Machine Learning for Outdoor Thermal Comfort Assessment and Optimization: Methods, Applications and Perspectives

1. Introduction

Urban environments are increasingly exposed to thermal stress due to the dual pressures of global climate change and localized phenomena such as the Urban Heat Island (UHI) effect [1,2,3]. These climatic burdens reduce outdoor livability, exacerbate public health risks, and elevate energy demand for indoor cooling. As cities strive to enhance resilience and sustainability, outdoor thermal comfort (OTC) has emerged as a key consideration in urban design and planning. OTC is defined as the physiological and psychological condition in which individuals perceive outdoor thermal conditions as acceptable [4,5,6,7,8,9,10], capturing the complex interplay between meteorological variables, urban geometry, and individual human attributes.

Poor outdoor thermal conditions reduce the usability of public spaces, discourage physical activity, and increase reliance on air-conditioned indoor environments—further amplifying energy consumption and greenhouse gas emissions [11,12]. For this reason, OTC has been increasingly integrated into climate-responsive design and urban adaptation strategies, guiding interventions such as street shading, landscape cooling, and the deployment of green infrastructure [13,14]. Enhancing OTC supports broader goals of public health, equity, and behavioral adaptation to heat extremes [15,16,17].

OTC is governed by multiple, interdependent factors spanning meteorology, urban morphology, and human physiology. Meteorological drivers—such as air temperature, humidity, wind speed, and solar radiation—modulate the energy exchange between the human body and the environment [18,19,20,21,22,23]. Urban form—including aspect ratio (AR), sky view factor (SVF), surface materials, albedo, and vegetation—regulates local microclimates through shading, thermal storage, and airflow [17,21,24,25,26]. For instance, while narrow canyons may reduce solar exposure, they can also inhibit ventilation; conversely, vegetated surfaces promote evapotranspiration and reduce mean radiant temperature (MRT) [27]. Human thermal perception is further shaped by physiological parameters, including metabolic rate, age, clothing insulation and acclimatization [21,28,29,30].

To quantify OTC, several bioclimatic indices have been developed, including the Physiological Equivalent Temperature (PET) [31] , Universal Thermal Climate Index (UTCI) [32], Standard Effective Temperature (SET)* (with the asterisk indicating that it is an equivalent temperature index adjusted for humidity, air movement, and clothing) [33] and Predicted Mean Vote (PMV) [34]. These indices integrate environmental and physiological variables into standardized measures, facilitating cross-context assessment of thermal comfort [35].

Over the past two decades, OTC research has employed experimental, empirical, and simulation-based approaches to assess outdoor thermal conditions and guide design interventions. Field campaigns and thermal perception surveys have linked subjective responses to in situ measurements [19]. Empirical models, primarily based on regression analysis, have identified statistical relationships between microclimatic inputs and thermal indices [36]. At the same time, simulation platforms such as ENVI-met [37], RayMan [38], and tools based on computational fluid dynamics (CFD) [39] have enabled detailed assessments of urban thermal behavior and the evaluation of mitigation strategies [40].

Despite these advances, traditional approaches exhibit notable limitations. High-fidelity simulations often demand complex input data and are computationally intensive—limiting their utility in real-time or large-scale design processes. Moreover, regression and simulation models frequently rely on simplified or steady-state assumptions that fail to capture the nonlinear, context-sensitive, and adaptive nature of human thermal perception in outdoor environments [36].

To address these limitations, machine learning (ML) has emerged as a promising complement to conventional modeling approaches [41,42,43]. ML algorithms can uncover complex patterns from large and diverse datasets—including environmental, physiological, and behavioral variables—without relying on predefined physical equations. When properly trained, these models deliver high predictive accuracy, adaptability to varying contexts, and scalability for real-time or city-scale applications [44,45,46]. They are already being used for tasks such as thermal index prediction, spatial microclimate mapping, and simulation emulation via surrogate models.

Nonetheless, significant research gaps remain. First, most ML applications in OTC are fragmented and narrow in scope, typically designed for single-purpose prediction tasks and rarely transferable across climatic or morphological contexts [41,42,43,47]. Second, thermal comfort is strongly influenced by cultural and behavioral factors, which complicates model calibration and limits generalizability [19,48]. Third, many studies continue to rely on static or semi-dynamic approaches, with limited capacity to capture adaptive behaviors, feedback loops, or real-time environmental interactions [41,42,49]. Finally, and most critically, there is a lack of integration between ML and optimization frameworks. While ML provides accurate predictions, these insights are seldom coupled with simulation-based optimization (SBO), surrogate modeling, or multi-objective algorithms capable of informing adaptive, performance-driven design. This disconnect constrains the translation of predictive insights into actionable urban interventions.

To address these gaps, this review provides a systematic and critical assessment of how ML techniques can be applied—and more importantly, integrated with optimization frameworks—to assess, predict, and enhance OTC in urban environments. Specifically, this paper aims to:

• Synthesize ML methodologies used in OTC research, across supervised, unsupervised, deep learning, (DL), semi-supervised, and reinforcement learning paradigms;
• Explore the integration of ML with SBO, surrogate modeling, and scenario-based design strategies;
• Propose a roadmap for scalable, intelligent, and thermally adaptive urban design informed by ML capabilities.

The key innovative contributions of this review are:

(a)

Interdisciplinary synthesis: This study bridges urban climatology, environmental modeling, and artificial intelligence to build a cohesive understanding of ML-driven OTC assessment.

(b)

Systematic classification of ML techniques: A structured taxonomy of ML paradigms is developed, covering theoretical principles, input-output variables, performance metrics, and their contextual application in real-world case studies.

(c)

Integration with optimization workflows: A central innovation of this review is its focus on ML-optimization coupling. Through selected case studies, it highlights how ML can enhance SBO, evolutionary algorithms, and surrogate modeling in support of multi-objective and adaptive design strategies.

(d)

Proposal of a practical framework for OTC optimization: Drawing from state-of-the-art applications, the review formulates a transferable, context-sensitive framework that links climate analysis, parametric modeling, and ML-based optimization. This framework offers guidance on selecting appropriate indices, algorithms, and design parameters to support computationally efficient and implementation-ready OTC interventions across different urban and climatic contexts.

To ensure methodological rigor and relevance, a systematic literature search was conducted using Scopus (including ScienceDirect), Web of Science, and Google Scholar, covering peer-reviewed articles published between 2010 and 2025. Selection criteria included:

(i)

Relevance: priority to studies addressing ML-driven prediction, classification, or optimization of OTC in diverse urban contexts;

(ii)

Quality: inclusion of rigorously validated, peer-reviewed studies;

(iii)

Diversity: coverage of a broad but representative range of ML algorithms and optimization techniques;

(iv)

Applicability: emphasis on case studies demonstrating practical implementation, scalability, or integration into design workflows.

Conference papers, non-English publications, and purely theoretical studies without methodological contributions were excluded.

The rest of the paper is organized as follows: Section 2 introduces and categorizes ML paradigms relevant to OTC; Section 3 synthesizes empirical applications and case studies; Section 4 explores the integration of ML with optimization methods; and Section 5 provides conclusions and perspectives. Together, these sections establish a comprehensive framework for applying ML to develop scalable, adaptive, and climate-responsive design solutions that enhance OTC in urban environments

2. Machine Learning Approaches

ML, a fundamental subfield of Artificial Intelligence (AI), is concerned with the development of data-driven models that can identify patterns and make predictions, without the need for explicitly defined instructions for each possible scenario [50,51]. ML mainly aims to extract patterns from historical data to generate accurate and generalizable predictions on new, unobserved data [52]. A typical ML process includes preparing the dataset, training a model on labeled or structured data, assessing its performance on a separate subset, and applying it to practical tasks such as forecasting, classification, or automated control systems [53].

ML methodologies are commonly grouped into five categories, depending on the nature of the data and the learning mechanism [50,51,52,53,54,55]: (1) supervised learning (SL), (2) unsupervised learning (UL), (3) semi-supervised learning (SSL), (4) deep learning (DL) and (5) reinforcement learning (RL). These paradigms offer diverse strategies for addressing a wide range of tasks, from identifying hidden structures in data to optimizing adaptive decision-making processes (Figure 1).

SL is among the most extensively used ML techniques, particularly when labeled data are available. In this learning framework, the model is trained on input–output pairs, aiming to approximate an underlying function that minimizes the discrepancy between predicted and actual values [52]. This approach is well-suited to both regression tasks, which involve estimating continuous values, and classification problems, where categorical outcomes are assigned to observations. Well-known algorithms in this domain include linear regression (LR), decision trees (DTs), support vector machines (SVM), and artificial neural networks (ANNs). These tools are widely applied in environmental research and human-centric studies for tasks such as modeling climate-related indicators or predicting comfort perception [41,43].

In contrast, UL deals with scenarios where the dataset lacks predefined labels or output values. The focus here is on discovering inherent patterns or relationships within the data, often through techniques like clustering or dimensionality reduction [56]. These methods are instrumental in exploratory analysis, enabling researchers to group similar data points through techniques such as k-means or hierarchical clustering, as well as to reduce the complexity of large, high-dimensional datasets while retaining critical features, employing methods like Principal Component Analysis (PCA). Their utility in environmental applications includes the identification of microclimatic zones or behavioral patterns in human comfort responses [57].

SSL offers a hybrid approach, drawing on both labeled and unlabeled data to train more robust models [56]. By leveraging the information present in a limited labeled subset and the broader context provided by unlabeled data, SSL can enhance prediction accuracy without requiring extensive manual annotation. This is particularly advantageous in outdoor environmental monitoring or thermal comfort studies, where obtaining labeled data (e.g., subjective thermal sensation votes) can be costly or time-consuming.

DL refers to a subset of ML approaches based on multi-layered neural networks capable of learning complex and abstract data representations [53]. Unlike traditional ML models, which often rely on feature engineering, DL architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) automatically learn hierarchical patterns from raw input. DL methods can operate under both supervised and unsupervised paradigms, but their distinguishing feature lies in their ability to handle unstructured and/or high-dimensional inputs such as images, time series, or spatial-temporal environmental data. DL has gained prominence in fields ranging from speech and image recognition to urban climate modeling and environmental sensing [58,59].

Finally, RL adopts a fundamentally different learning strategy. Rather than training on labeled datasets, RL involves an agent that learns by interacting with an environment, receiving evaluative feedback in the form of rewards or penalties [55]. Through trial and error, the agent refines its behavior to maximize cumulative rewards over time. RL is particularly powerful in applications requiring adaptive strategies, sequential decision-making, or dynamic control—ranging from robotics and game theory to emerging uses in real-time building operation and climate-responsive systems.

