Applications of Machine Learning to Wind Engineering

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 learning algorithms work and when these schemes succeed or fail. Moreover, critical challenges and prospects of ML applications in wind engineering are identified to facilitate future research efforts.


INTRODUCTION
Wind engineering is an interdisciplinary field to provide rational treatment of interaction between the atmospheric boundary-layer winds and human activities (Cermak 1975). There is a long and significant history for machine learning (ML) applications in several subfields involved in wind engineering, such as fluid mechanics (Brunton et al., 2020), meteorology  and mechanics of structures (Salehi and Burgueño 2018). The application of statistical learning to turbulence modeling in early 1940s (Kolmogorov 1941) and perceptron learning to structural design in late 1980s (Adeli and Yeh 1989) are representative examples. On the other hand, it seems similar passions have not been shared by researchers in the wind engineering community. Actually, ML-based wind engineering is still in its infancy stage and the full-capacity of ML has not been leveraged yet. However, the exceptional performance of ML to extract hidden informative features from data shows great promise for addressing unresolved complexities and issues originated from first principles investigations in the field of wind engineering. In addition, recent advances in performance-/resilience-based wind engineering have placed new demands on wind characterization, aerodynamics modeling and structural analysis that need powerful simulation tools such as ML to overcome the emerging challenges by simultaneously achieving high computational efficiency and accuracy. It is reasonable to expect the revitalization of ML within the wind engineering field that is fueled by 1) rapidly evolving learning algorithms and high-performance computational hardware, 2) unprecedented volume of data generated with improved wind engineering techniques and methodologies, and 3) urgent needs for more accurate and efficient learning and modeling of complex phenomena in wind-related problems.
As a key subfield of artificial intelligence (AI) [that together with natural intelligence plays a role of the computational part of the ability to achieve goals in the world (McCarthy 2007)], ML develops learning algorithms that use inputs from a sample generator and observations from a system to generate an approximation of its outputs (Cherkassky and Mulier 2007). The evolution of learning algorithms started when McCulloch and Pitts (1943) invented the first mathematical model of a neural network. In 1952, Arthur Samuel from IBM introduced the first self-learning computer program to play the game of checkers (Wiederhold et al., 1990). Then, Rosenblatt (1957) designed the first neural network for computers (the perceptron) that set the foundation of deep neural networks (DNNs). Kelley (1960) presented the method of gradients (or method of steepest descent) in his analytical development of flight performance optimization, which was used to develop the basics of a continuous backpropagation model for training feedforward neural networks (Rumelhart et al., 1986). On the other hand, Hopfield (1982) created a feedback neural network that was considered as the first recurrent neural network (RNN). LeCun et al. (1989) combined convolutional neural network (CNN) and backpropagation algorithm to recognize handwritten digits. Watkins (1989) introduced the concept of Q-learning based on Markov process to significantly enhance the practicability and feasibility of reinforcement learning. Later, Cortes and Vapnik (1995) designed a support-vector network considered as a new learning machine for two-group classification problems with high generalization ability. Hochreiter and Schmidhuber (1997) introduced a long short-term memory cell to address the longterm dependency issue in RNN. To overcome the learning difficulty in DNNs, Hinton et al. (2006) derived a fast, greedy algorithm that can learn deep, directed belief networks one layer at a time and hence facilitate the rapid development of deep learning. Recently, Goodfellow et al. (2014) proposed a generative adversarial network consisting of two models (i.e., generative and discriminative models) that compete with each other in a zero-sum game. The sophisticated ML algorithm needs the help of advanced computational hardware [e.g., graphics processing unit (GPU) and tensor processing unit (TPU)] to unlock its full potential (Berggren et al., 2020). For example, the great success of AlexNet (a deep CNN on GPU) is essentially attributed to its ability to leverage GPU for training (Krizhevsky et al., 2012).
Equipped with both sophisticated algorithms and advanced computational hardware, the learning machine (LM) is driven by data. Both the quantity (data rich and comprehensive) and quality of the training/testing data are important to ensure good performance of ML applications. Wind engineering by nature is a data-rich field (e.g., high spatial and temporal resolution), and it is rapidly becoming a data-comprehensive domain due to recent advances of analytical, numerical, experimental and fieldmeasurement methods Hangan et al., 2017). The data of spatiotemporally varying wind flows are extended from synoptic events measured by airport wind observation system with traditional anemometers to nonsynoptic events measured by several field campaigns with advanced doppler radars and Lidars (Light Detection and Ranging) [e.g., Verification of the Origins of Rotation in Tornadoes Experiment (VORTEX) and Radar Observations of Tornadoes and Thunderstorms Experiment (ROTATE) campaigns for tornado events and Severe Convective OUtflow in Thunderstorms (SCOUT) and Wind Ports and Sea (WPS) campaigns for thunderstorm downburst events]. Massive wind data over complex terrain/topography are collected by continuouswave short-range WindScanner systems (e.g., Berg et al., 2013). The low Reynolds-number, straight-line-wind, stationary aerodynamics data generated in conventional boundary-layer wind tunnels are extended to 1) high-Reynolds-number aerodynamics data resulting from recently built large-scale facilities [e.g., windstorm simulation facility at Insurance Institute for Business and Home Safety (IBHS), Wall of Wind (WOW) at Florida International University and Wind Engineering Energy and Environment (WindEEE) at Western University], 2) vortex-flow aerodynamics data produced by tornado simulators (e.g., tornado-like vortex simulator at Iowa State University and VorTECH at Texas Tech University), and 3) transient aerodynamics data generated in emerging actively controlled wind tunnels (e.g., individually-controlled multi-fan wind tunnels at Tongji University, University at Buffalo and University of Florida). Also, significant nonlinear and inelastic structural dynamics data under strong winds are being created in laboratories due to advances in performance-based wind design methodology (Abdullah et al., 2020). In addition to the experimental and field-measurement approaches the comprehensive data are further enriched by high-fidelity largescale simulation tools that are advanced by theoretical developments in wind engineering field (Blocken 2014;Kareem 2020), such as computational fluid dynamics/computational structural dynamics (CFS/CSI)-based hybrid modeling of transient structural response (Hao and Wu 2018) and statisticsbased synthesis of nonstationary wind field (Wang and Wu 2021). The Computational Modeling and Simulation Center (SimCenter) of the Natural Hazards Engineering Research Infrastructure (NHERI) program provides an effective way to integrate various simulation tools (Deierlein and Zsarnóczay 2021). Furthermore, novel real-time aerodynamics hybrid simulation techniques are emerging to effectively generate nonlinear and full-scale data in wind engineering by seamlessly stitching the numerical modeling in computer and physical testing in wind tunnel Wu and Song 2019). Data quality is essential to facilitate curation and reuse of the diverse and large datasets generated in the field of wind engineering. There are numerous methods and criteria specified by various wind engineering research groups/centers to ensure the high data quality, and the NHERI DesignSafe cyberinfrastructure platform recently suggested the best practices for detailed data quality assessment in terms of metadata quality, data content quality, data completeness and representation and data publications review (Rathje et al., 2017).
The improved understanding concerning the complex nature of wind fields (e.g., nonstationary and non-Gaussian features), the associated structural aerodynamics/aeroelasticity (e.g., transient and nonlinear features) and the resulting load effects (e.g., nonlinear and inelastic structural response), as well as the necessary shift from a prescriptive design approach to performance-based design methodology and further to resilience-based design philosophy (i.e., improving the rapidity, robustness, resourcefulness and redundancy), poses new challenges in wind engineering field. Hence, there is an urgent need of more accurate and efficient learning and modeling tools for effective solutions. The conventional stationary and linear analysis framework for windstructure interactions established by Robert H. Scanlan (1914-2001 and Alan G. Davenport (1932Davenport ( -2009 has been very successful due to its simplicity and applicability, however, its shortcomings have begun to surface since the underlying complexities associated with many wind engineering problems clearly show a departure from implicit assumptions of stationarity, Gaussianity and linear features. A number of semiempirical nonlinear reduced-order models have been developed in this context and improvement in their efficiency and robustness is a topic of cutting-edge research in the wind engineering community (Wu 2013). Unfortunately, these reduced-order models do not always have a satisfactory representation of the full nonlinear equations which govern the complex phenomena in wind-related problems. An alternate way is to utilize the CFD techniques, however, their computational effort is too high considering the three-dimensional nature of winds and associated bluff-body aerodynamics. While CFD plays a significant role in generating high-fidelity data of complex wind-structure interactions, its high computational cost makes it not easy to be used either in an informational mode to enhance wind hazard-related planning and development activities (e.g., risk mitigation that needs to quickly run thousands of scenarios at minimal computational expense) or in an operational mode to support emergency management and response associated with a wind hazard (e.g., decision making that needs real-time prediction capability under an uncertain environment). To address the emerging challenges, datadriven machine learning offers a promising approach that is capable of processing big data in wind engineering field as well as modeling associated complex phenomena with high computational efficiency and simulation accuracy.
With the rapid development of ML applications in wind engineering due to the confluence of advanced learning algorithms, high-performance computational hardware and big data, it is believed that a systematic review on this subject is important to suggest to what extend ML has been utilized in each of the topic areas within wind engineering and provide a comprehensive summary to improve understanding how learning algorithms work and when these schemes succeed or fail. Specifically, a total of 65 ML algorithms (Appendix A) are identified for their applications in the five topic areas of wind climate, terrain/topography, aerodynamics/aeroelasticity, structural dynamics and damage assessment, and mitigation and response. This review first presents technical background of typical ML approaches in terms of supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning (RL), followed by the state of research and practice of ML applications to each topic area within wind engineering field, and concluded with critical research gaps and future prospects. While ML can augment the analytical approaches [e.g., data-driven discovery of closure models (Raissi et al., 2019)], numerical schemes [e.g., data-driven turbulence modeling (Duraisamy et al., 2019)], experimental tests [e.g., data-driven active control of transient wind simulation (Li et al., 2021a)] and field measurements [e.g., data-driven sparse sensor placement (Manohar et al., 2018)] in wind engineering, the review only focuses on its role to complement existing methodologies and hence potentially extend/transform current lines of wind engineering research and practice.

