.1016/j.engstruct.2020.110259
Gray, W. M. (1979). “Hurricanes: Their Formation, Structure and Likely Role in Hu, G., and Kwok, K. C. S. (2020). Predicting Wind Pressures Around Circular
the Tropical Circulation. Meteorology over the Tropical Oceans,” in Cylinders Using Machine Learning Techniques. J. Wind Eng. Ind.
Meteorology over the Tropical Oceans. Editor D. B. Shaw (James Glaisher Aerodynamics 198, 104099. doi:10.1016/j.jweia.2020.104099
House, Grenville, Bracknell: Royal Meteorological Society), 155–218. Hu, G., Liu, L., Tao, D., Song, J., Tse, K. T., and Kwok, K. C. S. (2020). Deep
Haines, M., and Taylor, I. (2018). Numerical Investigation of the Flow Field Learning-Based Investigation of Wind Pressures on Tall Building under
Around Low Rise Buildings Due to a Downburst Event Using Large Eddy Interference Effects. J. Wind Eng. Ind. Aerodynamics 201, 104138. doi:10.
Simulation. J. Wind Eng. Ind. Aerodynamics 172, 12–30. doi:10.1016/j.jweia. 1016/j.jweia.2020.104138
2017.10.028 Huang, D., Shiqing, H., Xuhui, H., and Xue, Z. (2017). Prediction of Wind Loads on
Hall, T. M., and Jewson, S. (2007). Statistical Modelling of North Atlantic Tropical High-Rise Building Using a BP Neural Network Combined with POD. J. Wind
Cyclone Tracks. Tellus A: Dynamic Meteorology and Oceanography 59 (4), Eng. Ind. Aerodynamics 170, 1–17. doi:10.1016/j.jweia.2017.07.021
486–498. doi:10.1111/j.1600-0870.2007.00240.x Huang, G., He, H., Mehta, K. C., and Liu, X. (2015). Data-based Probabilistic
Hangan, H., Refan, M., Jubayer, C., Romanic, D., Parvu, D., LoTufo, J., et al. (2017). Damage Estimation for Asphalt Shingle Roofing. J. Struct. Eng. 141 (12),
Novel Techniques in Wind Engineering. J. Wind Eng. Ind. Aerodynamics 171, 04015065. doi:10.1061/(asce)st.1943-541x.0001300
12–33. doi:10.1016/j.jweia.2017.09.010 Huang, W. F., and Xu, Y. L. (2013). Prediction of Typhoon Design Wind Speed and
Hao, J., and Wu, T. (2018). Downburst-induced Transient Response of a Long- Profile over Complex Terrain. Struct. Eng. Mech. 45 (1), 1–18. doi:10.12989/
Span Bridge: A CFD-CSD-Based Hybrid Approach. J. Wind Eng. Ind. sem.2013.45.1.001
Aerodynamics 179, 273–286. doi:10.1016/j.jweia.2018.06.006 Huang, W. Y., and Lippmann, R. P. (1988). “Neural Net and Traditional
Hao, J., and Wu, T. (2017). Nonsynoptic Wind-Induced Transient Effects on Classifiers,” in Neural Information Processing Systems. Editor
Linear Bridge Aerodynamics. J. Eng. Mech. 143 (9), 04017092. doi:10.1061/ D. Z. Anderson (Berlin/Heidelberg, Germany: Springer Science & Business
(asce)em.1943-7889.0001313 Media), 387–396.
Hao, J., and Wu, T. (2020). Numerical Analysis of a Long-Span Bridge Response to Huo, S., Hemida, H., and Sterling, M. (2020). Numerical Study of Debris Flight in a
Tornado-like Winds. Wind and Structures 31 (5), 459–472. Tornado-like Vortex. J. Fluids Structures 99, 103134. doi:10.1016/j.jfluidstructs.
Hao, J., and Wu, T. “Tornado-induced Effects on Aerostatic and Aeroelastic 2020.103134
Behaviors of Long-Span Bridge,” in Proceeding of the 2016 World Congress on Iida, Y., and Uematsu, Y. (2019). Numerical Study of Wind Loads on Buildings
Advances in Civil Environmental & Materials Research, Jeju, Korea, September Induced by Downbursts. J. Wind Eng. Ind. Aerodynamics 191, 103–116. doi:10.
2016. 1016/j.jweia.2019.05.018
Hasegawa, K., Fukami, K., Murata, T., and Fukagata, K. “Data-driven Reduced Imran, M., Elbassuoni, S., Castillo, C., Diaz, F., and Meier, P. “Extracting
Order Modeling of Flows Around Two-Dimensional bluff Bodies of Various Information Nuggets from Disaster-Related Messages in Social media,” in
Shapes,” in Fluids Engineering Division Summer Meeting, American Society of Proceedings of the 10th International Conference on Information Systems
Mechanical Engineers, San Francisco, CA, July 28–August 1, 2019, for Crisis Response and Management, Information Systems for Crisis Response
V002T02A075. and Management, Baden-Baden, Germany, May 2013, 791–801.
Hasegawa, K., Fukami, K., Murata, T., and Fukagata, K. (2020). Machine-learning- Imran, M., Mitra, P., and Castillo, C. “Twitter as a Lifeline: Human-Annotated
based Reduced-Order Modeling for Unsteady Flows Around bluff Bodies of Twitter Corpora for NLP of Crisis-Related Messages,” in Proceedings of the
Various Shapes. Theor. Comput. Fluid Dyn. 34, 367–383. doi:10.1007/s00162- Tenth International Conference on Language Resources and Evaluation (LREC
020-00528-w 2016), European Language Resources Association (ELRA), Portorož, Slovenia,
Hawbecker, P. (2021). “Mesoscale, Microscale, and Numerical Models,” in The May 2016.
Oxford Handbook of Non-synoptic Wind Storms. Editors H. Hangan and Ishihara, T., Oh, S., and Tokuyama, Y. (2011). Numerical Study on Flow fields of
A. Kareem (Oxford, United Kingdom: Oxford University Press), 239. Tornado-like Vortices Using the LES Turbulence Model. J. Wind Eng. Ind.
Hayashi, K., and Ohsaki, M. (2020). Reinforcement Learning for Optimum Design Aerodynamics 99 (4), 239–248. doi:10.1016/j.jweia.2011.01.014
of a Plane Frame under Static Loads. Eng. Comput. 37, 1999–2011. doi:10.1007/ Jackson, P. S., and Hunt, J. C. R. (1975). Turbulent Wind Flow over a Low hill.
s00366-019-00926-7 Q.J R. Met. Soc. 101 (430), 929–955. doi:10.1002/qj.49710143015
He, Y. C., Li, Y. Z., Chan, P. W., Fu, J. Y., Wu, J. R., and Li, Q. S. (2019). A Height- Jesson, M., Sterling, M., Letchford, C., and Haines, M. (2015). Aerodynamic Forces
Resolving Model of Tropical Cyclone Pressure Field. J. Wind. Eng. Ind. on Generic Buildings Subject to Transient, Downburst-type Winds. J. wind Eng.