2.1. Supervised Learning

SL is one of the most established and frequently utilized ML approaches, especially when labeled datasets are available for model training. In this framework, an algorithm is trained on historical datasets comprising both input variables and their corresponding target values. Throughout the training process, the model refines its internal parameters iteratively to minimize a defined loss function that measures the discrepancy between predicted and actual values. Upon convergence, the trained model can be deployed to generate reliable predictions on unobserved data, provided that the training set adequately represents the underlying data distribution [50,51,60].

SL methodologies are broadly categorized into two functional domains: (1) regression, where the goal is to estimate continuous numerical responses, such as UTCI or PET; and (2) classification, which entails assigning discrete labels to input instances, such as thermal comfort categories). Both types are extensively employed in environmental informatics and human-centered thermal analysis, supporting the development of predictive models based on meteorological and subjective input variables.

In the following sub-sections, a selection of widely used regression and classification algorithms is presented. For each method, its theoretical framework is highlighted and its relevance to thermal comfort research is assessed, alongside a discussion of their advantages and limitations when applied to environmental datasets.

2.1.1. Regression Algorithms

Regression algorithms in SL are designed to predict continuous target variables by modeling the underlying relationships between input features and the corresponding outputs. These methods are particularly useful when the goal is to estimate scalar quantities, such as thermal indices, from complex environmental datasets. They form the basis of many predictive models in environmental and urban analytics due to their interpretability, scalability, and applicability to real-world datasets [61]. The most representative regression techniques include linear regression (LR), decision tree regression (DTR), support vector regression (SVR), and gradient boosting regression (GBR). More flexible methods, such as random forest (RF), artificial neural networks (ANNs), and extreme learning machines (ELMs), can also be applied to regression tasks, although they are capable of handling classification problems as well. All of these approaches are reviewed in the remainder of this section.

LR is a fundamental statistical learning technique employed to model the linear relationship between a continuous dependent variable and one or more independent predictors, under the assumption of additive and normally distributed errors. In the case of simple LR, the model is defined as [60,61]:

$$y_i={\beta }_0+{\beta }_1{x}_i+{\epsilon }_i$$

(1)

where $i=1,\dots ,n$ denotes the observations, ${(x}_i,y_i)$are the observed data pairs of the dependent variable $y$ and the single explanatory variable $x$, ${\beta }_0$ and ${\beta }_1$ are the regression coefficients, and $\epsilon$ is an error term with zero mean and constant variance.

When more than one independent variable is available, the model generalizes as follows:

$$y_i={\beta }_0+{\beta }_1{x}_{i1}+{\dots +\beta }_p{x}_{ip}+{\epsilon }_i$$

(2)

where $i=1,\dots ,n$ again denotes the observations. Model parameters are typically estimated using the ordinary least squares (OLS) method, which minimizes the sum of squared differences between the observed and predicted values of the dependent variable [56]. Model performance is commonly evaluated using metrics such as the coefficient of determination ($R^2$), adjusted $R^2$, root mean square error (RMSE), and information-theoretic criteria like the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC) [62,63].

While LR is valued for its simplicity and interpretability, it is based on several assumptions—linearity, independence of errors, homoscedasticity, and normally distributed residuals—that may not hold in complex real-world datasets. In the context of outdoor thermal comfort assessment, LR provides a fast and transparent baseline model, although its limited capacity to capture nonlinear relationships may constrain its predictive performance in highly variable environmental settings.

DTR is a non-parametric supervised learning technique used for predicting continuous outcomes by recursively partitioning a dataset into subsets that minimize intra-node variance [61,64]. At each internal decision node, a binary split of the input data is performed with the objective of minimizing the prediction error. The algorithm selects the feature and the corresponding split point that optimally reduces the prediction error. For a given node, the algorithm evaluates all possible splits across the input features and computes the weighted average variance of the resulting subsets. The most common criterion for splitting is the weighted mean squared error (MSE), defined as [56]:

$$\mathrm{MSE}\left(s\right)={\displaystyle \frac{N_L}{N}}{MSE}_L\left(s\right)+{\displaystyle \frac{N_R}{N}}{MSE}_R\left(s\right)$$

$$\mathrm{MSE}\left(s\right)={\displaystyle \frac{1}{N}}{\sum }_{i\in L}{\left(y_i-{\overline{y}}_L\right)}^2+{\displaystyle \frac{1}{N}}{\sum }_{i\in R}{\left(y_i-{\overline{y}}_R\right)}^2$$

(3)

where $L$ and $R$ denote the left and right child nodes resulting from split s, $y_i$are the observed values, ${\overline{y}}_L$ and ${\overline{y}}_R$ are the mean target values in each subset, and $N_L$ and $N_R$ represent the number of sample sizes in the left and right nodes, respectively. The split that minimizes the combined MSE is selected at each step. This recursive partitioning process continues until a stopping criterion is met, such as a minimum number of samples in a node or a threshold on variance reduction. The resulting DT consists of terminal leaves, each of which corresponds to a specific region of the input feature space. Within each of these regions, the predicted value is computed as the mean of the observed target values that fall into that region.

DTR is a rule-based learning algorithm that partitions data into increasingly homogeneous subsets by selecting optimal decision points, making it well-suited for modeling nonlinear patterns and intricate feature interactions. Unlike many other algorithms, DTR operates effectively without requiring normalization or scaling of inputs, and it can seamlessly manage both numerical and categorical variables. One of the strengths of DTR lies in its transparent decision structure, which facilitates interpretability and supports diagnostic analysis. It also shows resilience to outliers and multicollinearity, traits that are particularly valuable when working with real-world environmental datasets.

In the context of outdoor thermal comfort modeling, DTR has been employed to predict comfort-related indices such as PET and the UTCI by leveraging inputs such as air temperature, relative humidity, wind speed, and solar radiation [46,58]. Its ability to produce interpretable, tree-based decision rules makes it a practical and explainable modeling choice in environmental and urban climate studies [59,65,66].

SVM is a supervised ML technique derived from SVM, tailored for regression tasks. Unlike traditional regression methods that aim to minimize the prediction error directly, SVM’s primary objective is to identify a function $f\left(x\right)$ that approximates the relationship between input features and continuous target values with high generalization performance, while allowing for an acceptable error tolerance $\epsilon$ [64,67,68]. The SVR function is generally expressed as [64]:

$$f\left(x\right)=⟨w,\mathsf{\varphi }\left(x\right)⟩+b$$

(4)

where $\varphi \left(x\right)$ denotes a nonlinear mapping of the input $x$ into a higher-dimensional feature space, $w$ is the weight vector, $b$is the bias term, and $⟨\cdot ⟩$denotes the inner product.

A distinctive feature of SVR is the introduction of the $\epsilon$-insensitive loss function, which defines a margin of tolerance $\epsilon$ around the predicted function $f\left(x\right)$. Errors within this margin are not penalized, and only deviations larger than $\epsilon$ contribute to the optimization objective. The learning process involves minimizing both the model complexity, typically represented by ${‖w‖}^2$, and the prediction errors outside the $\epsilon$-margin.

To handle nonlinear patterns, SVR employs kernel functions—such as the radial basis function (RBF) or polynomial kernels—which enable implicit transformation of the input data into a higher-dimensional space where a LR model can be effectively constructed. Thanks to its balance between flexibility and generalization, SVR has been successfully used in outdoor thermal comfort studies for predicting indices such as PET and UTCI based on meteorological parameters, including air temperature, relative humidity, wind speed, and solar radiation [48,69].

GBR is an advanced ensemble ML strategy that constructs a strong predictive model through the iterative refinement of multiple weak learners—typically shallow DTs [70,71]. Unlike RFs, which employ parallel training via bootstrapped datasets, GBR takes a sequential approach. Each new tree is trained to address the prediction errors (residuals) made by the ensemble up to that point, thereby incrementally improving the model’s performance. The process begins with a baseline model, often a simple constant function such as the mean of the target variable. In each iteration, a new regression tree $h_t\left(x\right)$ is fitted to the residuals, which are derived from the negative gradients of a specified loss function with respect to the current predictions. This gradient-based optimization framework allows GBR to effectively minimize the loss function over time. The ensemble model is updated iteratively as follows [70]:

$$F_t\left(x\right)=F_{t-1}\left(x\right)+\nu h_t\left(x\right)$$

(5)

where $F_t\left(x\right)$ represents the model at iteration t, $h_t\left(x\right)$ is the newly fitted tree, and $v\in \left(0\right.,\left.1\right]$ is the learning rate that controls the contribution of each tree.

GBR provides considerable modeling flexibility by accommodating a wide range of loss functions—such as MSE for regression—and incorporating regularization mechanisms to control model complexity. A critical hyperparameter of GBR is the learning rate, since lower learning rate values (e.g., 0.01–0.1) typically lead to stable and generalized models, though they demand higher computational effort due to the increased number of boosting iterations required. Conversely, higher learning rates (e.g., 0.3–0.5) accelerate convergence but may compromise predictive performance by increasing the risk of overfitting. Techniques like shrinkage (via a learning rate) and constraints on tree depth help prevent overfitting, particularly when working with limited or noisy datasets.

In outdoor thermal comfort research, GBR has demonstrated strong predictive capabilities for indices such as the PET and UTCI [72]. Its ability to learn complex, nonlinear patterns and capture subtle interactions among meteorological inputs—such as air temperature, humidity, solar radiation, and wind speed—makes it especially valuable in applications involving diverse and heterogeneous environmental conditions. These attributes have positioned GBR as a reliable and high-performing method in climate-responsive predictive modeling.

Random forest regressor (RFR) is an ensemble-based ML technique that builds upon the principles of decision tree learning to improve predictive accuracy and model robustness. In this approach, multiple regression trees are trained in parallel, each on a different bootstrap sample drawn from the original dataset. To further diversify the ensemble, a randomly selected subset of input features is evaluated at each potential split within a tree [73,74,75,76]. This dual-randomization—both in data sampling and feature selection—serves to reduce the variance inherent in individual DTs, thereby enhancing the model’s generalization capabilities and resistance to overfitting. By aggregating the outputs of multiple trees, typically through averaging in the regression context, the final prediction benefits from reduced noise and improved stability. RFR is particularly effective in handling high-dimensional, nonlinear datasets with complex interactions and has been widely adopted in environmental modeling and climate-responsive prediction tasks [77,78].

Formally, for a dataset ${\left\{\left({\mathit{x}}_{\mathit{i}},y_i\right)\right\}}_{i=1}^N$, RFR constructs $T$ trees ${\left\{h_t\left(\mathit{x}\right)\right\}}_{i=1}^T$, and the predicted output $\widehat{y}$ for a new input $\mathit{x}$ is [17]:

$$\widehat{y}={\displaystyle \frac{1}{T}}{\sum }_{t=1}^T{h}_t\left(\mathit{x}\right)$$

(6)

where $h_t\left(\mathit{x}\right)$ denotes the output of the $t$-th DT. This averaging mechanism yields a smoother and more stable prediction compared to an individual tree, particularly in the presence of noisy data or high-dimensional input spaces.