BACKGROUND OF MACHINE LEARNING
Machine learning (ML) is a subclass of artificial intelligence (AI) that extracts the underlying pattern within a set of data (e.g., Murphy 2012; Goodfellow et al., 2016;Mohri et al., 2018). To acquire the hidden pattern and knowledge of a problem, the learning process involves in general five important steps, namely data collection, data preparation, training, evaluation and parameters tuning. Once the learning machine is trained based on the available data (usually retrieved from analytical solutions, numerical simulations, experimental tests or full-scale measurements), it can predict future or unseen events. Based on the data fed into the learning machine, ML algorithms can be classified into four categories, namely supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning (Figure 1).
To train the algorithm, the supervised learning fully depends on labeled data, the unsupervised learning relies purely on unlabeled data and the semi-supervised learning combines limited labeled data with a large amount of unlabeled data. For reinforcement learning (RL), there is essentially no predefined data. Although RL is occasionally treated as semisupervised learning considering the agent learns from its own experiences in terms of infrequent and partial rewards, it is classified here into separate category to highlight there is no explicit, external supervisory information provided to the learning agent. It is noted the kriging and polynomial chaos expansions as two widely-used, data-driven statistical interpolation approaches are not reviewed in this study.

Supervised Learning
Supervised learning models are a set of algorithms that learn the mapping, from given labeled training data, between known inputs and outputs. The trainable parameters of these models are determined based on the minimization of the loss function. Supervised learning models usually require a large amount of reliable and unbiased data for training which might not be always available. These algorithms can be employed for two important tasks, namely regression and classification.

Regression
Regression is a type of supervised learning in which the output is a numeric variable. Among many regression models, feed-forward 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 this study.
Deep neural networks (DNN) are also a type of FFNN characterized by a deep architecture equipped with multiple layers, and hence allows for better generalization and accuracy (Deng and Yu 2014;Pouyanfar et al., 2018). The convolutional neural networks (CNN) is another important FFNN with sparse convolutional matrices that are usually employed for pattern recognition and image classification (Krizhevsky et al., 2012;Goodfellow et al., 2016). Recurrent neural networks (RNN) are a class of feedback neural networks that allow previous outputs to be used as inputs while having hidden states and are suited to model time-dependent regression problems (e.g., Medsker and Jain 1999;Mandic and Chambers 2001). Long short-term memory (LSTM) are an advanced version of RNN to alleviate the gradient vanishing and exploding issue by only keeping necessary past information in future model states (Bengio et al., 1994).

Classification
Classification is another type of supervised learning in which the output is a categorical variable or a class. Support vector machines (SVM) (Scholkopf and Smola 2018) and random forest (RF) (Breiman 2001) are two classical examples of classification algorithms. SVM classifier identifies a hyperplane in a highdimensional space in which a simple linear classification can be performed. RF classifier, on the other hand, fits a number of decision tree classifiers on various sub-samples of the dataset, then averages the results to improve outcome accuracy [ Figure 3].

Unsupervised Learning
Unsupervised learning models draw inferences from datasets to describe hidden structures from unlabeled data based on   inherent characteristics (Russell and Norvig 2016). These models usually group instances of input data using a defined similarity index (global criterion). Clustering and dimensionality reduction are two standard examples of unsupervised learning applications.

Clustering
Clustering is an unsupervised learning task used for pattern recognition that automatically discovers natural groups or clusters in data. A cluster refers to a collection of data points aggregated together with similar features (Maulik and Bandyopadhyay 2002). The k-means clustering is one of the simplest unsupervised ML models. It is a centroid-based algorithm that partitions the data into k clusters. Mean-shift clustering is another unsupervised model with a slidingwindow-based algorithm to identify dense areas of data points. Other clustering algorithms such as the density-based spatial clustering of applications with noise, the expectation-maximization clustering using gaussian mixture models and the agglomerative hierarchical clustering are also popularly used for statistical data analysis.