Aerodyn. 186, 84–93. Ind. aerodynamics 137, 58–68. doi:10.1016/j.jweia.2014.12.003
Hinton, G. E., Osindero, S., and Teh, Y.-W. (2006). A Fast Learning Algorithm for Jiang, X., and Adeli, H. (2008). Dynamic Fuzzy Wavelet Neuroemulator for Non-
Deep Belief Nets. Neural Comput. 18 (7), 1527–1554. doi:10.1162/neco.2006.18. linear Control of Irregular Building Structures. Int. J. Numer. Meth. Engng 74
7.1527 (7), 1045–1066. doi:10.1002/nme.2195
Hochreiter, S., and Schmidhuber, J. (1997). Long Short-Term Memory. Neural Jiang, X., and Adeli, H. (2005). Dynamic Wavelet Neural Network for Nonlinear
Comput. 9 (8), 1735–1780. doi:10.1162/neco.1997.9.8.1735 Identification of Highrise Buildings. Comp-aided Civil Eng. 20 (5), 316–330.
Holmes, J. D. (1999). “Modeling of Extreme Thunderstorm Winds for Wind doi:10.1111/j.1467-8667.2005.00399.x
Loading of Structures and Risk Assessment,” in Wind Engineering into the 21st Jubayer, C., Elatar, A., and Hangan, H., “Pressure Distributions on a Low-Rise
Century, Proceedings of the Tenth International Conference on Wind Building in a Laboratory Simulated Downburst.” in Proceedings of the 8th

Frontiers in Built Environment | www.frontiersin.org 35 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

International Colloquium on Bluff Body Aerodynamics and Applications, Kutz, J. N. (2017). Deep Learning in Fluid Dynamics. J. Fluid Mech. 814, 1–4.
Boston, Massachusetts, USA. June 2016 doi:10.1017/jfm.2016.803
Junayed, C., Jubayer, C., Parvu, D., Romanic, D., and Hangan, H. (2019). Flow Field Lagerquist, R. A., Homeyer, C. R., McGovern, A., Potvin, C. K., Sandmael, T., and
Dynamics of Large-Scale Experimentally Produced Downburst Flows. J. Wind Smith, T. M. “Deep Learning for Real-Time Storm-Based Tornado Prediction,”
Eng. Ind. Aerodynamics 188, 61–79. doi:10.1016/j.jweia.2019.02.008 in Proceedings of the 29th Conference on Severe Local Storms, Stowe, VT,
Jung, S., Ghaboussi, J., and Kwon, S.-D. (2004). Estimation of Aeroelastic October 2018 (Boston, Massachusetts, United States: AMS).
Parameters of Bridge Decks Using Neural Networks. J. Eng. Mech. 130 (11), Lagerquist, R., McGovern, A., Homeyer, C. R., Gagne, D. J., and Smith, T. (2020).
1356–1364. doi:10.1061/(asce)0733-9399(2004)130:11(1356) Deep Learning on Three-Dimensional Multiscale Data for Next-Hour Tornado
Kamangir, H., Collins, W., Tissot, P., and King, S. A. (2020). Deep-learning Model Prediction. Monthly Weather Rev. 148, 2837–2861. doi:10.1175/mwr-d-19-
Used to Predict Thunderstorms within 400 Km2 of South Texas Domains. 0372.1
Meteorol. Appl. 27 (2), e1905. doi:10.1002/met.1905 Lagerquist, R., McGovern, A., and Smith, T. (2017). Machine Learning for Real-
Kareem, A. (2020). Emerging Frontiers in Wind Engineering: Computing, Time Prediction of Damaging Straight-Line Convective Wind. Weather
Stochastics, Machine Learning and beyond. J. Wind Eng. Ind. Aerodynamics Forecast. 32 (6), 2175–2193. doi:10.1175/waf-d-17-0038.1
206, 104320. doi:10.1016/j.jweia.2020.104320 Lahouar, A., and Slama, J. B. H. “Wind Speed and Direction Prediction for Wind
Kareem, A., and Wu, T. (2013). Wind-induced Effects on bluff Bodies in Turbulent Farms Using Support Vector Regression.” in Proceedings of the 2014 5th
Flows: Nonstationary, Non-gaussian and Nonlinear Features. J. Wind Eng. Ind. International Renewable Energy Congress (IREC). Hammamet, Tunisia. March
Aerodynamics 122, 21–37. doi:10.1016/j.jweia.2013.06.002 2014 (Piscataway, New Jersey, United States: IEEE), 1–6.
Kawaguchi, M., Tamura, T., and Kawai, H. (2019). Analysis of Tornado and Near- Lakshmanan, V., Stumpf, G., and Witt, A. “A Neural Network for Detecting and
Ground Turbulence Using a Hybrid Meteorological Model/engineering LES Diagnosing Tornadic Circulations Using the Mesocyclone Detection and Near
Method. Int. J. Heat Fluid Flow 80, 108464. doi:10.1016/j.ijheatfluidflow.2019. Storm Environment Algorithms,” in Proceedings of the AI Applications with a
108464 Nowcasting Flavor (Joint between the Fourth Conference on Artificial
Kelley, H. J. (1960). Gradient Theory of Optimal Flight Paths. Ars J. 30 (10), Intelligence and the 21st International Conference on Interactive
947–954. doi:10.2514/8.5282 Information and Processing Systems (IIPS) for Meteorology, Oceanography,
Khalatbarisoltani, A., Soleymani, M., and Khodadadi, M. (2019). Online Control of and Hydrology), San Diego, CA, USA, January 2005 (Boston, Massachusetts,
an Active Seismic System via Reinforcement Learning. Struct. Control. Health United States: AMS).
Monit. 26 (3), e2298. doi:10.1002/stc.2298 Le, V., and Caracoglia, L. (2020). A Neural Network Surrogate Model for the
Khanduri, A. C., Bédard, C., and Stathopoulos, T. (1997). Modelling Wind-Induced Performance Assessment of a Vertical Structure Subjected to Non-stationary,
Interference Effects Using Backpropagation Neural Networks. J. wind Eng. Ind. Tornadic Wind Loads. Comput. Structures 231, 106208. doi:10.1016/j.
aerodynamics 72, 71–79. doi:10.1016/s0167-6105(97)00259-6 compstruc.2020.106208
Khodabandehlou, H., Pekcan, G., Fadali, M. S., and Salem, M. M. A. (2018). Active LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W.,
Neural Predictive Control of Seismically Isolated Structures. Struct. Control. et al. (1989). Backpropagation Applied to Handwritten Zip Code Recognition.
Health Monit. 25 (1), e2061. doi:10.1002/stc.2061 Neural Comput. 1 (4), 541–551. doi:10.1162/neco.1989.1.4.541
Khosravi, A., Koury, R. N. N., Machado, L., and Pabon, J. J. G. (2018b). Letchford, C. W., Mans, C., and Chay, M. T. (2002). Thunderstorms—their
Prediction of Wind Speed and Wind Direction Using Artificial Neural Importance in Wind Engineering (A Case for the Next Generation Wind
Network, Support Vector Regression and Adaptive Neuro-Fuzzy Inference Tunnel). J. Wind Eng. Ind. Aerodynamics 90 (12-15), 1415–1433. doi:10.1016/
System. Sustainable Energ. Tech. Assessments 25, 146–160. doi:10.1016/j.seta. s0167-6105(02)00262-3
2018.01.001 Li, G., and Shi, J. (2010). On Comparing Three Artificial Neural Networks for
Khosravi, A., Machado, L., and Nunes, R. O. (2018a). Time-series Prediction of Wind Speed Forecasting. Appl. Energ. 87 (7), 2313–2320. doi:10.1016/j.