The ensemble averaging mechanism at the core of RFR contributes to its reliability by mitigating the sensitivity of individual DTs to data noise or variance. This collective decision-making approach enhances prediction stability, particularly in datasets characterized by high dimensionality or noisy measurements. A key advantage of RFR lies in its ability to model nonlinear relationships and variable interactions without extensive parameter tuning, while simultaneously offering tools for interpretability—such as feature importance rankings—which are valuable for understanding model behavior. Because of these strengths, RFR has gained traction in environmental and climate-related modeling, where data complexity and heterogeneity are common. Within the field of outdoor thermal comfort, RFR has been effectively used to estimate indices like PET and UTCI from meteorological predictors. Its consistent performance across varying climatic contexts [79] underscores its robustness and versatility as a predictive modeling tool in urban thermal analysis.

ANNs are algorithmic structures inspired by the architecture and functionality of the human brain, designed to simulate the way biological neurons transmit and process signals. These models have become central to many supervised learning applications, particularly those involving complex and highly nonlinear mappings between input features and target variables. One of the most widely used ANN architectures is the feedforward neural network (FFNN), also known as the multilayer perceptron (MLP) [80,81]. FFNNs are composed of layers of interconnected processing units, or neurons, through which information propagates in a single direction [52,53,81,82]. Figure 2 illustrates such an architecture, consisting of an input layer ($x_1,x_2,\dots ,x_n$), a number of hidden layers, two of which are shown ($h_1,h_2,\dots ,h_{n1}$ and $g,g_2,\dots ,g_{nl}$), and an output layer ($y_1,y_2,\dots ,y_m$).

In an ANN, each neuron operates by calculating a weighted sum of its inputs, adding a bias term, and then applying a nonlinear activation function to the result. For a single-layer perceptron, this operation is typically formulated as [75]:

$$y=\sigma \left({\sum }_{i=1}^n{w}_i{x}_i+b\right)$$

(7)

where, $y$ is the neuron’s output, $\sigma \left(\right)$ denotes the activation function, $x_i$ are the input features, $w_i$ are the corresponding weights, $b$ is the bias term, and $n$ is the number of input nodes.

Training an ANN involves adjusting the weights and biases to reduce the discrepancy between predicted and actual outcomes. This is commonly achieved using the backpropagation algorithm, which computes the gradient of the loss function with respect to each parameter and applies optimization techniques such as stochastic gradient descent to update the network iteratively [83]. Because of their flexible, layered structure and capacity to approximate complex nonlinear mappings, ANNs are especially suitable for predictive modeling in environmental systems. In outdoor thermal comfort studies, they have been successfully utilized to estimate thermal indices like PET and UTCI based on meteorological data [44,84,85,86]. Their adaptability and high predictive accuracy make them well-suited for use in diverse urban environments with variable climatic conditions.

Finally, ELMs represent a class of single-layer FFNNs designed for high-speed learning. Unlike traditional ANNs, ELMs assign input weights and biases randomly and compute output weights analytically, which eliminates iterative tuning and significantly reduces training time. Despite their simplicity, ELMs have shown strong performance in regression tasks, including predicting thermal sensation votes in outdoor environments, making them suitable for real-time applications [87].

2.1.2. Classification Algorithms

Classification is a core task in supervised learning that involves assigning input data to one of several predefined categories based on patterns extracted from labeled examples. This technique is particularly relevant in environmental modeling and human-centric studies, where outcomes such as thermal sensation categories (e.g., comfortable vs. uncomfortable), heat stress levels, or subjective perception classes are inferred from sensor measurements or survey responses [52,60].

Several well-established classification algorithms are commonly used in this context:

• Decision tree classification (DTC) is a non-parametric technique that segments the feature space by recursively applying decision rules. It is valued for its clarity and ability to accommodate both categorical and continuous input variables [88].
• SVM models seek to identify the hyperplane that best separates data points belonging to different classes by maximizing the margin between them. Through the use of kernel functions, SVMs are well-suited to problems with complex or nonlinear class boundaries [64,67].
• K-Nearest neighbors (KNN) is an intuitive, non-parametric classifier that assigns labels based on the majority class among the k closest data points in the feature space. While simple to implement, KNN is sensitive to noise and feature scaling, and can be computationally demanding for large datasets [89].
• Naive Bayes (NB) is a probabilistic classifier that applies Bayes’ theorem under the assumption of feature independence. Despite its simplistic assumptions, NB often performs competitively in high-dimensional or sparse data environments [90].
• Random forest classifier (RFC) is an ensemble learning approach that combines the predictions of multiple DTs through majority voting. RFC reduces variance and mitigates overfitting, offering improved stability over single-tree models [73].
• Finally, ANNs can also be applied to classification tasks, where they learn nonlinear decision boundaries through a layered structure of interconnected neurons. They are particularly effective in complex, high-dimensional classification problems where traditional models may underperform [53,81].

Each classification method offers distinct trade-offs in terms of interpretability, computational cost, scalability, and resilience to noise. The optimal choice often depends on the characteristics of the dataset and the specific classification objectives involved.

2.2. Unsupervised Learning

UL encompasses a class of ML methods that aim to uncover latent patterns, internal structure, or meaningful groupings within datasets that do not contain labeled outputs. Unlike supervised approaches, UL models operate solely on input data, identifying relationships or organizing data points based on their inherent similarities. These techniques are essential in exploratory data analysis, anomaly detection, and feature extraction, where interpretability and reduction of data complexity are often required. In environmental and human-centric studies, UL has proven valuable for tasks such as classifying occupant behavior, recognizing perceptual clusters in comfort surveys, or identifying distinct microclimatic regimes using sensor-based observations [53,75,91,92,93,94,95,96]. Several widely adopted ul techniques are discussed in the following paragraphs.

K-means clustering is a distance-based sequential method that partitions a dataset into k predefined non-overlapping clusters by minimizing the intra-cluster variance. The squared Euclidean distance is the most frequently used metric in the k-means algorithm; however, alternative distance measures such as Mahalanobis, Manhattan, and Chebyshev can also be utilized depending on the type of the data. The choice of distance metric can significantly influence the resulting clustering. Each observation is assigned to the cluster with the nearest mean (centroid), and the algorithm iteratively updates cluster assignments and centroids to reduce the following objective function [93]:

$$J={\sum }_{i=1}^k{\sum }_{x\in C_i}(x-{\mathsf{\mu }}_i{)}^2$$

(8)

where, $J$ is the total within-cluster sum of squares, $k$ is the number of clusters, $C_i$ is the set of points in cluster $i$, ${\mu }_i$ is the centroid of cluster $i$ and $x$ is a data point in cluster $i$. A common convergence criterion of the algorithm is that the cluster centroids remain unchanged across successive iterations.

Hierarchical clustering is an unsupervised learning technique that constructs a hierarchy of clusters without requiring the number of clusters to be predefined. It has two main forms: agglomerative (bottom-up), where each observation starts as its own cluster and clusters are merged iteratively based on similarity; and divisive (top-down), which begins with all observations in a single cluster and recursively splits them. The method relies on distance metrics such as Euclidean, Manhattan, or Mahalanobis distance for calculating the distances between observations, and linkage criteria such as simple (nearest neighbor), complete (farthest neighbor), average, Centroid, or Ward’s linkage to determine the proximity between clusters. The result is a dendrogram that illustrates cluster hierarchy at multiple scales. This method is particularly insightful in climate analytics, where multi-scale structures—such as regional comfort patterns or temporal behavioral trends—must be explored [97].

PCA is a method of identifying patterns in data and re-expressing the data in a way that highlights the underlying similarities and distinctions among observations. It is a linear technique for dimensionality reduction that projects high-dimensional data into a new coordinate system defined by uncorrelated principal components. These components are linear combinations of the initial variables, ranked by the proportion of variance they capture. The transformation is expressed as [95]:

$$Z=XW$$

(9)

where, $X$ is the mean-centered data matrix, $W$ contains the eigenvectors of the covariance matrix, and $Z$ represents the transformed data. PCA is widely applied to reduce complexity in meteorological and comfort datasets, while preserving most of the variance [98].

T-distributed stochastic neighbor embedding (t-SNE) is a nonlinear dimensionality reduction algorithm aimed at preserving local data structure when projecting high-dimensional data into two or three dimensions. It is particularly effective for visualizing complex datasets where traditional linear methods like PCA may fail to preserve neighborhood information. In comfort studies, t-SNE has been utilized to reveal latent clusters in perception surveys or to explore spatial patterns in urban microclimates [53,94].

Autoencoders are a type of neural network trained in an unsupervised manner to compress and reconstruct input data. They consist of an encoder that maps the input $x$ to a latent representation $z$, and a decoder that reconstructs the input from this compressed form. The overall transformation can be represented as: [92,96]:

$${\mathit{x}}^{\prime }=f_{\mathrm{decoder}}\left(f_{\mathrm{encoder}}\left(\mathit{x}\right)\right)$$

(10)

where $\mathit{x}$ is the original input data vector, $f_{\mathrm{encoder}}$ is the encoder function that compresses the input into a latent representation, $f_{\mathrm{decoder}}$ is the decoder function that reconstructs the input from the latent code, and ${\mathit{x}}^{\prime }$ is the reconstructed approximation of the input. During training, the network learns to retain the most relevant features of the input by optimizing a loss function such as the MSE between $\mathit{x}$ and $\mathit{x}\prime$. Autoencoders are particularly valuable for dimensionality reduction, denoising, and unsupervised feature extraction. In microclimate modeling and environmental studies, autoencoders have been employed to generate compact representations of complex sensor data, enabling improved clustering, pattern recognition, or subsequent predictive modeling [99].

2.3. Semi-Supervised Learning

SSL occupies a middle ground between supervised and unsupervised learning, leveraging a limited set of labeled examples together with a much larger collection of unlabeled data to train more effective models. This hybrid approach is particularly advantageous in fields such as environmental sensing, urban microclimate analysis, and human thermal comfort studies, where collecting raw data is relatively straightforward, but obtaining ground truth labels is time-consuming or resource-intensive. By utilizing the latent structure within the data, SSL models can achieve better generalization performance compared to supervised methods trained solely on small labeled datasets [54,100].

A variety of SSL methodologies have been proposed. Some of the most commonly applied include self-training, co-training, graph-based approaches, consistency regularization, and generative models. In self-training, a supervised model is initially trained on the available labeled data. It then generates pseudo-labels for the unlabeled samples, and those predictions with the highest confidence are gradually added to the training set. This iterative process refines the model with increasingly enriched data [101]. Co-training involves training multiple classifiers on different, ideally conditionally independent subsets of features. Each model assigns labels to unlabeled examples, which are then used to train the other classifiers. This mutual reinforcement encourages complementary learning and can improve robustness [102]. In graph-based approaches, data points are represented as nodes in a graph, with edges weighted according to similarity metrics. Label information is propagated from labeled to unlabeled nodes using principles of transductive learning, which focuses on predicting labels only for the given unlabeled data rather than learning a general model, enabling label inference through relational structure [103]. Consistency regularization encourages the model to produce consistent predictions even when small perturbations (such as noise, augmentation, or dropout) are applied to the input data. By enforcing smooth decision boundaries, the algorithm becomes more robust to data variability [104]. Finally, in generative models, semi-supervised variational autoencoders (VAEs) and related generative techniques integrate unsupervised representation learning with supervised objectives, using latent variables to encode data structure in ways that benefit classification or regression tasks [105].