Dimensionality Reduction
Dimensionality reduction aims to find the most important features within the dataset by identifying lower-dimensional representations for high-dimensional data. It minimizes the storage space, reduces the computation time and avoids overfitting. The ML-based dimensionality reduction can be divided into linear and nonlinear algorithms. The principal component analysis (PCA) is a commonly used linear technique that can be regarded as a two-layer neural network with a linear activation function. It essentially provides new uncorrelated variables, also denoted as principal components, which maximize the variance. The nonlinear autoencoder is a specific type of FFNN that compresses the initial input space into a reduced dimensional space using the encoder and then decompresses the obtained latent space back to the original input space using the decoder. Accordingly, deep autoencoders have a "bottleneck" architecture designed for extraction of representative features [ Figure 4]. The autoencoder algorithm has been attracting attention in fluid mechanics community for efficient development of reduced-order models.

Semi-Supervised Learning
Semi-supervised learning models operate based on limited labeled data with a large amount of unlabeled data. Hence, they can be regarded as combination results of supervised learning and unsupervised learning algorithms. The generative adversarial network (GAN) is a well-known semi-supervised learning algorithm for estimating generative models via an adversarial process. One important feature of semi-supervised learning algorithms is their labelled-data efficiency. To this end, it may be reasonable to consider the physics-informed deep learning (PIDL) as a semi-supervised model that leverages physics-based equations in the augmented loss function to significantly reduce the data demand during training process.

Generative Adversarial Network
The GAN model consists of two competing neural networks, namely the generator and the discriminator (Goodfellow et al., 2014). It generates new data based on a probability distribution that approximately represents the training data (true or labelled data). Specifically, the generator produces fake samples to imitate the distribution of a real dataset, then the discriminator tries to distinguish (through a classification process) between the real samples and fake ones (from the generator). The GAN model is trained such that the new generated samples accurately represent the underlying mechanisms of the studied system. The architecture of a typical GAN model is illustrated in Figure 5.

Physics-Informed Deep Learning
The concept of PIDL models was originally proposed several decades ago (Psichogios and Ungar 1992;Dissanayake and Phan-Thien 1994) in which prior knowledge (in terms of the physicsbased governing equations) is integrated within the neural networks to reduce the high-volume of required training data.
Typically, a small amount of labelled data along with a large number of unlabeled data that satisfy the underlying physics of the system of interest (also denoted as collocations points) are used to train these models. Hence, self-supervision plays a significant role in PIDL models. Recently, Raissi et al. (2017a, FIGURE 5 | Architecture of a typical GAN model. b) advanced the PIDL models by leveraging the automatic differentiation technique to solve partial differential equations. The architecture of a typical PIDL model is presented in Figure 6.

Reinforcement Learning
RL algorithm is usually formulated based on Markov decision process (Sutton and Barto, 2018). The core part of RL is its agent that interacts with its environment. Accordingly, the agent learns a policy that maps the states to the actions by maximizing the expected cumulative reward using an automated trial-and-error process (e.g., Mnih et al., 2015;Silver et al., 2017). Typical reinforcement learning models include value-based models (e.g., Q-learning or deep Q-learning) (Watkins and Dayan 1992), policy-based models (e.g., deep deterministic policy gradient) (Lillicrap et al., 2015) and hybrid models (e.g., actor-critic) (Williams 1992). Recently, the deep RL (with DNN-based policy) has been gaining attention in wind engineering community as an efficient way for dynamic control and shape optimization (Li et al., 2021a;2021b). The architecture of a typical deep RL is depicted in Figure 7.

APPLICATIONS OF MACHINE LEARNING TO WIND ENGINEERING
This section provides a comprehensive review of 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). 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.

Wind Climate
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 increasingly available datasets (e.g., satellite data), ML has become a supporting tool or even a reliable competitor of classical approaches for wind climate modeling (e.g., CFD). Most reviewed articles employed ML algorithms as a regression (e.g., long-term prediction of surface wind speed) or a classification (e.g., downburst occurrence prediction) tool. The selected metrics to evaluate the performance of ML algorithms included the root mean square (RMS), coefficient of correlation, mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination (R2), among others.

Classical Boundary-Layer Winds
Air movement in the planetary boundary layer plays a fundamental role in current wind design of structures and infrastructure. Although a detailed universal description of flow characteristics in the boundary-layer region has not been possible, the classical boundary-layer winds in gales from large 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 spectrum for turbulent fluctuations (Panofsky and McCormick 1960)]. The major research efforts have been focused on the accurate estimate of design wind speed in a statistical analysis framework (Simiu and Scanlan 1978). Specifically, long-term wind data from meteorological observations are analyzed based on extreme value theory to obtain the design wind speed at each location. However, the accurate forecast of  classical boundary-layer winds is very challenging since it involves a large range of various temporal and spatial scales (e.g., from fractions of a meter to several thousand kilometers for spatial scale and from fractions of seconds to several years for time scales). Usually, the temporal and spatial resolutions from the state-of-the-art weather forecast models [e.g., global forecast system from National Oceanic and Atmospheric Administration (NOAA)] are not sufficient for wind engineering purpose. On the other hand, the unprecedented volumes of data from field measurements (e.g., weather station and satellite) provide a solid foundation to advance ML applications for classical boundary-layer winds. Table 1 presents the reviewed applications of ML for classical boundary-layer winds, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from field measurements. From Table 1, it can be concluded that most applications used ML as a regression model for prediction of mean wind speed (averaging time ranged from minutes to months), while the short-term prediction of turbulent fluctuations that are very important to structural dynamics is very limited. In many applications, the selection of ML models is simply based on gut feeling or past experience. Although several researchers conducted comparison studies to select good ML models for their specific applications, it might be very challenging to generalize the obtained results to other applications due to a lack of a systematic comparison framework.

Tropical Cyclones
Tropical cyclones (TCs), also commonly known as hurricanes in North Atlantic, typhoons in western North Pacific and cyclones in Australia, are low-pressure storms that form over a warm ocean surface (Holton and Hakim 2013). With an average of 90 events reported annually (Zhao et al., 2012), TCs and their cascading hazards (e.g., wind, rain, storm surge and wave) pose a serious threat to public safety, livelihoods and local economies in many coastal regions around the globe. Hence, significant efforts have been made in modeling and predicting TCs and relatively wellestablished mesoscale numerical weather prediction frameworks [e.g., Weather Research and Forecasting (WRF) model] are available for high-fidelity simulations. However, the high-fidelity computationally expensive models might not be always appropriate for planning activities in an uncertain environment where Monte Carlo simulations are needed or emergency managements where real-time or near-real-time predictions are required. The high demand for a rapid and reliable technique used to assist decision-makers and planers results in many ML models for efficient simulations of key stages in the life cycle of a TC. These ML applications to TCs are fueled by increasingly available remotely-sensed and high-fidelity numerical data. The review in this section is organized following the four important components of full track of a TC, namely genesis, translation, intensity and wind field.