Wind Speed Using Machine Learning Algorithms: A Case Study Osorio Wind apenergy.2009.12.013
Farm, Brazil. Appl. Energ. 224, 550–566. doi:10.1016/j.apenergy.2018.05.043 Li, S., Laima, S., and Li, H. (2018). Data-driven Modeling of Vortex-Induced
Kim, J., and Lee, C. (2020). Deep Unsupervised Learning of Turbulence for Inflow Vibration of a Long-Span Suspension Bridge Using Decision Tree Learning and
Generation at Various Reynolds Numbers. J. Comput. Phys. 406, 109216. doi:10. Support Vector Regression. J. Wind Eng. Ind. Aerodynamics 172, 196–211.
1016/j.jcp.2019.109216 doi:10.1016/j.jweia.2017.10.022
Kim, M., Park, M.-S., Im, J., Park, S., and Lee, M.-I. (2019). Machine Learning Li, S., Snaiki, R., and Wu, T. (2021a). A Knowledge-enhanced Deep Reinforcement
Approaches for Detecting Tropical Cyclone Formation Using Satellite Data. Learning-based Shape Optimizer for Aerodynamic Mitigation of Wind-
Remote Sensing 11 (10), 1195. doi:10.3390/rs11101195 sensitive Structures. Computer-Aided Civil Infrastructure Eng. 36 (6),
Kim, S., Kim, H., Lee, J., Yoon, S., Kahou, S. E., Kashinath, K., et al. “Deep- 733–746. doi:10.1111/mice.12655
hurricane-tracker: Tracking and Forecasting Extreme Climate Events,” in Li, S., Snaiki, R., and Wu, T. (2021b). Active Simulation of Transient Wind Field in
Proceedings of the 2019 IEEE Winter Conference on Applications of a Multiple-Fan Wind Tunnel via Deep Reinforcement Learning. J. Eng. Mech.
Computer Vision (WACV), Waikoloa, HI, USA, January 2019 (Piscataway, 147 (9), 04021056. doi:10.1061/(asce)em.1943-7889.0001967
New Jersey, United States: IEEE), 1761–1769. Li, T., Wu, T., and Liu, Z. (2020). Nonlinear Unsteady Bridge Aerodynamics:
Kolmogorov, A. N. (1941). The Local Structure of Turbulence in Incompressible Reduced-Order Modeling Based on Deep LSTM Networks. J. Wind Eng. Ind.
Viscous Fluid for Very Large Reynolds Numbers. Cr Acad. Sci. URSS 30, Aerodynamics 198, 104116. doi:10.1016/j.jweia.2020.104116
301–305. Li, Z., and Li, C. (2018). “Selection of Kernel Function for Least Squares Support
Križan, J., Gašparac, G., Kozmar, H., Antonić, O., and Grisogono, B. (2015). Vector Machines in Downburst Wind Speed Forecasting,” in Proceedings of the
Designing Laboratory Wind Simulations Using Artificial Neural Networks. 2018 11th International Symposium on Computational Intelligence and Design
Theor. Appl. climatology 120 (3-4), 723–736. doi:10.1007/s00704-014- (ISCID), Hangzhou, China (Piscataway, New Jersey, United States: IEEE),
1201-4 337–341. doi:10.1109/iscid.2018.10178
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet Classification with Liang, X. (2019). Image-based post-disaster Inspection of Reinforced concrete
Deep Convolutional Neural Networks. Adv. Neural Inf. Process. Syst. 60, 84–90. Bridge Systems Using Deep Learning with Bayesian Optimization. Computer-
doi:10.1145/3065386 Aided Civil Infrastructure Eng. 34 (5), 415–430. doi:10.1111/mice.12425
Kuai, L., Haan, F. L. J., Jr., Gallus, W. A. J., Jr., and Sarkar, P. P. (2008). CFD Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., et al. (2015).
Simulations of the Flow Field of a Laboratory-Simulated Tornado for Parameter Continuous Control with Deep Reinforcement Learning. arXiv. arXiv:
Sensitivity Studies and Comparison with Field Measurements. Wind and 1509.02971.
Structures 11 (2), 75–96. doi:10.12989/was.2008.11.2.075 Lin, T. Y., Goyal, P., Girshick, R., He, K., and Dollár, P. “Focal Loss for Dense
Kumar, G., and Malik, H. (2016). Generalized Regression Neural Network Based Object Detection,” in Proceedings of the IEEE international conference on
Wind Speed Prediction Model for Western Region of India. Proced. Comput. computer vision, Venice, Italy, October 2017, 2980–2988. doi:10.1109/iccv.
Sci. 93, 26–32. doi:10.1016/j.procs.2016.07.177 2017.324

Frontiers in Built Environment | www.frontiersin.org 36 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Litta, A. J., Idicula, S. M., and Francis, C. N. (2012). Artificial Neural Network McCulloch, W. S., and Pitts, W. (1943). A Logical Calculus of the Ideas Immanent
Model for the Prediction of Thunderstorms over kolkata. Int. J. Comput. Appl. in Nervous Activity. Bull. Math. Biophys. 5 (4), 115–133. doi:10.1007/
50 (11), 50–55. doi:10.5120/7819-1135 bf02478259
Liu, Z., Cao, Y., Cao, J., Wang, Y., and Cao, S. (2021). Numerical Study of Medina, B., Carey, L., Amiot, C., Mecikalski, R., Roeder, W., McNamara, T., et al.
Tornado-Borne Debris on a Low-Rise Building through Large Eddy (2019). A Random forest Method to Forecast Downbursts Based on Dual-
Simulation. J. Fluids Structures 106, 103379. doi:10.1016/j.jfluidstructs. Polarization Radar Signatures. Remote Sensing 11 (7), 826. doi:10.3390/
2021.103379 rs11070826
Liu, Z., and Ishihara, T. (2015). Numerical Study of Turbulent Flow fields and the Medsker, L., and Jain, L. C. (1999). Recurrent Neural Networks: Design and
Similarity of Tornado Vortices Using Large-Eddy Simulations. J. Wind Eng. Applications. Boca Raton, Florida, United States: CRC Press.
Ind. Aerodynamics 145, 42–60. doi:10.1016/j.jweia.2015.05.008 Michael, B. (20172017). Tropical Cyclone Genesis Forecasting and Pre-genesis
López, P., Velo, R., and Maseda, F. (2008). Effect of Direction on Wind Speed Forecasts Report. Miami, FL, USA: National Hurricane Center.
Estimation in Complex Terrain Using Neural Networks. Renew. Energ. 33 (10), Micheli, L., Hong, J., Laflamme, S., and Alipour, A. (2020). Surrogate Models for
2266–2272. doi:10.1016/j.renene.2007.12.020 High Performance Control Systems in Wind-Excited Tall Buildings. Appl. Soft
Lute, V., Upadhyay, A., and Singh, K. K. (2009). Support Vector Machine Based Comput. 90, 106133. doi:10.1016/j.asoc.2020.106133
Aerodynamic Analysis of cable Stayed Bridges. Adv. Eng. Softw. 40 (9), Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., et al.
830–835. doi:10.1016/j.advengsoft.2009.01.008 (2015). Human-level Control through Deep Reinforcement Learning. nature
Mandic, D., and Chambers, J. (2001). Recurrent Neural Networks for Prediction: 518 (7540), 529–533. doi:10.1038/nature14236
Learning Algorithms, Architectures and Stability. Hoboken, New Jersey, Mohandes, M. A., Halawani, T. O., Rehman, S., and Hussain, A. A. (2004). Support
United States: John Wiley & Sons. Vector Machines for Wind Speed Prediction. Renew. Energ. 29 (6), 939–947.