Although no current OTC-specific studies have adopted SSL frameworks, the method’s ability to reduce labeling requirements while maintaining predictive accuracy renders it highly promising for future applications. As cities deploy dense sensor networks and physiological data collection become more accessible via wearable devices, SSL is likely to play a pivotal role in advancing scalable, data-efficient, and context-aware thermal comfort modeling.

2.4. Deep Learning

DL refers to a class of ML techniques grounded in ANNs with multiple hidden layers. These multilayered structures can learn directly from raw data by extracting progressively more abstract features at each layer. DL models are particularly effective for high-dimensional, nonlinear problems, as they enable automatic feature learning and hierarchical representation of data [53,106,107]. Several prominent DL architectures have been applied in the environmental sciences and outdoor thermal comfort domain, each suited to different types of data and learning objectives [107], and are briefly described below.

MLPs are the most straightforward DL models, comprising fully connected layers where each neuron is connected to all neurons in the subsequent layer. They are suitable for tasks such as classification or regression when the dataset lacks spatial or sequential structure. Despite their simplicity, MLPs can approximate complex functions and are widely used for predicting thermal indices when the feature space is moderate [52,53,107].

CNNs are specialized for data with spatial structure, such as images or gridded environmental data. They use localized filters (kernels) that slide across the input to extract features like edges, gradients, or textures. This makes CNNs especially powerful for spatial prediction tasks, including urban microclimate mapping. The convolution operation is typically defined as [107,108]:

$$F_{i,j}={\sum }_m{\sum }_n\left(I_{i+m,j+n}·K_{m,n}\right)+b$$

(11)

where $F_{i,j}$ is the output at spatial location $\left(i,j\right)$, $I$ represents the input image or feature map, $K$ is the convolutional kernel (or filter) of size $m\times n$, $I_{i+m,j+n}$ denotes the region of the input over which the filter is convolved, $b$ is the bias, and ${\Sigma }_m{\Sigma }_n$ denotes summation over the dimensions of the kernel.

Recurrent neural networks (RNNs) are designed to process sequential data and capture temporal dependencies, making them ideal for time-series predictions in microclimatic and energy demand applications [53,107,109]. A mathematical expression representing the recurrent relationship that defines how the hidden state $h_t$ is updated at each time step $t$, based on the input $x_t$ and the previous hidden state $h_{t-1}$, can be written as follows [53,107]:

$$h_t=f\left(W_h{h}_{t-1}+W_x{x}_t+b_h\right)$$

(12)

where $h_t$is the hidden state at time $t$, $W_h$ is the weight matrix applied to the previous hidden state, $W_x$ is the weight matrix applied to the current input $x_t$, $b_h$ is the bias term for the hidden layer, and $f$ is the activation function. In contrast to FFNNs, which process inputs independently, RNNs utilize feedback loops that allow them to capture temporal dependencies. This makes them essential in domains involving climate dynamics and behavioral sequences [83,107]. Several types of RNNs have been developed [110], such as Elman RNN [111], Jordan RNN [112], LSTM [107,110,113] and the Gated Recurrent Unit (GRU) [114].

Deep belief networks (DBNs) are composed of multiple layers of Restricted Boltzmann Machines [115], trained in a layer-wise unsupervised manner followed by supervised fine-tuning. Each layer extracts increasingly abstract representations of the data, enabling DBNs to capture intricate patterns. Their hybrid training process makes them particularly effective in scenarios with limited labeled data but large volumes of raw inputs. In environmental modeling, DBNs have been used for feature reduction, anomaly detection, and preprocessing sensor data [116,117].

Finally, generative adversarial networks (GANs) are a relatively recent innovation in the field of DL, particularly suited for unsupervised and semi-supervised tasks. They comprise two neural networks, the generator and the discriminator, that are trained simultaneously in a competitive framework. The generator is tasked with producing synthetic data that resemble real observations, while the discriminator attempts to distinguish between genuine data samples and those generated artificially. This adversarial interplay drives both models to improve over time, resulting in the generator learning to approximate the underlying data distribution effectively. The training objective is structured as a minimax game, where the generator aims to minimize the probability of the discriminator correctly identifying fake samples, and the discriminator seeks to maximize its accuracy. Through this iterative optimization, GANs are capable of producing highly realistic data outputs even in complex, high-dimensional spaces [118].

2.5. Reinforcement Learning

RL represents a distinctive branch of ML focused on training agents to make optimal decisions through trial-and-error interactions within an environment. Unlike supervised learning, which relies on annotated datasets, RL operates without direct instruction. Instead, the agent autonomously navigates the environment by selecting actions and receiving evaluative feedback in the form of rewards or penalties based on the consequences of those actions. Over time, the agent aims to refine its behavior by discovering a policy—a mapping from states to actions—that yields the maximum cumulative reward. This objective often involves balancing two competing strategies: exploration (testing new actions to gather information) and exploitation (choosing actions known to produce high rewards) [55].

RL is particularly effective for tackling problems that require sequential decision-making in environments where the system dynamics may be uncertain or change over time. Its ability to learn optimal actions based on feedback makes it suitable for complex, real-time scenarios. Beyond its established use in robotics and autonomous control systems [119,120] RL is gaining momentum in applied environmental and urban research [121,122]. It has been explored for smart building management, including dynamic HVAC control, adaptive regulation of outdoor thermal comfort, and optimization of microclimatic conditions in urban spaces [123,124]. These applications benefit from RL’s capacity to continuously adjust strategies based on evolving environmental data, enabling more responsive and energy-efficient interventions.

Formally, RL problems are often modeled as Markov Decision Processes [55], defined by a tuple $(S,A,P,R,\gamma )$ where $S$ is the set of states; $A$ is the set of actions; $P\left(s^{\prime }\right|s,a)$ is the state transition probability from state $s$ to $s\prime$ given action $a\in A$ and $s$, $s\prime \in S$; $R(s,a)$ is the reward function; and $\gamma \in \left[\mathrm{0,1}\right]$ is the discount factor that balances immediate and future rewards [125]. At each time step $t$, the agent perceives its current state $s_t$ taking an action $aₜ$, moving to a subsequent state $sₜ₊₁$, and receiving a reward signal $rₜ$. Using this sequence of experiences, the agent continually adjusts its policy $\pi \left(a\right|s)$, which defines the probability or strategy of choosing action $a$ when in state s.

RL algorithms are broadly categorized into two main classes [55,119]. Value-based methods, such as Q-learning and deep Q-Networks (DQN), estimate a value function $Q(s,a)$ that represents the expected return of performing action $a$ in state $s$. The agent then chooses actions that maximize this estimated value. Policy-based methods, such as REINFORCE and actor–critic algorithms, directly optimize the policy $\pi \left(a\right|s)$ without explicitly computing a value function. These approaches are especially suited for continuous or high-dimensional action spaces.

Recent studies have shown RL’s potential in climate-responsive applications, especially where adaptation and automation are key [121]. For example, RL has been deployed to optimize thermal control in smart buildings [123,124], adaptive shading systems [126,127], and real-time energy management frameworks [128,129]. Although its application in OTC remains largely theoretical, the method’s capacity to autonomously learn context-specific control strategies from continuous environmental feedback makes it a compelling candidate for outdoor thermal regulation. Nonetheless, several challenges persist. RL algorithms are often sample-inefficient, requiring extensive interactions to converge on optimal policies. They can also be sensitive to hyperparameter tuning and may exhibit instability in non-stationary environments. Despite these limitations, RLs ability to improve decision-making in complex, real-world systems underscores its promise as a foundation for intelligent outdoor comfort optimization.

Table 1 summarizes the main advantages and disadvantages of commonly used ML techniques. Broadly speaking, the key challenges associated with ML techniques can be grouped into two principal categories: (a) the complexity of algorithmic assumptions and model interpretability, and (b) the computational cost and data requirements associated with training and validation. These trade-offs become particularly important when selecting appropriate algorithms for environmental modeling tasks, where accuracy, scalability, and transparency are often competing priorities.

3. Real-World Applications

Over the last decade, the integration of ML techniques into OTC research has advanced rapidly, presenting robust alternatives to conventional thermal indices and static empirical formulations [41,131]. Traditional models such as the PMV, PET, and UTCI have long provided standardized methods for evaluating thermal comfort in outdoor settings [21,32]. Nevertheless, these indices often rely on steady-state assumptions, limited physiological representations, and predefined exposure scenarios, which may not fully account for complex environmental interactions or inter-individual variability in thermal perception [36,132]. Furthermore, their applicability under rapidly changing urban microclimates or diverse demographic contexts remains constrained. In contrast, ML models—particularly those employing SL and DL architectures—offer adaptive, nonlinear, and scalable frameworks capable of learning from empirical data and integrating multifactorial influences such as urban morphology, behavioral patterns, and real-time weather dynamics [36,41,42,43,132,133,134]

ML models are uniquely suited to unravel the intricate, nonlinear interdependencies among environmental variables, human physiology, and behavioral adaptation, thereby enabling real-time, high-resolution predictions of OTC across diverse urban contexts [48,86,135,136]. The growing proliferation of environmental sensor networks, wearable biometric devices, remote sensing platforms, and three-dimensional urban morphometric datasets has further enhanced the applicability of ML in OTC modeling. Among SL techniques, algorithms such as RF [44,45,47,59,137,138], SVM [139,140], Gradient Boosting Machines (GBMs) [46,47,59] have demonstrated strong predictive performance in classifying thermal sensation categories and estimating index-based comfort metrics such as PET and UTCI. In parallel, DL architectures—including ANNs with multiple hidden layers [59,141,142], CNNs [58,143], and hybrid models [138]—have shown increasing utility in capturing spatiotemporal dynamics and learning abstract representations of thermal conditions from high-dimensional, multimodal datasets. These advancements highlight the capacity of ML to support individualized, context-aware assessments within complex outdoor environments.

Although ML models have achieved commendable predictive accuracy in OTC applications, their adoption in this domain is still emerging when compared to the more established field of indoor environmental modeling. Most research efforts to date have focused on SL and DL, often adapted from methods originally developed for indoor conditions or environmental forecasting. Notably, RL—which has demonstrated substantial potential in building automation and adaptive HVAC control—remains underexplored in outdoor contexts, where dynamic and less-controllable environmental variables present additional modeling challenges.