Tropical Cyclone Genesis
TC genesis requires several necessary environmental conditions (e.g., existence of low-pressure area and sea surface temperature of at least 26°C), however, the exact mechanisms of TC formation are still not well understood (Gray 1968(Gray , 1979Emanuel 2003;Holton and Hakim 2013). To predict the TC genesis, both numerical and statistical models were developed. The numerical models (e.g., global forecast system) are essentially based on the physical principles and their performance heavily depends on improved understanding of TC genesis mechanism. The statistical models (e.g., Michael 2017; Chen and Duan 2018; Cui and Caracoglia 2019) linearly relate the TC genesis to a few selected environmental factors, and hence show poor interpolation and limited predictability. The lack of a deep understanding of underlying mechanisms stimulated datadriven techniques for TC genesis simulations. As a result, increasing ML applications are available to accurately predict TC genesis. Table 2i presents the reviewed applications of ML for TC genesis, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from satellite measurements along with reanalysis results. It is expected the improved spatial resolution of currently available datasets will further enhance simulation results of ML models. From Table 2i, it can be concluded that most applications used ML as a classification model for either short-term or long-term forecasting of TC genesis. Although more dynamic and thermodynamic environmental factors can be retrieved using advanced remote sensing technologies in recent years, the identification of the most appropriate set of inputs to ML models (predictors) is still very challenging.

Tropical Cyclone Translation
Numerical forecast models have been successfully applied in forecasting normal TC trajectories, but they are computationally expensive. Although several statistical models were also developed based on a large amount of historical TC path records (e.g., Vickery et al., 2000Vickery et al., ,2009Emanuel et al., 2006;Hall and Jewson 2007;Chen and Duan 2018;Snaiki and Wu 2020a;Snaiki and Wu 2020b), their linear nature makes them incapable of capturing the inherent nonlinearities in such a complex dynamic system (Zhang and Nishijima 2012). Both numerical and statistical models or their combinations (statistical-dynamics models) show poor performance in forecasting sudden speed change, recurvature and stagnation in TC movement . To satisfy both simulation accuracy and efficiency, increasing ML applications emerged for TC path prediction. Table 2ii presents the reviewed applications of ML for TC translation, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from meteorological databases (e.g., satellite data) and reanalysis results. Typically, the TC track information is available only at each 6-h interval. From Table 2ii, it can be concluded that most applications used ML as a regression model for TC path prediction. Since the forecast of TC track can be regarded as a  time series prediction problem, the feedback neural networks such as RNNs and LSTMs are preferred and lead to good performance. However, their performance within each 6-h interval is unknown due to the sampling limitation in the training data.

Tropical Cyclone Intensity
The TC intensity (over ocean or land) can be measured in terms of central pressure or maximum sustained wind speed. It is impacted by several complicated physical phenomena (e.g., atmosphere-ocean interaction and vertical wind shear), and hence remains one of the most challenging issues in TC forecasting especially for rapid intensification prediction. To avoid the high computational cost of numerical forecast models, both statistics-based (e.g., Vickey et al., 2000;DeMaria et al., 2005;Hall and Jewson 2007;Vickey et al., 2009) and physics-based (e.g., Snaiki and Wu 2020a) tools were developed for fast prediction of TC intensity. However, neither statistical nor physical models guarantee prediction accuracy of TC intensity due essentially to the over-simplification of such a complicated dynamic system. To improve simulation accuracy while keeping a high efficiency, increasing ML applications are available for TC intensity prediction. Table 2iii presents the reviewed applications of ML for TC intensity, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from meteorological databased (e.g., satellite data) and reanalysis results. From Table 2iii, it can be concluded that most applications used ML as a regression (or a classification) model for estimation of intensity time series (or levels). Although encouraging simulation results indicate a good performance of ML models in predicting TC intensity for their specific applications, the selection of the most appropriate set of inputs (including the number of predictors and previous time steps) is still very challenging. In addition, it is not easy to conduct a systematic comparison among reviewed ML models since the used performance metrics differ substantially from one application to another.

Tropical Cyclone Wind Field
TC wind hazard is of great significance since it (directly) induces significant damage to life and property and (indirectly) triggers other TC-induced hazards (e.g., storm surge and waves). Substantial research efforts have been made for development of numerical models (e.g., WRF) or analytical models (e.g., Snaiki and Wu 2017a;Snaiki and Wu 2017b;Snaiki and Wu 2018;Snaiki and Wu 2020c;Fang et al., 2018;He et al., 2019) to simulate the boundary-layer wind field. However, none of these models can simultaneously achieve simulation accuracy and efficiency. To address this issue, increasing ML applications emerged for TC boundary-layer wind field simulation. Table 2iv presents the reviewed applications of ML for TC wind field, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from meteorological databases (e.g., satellite data) and high-fidelity simulations. It is expected the improved spatial resolution of currently available datasets will further enhance simulation results of ML models. From Table 2iv, it can be concluded that most applications use ML as a regression model for prediction of surface wind speed. Since these ML models were often trained and fine-tuned to predict the TC wind field at a specific region, it might be very challenging to generalize the obtained results to other locations. It is noted that only wind field at a certain altitude is available in most ML applications due essentially to training data sparsity issue in vertical dimension. The widely-used logarithmic or power-law profiles are typically employed to obtain the TC boundary-layer winds. Accordingly, the supergradient winds that may have significant implications to the wind design of tall buildings is not captured (Snaiki and Wu 2020c).

Non-synoptic Winds
Unlike synoptic winds that are associated with large-scale meteorological systems characterized by horizontal scales of thousands of kilometers and time scales of days, the nonsynoptic wind systems are local phenomena (e.g., a horizontal scale of several hundreds of meters) and short lived (e.g., a time scale of a few minutes) (Chowdhury and Wu 2021). Furthermore, the transient nature of non-synoptic winds makes them exhibit time-varying mean wind speeds and nonstationary/non-Gaussian fluctuations. Accordingly, the detection, measurement, and modeling of non-synoptic wind systems lag behind those of synoptic winds. However, numerous studies have demonstrated the importance of the non-synoptic wind events on the structural design (e.g., Holmes 1999;Letchford et al., 2002;Hao and Wu 2017). For example, the design wind speeds with relatively high return periods are usually dominated by the thunderstorm downbursts (Twisdale and Vickery 1992; Solari  , 2015) and the ASCE 7-22 includes the first-ever criteria for tornado-resistant design (ASCE, 2021). Recently, there is a rapid development of field-measurement networks (e.g., THUNDERR project at University of Genova) and laboratory facilities (e.g., WindEEE at Western University) for improved understanding of non-synoptic wind systems. These advances offer an unprecedented volume of data, and hence provide an opportunity to facilitate ML applications to non-synoptic winds. Although the non-synoptic wind systems can be originated from various mechanisms (e.g., convective storm, gravity wave or negative buoyancy) (Bluestein 2021), the review only focuses on those associated with convective storms. Specifically, ML applications to thunderstorms (subsynoptic-scale weather system) are first presented, followed by detailed reviews of its applications to two important types of non-synoptic wind events associated with thunderstorms, namely downbursts and tornadoes.