Manna, S., and Nakai, H. “Effectiveness of Word Embeddings on Classifiers: A doi:10.1016/j.renene.2003.11.009
Case Study with Tweets,” in Proceedings of the 2019 IEEE 13th Mohri, M., Rostamizadeh, A., and Talwalkar, A. (2018). Foundations of Machine
International Conference on Semantic Computing (ICSC), Newport Learning. Cambridge, Massachusetts, United States: MIT press.
Beach, CA, USA, January 2019 (Piscataway, New Jersey, United States: Moradi Kordmahalleh, M., Gorji Sefidmazgi, M., and Homaifar, A. (2016). “A
IEEE), 158–161. Sparse Recurrent Neural Network for Trajectory Prediction of atlantic
Manohar, K., Brunton, B. W., Kutz, J. N., and Brunton, S. L. (2018). Data-driven Hurricanes,” in Proceedings of the Genetic and Evolutionary Computation
Sparse Sensor Placement for Reconstruction: Demonstrating the Benefits of Conference 2016. Editor F. Neumann (New York, United States: Association
Exploiting Known Patterns. IEEE Control. Syst. Mag. 38 (3), 63–86. doi:10. for Computing Machinery), 957–964.
1109/MCS.2018.2810460 More, A., and Deo, M. C. (2003). Forecasting Wind with Neural Networks. Mar.
Martínez-Vázquez, P., and Rodríguez-Cuevas, N. (2007). Wind Field structures 16 (1), 35–49. doi:10.1016/s0951-8339(02)00053-9
Reproduction Using Neural Networks and Conditional Simulation. Eng. Murphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. Cambridge,
structures 29 (7), 1442–1449. doi:10.1016/j.engstruct.2006.08.024 Massachusetts, United States: MIT press.
Marzban, C. (2000). A Neural Network for Tornado Diagnosis: Managing Local Nikose, T. J., and Sonparote, R. S. (2020). Computing Dynamic Across-Wind
Minima. Neural Comput. Appl. 9, 133–141. doi:10.1007/s005210070024 Response of Tall Buildings Using Artificial Neural Network. J. Supercomput 76
Marzban, C., Paik, H., and Stumpf, G. J. (1997). Neural Networks vs. Gaussian (5), 3788–3813. doi:10.1007/s11227-018-2708-8
Discriminant Analysis. AI Appl. 11 (1), 49–58. Nikose, T. J., and Sonparote, R. S. (2019a). Dynamic along Wind Response of Tall
Marzban, C., and Stumpf, G. J. (1998). A Neural Network for Damaging Wind Buildings Using Artificial Neural Network. Cluster Comput. 22 (2), 3231–3246.
Prediction. Wea. Forecast. 13, 151–163. doi:10.1175/1520-0434(1998) doi:10.1007/s10586-018-2027-0
013<0151:annfdw>2.0.co;2 Nikose, T. J., and Sonparote, R. S. (2019b). Dynamic Wind Response of Tall
Marzban, C., and Stumpf, G. J. (1996). A Neural Network for Tornado Prediction Buildings Using Artificial Neural Network. The Struct. Des. Tall Spec. Buildings
Based on Doppler Radar-Derived Attributes. J. Appl. Meteorol. 35, 617–626. 28 (13), e1657. doi:10.1002/tal.1657
doi:10.1175/1520-0450(1996)035<0617:annftp>2.0.co;2 Oh, B. K., Glisic, B., Kim, Y., and Park, H. S. (2019). Convolutional Neural
Maskey, M., Ramachandran, R., Ramasubramanian, M., Gurung, I., Freitag, Network-Based Wind Induced Response Estimation Model for Tall
B., Kaulfus, A., et al. (2020). Deepti: Deep-Learning-Based Tropical Buildings. Comput. Aided Civ. Infrastruct. Eng. 34, 843–858. doi:10.
Cyclone Intensity Estimation System. IEEE J. Sel. Top. Appl. Earth 1111/mice.12476
Observations Remote Sensing 13, 4271–4281. doi:10.1109/jstars.2020. Oh, B. K., Kim, K. J., Kim, Y., Park, H. S., and Adeli, H. (2017). Evolutionary
3011907 Learning Based Sustainable Strain Sensing Model for Structural Health
Mason, M. S., Wood, G. S., and Fletcher, D. F. (2009). Numerical Simulation of Monitoring of High-Rise Buildings. Appl. Soft Comput. 58, 576–585. doi:10.
Downburst Winds. J. Wind Eng. Ind. Aerodynamics 97 (11-12), 523–539. 1016/j.asoc.2017.05.029
doi:10.1016/j.jweia.2009.07.010 O’Neal, A., Rodgers, B., Segler, J., Murthy, D., Lakuduva, N., Johnson, M., et al.
Masri, S. F., Chassiakos, A. G., and Caughey, T. K. (1993). Identification of “Training an Emergency-Response Image Classifier on Signal Data,” in
Nonlinear Dynamic Systems Using Neural Networks. J. Appl. Mech. 60 (1), Proceedings of the 2018 17th IEEE International Conference on Machine
123–133. doi:10.1115/1.2900734 Learning and Applications (ICMLA), Orlando, FL, USA, December 2018
Matsumoto, M., Kobayashi, Y., and Shirato, H. (1996). The Influence of (Piscataway, New Jersey, United States: IEEE), 751–756.
Aerodynamic Derivatives on Flutter. J. Wind Eng. Ind. Aerodynamics 60, Oreskovic, C., Orf, L. G., and Savory, E. (2018). A Parametric Study of Downbursts
227–239. doi:10.1016/0167-6105(96)00036-0 Using a Full-Scale Cooling Source Model. J. Wind Eng. Ind. Aerodynamics 180,
Maulik, U., and Bandyopadhyay, S. (2002). Performance Evaluation of Some 168–181. doi:10.1016/j.jweia.2018.07.020
Clustering Algorithms and Validity Indices. IEEE Trans. Pattern Anal. Machine Oreskovic, C., and Savory, E. (2018). Evolution and Scaling of a Simulated
Intell. 24 (12), 1650–1654. doi:10.1109/tpami.2002.1114856 Downburst-Producing Thunderstorm Outflow. Wind and Structures 26 (3),
Mayo, M., Wakes, S., and Anderson, C. “Neural Networks for Predicting the 147–161. doi:10.12989/was.2018.26.3.147
Output of Wind Flow Simulations over Complex Topographies,” in Pan, B., Xu, X., and Shi, Z. (2019). Tropical Cyclone Intensity Prediction Based on
Proceedings of the 2018 IEEE International Conference on Big Knowledge Recurrent Neural Networks. Electron. Lett. 55 (7), 413–415. doi:10.1049/el.2018.8178
(ICBK), Singapore, November 2018 (Piscataway, New Jersey, United States: Panofsky, H. A., and McCormick, R. A. (1960). The Spectrum of Vertical Velocity
IEEE), 184–191. Near the Surface. Q.J R. Met. Soc. 86 (370), 495–503. doi:10.1002/qj.