Table 2 (for SL), 3 (for UL), and 4 (for DL) consolidate a comprehensive overview of recent studies that apply supervised, unsupervised, and DL algorithms to the prediction and classification of OTC conditions.

The reviewed SL studies in Table 2 were conducted across diverse regions, including Europe (Germany, Spain, Switzerland, Serbia, Cyprus), North America (United States), the Middle East (Iran), and Asia (China, Hong Kong, India, Taiwan, Singapore, Japan). Diverse urban contexts were examined, including high-density cities and metropolitan areas, public squares, pedestrian streets, urban canyons, various types of urban parks and open spaces, shaded areas, green sidewalks, and courtyards.

A wide range of methods was applied, including tree-based models (RF and its extensions such as quantile regression forest, DT, XGBoost, AdaBoost, LightGBM, CatBoost with Bayesian optimization, GB, GBT), linear and probabilistic models (LR, multinomial logit and ordered probability models, Bayesian ridge regression, least angle regression, logistic regression, NB, regression kriging), neural networks (ANN, including MLP, FFNN, BPNN), DL approaches, kernel-based methods (SVM, SVR), instance-based methods (KNN), and dimensionality reduction and clustering techniques (PCA, k-means), occasionally combined with SHAP analysis for improved model interpretability. SL algorithms—most notably RF, SVM, and XGBoost—have emerged as dominant tools in OTC modeling, owing to their robustness, interpretability, and efficiency in handling structured environmental datasets. These models are frequently employed to predict subjective thermal responses such as TSV, TCV, and TA, as well as to estimate objective comfort indices including the PET and UTCI.

The reviewed studies employed diverse input parameters, encompassing demographic and socioeconomic (age, gender), meteorological and environmental variables (time of day, season, air temperature, relative humidity, wind speed, solar and globe radiation, mean radiant temperature, PMV, PET), urban form and geometry metrics (elevation, slope, AR, orientation, sky view factor, canyon surface materials, tree metrics, green coverage, building density, albedo, emissivity), remote sensing indices (NDVI, NDWI, NDBI, LULC categories), and other personal factors (clothing insulation, metabolic rate, activity level, BMI, emotional state). Some studies incorporated physiological measurements (skin and ear temperature, heart rate) using sensors or infrared thermography, smartphone-based data (temperature readings, GPS), and questionnaire-derived variables (reason for visiting, frequency of attendance). Several studies also used simulated or modeled data (ENVI-met, Rayman) to compute thermal comfort indices and enhance the interpretation of field measurements.

Outputs included widely recognized thermal comfort indices—most notably PMV and UTCI, which are considered gold-standard metrics—together with PET, mPET, NET, and THI, representing additional physiologically based or empirical indicators of thermal stress. Environmental variables such as MRT, LST, and high-resolution Ta and RH fields (100–200 m) were also extracted to support detailed microclimatic analysis. Subjective responses were captured primarily through TSV, the benchmark ASHRAE −3 to +3 thermal sensation scale, complemented by TCV and TAV, which assess perceived comfort and acceptability, respectively. For statistical analysis, these responses were occasionally recoded into binary comfort classes (comfortable/uncomfortable) or aggregated into three broader categories—cool discomfort, neutral comfort, and warm discomfort. These subjective metrics served as dependent variables in ML models designed to evaluate outdoor and courtyard thermal performance.

Turning to different modeling approaches, model performance was consistently high across the reviewed studies, with reported classification accuracies often exceeding 85%, and regression metrics demonstrating strong predictive reliability—e.g., coefficient of determination (R²) values above 0.90, and MAPE below 5%. ML models, including ELM, RF, SVM, ANN, and XGBoost, consistently outperformed classical approaches in predicting TSV, PET, UTCI, and MRT. ELM achieved highest accuracy for MRT (R²=0.99) and TSV (R²=0.94), while SVM effectively predicted thermal sensation categories using local skin temperatures. Notably, several models retained high performance even when trained with reduced or simplified input feature sets, underscoring the efficiency and adaptability of these algorithms in real-world urban microclimatic applications. RF highlighted key physiological and seasonal predictors, and XGBoost often provided top OTC prediction accuracy, with SHAP analyses revealing nonlinear impacts of environmental, physiological, and design factors. ANN and DL models delivered accurate hourly PET and UTCI predictions, supporting applications with limited microclimatic inputs. As regards statistical and classical models, ordered probability and multinomial logit models outperformed LR in TSV prediction, with multinomial logit better for individual predictions and ordered probability capturing distributions. GA-trained binary-linear classifiers successfully predicted thermal comfort using PET thresholds, and regression kriging with MLR achieved high urban air temperature accuracy, although performance decreased under high variability.

Furthermore, including physiological data such as heart rate, skin temperature, and facial expressions improved OTC predictions, with children’s TSV predicted at 97.1% and real-time preferences at 66% accuracy. Individual factors like BMI, respiratory rate, and metabolic rate influenced comfort outcomes, particularly under varying shading, exposure duration, and activity levels. Urban landscape and environmental design strongly affected LST and thermal comfort. Built-up and vegetated areas reduced LST at boundaries, while urban expansion intensified UHI. Courtyard and sunken square parameters (including SVF, slope, AR, orientation, and vegetation) significantly influenced PET, with greening often more effective than material choices. Green exposure in subtropical environments further enhanced OTC, especially under moderate air velocity and SVF. High-resolution datasets and crowd-sourced sensors enabled accurate urban temperature and OTC mapping at city-wide and one-meter resolutions, capturing fine spatial–temporal variations and nighttime heat stress. Simplified sensor inputs retained predictive power, demonstrating scalable, cost-efficient alternatives to CFD simulations for urban planning and heat mitigation.

Finally, regarding index performance and predictive insights, PET and UTCI remained widely used indices, with ML improving prediction over PET-based methods, particularly for extreme and complex conditions. TSV predictions were generally reliable, whereas TCV showed greater variability due to subjectivity. Shading, exposure duration, and physiological adaptation strongly influenced accuracy. Regarding computational efficiency and model optimization, using GA, PCA, k-means clustering, and Bayesian methods enhanced ML performance while reducing computational demands, decreasing training dataset sizes by up to 70% without losing accuracy (MAPE <1% for MRT). ML approaches consistently captured complex spatial, temporal, and extreme thermal patterns more effectively than classical methods, enabling practical, human-centered, climate-responsive urban design.

The two UL studies presented in Table 3 were conducted in Tehran, Iran, using satellite imagery and ground-based climate data, and in Taipei, Taiwan, using meteorological big data from a dense urban area over summer months (2011–2020). PCA was used in Tehran to reduce dimensionality and inter-correlation among surface and meteorological features, while k-means clustering was applied in Taipei to categorize rainfall patterns and assess their effects on AT and PET. Input data included Landsat-derived LST, NDVI, NDWI, land cover, DEM, and in-situ meteorological measurements in Tehran, and hourly temperature, relative humidity, wind speed, precipitation, and cloud cover in Taipei. Outputs comprised a composite comfort index based on principal components calibrated to the DI in Tehran, and AT and PET under varying rainfall scenarios in Taipei. The PCA model in Tehran showed strong predictive performance, with DI higher in the warm season and bare land most uncomfortable, while in Taipei, summer rainfall influenced urban outdoor thermal comfort, improving it in just over one-third of events and worsening it in just under two-thirds.

Table 4. DL applications in OTC studies.

Reference	Location/Investigation method	DL Method	Input Parameters	Output Parameters	Main Results
[143]	Simulation-based study using 2,400 synthetic residential block configurations generated in Grasshopper, with building layouts inspired by Shenzhen planning guidelines (China)	Deep CNNs were tested in 12 model variations with different layer depths and kernel sizes, with the best model using six convolutional layers (3×3 kernels) followed by fully connected layers	Building geometry and meteorological data, including Ta, RH, and wind speed	UTCI	The best model (6-layer CNN, Group B) achieved R2 = 0.960 and MSE = 0.022, significantly outperforming ANNs; spatially invariant convolutional layers enhanced generalization, and hidden-layer visualization revealed how spatial features were extracted for UTCI prediction
[141]	Tianjin, China, where field measurements and questionnaire surveys were conducted in four Urban Blue-Green Infrastructure (UBGI) spaces during the winter	MLP with feed-forward back-propagation using hyperbolic tangent activation in the hidden layer and softmax activation in the output layer	Time of day, distance from water, and air temperature at waterfront greenery sites, based on 3,848 samples collected from eight sites on December 19, 2021	Heat Island Intensity (HII), Coupling Effect Intensity (CEI), and UTCI	The ANN model showed high predictive performance, with relative area under the ROC curve 95.3–99.2% for HII and 72.0–95.9% for CEI; air temperature was the most influential factor for HII and time of day for CEI, supporting the development of a Microclimate Influence model and identifying an optimal thermal comfort range of Ta=9.07-14.75 °C
[58]	Xi’an, China, where real-time field measurements, thermal imaging, and questionnaires were conducted in an urban park with 405 children (aged 7–14) participating under three thermal stress levels and three activity intensities	Deep CNN was used for facial expression recognition and combined with RFR to predict children’s TCV	Facial expression images captured via real-time video, individual characteristics (age, gender), environmental parameters (air temperature, globe temperature, relative humidity, wind speed), physical activity levels, and clothing insulation	Children’s TCVs on a 7-point scale (−3=cold to +3=hot), used as ground truth for model prediction and validation	The CNN effectively extracted emotional features from children’s facial expressions, which, when combined with RF, yielded R2 = 0.825 for TCV prediction; the DL model outperformed manual scoring methods and traditional ML classifiers, demonstrating potential for real-time, non-contact outdoor thermal comfort assessment based on facial images
[59]	Shiraz, Iran, where a spatiotemporal analysis of the UHI was conducted using 2006–2021 Landsat imagery and urban configuration metrics, with LST predictions across built-up, soil, and vegetation areas	A Deep Neural Network (DNN) with three hidden layers was trained on geospatial features to predict LST, with performance compared against five other ML models, including RF, XGBoost, and KNN	Normalized indices (NDVI, NDBI, NDWI), LST, LULC, elevation, slope, aspect, proximity to roads and water bodies, and climatic parameters (Ta, humidity, and wind speed)	LST	Among six ML models tested, the DNN achieved the highest predictive performance for LST (R2 = 0.94, RMSE = 1.71 °C), accurately capturing spatial LST distributions across urban surfaces, with vegetation and NDVI as the most influential features, supporting DL-based urban heat modeling for climate-resilient planning
[138]	Freiburg, Germany, where citywide thermal comfort was modeled using urban sensor network data, high-resolution GIS layers, and ML emulation of numerical climate models over 2018–2022	A hybrid DL human thermal comfort neural network (HTC-NN) combined two MLPs (each with three hidden layers) for air temperature and relative humidity, a U-Net CNN for MRT, and RF for wind speed estimation	Meteorological parameters (Ta, RH, wind speed, radiation, precipitation), geospatial data (land cover, building and vegetation height), and surface characteristics (SVF, albedo, population density)	UTCI at 1×1 m spatial resolution, derived from ML-predicted Ta, RH, wind speed (U), and MRT fields	HTC-NN achieved UTCI prediction accuracy of RMSE = 3.0 K and R2 = 0.92 against street-level sensor data, outperforming the numerical model
[142]	Hong Kong, where in situ microclimatic monitoring and questionnaire surveys were conducted in three urban parks across summer and winter, and neural network models were trained on subjective and objective variables to predict thermal comfort	ANN with two hidden layers, optimized via the Levenberg–Marquardt algorithm, and trained separately for summer (17 input neurons) and winter (14 input neurons)	Microclimatic variables (Ta, MRT, wind speed, RH, solar radiation), physiological variables (clo, MET), psychological inputs (thermal, solar, wind, humidity sensations), perceptions of park features bodies(trees, shade, water ), and personal traits (age, purpose of visit, thermal sensitivity)	Thermal Comfort Evaluation: self-rated comfort score from a field survey on a continuous scale	The optimized ANN with two hidden layers achieved R2 = 0.653 (summer) and R2 = 0.771 (winter) on the validation set; the DL model significantly outperformed PMV and PET models, with inclusion of perception and psychological variables improving prediction accuracy by over 30%
[151]	A multi-site analysis using data from 43 previously published OTC studies conducted across diverse climate zones worldwide	ANN models with three hidden layers, trained using backpropagation, were applied to predict TSV based on microclimatic and macroclimatic variables across global datasets	Meteorological parameters (Ta, RH, wind speed, globe temperature), subjective variables (clothing insulation, metabolic rate, gender), microclimatic modifiers (shade, surface material, tree coverage), and macroclimatic classification (climate zone index)	PET	ANN models with three hidden layers successfully predicted PET based on macro- and microclimatic variables across diverse urban settings; analysis showed that macroclimate factors (latitude, altitude, distance from the sea) and microclimate features (albedo, H/W, SVF, LAI) contributed similarly to OTC, with PET in equatorial regions ~13 °C higher than in polar regions during the summer, highlighting the need to integrate microclimate-sensitive design in cities with unfavorable macroclimates
[152]	Tehran, Iran, where image-based DL models were developed using ENVI-met simulation data	DL using a conditional GAN (pix2pix)	Urban geometry, greening configuration, and façade materials	UTCI map simulated by ENVI-met and predicted using a cGAN	The cGAN model accurately predicted UTCI maps from urban configuration images with a Structural Similarity Index (SSIM) of 96%, generating results in ~3 seconds compared to 30 minutes for ENVI-met, thus providing a fast and reliable alternative for urban design evaluations
[153]	Tallinn, Estonia, where CFD-based thermal and wind simulations were combined with deep generative surrogate modeling	DL using a generative surrogate model	Urban geometry, building shape and orientation, surface material properties, and boundary wind conditions	UTCI and wind comfort maps	A predictive and generative ML model, trained on CFD simulation data, reliably classified and generated outdoor thermal and wind comfort indicators across various configurations, offering a fast and scalable alternative for urban planning and design