Thunderstorms
A thunderstorm is short-lived atmospheric weather system accompanied by lightning and thunder, gusty winds, heavy rain, and sometimes hail (Solari 2020). The life cycle of a thunderstorm usually consists of cumulus stage, mature stage and dissipative stage, and it typically lasts around 30 min. Both mesoscale and microscale numerical models have been developed for simulation of thunderstorms (Hawbecker 2021). Mesoscale modeling covers a large-scale computational domain (and hence fully considers physics involved), however, it is limited to a low spatiotemporal resolution. Microscale modeling utilizes a high spatiotemporal resolution (and hence obtains important small-scale features in the simulation of winds), however, it is limited to a relatively small-scale computational domain resulting in insufficiently reliable boundary conditions. To avoid shortcomings of currently available numerical models, ML models may provide a promising approach for efficient and accurate simulation of key stages in the life cycle of a thunderstorm. Table 3i presents the reviewed applications of ML thunderstorms, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from meteorological databases and reanalysis results. From Table 3i, it can be concluded that most applications used ML as either a classification or a regression model for prediction of thunderstorm occurrence. Obviously, there is still room for more comprehensive applications of ML in terms of modeling and forecasting each aspect of the thunderstorm from formation to dissipation. In addition, most ML applications to thunderstorm were limited to simple models with standard algorithms (e.g., ANN with backpropagation).

Downbursts
Downbursts are one of the most spectacular and dangerous events resulting from thunderstorms (Solari 2020). Their radial outflows and ring vortices after touchdown produce strong wind gusts very close to the ground and therefore lead to substantial structural damages (e.g., Yang et al., 2018). Downbursts are typically simulated numerically using CFD (e.g., Mason et al., 2009;Aboshosha et al., 2015;Haines and Taylor 2018;Hao and Wu 2018;Iida and Uematsu 2019) or experimentally using wind tunnels (e.g., Jesson et al., 2015;Jubayer et al., 2016;Hoshino et al., 2018;Aboutabikh et al., 2019;Asano et al., 2019;Junayed et al., 2019;Romanic et al., 2019). Both numerical and experimental approaches to obtain wind fields associated with downbursts are very time consuming (either computational expensive or labor intensive). This shortcoming motivated increasing use of ML tools for efficient and accurate simulations of downbursts. Table 3ii presents the reviewed applications of ML for downbursts, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from field measurement. From Table 3ii, it can be concluded that most applications used ML as a classification model for prediction of the occurrence of downburst or probability of damaging wind. There are a very limited number of ML applications for modeling and forecasting the downburst wind field, hence more research efforts are needed in this aspect. It is noted that the reviewed ML applications usually involved a high number of predictors. The employment of relatively high number of input variables may be necessary due to the complexity of downburst prediction. However, it makes the ML models not easy to use since these input variables might not be always available.

Tornadoes
Tornadoes are characterized by a rotating column of air descending from supercell thunderstorms lasting from several minutes to few hours. They are the most intense of all nonsynoptic wind events, and hence result in significant damage and collapse of structures Wu 2016, 2020). Several analytical and empirical models have been developed to simulate the vertical and radial wind profiles of tornado-like vortices (e.g., Wen and Chu 1973;Baker and Sterling 2017). These models are clearly over-simplified. The tornado wind fields are also modeled using CFD simulations (e.g., Kuai et al., 2008;Ishihara et al., 2011;Liu and Ishihara 2015;Eguchi et al., 2018;Gairola and Bitsuamlak 2019;Kawaguchi et al., 2019;Huo et al., 2020;Liu et al., 2021) or laboratory tests (e.g., Sarkar et al., 2006;Refan and Hangan 2016;Razavi and Sarkar 2018;Tang et al., 2018;Ashton et al., 2019;Gillmeier et al., 2019;Hou and Sarkar 2020;Razavi and Sarkar 2021). However, CFD simulations of tornadoes are computational expensive while the laboratory tests are labor intensive. These shortcomings motivated increasing use of ML tools for efficient and accurate modeling of tornadoes. Table 3iii presents the reviewed applications of ML for tornadoes, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from meteorological datasets (e.g., Radio-based data). From Table 3iii, it can be concluded that most applications use ML as a classification or a regression model for prediction of tornado occurrence. Obviously, there is still room for more comprehensive applications of ML in terms of simulation of the full track of a tornado (including its intensity and associated       wind field). Just like ML applications to downbursts, a high number of input variables (predictors) were utilized for the reviewed ML models. The identification of the most appropriate set of predictors is still very challenging, and a trail-and error approach was typically employed. In addition, it is not easy to conduct a systematic comparison among reviewed ML models since the used performance metrics differ substantially from one application to another.

Terrain and Topography
Wind characteristics including mean wind speeds and turbulent fluctuations are much affected by the surrounding terrain and topography. As a consequence, careful consideration of local terrain roughness and topographic features as well as surrounding obstacles is vital to the accurate determination of wind pressures on structures and pedestrian level winds. Wind codes and standards consider the terrain effects corresponding to limited (and simplified) terrain geometries (e.g., escarpment and single hill) through correction factors. To examine the effects of complex terrain condition on wind fields, wind tunnel tests are usually employed with a very small geometric scale (e.g., 1:500). Alternatively, numerical schemes such as the mass-conservation or momentum-conservation model can be used to capture the terrain effects on oncoming wind fields. Although the topographic effects can be well simulated based on momentum-conservation models (e.g., using Reynoldsaveraged Navier-Stokes equations), the needed computational time makes it impractical for use as a real-time decision support tool. The mass-conservation model computes wind fields over complex terrain in seconds to a few minutes (Forthofer et al., 2014a;2014b), but the accuracy of simulation may be poor because nonlinear momentum effects are not considered (Jackson and Hunt 1975). Considering the complex terrainwind data from high-fidelity CFD simulations, wind tunnel tests and field measurements are increasingly available, ML tools can be utilized (as computationally efficient reduced- order models that possess high simulation accuracy of complex nonlinear systems) to provide rapid estimation of wind flows over various terrain conditions. However, ML development for terrain and topographic considerations is still at an early stage with a limited number of studies reported in the literature. Table 4 presents the reviewed applications of ML for terrain and topography, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from either CFD simulations or wind tunnel tests. From Table 4, it can be concluded that most applications used ML as a regression model for prediction of wind fields over various terrain conditions and topographic configurations. There are a few studies that applied ML techniques to assist in efficient search for a correct layout of passive flow altering devices (e.g., spires and roughness elements) in the boundary-layer wind tunnel. It is noted that the current ML applications to consider topographic effects on wind fields are usually limited to terrain configurations that can be characterized by several parameters, hence, the employed ML models and training schemes are simple and standard (e.g., ANN with backpropagation). However, several advanced ML models such as autoencoder (e.g., Fukami et al., 2019) and GAN (Kim and Lee 2020) have been utilized to assist in the generation of turbulent inflow (as a realistic inlet boundary condition of CFD simulations).