McCann, D. W. (1992). A Neural Network Short-Term Forecast of Significant 49708637006
Thunderstorms. Wea. Forecast. 7 (3), 525–534. doi:10.1175/1520-0434(1992) Park, M.-S., Kim, M., Lee, M.-I., Im, J., and Park, S. (2016). Detection of Tropical
007<0525:annstf>2.0.co;2 Cyclone Genesis via Quantitative Satellite Ocean Surface Wind Pattern and
McCarthy, J. (2007). From Here to Human-Level AI. Artif. Intelligence 171 (18), Intensity Analyses Using Decision Trees. Remote sensing Environ. 183,
1174–1182. doi:10.1016/j.artint.2007.10.009 205–214. doi:10.1016/j.rse.2016.06.006

Frontiers in Built Environment | www.frontiersin.org 37 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Pei, J. S., Wright, J. P., and Smyth, A. W. (2005). Mapping Polynomial Fitting into Rüttgers, M., Lee, S., Jeon, S., and You, D. (2019). Prediction of a Typhoon Track
Feedforward Neural Networks for Modeling Nonlinear Dynamic Systems and Using a Generative Adversarial Network and Satellite Images. Sci. Rep. 9 (1),
beyond. Comput. Methods Appl. Mech. Eng. 194 (42-44), 4481–4505. doi:10. 1–15. doi:10.1038/s41598-019-42339-y
1016/j.cma.2004.12.010 Salehi, H., and Burgueño, R. (2018). Emerging Artificial Intelligence Methods in
Pi, Y., Nath, N. D., and Behzadan, A. H. (2020). Convolutional Neural Networks Structural Engineering. Eng. structures 171, 170–189. doi:10.1016/j.engstruct.
for Object Detection in Aerial Imagery for Disaster Response and Recovery. 2018.05.084
Adv. Eng. Inform. 43, 101009. doi:10.1016/j.aei.2019.101009 Santosa, B. (2007). Feature Selection with Support Vector Machines Applied on
Potter, C. W., and Negnevitsky, M. (2006). Very Short-Term Wind Forecasting for Tornado Detection. IPTEK J. Technol. Sci. 18 (1). doi:10.12962/j20882033.
Tasmanian Power Generation. IEEE Trans. Power Syst. 21 (2), 965–972. doi:10. v18i1.178
1109/tpwrs.2006.873421 Sarkar, P. P., Haan, F. L., Jr., Balaramudu, V., and Sengupta, A., “Laboratory
Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M. P., et al. (2018). A Simulation of Tornado and Microburst to Assess Wind Loads on Buildings.” in
Survey on Deep Learning: Algorithms, Techniques, and Applications. ACM Proceedings of the Structures Congress 2006: Structural Engineering and Public
Comput. Surv. (Csur) 51 (5), 1–36. doi:10.1145/3234150 Safety. St. Louis, Missouri, United States. May 2006. 1–10. doi:10.1061/
Psichogios, D. C., and Ungar, L. H. (1992). A Hybrid Neural Network-First 40889(201)11
Principles Approach to Process Modeling. Aiche J. 38 (10), 1499–1511. Saunders, J. W., and Melbourne, W. H. (1980). “Buffeting Effects of Upstream
doi:10.1002/aic.690381003 Buildings,” in Wind Engineering. Editor J. W. Saunders (Pergamon: Pergamon
Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2017a). Physics Informed Deep Press), 593–606. doi:10.1016/b978-1-4832-8367-8.50059-0
Learning (Part I): Data-Driven Solutions of Nonlinear Partial Differential Scholkopf, B., and Smola, A. J. (2018). Learning with Kernels: Support Vector
Equations. arXiv preprint. arXiv:1711.10561. Machines, Regularization, Optimization, and beyond. Adaptive Computation
Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2017b). Physics Informed Deep and Machine Learning Series. Cambridge, Massachusetts, United States: MIT
Learning (Part II): Data-Driven Discovery of Nonlinear Partial Differential Press.
Equations. arXiv 1711, 10566. Sfetsos, A. (2000). A Comparison of Various Forecasting Techniques Applied to
Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2019). Physics-informed Neural Mean Hourly Wind Speed Time Series. Renew. Energ. 21 (1), 23–35. doi:10.
Networks: A Deep Learning Framework for Solving Forward and Inverse 1016/s0960-1481(99)00125-1
Problems Involving Nonlinear Partial Differential Equations. J. Comput. Sharma, R., Shikhola, T., and Kohli, J. K. (2020). Modified Fuzzy Q-Learning Based
Phys. 378, 686–707. doi:10.1016/j.jcp.2018.10.045 Wind Speed Prediction. J. Wind Eng. Ind. Aerodynamics 206, 104361. doi:10.
Rathje, E. M., Dawson, C., Padgett, J. E., Pinelli, J.-P., Stanzione, D., Adair, A., 1016/j.jweia.2020.104361
et al. (2017). DesignSafe: New Cyberinfrastructure for Natural Hazards Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., et al.
Engineering. Nat. Hazards Rev. 18 (3), 06017001. doi:10.1061/(asce)nh. (2017). Mastering the Game of Go without Human Knowledge. nature 550
1527-6996.0000246 (7676), 354–359. doi:10.1038/nature24270
Razavi, A., and Sarkar, P. P. (2021). Effects of Roof Geometry on Tornado-Induced Simiu, E., and Scanlan, R. H. (1978). Wind Effects on Structures. Hoboken, New
Structural Actions of a Low-Rise Building. Eng. structures 226, 111367. doi:10. Jersey, United States: Wiley.
1016/j.engstruct.2020.111367 Smith, T. M., Elmore, K. L., and Dulin, S. A. (2004). A Damaging Downburst
Razavi, A., and Sarkar, P. P. (2018). Laboratory Study of Topographic Effects on the Prediction and Detection Algorithm for the WSR-88D. Wea. Forecast. 19 (2),
Near-Surface Tornado Flow Field. Boundary-layer Meteorol. 168 (2), 189–212. 240–250. doi:10.1175/1520-0434(2004)019<0240:addpad>2.0.co;2
doi:10.1007/s10546-018-0347-5 Snaiki, R., and Wu, T. (2017b). A Linear Height-Resolving Wind Field Model for
Refan, M., and Hangan, H. (2016). Characterization of Tornado-like Flow fields in Tropical Cyclone Boundary Layer. J. Wind Eng. Ind. Aerodynamics 171,
a New Model Scale Wind Testing Chamber. J. Wind Eng. Ind. Aerodynamics 248–260. doi:10.1016/j.jweia.2017.10.008
151, 107–121. doi:10.1016/j.jweia.2016.02.002 Snaiki, R., and Wu, T. (2018). A Semi-empirical Model for Mean Wind Velocity
Richman, M. B., and Leslie, L. M. (2012). Adaptive Machine Learning Approaches Profile of Landfalling hurricane Boundary Layers. J. Wind Eng. Ind.
to Seasonal Prediction of Tropical Cyclones. Proced. Comput. Sci. 12, 276–281. Aerodynamics 180, 249–261. doi:10.1016/j.jweia.2018.08.004
doi:10.1016/j.procs.2012.09.069 Snaiki, R., and Wu, T. (2020c). An Analytical Model for Rapid Estimation of
Richman, M. B., Leslie, L. M., Ramsay, H. A., and Klotzbach, P. J. (2017). Reducing hurricane Supergradient Winds. J. Wind Eng. Ind. Aerodynamics 201, 104175.