Table 4. DL applications in OTC studies.

Reference	Location/Investigation method	DL Method	Input Parameters	Output Parameters	Main Results
[143]	Simulation-based study using 2,400 synthetic residential block configurations generated in Grasshopper, with building layouts inspired by Shenzhen planning guidelines (China)	Deep CNNs were tested in 12 model variations with different layer depths and kernel sizes, with the best model using six convolutional layers (3×3 kernels) followed by fully connected layers	Building geometry and meteorological data, including Ta, RH, and wind speed	UTCI	The best model (6-layer CNN, Group B) achieved R2 = 0.960 and MSE = 0.022, significantly outperforming ANNs; spatially invariant convolutional layers enhanced generalization, and hidden-layer visualization revealed how spatial features were extracted for UTCI prediction
[141]	Tianjin, China, where field measurements and questionnaire surveys were conducted in four Urban Blue-Green Infrastructure (UBGI) spaces during the winter	MLP with feed-forward back-propagation using hyperbolic tangent activation in the hidden layer and softmax activation in the output layer	Time of day, distance from water, and air temperature at waterfront greenery sites, based on 3,848 samples collected from eight sites on December 19, 2021	Heat Island Intensity (HII), Coupling Effect Intensity (CEI), and UTCI	The ANN model showed high predictive performance, with relative area under the ROC curve 95.3–99.2% for HII and 72.0–95.9% for CEI; air temperature was the most influential factor for HII and time of day for CEI, supporting the development of a Microclimate Influence model and identifying an optimal thermal comfort range of Ta=9.07-14.75 °C
[58]	Xi’an, China, where real-time field measurements, thermal imaging, and questionnaires were conducted in an urban park with 405 children (aged 7–14) participating under three thermal stress levels and three activity intensities	Deep CNN was used for facial expression recognition and combined with RFR to predict children’s TCV	Facial expression images captured via real-time video, individual characteristics (age, gender), environmental parameters (air temperature, globe temperature, relative humidity, wind speed), physical activity levels, and clothing insulation	Children’s TCVs on a 7-point scale (−3=cold to +3=hot), used as ground truth for model prediction and validation	The CNN effectively extracted emotional features from children’s facial expressions, which, when combined with RF, yielded R2 = 0.825 for TCV prediction; the DL model outperformed manual scoring methods and traditional ML classifiers, demonstrating potential for real-time, non-contact outdoor thermal comfort assessment based on facial images
[59]	Shiraz, Iran, where a spatiotemporal analysis of the UHI was conducted using 2006–2021 Landsat imagery and urban configuration metrics, with LST predictions across built-up, soil, and vegetation areas	A Deep Neural Network (DNN) with three hidden layers was trained on geospatial features to predict LST, with performance compared against five other ML models, including RF, XGBoost, and KNN	Normalized indices (NDVI, NDBI, NDWI), LST, LULC, elevation, slope, aspect, proximity to roads and water bodies, and climatic parameters (Ta, humidity, and wind speed)	LST	Among six ML models tested, the DNN achieved the highest predictive performance for LST (R2 = 0.94, RMSE = 1.71 °C), accurately capturing spatial LST distributions across urban surfaces, with vegetation and NDVI as the most influential features, supporting DL-based urban heat modeling for climate-resilient planning
[138]	Freiburg, Germany, where citywide thermal comfort was modeled using urban sensor network data, high-resolution GIS layers, and ML emulation of numerical climate models over 2018–2022	A hybrid DL human thermal comfort neural network (HTC-NN) combined two MLPs (each with three hidden layers) for air temperature and relative humidity, a U-Net CNN for MRT, and RF for wind speed estimation	Meteorological parameters (Ta, RH, wind speed, radiation, precipitation), geospatial data (land cover, building and vegetation height), and surface characteristics (SVF, albedo, population density)	UTCI at 1×1 m spatial resolution, derived from ML-predicted Ta, RH, wind speed (U), and MRT fields	HTC-NN achieved UTCI prediction accuracy of RMSE = 3.0 K and R2 = 0.92 against street-level sensor data, outperforming the numerical model
[142]	Hong Kong, where in situ microclimatic monitoring and questionnaire surveys were conducted in three urban parks across summer and winter, and neural network models were trained on subjective and objective variables to predict thermal comfort	ANN with two hidden layers, optimized via the Levenberg–Marquardt algorithm, and trained separately for summer (17 input neurons) and winter (14 input neurons)	Microclimatic variables (Ta, MRT, wind speed, RH, solar radiation), physiological variables (clo, MET), psychological inputs (thermal, solar, wind, humidity sensations), perceptions of park features bodies(trees, shade, water ), and personal traits (age, purpose of visit, thermal sensitivity)	Thermal Comfort Evaluation: self-rated comfort score from a field survey on a continuous scale	The optimized ANN with two hidden layers achieved R2 = 0.653 (summer) and R2 = 0.771 (winter) on the validation set; the DL model significantly outperformed PMV and PET models, with inclusion of perception and psychological variables improving prediction accuracy by over 30%
[151]	A multi-site analysis using data from 43 previously published OTC studies conducted across diverse climate zones worldwide	ANN models with three hidden layers, trained using backpropagation, were applied to predict TSV based on microclimatic and macroclimatic variables across global datasets	Meteorological parameters (Ta, RH, wind speed, globe temperature), subjective variables (clothing insulation, metabolic rate, gender), microclimatic modifiers (shade, surface material, tree coverage), and macroclimatic classification (climate zone index)	PET	ANN models with three hidden layers successfully predicted PET based on macro- and microclimatic variables across diverse urban settings; analysis showed that macroclimate factors (latitude, altitude, distance from the sea) and microclimate features (albedo, H/W, SVF, LAI) contributed similarly to OTC, with PET in equatorial regions ~13 °C higher than in polar regions during the summer, highlighting the need to integrate microclimate-sensitive design in cities with unfavorable macroclimates
[152]	Tehran, Iran, where image-based DL models were developed using ENVI-met simulation data	DL using a conditional GAN (pix2pix)	Urban geometry, greening configuration, and façade materials	UTCI map simulated by ENVI-met and predicted using a cGAN	The cGAN model accurately predicted UTCI maps from urban configuration images with a Structural Similarity Index (SSIM) of 96%, generating results in ~3 seconds compared to 30 minutes for ENVI-met, thus providing a fast and reliable alternative for urban design evaluations
[153]	Tallinn, Estonia, where CFD-based thermal and wind simulations were combined with deep generative surrogate modeling	DL using a generative surrogate model	Urban geometry, building shape and orientation, surface material properties, and boundary wind conditions	UTCI and wind comfort maps	A predictive and generative ML model, trained on CFD simulation data, reliably classified and generated outdoor thermal and wind comfort indicators across various configurations, offering a fast and scalable alternative for urban planning and design

The reviewed studies span diverse geographic and climatic contexts, from high-density Asian cities (Shenzhen, Tianjin, Xi’an, Hong Kong) to European settings (Freiburg, Tallinn) and Middle Eastern climates (Shiraz, Tehran). Methodologically, investigations range from simulation-based analyses using parametric design tools, CFD, and ENVI-met, to field-based approaches incorporating in situ microclimatic monitoring, thermal imaging, and questionnaire surveys. Remote sensing and spatiotemporal analyses of urban heat islands were also employed, notably in Shiraz, while large-scale comparative studies synthesized findings from dozens of prior OTC investigations worldwide. Increasingly, hybrid approaches integrate empirical measurements with machine learning emulation or surrogate modeling, reflecting a methodological shift toward scalable, data-driven frameworks for assessing OTC.