Aerodynamics and Aeroelasticity
The bluff-body aerodynamics and aeroelasticity play a critical role in the safe and cost-effective design of wind-sensitive structures, and their considerations rely heavily on boundarylayer wind tunnels. In addition to the Reynolds number effects (due to very small model scales), wind tunnel tests are very time consuming and labor intensive. To this end, CFD techniques have been rapidly developed for simulations of structural aerodynamics (gust-induced effects) and aeroelasticity (motion-induced effects). The purpose is to make CFD simulations serve as a complementary or even alternative approach to wind tunnel tests. Despite significant advances of hardware and algorithms, the reliable CFD simulations of windstructure interactions are still computationally very expensive due to three-dimensional nature of wakes and intensive flow separations from structures. Hence, a number of reducedorder models have been developed to efficiently model structural aerodynamics and aeroelasticity . Unfortunately, these reduced-order models do not always have a satisfactory representation of the full nonlinear equations that govern the wind-structure interactions. Specifically, modern bridge decks and super tall buildings with unusually geometries all exhibit nonlinear unsteady aerodynamics and aeroelasticity that limit the applicability of the state-of-the-art reduced-order modeling methodologies. On the other hand, the Kolmogorov Neural Network existence theorem offers mathematical foundation for applying multilayer neural networks to approximate arbitrary nonlinear systems with any precision (Huang and Lippmann 1988;Hornik, 1991). With high-fidelity data and advanced algorithms, ML models can simultaneously achieve great simulation efficiency and accuracy. It is noted that there are numerous ML applications to aerodynamics and aeroelasticity of both bluff bodies (e.g., circular cylinder) and streamlined bodies (e.g., airfoil) in fluid mechanics community (e.g., Kutz 2017;Brunton et al., 2020), however, they are not discussed here. The review in this section only covers wind-sensitive structures in civil engineering. The ML applications for bridge aerodynamics and aeroelasticity are first reviewed in Table 5i and then followed by buildings and other structures in Table 5ii, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from either CFD simulations or wind tunnel tests. From Table 5, it can be concluded that most applications used ML as a regression model for prediction of steady-state force coefficients, flutter derivatives and vortex-induced vibrations (VIV) of various bridges and for modeling of wind pressure coefficients of various buildings (as well as estimation of the interference factors for adjacent buildings). The different aerodynamic representations in bridges (mainly using global quantities such as force coefficients) and buildings (mainly using local quantities such as pressure coefficients) are partially due to available data types from wind tunnel tests. Although satisfactory ML simulation results have been obtained (in terms of interpolations), most reviewed applications do not necessarily have good performance in terms of extrapolations outside the training datasets. It is noted that the currently available ML models of aerodynamics and aeroelasticity are developed for the main purpose of being used as preliminary design tools to avoid the high-cost wind tunnel tests in the early design stage. There is a lack of systematic comparison among various ML models, hence, their selection for specific applications is rather rudimentary.

Structural Dynamics and Damage Assessment
Due to the computational complexity of numerical techniques (e.g., finite element method) for solving wind-induced nonlinear structural response, reduced-order models (e.g., ANN) have been developed to alleviate the computational cost of the high-fidelity models. The ML models have been used for structural dynamics and damage assessment for several decades mainly in the field of earthquake engineering (e.g., Wu et al., 1992;Masri et al., 1993;Jiang and Adeli 2005;Pei et al., 2005;Gholizadeh et al., 2009;Facchini et al., 2014;Derkevorkian et al., 2015;Liang 2019;Wu and Jahanshahi 2019;Yu et al., 2020). However, similar applications have not emerged in wind engineering community until recently due essentially to the linear consideration of the wind-induced structural response [ASCE 7-16 (ASCE, 2017)]. Recent advances of performance-based wind design methodology have placed increasing importance on effective simulations of nonlinear, inelastic structural dynamics response under strong winds. The numerical estimation of windinduced nonlinear structural response using a high-fidelity finite element model is computationally very expensive due to its small time-step size and long simulation duration. Accordingly, several ML applications to wind-induced structural dynamics have been     developed in recent years for simultaneously achieving high simulation accuracy and efficiency. The performance-based (and further resilience-ba sed) wind design philosophies also require accurate damage assessment of structures and infrastructure under extreme storms. The structural damages under winds depend on numerous factors including wind features (e.g., wind speed/direction and topography) and built environment characteristics (e.g., building opening and roof slope), hence its assessment and quantification are extremely challenging. On the other hand, increasingly available fieldmeasurement data characterizing structural damages under strong wind events [e.g., resulting from post-disaster reconnaissance activities such NHERI Natural Hazards Reconnaissance (RAPID) Facility and NSF Structural Extreme Events Reconnaissance (StEER) Network] provide a great opportunity to learn from data by using various ML models. The ML applications for structural dynamics are first reviewed in Table 6i and then followed by damage assessment in Table 6ii, where the ML model, training scheme, input data, output data, data source and performance metric are summarized for each application. The training/testing data were essentially retrieved from numerical simulations, wind tunnel tests and field measurements. From Table 6, it can be concluded that most applications used ML as a regression model for modeling structural dynamics and as a regression or a classification model for structural damage assessment. While many applications employed simple ML models and standard training schemes (e.g., ANN with backpropagation), some advanced schemes such as knowledge-enhanced LSTM have been successfully applied to predict time series of wind-induced nonlinear structural response. It is noted that the selection of the most appropriate set of inputs to ML models for damage assessment (predictors or features) is still very challenging.

Mitigation and Response
Both long-term and short-term strategies are needed to enhance resilience of individual structures or communities to withstand wind-related hazards. One important long-term consideration is to mitigate structural response/vibration subjected to winds through structural optimization and/or control. For structural optimization under winds, the shape optimization is probably the most effective approach to reduce aerodynamic loading. For wind-induced vibration control, both aerodynamic and mechanical measures are well recognized in wind engineering community. Although the structural performance evaluation under winds is typically a very complicated task, the corresponding simulations during optimization or (active) control process is required to be efficient and accurate because they need to be conducted either repeatedly for numerous scenarios or in a (near) real-time sense. As noted earlier, the ML models are very promising to simultaneously achieve the high simulation efficiency and accuracy goal. In addition, the RL models that have gained increasing popularity in recent years can be used as very effective optimization or control algorithms compared to conventional approaches (Silver et al., 2017). In the consideration of short-term actions, efficient management strategies are critically important. Although the ML models used in the disaster (including wind-related hazard) management framework (i.e., covering preparedness, response and recovery) have recently been systematically reviewed (e.g., Sun et al., 2020), its applications to social media-informed response are still discussed here since the unprecedentedly   from social media platforms for disaster response. From Table 7i, it can be concluded that the structural performance evaluations in mitigation applications usually used ML as a regression model while RL was typically utilized as an effective optimization or control algorithm. It is noted that relatively few ML applications for structural optimization and control under winds have been generated compared to those in earthquake engineering community (e.g., Ghaboussi and Joghataie 1995;Adam and Smith 2008;Jiang and Adeli 2008;Yakut and Alli 2011;Subasri et al., 2014;Khodabandehlou et al., 2018;Khalatbarisoltani et al., 2019;Hayashi and Ohsaki 2020). From Table 7ii, it can be concluded that most social media-informed response applications used ML as a classification model for disaster rescue and relief information dissemination. Although these ML applications present promising results in terms of effectively supporting timely decision-making, there is a concern of using information from social media platforms due to a lack of data quality control.