Tropical Cyclone Prediction Errors Using Machine Learning Approaches. doi:10.1016/j.jweia.2020.104175
Proced. Comput. Sci. 114, 314–323. doi:10.1016/j.procs.2017.09.048 Snaiki, R., and Wu, T. (2020b). Hurricane hazard Assessment along the
Riedmiller, M., and Braun, H. “A Direct Adaptive Method for Faster United States Northeastern Coast: Surface Wind and Rain fields under
Backpropagation Learning: The RPROP Algorithm,” in Proceedings of Changing Climate. Front. Built Environ. 6, 573054. doi:10.3389/fbuil.2020.
the IEEE international conference on neural networks, San Francisco, 573054
CA, USA, March 1993 (Piscataway, New Jersey, United States: IEEE), Snaiki, R., and Wu, T. (2019). Knowledge-enhanced Deep Learning for Simulation
586–591. of Tropical Cyclone Boundary-Layer Winds. J. Wind Eng. Ind. Aerodynamics
Rizzo, F., and Caracoglia, L. (2020). Artificial Neural Network Model to Predict the 194, 103983. doi:10.1016/j.jweia.2019.103983
Flutter Velocity of Suspension Bridges. Comput. Structures 233, 106236. doi:10. Snaiki, R., and Wu, T. (2017a). Modeling Tropical Cyclone Boundary Layer:
1016/j.compstruc.2020.106236 Height-Resolving Pressure and Wind fields. J. Wind Eng. Ind. Aerodynamics
Robertson, B. W., Johnson, M., Murthy, D., Smith, W. R., and Stephens, K. K. 170, 18–27. doi:10.1016/j.jweia.2017.08.005
(2019). Using a Combination of Human Insights and ‘deep Learning’ for Real- Snaiki, R., and Wu, T. (2020a). Revisiting hurricane Track Model for Wind Risk
Time Disaster Communication. Prog. Disaster Sci. 2, 100030. doi:10.1016/j. Assessment. Struct. Saf. 87, 102003. doi:10.1016/j.strusafe.2020.102003
pdisas.2019.100030 Solari, G., Burlando, M., De Gaetano, P., and Repetto, M. P. (2015). Characteristics
Romanic, D., LoTufo, J., and Hangan, H. (2019). Transient Behavior in Impinging of Thunderstorms Relevant to the Wind Loading of Structures. Wind and
Jets in Crossflow with Application to Downburst Flows. J. Wind Eng. Ind. Structures 20 (6), 763–791. doi:10.12989/was.2015.20.6.763
Aerodynamics 184, 209–227. doi:10.1016/j.jweia.2018.11.020 Solari, G. (2020). Thunderstorm Downbursts and Wind Loading of Structures:
Rosenblatt, F. (1957). The Perceptron, a Perceiving and Recognizing Automaton Progress and prospect. Front. Built Environ. 6, 63. doi:10.3389/fbuil.2020.00063
Project Para. Buffalo, New York, United States: Cornell Aeronautical Stiles, B. W., Danielson, R. E., Poulsen, W. L., Brennan, M. J., Hristova-Veleva, S.,
Laboratory. Tsae-Pyng Shen, T. P., et al. (2014). Optimized Tropical Cyclone Winds from
Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning QuikSCAT: A Neural Network Approach. IEEE Trans. Geosci. Remote Sensing
Representations by Back-Propagating Errors. nature 323 (6088), 533–536. 52 (11), 7418–7434. doi:10.1109/tgrs.2014.2312333
doi:10.1038/323533a0 Subasri, R., Suresh, S., and Natarajan, A. M. (2014). Discrete Direct Adaptive ELM
Russell, S., and Norvig, P. (2016). Artificial Intelligence: A Modern Approach. Controller for Active Vibration Control of Nonlinear Base Isolation Buildings.
London, United Kingdom: Pearson Education Limited. Neurocomputing 129, 246–256. doi:10.1016/j.neucom.2013.09.035

Frontiers in Built Environment | www.frontiersin.org 38 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Subramanian, D., Salazar, J., Duenas-Osorio, L., and Stein, R. “Constructing and Wang, H., Zhang, Y.-M., Mao, J.-X., and Wan, H.-P. (2020). A Probabilistic
Validating Geographically Refined HAZUS-MH4 hurricane Wind Risk Models: Approach for Short-Term Prediction of Wind Gust Speed Using Ensemble
A Machine Learning Approach,” in Proceedings of the Advances in hurricane Learning. J. Wind Eng. Ind. Aerodynamics 202, 104198. doi:10.1016/j.jweia.
engineering: Learning from our past, Miami, Florida, United States, October 2020.104198
2013, 1056–1066. doi:10.1061/9780784412626.092 Wang, J., and Cheng, C. M. “Aero-Data Based Wind Resistant Design of
Sun, W., Bocchini, P., and Davison, B. D. (2020). Applications of Artificial Rectangular Shaped Tall Buildings,” in Proceedings of the International
Intelligence for Disaster Management. Nat. Hazards 103, 2631–2689. doi:10. Conference on Innovations in Civil and Structural Engineering (ICICSE’15),
1007/s11069-020-04124-3 Istanbul, Turkey, June 2015, 148–154.
Sun, Y., Yang, L., and Wu, Y. “Wind Load Prediction of Large-Span Dry Coal Sheds Wang, J., Cheng, C. M., and Chen, C. H. “The Study of Wind Force Coefficient
Based on GRNN and its Application,” in Proceedings of International Predictions for Rectangular High-Rise Buildings,” in Proceedings of The eighth
Structural Engineering and Construction, Valencia, Spain, July, 2017. doi:10. Asia-Pacific conference on wind engineering, Chennai, India, December 2013,
14455/isec.res.2017.189 10–14.
Sutton, R. S., and Barto, A. G. (2018). Reinforcement Learning: An Introduction. Wang, J., and Cheng, C. M. (2017). Formulation of Estimation Models for Wind
Cambridge, Massachusetts, United States: MIT press. Force Coefficients of Rectangular Shaped Buildings. J. Appl. Sci. Eng. 20 (1),
Tagliaferri, F., Viola, I. M., and Flay, R. G. J. (2015). Wind Direction Forecasting 55–62. doi:10.6180/jase.2017.20.1.07
with Artificial Neural Networks and Support Vector Machines. Ocean Eng. 97, Wang, Y., Zhang, W., and Fu, W. “Back Propogation (BP)-neural Network for
65–73. doi:10.1016/j.oceaneng.2014.12.026 Tropical Cyclone Track Forecast,” in Proceedings of the 2011 19th
Tang, Z., Feng, C., Wu, L., Zuo, D., and James, D. L. (2018). Characteristics of International Conference on Geoinformatics, Shanghai, China, June 2011
Tornado-like Vortices Simulated in a Large-Scale ward-type Simulator. (Piscataway, New Jersey, United States: IEEE), 1–4.
Boundary-layer Meteorol. 166 (2), 327–350. doi:10.1007/s10546-017-0305-7 Watkins, C. J. C. H. (1989). Learning from Delayed Rewards. Cambridge, UK:
Taniike, Y., and Inaoka, H. (1988). “Aeroelastic Behavior of Tall Buildings in King’s College.
Wakes,” in Advances in Wind Engineering. Editors C. Kramer and Watkins, C. J., and Dayan, P. (1992). Q-learning. Machine Learn. 8 (3-4), 279–292.
H. J. Gerhardt (Amsterdam, Netherlands: Elsevier), 317–327. doi:10.1016/ doi:10.1023/a:1022676722315
b978-0-444-87156-5.50043-6 Wei, C.-C. (2015). Forecasting Surface Wind Speeds over Offshore Islands Near
Tian, J., Gurley, K. R., Diaz, M. T., Fernández-Cabán, P. L., Masters, F. J., and Fang, Taiwan during Tropical Cyclones: Comparisons of Data-Driven Algorithms
R. (2020). Low-rise Gable Roof Buildings Pressure Prediction Using Deep and Parametric Wind Representations. J. Geophys. Res. Atmos. 120 (5),
Neural Networks. J. Wind Eng. Ind. Aerodynamics 196, 104026. doi:10.1016/j. 1826–1847. doi:10.1002/2014jd022568
jweia.2019.104026 Wei, C.-C. (2019). Study on Wind Simulations Using Deep Learning Techniques
Tian, W., Huang, W., Yi, L., Wu, L., and Wang, C. (2020). A CNN-Based Hybrid during Typhoons: a Case Study of Northern Taiwan. Atmosphere 10 (11), 684.