The reviewed studies applied a wide range of DL architectures. Deep CNNs were tested in multiple variations, with the most effective configuration comprising six convolutional layers with 3×3 kernels and fully connected layers. Other approaches included MLPs with back-propagation, hyperbolic tangent hidden activations, and softmax outputs, as well as ANN models with two to three hidden layers optimized via backpropagation or the Levenberg–Marquardt algorithm. Hybrid architectures were also explored, such as an HTC-NN combining dual MLPs, a U-Net CNN, and RF components for multi-variable estimation. In specific applications, deep CNNs were coupled with RFR for children’s thermal comfort prediction, while DNNs with three hidden layers were benchmarked against RF, XGBoost, and KNN for LST forecasting. More advanced generative methods were employed in selected cases, including conditional GANs (pix2pix) and deep generative surrogate modeling, to emulate simulation data and generate synthetic comfort scenarios.

The studies incorporated a diverse range of environmental, physiological, and personal input parameters to capture OTC. Building geometry and meteorological variables such as Ta, RH, and wind speed were commonly used, while some studies extended inputs to include land surface indices (NDVI, NDBI, NDWI), LST, land use/cover, elevation, slope, and proximity to water or roads. Human-centric variables were also considered, including facial expressions, clothing insulation, metabolic rate, age, gender, and thermal sensitivity, often combined with perceptions of microclimatic features such as shade, vegetation, and water bodies. Several investigations integrated urban geometry, façade materials, greening configurations, and boundary wind conditions, reflecting a comprehensive approach to linking environmental context with individual comfort responses across diverse climatic and morphological settings.

The studies reported a variety of output metrics to quantify OTC. UTCI was the most frequently used indicator, often complemented by HII, CEI, or PET. In human-centric assessments, subjective measures such as children’s TCVs on a 7-point scale or self-reported comfort scores from field surveys were employed as ground truth for model validation. Other outputs included high-resolution maps of LST, UTCI, and wind comfort, with several studies leveraging ML or deep generative models (such as cGANs) to predict spatially detailed comfort fields derived from meteorological and urban parameters. These outputs collectively enabled both quantitative evaluation and spatial visualization of urban thermal environments.

The results of the reviewed studies collectively demonstrated that DL and hybrid ML approaches can achieve high predictive accuracy for OTC across diverse urban and climatic contexts. CNN- and DNN-based models consistently outperformed traditional ANNs and standard indices, with R² values often exceeding 0.9 for both UTCI and LST predictions, while hybrid architectures incorporating RF, MLPs, or U-Net components further enhanced performance for multi-variable estimation. Key drivers of thermal comfort were consistently identified, including Ta, vegetation cover, NDVI, urban geometry, and macroclimatic factors, with perception- and behavior-related variables improving predictive accuracy when integrated. Generative and surrogate modeling approaches, such as cGANs and ML emulators of CFD simulations, enabled rapid, high-fidelity mapping of spatial comfort fields, reducing computational time from tens of minutes to seconds.

Although less widely applied than traditional SL methods, DL approaches have shown strong potential for handling complex, high-dimensional datasets, especially when fine spatial or temporal resolution is required. However, their broader adoption is hindered by the need for large, labeled datasets and limited interpretability—the so-called “black box” issue. To mitigate these challenges, studies have incorporated explainable AI techniques like SHAP [154], to improve model transparency and employed methods such as PCA and K-means clustering to reduce complexity, enhance generalization, and maintain predictive accuracy.

Overall, Table 2 and Table 3, and Table 4 underscore the ability of ML models to deliver high-accuracy predictions of OTC by integrating diverse environmental, physiological, and behavioral inputs. These data-driven approaches mark a methodological leap beyond traditional index-based models, enabling more flexible and nuanced representations of human–environment interactions. When applied to both subjective thermal votes and bioclimatic indices, ML frameworks—especially those enhanced with feature selection, dimensionality reduction, or interpretability tools—exhibit strong generalization across varied climates and urban forms. Collectively, these results highlight the potential of data-driven, scalable frameworks for informing climate-sensitive urban design, enabling both real-time assessment and scenario-based planning while capturing the combined effects of environmental, morphological, and human factors.

The reviewed studies highlight the growing relevance of ML in OTC assessment, particularly for its predictive accuracy and adaptability across diverse urban and climatic contexts. A key advantage of ML lies in its ability to integrate heterogeneous inputs—ranging from meteorological and morphological to physiological and behavioral data—thus addressing the limitations of traditional indices based on steady-state assumptions. Among the techniques reviewed, SL algorithms like RF, SVM, and XGBoost offer an effective balance between accuracy, efficiency, and interpretability, making them well-suited for both research and practical applications in urban design, planning, and real-time monitoring.

Despite notable advancements, several limitations hinder the broader application of ML in OTC research. First, most models are trained on location-specific datasets, limiting their generalizability across different urban morphologies and cultural contexts. Second, the subjective and culturally influenced nature of thermal comfort complicates model calibration and reduces transferability. Third, many existing models are static or semi-dynamic, lacking integration with adaptive behaviors, feedback loops, or real-time environmental interactions—elements essential for responsive, human-centric urban design

4. A Practical Framework for Optimizing Outdoor Thermal Comfort

Optimizing OTC has emerged as a critical decision-making procedure in climate-responsive urban design. In dense urban settings, achieving thermally comfortable outdoor conditions is inherently complex due to the diversity and the nonlinear interplay of multiple variables—including urban geometry, material properties, vegetation cover, wind dynamics, and solar exposure. A systematic approach to decision-making is of crucial importance for designing responsive, efficient and sustainable thermal comfort strategies. The supporting methodologies must be aptly selected and robustly combined to generate realistic and viable solutions that ensure accountability, transparency and integrity, as well as broader acceptance from stakeholders. Any approach that relies exclusively on the cumulative experience, trial-and-error, an intuitive style, empirical concepts or rule-based patterns, is highly likely to lead to controversial solutions with increased uncertainty and difficulty in identifying configurations that meet performance goals. Optimization provides an appropriate framework as well as reliable methods and tools for systematically navigating these high-dimensional design spaces and telling the decision maker which are the best ways to solve the puzzle of OTC. The core of optimization is to maximize or minimize one or more objective functions—quantifiable expressions such as energy use, cost, or thermal comfort—by calculating the optimum values of a defined set of design variables under a bundle of specified technological, economic, environmental, and logical constraints. These objectives may encompass both quantitative goals (e.g., minimizing energy consumption or construction cost) and qualitative targets (e.g., enhancing visual amenity, environmental quality, or thermal comfort) which often are conflicting in nature, resulting to an intriguing tug of war and interesting trade-off expressions [155,156]. A typical multi-objective optimization (mathematical programming) problem can be defined as follows:

$$\begin{array}{c}\mathrm{M}\mathrm{i}\mathrm{n}/\mathrm{m}\mathrm{a}\mathrm{x}\left\{\begin{array}{c}{f}_1\left(x\right)=c_1^T·x\\ f_2\left(x\right)=c_2^T·x\\ ⋮⋮⋮\\ f_k\left(x\right)=c_k^T·x\end{array}\right\}\\ \mathrm{subject} \mathrm{to}x\in S,\mathrm{where}S=\left\{x\in \left.{\mathbb{R}}^n\right|A·x\begin{array}{c}\le \\ =\\ \ge \end{array}b,x\ge 0,b\in {\mathbb{R}}^m\right\}\end{array}$$

(13)

where, k is the number of objective functions, n is the number of decision variables, m is the number of constraints, fk(x) is the k-st objective function, x is n-dimensional vector of the decision variables, ck is the vector of coefficients of the objective function, A is the matrix of the technological coefficients, b is the m-dimensional vector of right-hand side parameters and S is is the feasible region.

Multi-objective optimization supports the identification of comprehensive solutions that balance competing targets such as solar access, shading, ventilation, and thermal regulation. Within the OTC domain—where interactions among urban form, materiality, vegetation, and microclimate critically influence human thermal experience—the concept of multi-criteria decision making is not just advantageous but essential. It enables the formulation of design strategies that reconcile competing performance criteria (e.g., maximizing shading while preserving daylight access, or enhancing airflow without increasing discomfort), thereby supporting more adaptive, resilient, and human-centric urban environments [155]. Moreover, as cities confront escalating climate change impacts, the capacity to optimize OTC across seasonal and meteorological variations is increasingly vital for adaptive and resilient urban environments. To address the often-conflicting interactions between environmental drivers, architectural morphology, and human thermal perception, optimization techniques have gained substantial momentum in research and practice [157,158,159]. However, the multi-objective nature of OTC problems can significantly hinder the timely generation of design strategies due to their large computational complexity and time-consuming solution processes. These problems can become prohibitive as the number of conflicting objectives, and the size of datasets increase. One major reason for this large computational burden is the determination of the so-called nadir points when calculating the range of each objective function in order to construct the pay-off table (lexicographic optimization) using the advantageous ε-constraint method [160,161,162,163].

Optimization techniques for OTC can be broadly categorized into two main approaches: SBO and Surrogate Modeling. SBO integrates parametric modeling, dynamic simulation engines, and formal optimization algorithms to systematically investigate complex, high-dimensional design spaces [156,164,165].

. In the context of thermal comfort, SBO facilitates the exploration of urban design alternatives—such as variations in geometry, orientation, vegetation, or materiality—that minimize thermal stress while adhering to contextual and regulatory constraints [157,158]. Widely used tools such as EnergyPlus [166], ENVI-met [37], and Urban Weather Generator (UWG) [167] are often coupled with optimization algorithms like GA [168,169] or PSO [169,170], enabling iterative performance assessments based on objective functions such as UTCI or PET. In contrast to rule-of-thumb or trial-and-error methods, SBO supports multi-objective, multi-variable optimization, revealing trade-offs among key urban design goals such as reducing cooling energy demand, ensuring solar access, and enhancing pedestrian thermal comfort. This data-driven methodology is particularly valuable in dense or climatically challenging urban contexts, where nonlinear interactions between form, climate, and human exposure demand high-fidelity simulations and comprehensive optimization strategies [171,172,173].

In contrast, surrogate modeling—also known as meta-modeling—seeks to replicate the output behavior of computationally intensive simulation or optimization structures through data-driven ML algorithms, including ANNs [107,174], SVMs [107,175,176,177,178], and GPR models [110,179]. These surrogates serve as computationally efficient proxies, emulating the complex input–output relationships of high-fidelity modeling systems and enabling rapid performance evaluations across large, high-dimensional design spaces [156,164,180,181]. Once trained on representative datasets, surrogate models drastically reduce computational demands while preserving acceptable levels of predictive accuracy. This efficiency makes them ideal for supporting advanced tasks such as multi-objective optimization, global sensitivity analysis, and uncertainty quantification—processes that would be otherwise prohibitively time-consuming using direct optimization alone [182].

The surrogate modeling workflow typically involves four sequential steps [180,182]: (i) sampling the design space, often using Latin Hypercube Sampling or Sobol sequences [180,182]; (ii) generating simulation data through high-fidelity models [177,180]; (iii) training a predictive ML model [64,178]; and (iv) validating accuracy using metrics such as RMSE or MAE [179,183].