Summary
The ML applications in each topical area of wind engineering are summarized in Figure 8. As shown in the figure, ML models are unevenly distributed among these areas. The wind climate area has the most ML applications followed by the aerodynamics and aeroelasticity area, and they are respectively contributed by wind engineering-related fields of meteorology and fluid mechanics. 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 structural design for winds, their developments are in a very preliminary stage and there is still a long way to go before they can complement or even replace existing approaches of wind tunnel tests and CFD simulations. In general, the supervised learning dominates the ML applications in wind engineering with the podium position attributed to simple models with standard algorithms (e.g., ANN with backpropagation). Actually, the selection of various ML models is rather rudimentary since there is a lack of systematic comparison among them (e.g., in terms of model complexity and performance). It is noted that the great potential of semi-supervised learning and unsupervised learning (as well as RL) with little or no labelled data is not leveraged yet. Accordingly, the current ML developments in wind engineering heavily rely on available labelled data. For example, the ML applications to non-synoptic winds are much less than those of synoptic winds due essentially to the difficulty in obtaining the data of local and short-lived storms.
On the other hand, the recent emergence of numerous ML applications to social media-informed disaster response is due mainly to the unprecedentedly abundant data from various powerful communication tools. For the reviewed ML applications, the training/testing data are retrieved from several major sources (e.g., field measurements, wind tunnel tests, numerical simulations and social media platforms). In the determination of ML model inputs and outputs, a good understanding of underlying physics of each application is critical to effectively select an appropriate set of predictors (ML inputs) while the output types heavily depend on the needs of traditional analysis procedure in each application (e.g., local wind pressures for building design and global wind forces for bridge design).

CHALLENGES AND PROSPECTS
The rapidly increasing ML applications to wind engineering have generated a large volume of datasets associated with a large set of domain-specific algorithms. It is strongly believed that the platforms encouraging open sharing of these datasets and algorithms would greatly benefit the ML research progress in wind engineering. The openly available wind engineering datasets will greatly reduce efforts for their creation/collection and preprocessing, and open-source ML algorithms will save significant time for their re-implementation. The reduced need of time and effort to use the state-of-the-art or latest developed ML tools under such a culture of openness would spur interests among researchers in wind engineering, and hence result in more related ML applications. Moreover, the developed cyberinfrastructure to store and share data usually has a systematic curation procedure to ensure the high quality of its standardized benchmark datasets.
Also, the open-source software allows the hidden bugs/tricks of ML algorithms to be easily uncovered and accordingly makes them more robust. In addition to availability, the reproducibility and testability of wind engineering data and domain-specific algorithms due to a culture of openness would also facilitate the adoption of the obtained transparent and trustworthy ML tools in real-world problems. Although the wind engineering community has started to embrace the prevalent openness of ML community (e.g., NHERI DesignSafe platform), the culture of openness is still in its early stage. It is expected that more incentives based on the existing reward system (e.g., a digital object identifier for each dataset or algorithm published by the platform) are needed to motivate the ML wind engineering community towards open science. Given a potential openscience environment with openly available datasets and opensource algorithms (supported by open-access scientific publications), some remaining challenges and future prospects are discussed in terms of data in wind engineering and algorithms in ML. It is noted that both challenge and prospect lists are not exhaustive.

Challenges and Research Gaps
The reviewed various ML models for a wide range of topics in wind engineering suggests that their cross field has recently attracted much interest. However, there are still numerous challenges to advance ML applications to wind engineering from conception and research into practice. These remaining challenges of data in wind engineering and algorithms in ML are discussed in this sub-section.

Wind Engineering Data Challenges
Wind engineering data could be rich in some dimensions but may be poor in others. For example, a large volume of flow data or pressure data could be obtained by one wind tunnel test (using advanced measurement systems with high resolution in space and high sampling rate in time), however, all these data would be located at a point in the Reynolds number dimension. For structural response under winds, most of the data are located in the linear elastic domain, while very limited nonlinear inelastic data needed to advance implementation of performance-based wind design are available. Another example is that the anemometric monitoring network typically generates abundant data in time dimension but sparse data in space. More importantly, it is usually very challenging or expensive to create extra points in currently data-scarce dimensions. Wind engineering data could be short in time span of their collection. For example, the climate changing impacts are not easy to be considered based on the currently available wind data since their record period is much shorter than the time scale of climate changing. Also, few structural performance data under winds are long enough to take the life-span deterioration behaviors into account. Essentially, the learning machine based on current windstructure interaction data cannot be used for accurately predicting future long-term behaviors of the same windstructure system. Wind engineering data could be highly heterogeneous for collaborative or large-scale ML applications. Many complex tasks (e.g., life-cycle performance evaluation of structures under winds) and/or real-world problems (e.g., hurricane resilience assessment of coastal communities) in wind engineering need collaborative efforts and/or large-scale implementations. The datasets generated from these activities may result from various CFD simulation tools or field measurement devices, and they are typically interpreted by different entities before sent to a central processing platform. Accordingly, significant processing efforts (e.g., data cleaning, data aggregation, dimension reduction and data standardization) are needed for these heterogeneous datasets with high variability of data types and formats (e.g., mixtures of structured, semistructured and unstructured data). In addition, advanced powerful learning machines are necessary to generate new knowledge from large, heterogeneous sets of wind engineering data.

Machine Learning Algorithm Challenges
ML algorithms commonly-used in wind engineering are standard ones designed for solving problems in other fields (e.g., handwriting recognition or computer vision). While these classical algorithms (e.g., ANN with backpropagation) achieved great success for simple wind engineering applications, they are not necessarily concise and efficient. More importantly, the immediate applications of these popular algorithms to modern wind engineering (involving nonstationary and non-Gaussian wind flow, transient and nonlinear aerodynamics, nonlinear and inelastic structural dynamics, or time-variant wind-structure system under a changing climate) may be very challenging. On the other hand, the newly developed ML algorithms (e.g., advanced LSTM and GAN) need to be carefully scrutinized for their applicability to these complex problems. ML algorithms commonly-used in wind engineering are supervised ones that need a significant amount of labelled data. Although the cost of obtaining/collecting the data from various sources (e.g., numerical simulations, wind tunnel tests, or field measurements) is greatly reduced and accordingly unprecedented volume of data are increasingly available, these datasets may be limited to unlabeled due to a lack of sufficient human resources (with expert knowledge) for data labeling. ML algorithms commonly-used in wind engineering are purely datadriven ones that are usually consider as black boxes. Furthermore, currently available ML models usually present a conflict between their advances (and hence performance) and explainability. One important feature of human intelligence is the ability to explain the rationale behind its decisions to others, hence, the explainability of learning machines is often an essential prerequisite for establishing a trust relationship between human intelligence and artificial intelligence. The highly non-transparent nature of ML algorithms may be acceptable for some applications in wind engineering (e.g., a CNN mapping the oncoming winds to pressure fields on or velocity fields around various bridge decks), however, it may be a clear drawback for many high-stake applications (e.g., evacuation planning or transportation infrastructure management under a landfalling hurricane) since any error in prediction may have catastrophic consequences. It is noted that the high-stake applications also place a high demand for quantification of uncertainties involved in ML algorithm selection, training and performance evaluation (along with data collection), whereas the formalization of uncertainty quantification for purely datadriven approaches is very challenging and not well established yet. ML algorithms commonly-used in wind engineering are typically selected based on past experience (or simply by "gut feeling") and the associated model hyperparameters (e.g., layer and neuron numbers, activation function and learning rate) are usually obtained by extensive trial and error. While the selected ML algorithms present good performance for the particular applications of interest, they are not necessarily an optimal choice. A systematic approach to identify the most appropriate ML model and associated best hyperparameters essentially needs a global optimization within a high dimensional space, and is currently very challenging for wind engineering applications.