Model for Tropical Cyclone Intensity Estimation in Meteorological Industry. doi:10.3390/atmos10110684
IEEE Access 8, 59158–59168. doi:10.1109/access.2020.2982772 Wen, Y.-K., and Chu, S.-L. (1973). Tornado Risks and Design Wind Speed.
Tian, Z., Ren, Y., and Wang, G. (2020). An Application of Backtracking Search J. Struct. Div. 99 (12), 2409–2421. doi:10.1061/jsdeag.0003666
Optimization-Based Least Squares Support Vector Machine for Prediction of Wiederhold, G., McCarthy, J., and Feigenbaum, E. (1990). Arthur Samuel: pioneer
Short-Term Wind Speed. Wind Eng. 44 (3), 266–281. doi:10.1177/ in Machine Learning. Commun. ACM 33 (11), 137–139.
0309524x19849843 Wijnands, J. S., Qian, G., and Kuleshov, Y. (2016). Variable Selection for Tropical
Trafalis, T. B., Adrianto, I., Richman, M. B., and Lakshmivarahan, S. (2014). Cyclogenesis Predictive Modeling. Monthly Weather Rev. 144 (12), 4605–4619.
Machine-learning Classifiers for Imbalanced Tornado Data. Comput. Manag. doi:10.1175/mwr-d-16-0166.1
Sci. 11 (4), 403–418. doi:10.1007/s10287-013-0174-6 Wijnands, J. S., Shelton, K., and Kuleshov, Y. (2014). Improving the Operational
Turkkan, N., and Srivastava, N. K. (1995). Prediction of Wind Load Distribution Methodology of Tropical Cyclone Seasonal Prediction in the Australian and the
for Air-Supported Structures Using Neural Networks. Can. J. Civ. Eng. 22 (3), South Pacific Ocean Regions. Adv. Meteorology 2014, 838746. doi:10.1155/
453–461. doi:10.1139/l95-053 2014/838746
Twisdale, L. A., and Vickery, P. J. (1992). Research on Thunderstorm Wind Design Williams, R. J. (1992). Simple Statistical Gradient-Following Algorithms for
Parameters. J. Wind Eng. Ind. Aerodynamics 41 (1-3), 545–556. doi:10.1016/ Connectionist Reinforcement Learning. Machine Learn. 8 (3-4), 229–256.
0167-6105(92)90461-i doi:10.1007/bf00992696
Uematsu, Y., and Tsuruishi, R. (2008). Wind Load Evaluation System for the Wu, R.-T., and Jahanshahi, M. R. (2019). Deep Convolutional Neural Network for
Design of Roof Cladding of Spherical Domes. J. wind Eng. Ind. aerodynamics 96 Structural Dynamic Response Estimation and System Identification. J. Eng.
(10-11), 2054–2066. doi:10.1016/j.jweia.2008.02.051 Mech. 145 (1), 04018125. doi:10.1061/(asce)em.1943-7889.0001556
Ukkonen, P., Manzato, A., and Mäkelä, A. (2017). Evaluation of Thunderstorm Wu, T., and Kareem, A. (2013). Bridge Aerodynamics and Aeroelasticity: A
Predictors for Finland Using Reanalyses and Neural Networks. J. Appl. Comparison of Modeling Schemes. J. Fluids Structures 43, 347–370. doi:10.
Meteorology Climatology 56 (8), 2335–2352. doi:10.1175/jamc-d-16-0361.1 1016/j.jfluidstructs.2013.09.015
Varshney, K., and Poddar, K. (2012). Prediction of Wind Properties in Urban Wu, T., and Kareem, A. (2011). Modeling Hysteretic Nonlinear Behavior of Bridge
Environments Using Artificial Neural Network. Theor. Appl. Climatology 107 Aerodynamics via Cellular Automata Nested Neural Network. J. Wind Eng. Ind.
(3-4), 579–590. doi:10.1007/s00704-011-0506-9 Aerodynamics 99 (4), 378–388. doi:10.1016/j.jweia.2010.12.011
Vickery, P. J., Skerlj, P. F., and Twisdale, L. A. (2000). Simulation of Hurricane Risk Wu, T., Li, S., and Sivaselvan, M. (2019). Real-time Aerodynamics Hybrid
in the U.S. Using Empirical Track Model. J. Struct. Eng. 126 (10), 1222–1237. Simulation: a Novel Wind-Tunnel Model for Flexible Bridges. J. Eng. Mech.
doi:10.1061/(asce)0733-9445(2000)126:10(1222) 145 (9), 04019061. doi:10.1061/(asce)em.1943-7889.0001649
Vickery, P. J., Wadhera, D., Twisdale, L. A., Jr., and Lavelle, F. M. (2009). U.S. Wu, T. (2013). “Nonlinear bluff-body Aerodynamics,” (Indiana, USA: University
Hurricane Wind Speed Risk and Uncertainty. J. Struct. Eng. 135 (3), 301–320. of Notre Dame). Doctoral dissertation.
doi:10.1061/(asce)0733-9445(2009)135:3(301) Wu, T., and Song, W. (2019). Real-time Aerodynamics Hybrid Simulation: Wind-
Vyavahare, A. Y., Godbole, P. N., and Nikose, T. (2012). Analysis of Tall Building Induced Effects on a Reduced-Scale Building Equipped with Full-Scale
for across Wind Response. Int. J. Civil Struct. Eng. 2 (3), 679–986. Dampers. J. Wind Eng. Ind. Aerodynamics 190, 1–9. doi:10.1016/j.jweia.
Wang, H., and Wu, T. (2021). Fast Hilbert-Wavelet Simulation of 2019.04.005
Nonstationary Wind Field Using Noniterative Simultaneous Matrix Wu, X., Ghaboussi, J., and Garrett, J. H., Jr. (1992). Use of Neural Networks in
Diagonalization. J. Eng. Mech. 147 (3), 04020153. doi:10.1061/(asce)em. Detection of Structural Damage. Comput. Structures 42 (4), 649–659. doi:10.