Among ML algorithms, ANNs are widely favored for capturing complex nonlinear dependencies [107,174]; SVMs offer robust generalization when data are limited [107,175,176]; and more advanced architectures—such as CNNs [107,109] and LSTM networks [107,110,184] —are increasingly employed for spatial and temporal modeling at urban scales. Recent OTC research shows a growing trend toward hybrid workflows that combine simulation-based optimization with ML surrogates. These integrated approaches effectively balance modeling fidelity and computational efficiency, particularly in contexts where urban morphology, vegetation, materials, and climatic variables interact in nonlinear and dynamic ways. As such, surrogate modeling emerges as a cornerstone of next-generation, performance-driven urban design workflows that center on thermal comfort without sacrificing scalability or speed [136,157,185].

Building on these foundations, Table 5 synthesizes recent real-world applications of optimization in OTC. The cases span diverse climates and design challenges, each defined by its design variables, tools, algorithms, and objective functions. Most studies employ SBO with evolutionary algorithms to navigate the interactions among form, materiality, vegetation, and microclimate. However, ML integration remains limited. Only three studies— [136,185,186]—use surrogate models, and just one [186] combines them with multi-objective optimization. This underutilization signals a significant opportunity: the full capabilities of ML—particularly for accelerating simulations, handling high-dimensional inputs, and modeling spatiotemporal dynamics—remain largely untapped in OTC optimization.

The key insights from the case studies presented in Table 5 could be synthesized as follows:

• From geometric tuning to high-dimensional optimization: Early studies focused on adjusting geometric variables (e.g., building height, spacing, orientation) to reduce PET or UTCI [172,173,190]. More recent works [194,195,196] adopt multi-parametric optimization strategies incorporating energy efficiency, daylight, and visual comfort—marking a shift toward data-driven, performance-oriented design.
• b. Hybrid and surrogate-based methodologies as emerging standards: Studies [136,185,186] showcase hybrid workflows combining surrogate models (e.g., ANN) with formal optimizers (e.g., GA, NSGA-II). Study [186] reported 89.9% accuracy across 150 configurations, demonstrating significant speed and accuracy benefits.
• c. Optimization as a multi-scalar design instrument: Optimization has been applied at multiple scales—from vegetation layout [197] and courtyard form [195] at the micro level, to block morphology and land use [187,188,190] at the meso level—bridging architectural detailing and urban planning.
• d. Beyond thermal comfort and toward a multi-criteria urban performance: Many studies now include daylight access, SVF, wind protection, and energy demand as co-objectives [171,194,196]. For instance, study [196] simultaneously optimized PET and heating/cooling loads to balance comfort and energy use. (e) Sensitivity to climatic and socio-spatial contexts: Design objectives vary by climate: Guangzhou [188] emphasized shading and vegetation; Sydney [196] balanced summer cooling and winter sunlight; Shenyang and Beijing [186,191] prioritized solar gain and wind protection. Study [189] tailored optimization to user groups (e.g., hospital patients vs. healthy users), underscoring the need for inclusive design.
• e. Demonstrated performance gains and implementation potential: Comfort gains ranged from 2 °C to 6 °C in PET or UTCI—improvements with tangible public health and livability implications. Tools like Grasshopper [198], Ladybug [199], and Galapagos [200] enable integration into design workflows, even at early planning stages [194,197], enhancing regulatory and participatory potential.
• f. The reviewed literature demonstrates that optimization applications have matured from simple geometric manipulations into a robust, multi-objective and multi-scalar design approaches. Although SBO is currently dominant, ML-assisted methods—particularly surrogate modeling—hold considerable promise for future development. As urban areas grapple with climatic stressors, the convergence of optimization algorithms, parametric design tools, and data-driven surrogates offers a scalable and effective pathway toward thermally comfortable, climate-resilient urban environments.

Based on this body of evidence, a practical framework for optimizing outdoor thermal comfort can be articulated as a sequence of six interdependent stages that collectively link climate diagnosis, performance modeling, and design decision-making. First, the process begins with a systematic assessment of local climatic conditions, thermal exposure patterns, and user characteristics, which together inform the selection of appropriate thermal comfort indices and the definition of optimization objectives. Second, this diagnostic stage guides the parametric formulation of the urban design space, encompassing geometric configuration, material properties, vegetation characteristics, and spatial organization. Third, high-fidelity simulation tools are employed to generate performance datasets that capture the nonlinear interactions between urban form, microclimate, and human thermal exposure. Fourth, machine learning and surrogate modeling techniques are introduced to approximate simulation outputs, substantially reducing computational demands while enabling sensitivity analysis and uncertainty quantification. Fifth, multi-objective optimization algorithms are applied to systematically explore trade-offs among thermal comfort, daylight access, energy demand, and other competing constraints, yielding Pareto-optimal design solutions. Finally, these solution sets are translated into decision-support information, facilitating context-sensitive, scalable, and implementation-ready urban design strategies.

Across scales and contexts, the most impactful frameworks balance thermal comfort with co-constraints such as daylight availability, energy use, and social inclusion. In this way, optimization-driven workflows do not merely target thermal indices but provide a transparent, computationally efficient, and reproducible pathway for climate-responsive urban design.

5. Discussion and Conclusions

This review paper aimed to (1) synthesize ML methodologies used in OTC research, across learning paradigms, (2) explore the integration of ML with SBO, surrogate modeling, and scenario-based design strategies, and (3) propose a roadmap for scalable, intelligent, and thermally adaptive urban design informed by ML capabilities. Furthermore, this review aimed to propose a practical framework for OTC optimization, offering guidance on selecting appropriate indices, algorithms, and design parameters to support computationally efficient and implementation-ready OTC interventions across different urban and climatic contexts.

The reviewed literature spanned a broad geographic range—including urban contexts across Europe, Asia, and the Middle East—and encompasses diverse climatic zones such as Mediterranean, humid subtropical, arid, and tropical rainforest environments. While the majority of investigations rely on SL and DL techniques—often trained on labeled datasets such as thermal sensation votes or bioclimatic indices—several studies also incorporate unsupervised approaches, including clustering methods and PCA, to identify latent structures in thermal comfort data. Despite variability in data sources (e.g., in situ microclimatic monitoring, wearable physiological sensors, remote sensing), modeling strategies, and target variables (e.g., UTCI, PET, or subjective thermal responses), the collective findings emphasize the versatility and predictive strength of ML-based frameworks in modeling the complex, nonlinear dynamics that characterize outdoor thermal environments.

Among the reviewed studies, SL was by far the most commonly applied approach (28 applications), reflecting the widespread availability of labeled data such as TSV, PET, UTCI, and DI, which are well-suited for regression and classification tasks using models like RF, XGBoost, SVM, and ANN. DL was the second most frequent approach, with nine studies, as recent research has leveraged larger and higher-resolution datasets—including satellite imagery, high-frequency meteorological measurements, and crowdsourced sensor data—where the feature extraction and nonlinear modeling capabilities of MLPs, CNNs, and LSTMs provide clear advantages. In contrast, UL was rarely applied, with only two identified studies, primarily for clustering rainfall events or urban microclimate typologies; this reflects the field’s predominant focus on predicting target variables rather than exploring latent structures in the data. The disparity in application frequency among these paradigms is therefore reasonable, yet it also highlights a potential opportunity for more exploratory and pattern-discovery studies using unsupervised or hybrid methods to uncover hidden drivers of outdoor thermal comfort. Notably, none of the reviewed studies have explored RL or SSL, despite their success in related domains such as indoor environmental control [126,127]. RL, with its ability to learn optimal actions through environmental feedback, could support adaptive outdoor systems like dynamic shading or misting. SSL, by leveraging both labeled and unlabeled data, offers a solution to the limited availability of subjective comfort datasets. Coupling RL and SSL with wearable sensors, mobile data streams, and edge computing could enable intelligent, scalable, and personalized OTC management strategies in real time.

The reviewed optimization studies show that outdoor thermal comfort (OTC) optimization has advanced from simple geometric adjustments to sophisticated, multi-objective, and multi-scalar approaches that integrate thermal, energy, daylight, and social dimensions. Hybrid and surrogate-based methods are emerging as standards, offering high accuracy and reduced computational cost, while performance gains of 2–6 °C in PET/UTCI demonstrate tangible health and livability benefits. Effective frameworks are site-specific, adaptive to climatic and socio-spatial contexts, and increasingly supported by accessible parametric tools, making optimization a practical and scalable pathway toward climate-resilient urban design.

ML has emerged as a transformative tool for OTC assessment, capable of addressing the nonlinearities, contextual dependencies, and data complexities that limit conventional empirical and simulation approaches. A synthesis of evidence from the literature has demonstrated that SL, UL, and DL methods provide robust predictions for both comfort indices and subjective responses, while enabling applications such as city-scale mapping and fine-grained spatiotemporal modeling. In parallel, optimization methodologies are increasingly integrated with ML to translate predictive insights into design solutions. SBO offers direct integration with detailed microclimate simulations, while surrogate modeling replaces high-fidelity solvers with approximate models to drastically reduce computational demand. When combined with traditional or ML-based prediction and estimation techniques, both SBO and surrogate modeling enable efficient exploration of multi-objective design spaces, supporting trade-offs between thermal comfort, energy use, daylight, and other performance metrics. Reviewed applications have confirmed that these approaches can deliver tangible improvements in outdoor comfort and broaden the range of feasible interventions.

In conclusion, ML-based approaches for OTC modeling offer a flexible, accurate, and adaptable alternative to traditional methods. Future research should address three critical frontiers. First, improving the transferability of ML models across diverse urban morphologies and climates remains essential to achieving generalizable solutions. Second, greater exploration of advanced paradigms—such as semi-supervised learning, reinforcement learning, and physics-informed models—could enhance adaptability, reduce data requirements, and embed physical realism. Third, integrating ML and optimization within real-time sensing systems, digital twins, and participatory planning platforms offers a pathway to actionable, citizen-centered interventions. Future research should also focus on integrating multimodal data—including meteorological, morphological, physiological, and behavioral inputs—while advancing interpretable and scalable model architectures. Embracing adaptive learning frameworks that address uncertainty, individual variability, and dynamic feedback will be critical for translating research into practical, user-centered comfort optimization tools for complex urban environments.

By consolidating methodological advances and outlining a practice-oriented optimization framework, this work establishes the basis for a roadmap guiding the next generation of OTC research. Future directions should prioritize the integration of ML for predictive accuracy with SBO and surrogate modeling for computational efficiency, enabling context-sensitive, multi-objective, and scalable solutions. Such convergence can shift OTC assessment from a primarily diagnostic exercise to a proactive design instrument, advancing resilient, human-centered, and sustainable urban environments.