Prospects and Future Directions
The remaining challenges, while not trivial, provide new research opportunities for the development of more effective ML tools. The identified prospects of data in wind engineering and algorithms in ML are discussed in this sub-section.

Wind Engineering Data Prospects
To generate/collect wind engineering data that are scarce in certain dimensions, advanced full-scale/laboratory/numerical tools and technologies need to be utilized or developed. In addition to large-scale facilities (e.g., WindEEE), various highfidelity and efficient modern CFD techniques (e.g., hybrid large eddy simulation/Reynolds-averaged Navier-Stokes schemes) should be exploited to generate data of high-Reynolds number scenarios. The rational loading protocols for extreme wind performance cyclic testing of deformation-controlled MWFRS (Main Wind Force Resisting System) members need to be designed to generate the wind-induced nonlinear inelastic structural response data. Also, data reconstructions using linear/nonlinear dimensionality reduction techniques (e.g., singular value decomposition/autoencoder) should be employed to enhance spatial resolution of full-scale measurements. To generate/collect wind engineering data that cover sufficiently-long time span of structural behaviors, more reliable long-term structural health monitoring systems should be established in addition to high-fidelity modeling of aging and deterioration of wind-sensitive structures. For the consideration of wind engineering data under a changing climate, synthesized wind fields (resulting from tropical cyclones, extratropical cyclones or local non-synoptic storms) need to be generated by global climate models coupled with accurate and efficient downscaling exercises under projected climate conditions [e.g., various RCP (Representative Concentration Pathway) scenarios]. To effectively learn from heterogeneous data that need to be first unified, they can be efficiently processed by advanced big data analytics. For example, unsupervised or semi-supervised clustering techniques could be used for data cleaning, data fusion techniques of Kalman filters could be used for data aggregation, and linear principal component analysis or nonlinear self-organizing map could be used for dimensional reduction.

ML Algorithm Prospects
To facilitate ML applications to complex wind engineering problems, the state-of-the-art or latest algorithms emerging in ML community could be leveraged. For example, the GAN could be used for effectively generating nonstationary and non-Gaussian wind flow through its two competing sub-networks, the CNN could be employed for efficiently mapping oncoming winds to pressure fields (characterizing transient and nonlinear aerodynamics) on structures with an arbitrary shape because it is particularly good at handling input-output data with a known grid-like topology, the LSTM could be utilized for accurately simulating nonlinear and inelastic structural dynamics since its forget gates ensure a reliable consideration of long-term dependencies (where the structural response at the current time depends on not only the current wind load but also the load history), and the lifelong learning networks should be explored for adaptively modeling time-variant wind-structure system assuming their underlying parameters can be continuously modified to accommodate new data inputs. The direct or immediate applications of the advanced ML algorithms to complex wind engineering problems may not necessarily result in parsimonious models that may need specialized customization for each application. To reduce the demand for labelled data in ML applications to wind engineering, both unsupervised learning and semi-supervised learning (including physics-informed machine learning) are promising alternatives to popularly used supervised learning. In addition, advanced ML algorithms have been emerging (e.g., reservoir computing) for processing information generated by complicated dynamical systems using very small training datasets. To open the ML black box, model explainability and interpretability in wind engineering applications needs to be enhanced. Various general techniques have been developed to improve understanding of the ML model predictions, such as sensitivity analysis and layer-wise relevance propagation. On the other hand, the definitions of explainability and interpretability are typically domain dependent, hence, the domain knowledge in wind engineering should be leveraged for enhanced explainability/interpretability of each ML application. It is expected that the explainability/interpretability analysis (along with uncertain quantification) will likely become a fundamental building block for bounding the overall confidence in ML applications in wind engineering (parallel to verification and validation in CFD simulations). To enable an automatic search of ML model hyperparameters in wind engineering applications, increasingly available optimization schemes with improved efficiency and accuracy (e.g., grid search, random search, Bayesian optimization and population-based training) can be utilized to find the best configuration for each task. On the other hand, it is believed that a practical guide to selection of ML models in wind engineering applications will greatly facilitate their appropriate use. The best practices for model selection in each application are essentially consistent with the principle of Ockham's razo by first testing simple linear ML models (due to their easy to implement and high model explainability), and then followed by more complex 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 needed in a purely performance-driven ML model selection, the domain knowledge is suggested to be utilized for a more effective search process.

Knowledge-Enhanced Machine Learning
As discussed in preceding sections, domain knowledge could be leveraged for improved selection of ML model and its inputs and outputs in wind engineering applications. Hence, a good understanding of fundamental physics and other types of domain knowledge underlying each subfield of wind engineering would enable more effective use of ML tools. It is noted that the fundamental physics in terms of governing equations is a special type of domain knowledge, and recent studies have demonstrated that the required labelled datasets could be significantly reduced by incorporating the underlying physics into training process (and hence enhancing the regularization mechanism) (e.g., Raissi et al., 2017a;. Other equation-based domain knowledge such as empirical/ semi-empirical formulas were also employed as part of the loss function in deep learning to provide machine-readable prior knowledge that facilitates the effective regularization of the neural networks for simulations of tropical cyclone winds (Snaiki and Wu 2019) and nonlinear structural dynamics . In addition, the equation-free domain knowledge has been integrated into a deep RL-based aerodynamic shape optimizer (via the transfer-learning and 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 scheme of knowledge-enhanced machine learning (KEML) could significantly enhance ML applications to wind engineering. To fully embrace the promising potential of KEML, systematic research efforts are needed to efficiently identify knowledge representations (invariances, physics equations, empirical formulas, probabilistic relations, logic rules, simulation results, field observations, human feedback, and others) in various subfields of wind engineering and then to effectively integrate them into each module of machine learning pipeline (data preparation, model selection, model training, and others). While domain knowledge could be employed to enhance purely data-driven ML tools, it is expected that learning machines could be utilized for harnessing data to discover new knowledge in wind engineering (e.g., governing laws characterizing transport of turbulence quantities or optimization of wind-structure system).

CONCLUDING REMARKS
A total of 65 machine learning (ML) algorithms were reviewed in terms of their applications to each topical area of wind engineering, namely wind climate, terrain/topography, aerodynamics/aeroelasticity, structural dynamics, wind damage assessment and wind-related hazard mitigation and response. The most ML applications were found in wind climate area, while the terrain/topography area had the least applications of ML. Although the ML-based wind engineering is fueled by the unprecedented volume of analytical, numerical, experimental and field-measurement data together with rapidly evolving learning algorithms and highperformance computational hardware, it is still at an early stage of development. Most of wind engineering applications employed supervised learning with standard ML models designed for solving problems in other fields, and the promising unsupervised and semi-supervised learning tools were rarely used to reduce the high demand of labelled data. For the selection of ML models and associated hyperparameters in wind engineering applications, it was typically based on expertise and extensive trial and error. In this review, the culture of openness, explainability/interpretability and uncertainty quantification were identified as important research gaps that need to be addressed in ML-based wind engineering community. Furthermore, the knowledge-enhanced machine learning was considered as a very promising scheme to enhance ML applications to wind engineering.