1943-7889.0001897 1016/0045-7949(92)90132-j
Wang, H., and Wu, T. (2020). Knowledge-Enhanced Deep Learning for Wind- Yakut, O., and Alli, H. (2011). Neural Based Sliding-Mode Control with Moving
Induced Nonlinear Structural Dynamic Analysis. J. Struct. Eng. 146 (11), Sliding Surface for the Seismic Isolation of Structures. J. Vibration Control. 17
04020235. doi:10.1061/(asce)st.1943-541x.0002802 (14), 2103–2116. doi:10.1177/1077546310395964

Frontiers in Built Environment | www.frontiersin.org 39 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

Yang, Q., Gao, R., Bai, F., Li, T., and Tamura, Y. (2018). Damage to Buildings and Machine Learning. Weather Forecast. 34 (4), 1035–1049. doi:10.1175/waf-
Structures Due to Recent Devastating Wind Hazards in East Asia. Nat. Hazards d-18-0201.1
92 (3), 1321–1353. doi:10.1007/s11069-018-3253-8 Zhang, W., Fu, B., Peng, M. S., and Li, T. (2015). Discriminating Developing versus
Yasen, M. Z. Y., Al-Jundi, R. A. S., and Al-Madi, N. S. A. “Optimized ANN-ABC Nondeveloping Tropical Disturbances in the Western North Pacific through
for Thunderstorms Prediction,” in Proceedings of the 2017 International Decision Tree Analysis. Weather Forecast. 30 (2), 446–454. doi:10.1175/waf-d-
Conference on New Trends in Computing Sciences (ICTCS), Amman, 14-00023.1
Jordan, October 2017 (Piscataway, New Jersey, United States: IEEE), 98–103. Zhang, Y., Chandra, R., and Gao, J. “Cyclone Track Prediction with Matrix Neural
Yu, C., Li, Y., Xiang, H., and Zhang, M. (2018). Data Mining-Assisted Short-Term Networks,” in Proceedings of the 2018 International Joint Conference on
Wind Speed Forecasting by Wavelet Packet Decomposition and Elman Neural Neural Networks (IJCNN), Rio de Janeiro, Brazil, July 2018 (Piscataway,
Network. J. Wind Eng. Ind. Aerodynamics 175, 136–143. doi:10.1016/j.jweia. New Jersey, United States: IEEE), 1–8.
2018.01.020 Zhao, M., Held, I. M., and Lin, S.-J. (2012). Some Counterintuitive Dependencies of
Yu, M., Huang, Q., Qin, H., Scheele, C., and Yang, C. (2019). Deep Learning for Tropical Cyclone Frequency on Parameters in a GCM. J. Atmos. Sci. 69 (7),
Real-Time Social media Text Classification for Situation Awareness - Using 2272–2283. doi:10.1175/jas-d-11-0238.1
Hurricanes Sandy, Harvey, and Irma as Case Studies. Int. J. Digital Earth 12 Zhu, J., and Zhang, W. (2018). Probabilistic Fatigue Damage Assessment of Coastal
(11), 1230–1247. doi:10.1080/17538947.2019.1574316 Slender Bridges under Coupled Dynamic Loads. Eng. Structures 166, 274–285.
Yu, Y., Yao, H., and Liu, Y. (2020). Structural Dynamics Simulation Using a Novel doi:10.1016/j.engstruct.2018.03.073
Physics-Guided Machine Learning Method. Eng. Appl. Artif. Intelligence 96,
103947. doi:10.1016/j.engappai.2020.103947 Conflict of Interest: The authors declare that the research was conducted in the
Zambrano, T. G., and Peterka, J. A. (1978). Wind Load Interaction on an Adjacent absence of any commercial or financial relationships that could be construed as a
Building. CER 77, 78–26. potential conflict of interest.
Zhang, A., and Zhang, L. (2004). RBF Neural Networks for the Prediction of
Building Interference Effects. Comput. Structures 82 (27), 2333–2339. doi:10. Publisher’s Note: All claims expressed in this article are solely those of the authors
1016/j.compstruc.2004.05.014 and do not necessarily represent those of their affiliated organizations, or those of
Zhang, C., Dai, L., Ma, L., Qian, J., and Yang, B. (2017). Objective Estimation the publisher, the editors and the reviewers. Any product that may be evaluated in
of Tropical Cyclone Innercore Surface Wind Structure Using Infrared this article, or claim that may be made by its manufacturer, is not guaranteed or
Satellite Images. J. Appl. Remote Sensing 11 (4), 046030. doi:10.1117/1. endorsed by the publisher.
jrs.11.046030
Zhang, S., and Nishijima, K. “Statistics-based Investigation on Typhoon Transition Copyright © 2022 Wu and Snaiki. This is an open-access article distributed under
Modeling,” in Proceedingsof the Seventh International Colloquium on Bluff the terms of the Creative Commons Attribution License (CC BY). The use,
Body Aerodynamics and Application, Shanghai, China, September 2012, distribution or reproduction in other forums is permitted, provided the original
364–373. author(s) and the copyright owner(s) are credited and that the original publication
Zhang, T., Lin, W., Lin, Y., Zhang, M., Yu, H., Cao, K., et al. (2019). Prediction of in this journal is cited, in accordance with accepted academic practice. No use,
Tropical Cyclone Genesis from Mesoscale Convective Systems Using distribution or reproduction is permitted which does not comply with these terms.

Frontiers in Built Environment | www.frontiersin.org 40 March 2022 | Volume 8 | Article 811460

Wu and Snaiki ML Applications to Wind Engineering

APPENDIX A: LIST OF REVIEWED KEDL* knowledge-enhanced deep learning

MACHINE LEARNING ALGORITHMS KE-DRL* knowledge-enhanced deep reinforcement learning
(NOTE: ACRONYMS WITH * REPRESENT KE-LSTM* knowledge enhanced long short-term memory
THOSE REVIEWED IN THIS KM k-means
CONTRIBUTION). KNN* k-nearest neighbors
LAFMN* local activation feedback multilayer network
A2C advantage actor critic
LDA* linear discriminant analysis
AdaBoost* adaptive boosting
LNN* linear neural network
AE* autoencoder
LR* logistic regression
ALEN* adaptive linear element network
LSSVM* least squares support vector machine
ANFIS* adaptive neuro-fuzzy inference system
LSTM* long short-term memory
ANN* artificial neural network
MC* multiple correlation
AWN* adaptive wavelet network
MFQL* modified fuzzy Q-learning
BNB* Bernoulli naive Bayes
MLR* multiple linear regression
BNN* Bayesian neural network
MNB* multinomial naive Bayes
CGAN conditional GAN
MNN* matrix neural network
CNN* convolutional neural network
MSC mean-shift clustering
CNN-AE* convolutional neural network-based autoencoder
NB* naïve Bayes
ConvLSTM* convolutional Long Short-Term Memory
NESN* nonlinear echo state networks
DCGAN deep convolutional GAN
NLN* neural logic network
DDPG* deep deterministic policy gradient
OLR* ordinary linear regression
DDQN double deep Q-network
PCA principal component analysis
DNN* deep neural network
PI* Physics-informed
DQN deep Q-network
PPO proximal policy optimization
DRL* deep reinforcement learning
QDA* quadratic discriminant analysis
DRNN* diagonal recurrent neural networks
QL* Q-learning
DT* decision tree
RBF* radial basis function
ENN* Elman neural network
RBFNN* radial basis function neural network
ERBFN* radial basis function neural network
RF* random forest
ERNN* Elman recurrent neural networks
RL* reinforcement learning
FIS* fuzzy inference system
RNN* recurrent neural networks
FNN* fuzzy neural network
RR* ridge regression
GAN* generative adversarial network
SC spectral clustering
GBRT* gradient boosted regression trees
SD-AE* stacked denoising autoencoder
GBTE* gradient-boosted tree ensembles
SLDA* supervised latent Dirichlet Allocation
GMDH* group method of data handling
SGD* stochastic gradient descent
GNB* gaussian naïve Bayes
SVM* support vector machines
GPR* gaussian process regression
SVR* support vector regression
GRU* gated recurrent unit network
TRPO trust region policy optimization
GRNN* generalized regression neural network
WGAN Wasserstein GAN
ICA independent component analysis
XGBoost* extreme gradient boosting
IIRANN* infinite impulse response artificial neural network
JRNN* Jordan recurrent neural networks

Frontiers in Built Environment | www.frontiersin.org 41 March 2022 | Volume 8 | Article 811460