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

Abstract

Urban environments face increasing thermal stress from climate change and the Urban Heat Island effect, with significant implications for livability, public health, and energy sustainability. Outdoor thermal comfort is defined as the state in which conditions are perceived as acceptable, depends on interactions among meteorological, morphological, physiological, and behavioral factors. This review synthesizes the application of machine learning (ML) to outdoor thermal comfort assessment into a practice-oriented taxonomy. Research spans diverse climates and urban forms, using inputs across environmental and human domains. Supervised learning dominates. Regression approaches (linear regression, support vector regression, random forest, gradient boosting) and classification algorithms (decision trees, support vector machines, K-nearest neighbors, Naïve Bayes, random forest classifiers) are widely used to predict thermal indices such as the Physiological Equivalent Temperature and Universal Thermal Climate Index, or to classify subjective responses including thermal sensation, comfort, and acceptability. Unsupervised learning (clustering, principal component analysis) supports identification of microclimatic zones and perceptual clusters, while deep learning (multilayer perceptrons, convolutional and recurrent neural networks, generative adversarial networks) achieves superior accuracy for complex, high-dimensional, and spatiotemporal data. Algorithms such as random forests, support vector machines, and gradient boosting consistently show strong performance for both indices and subjective responses when integrating multi-domain inputs. Semi-supervised and reinforcement learning remain underexplored but offer promise for leveraging large-scale sensor data and enabling adaptive, real-time comfort management. The review concludes with a roadmap emphasizing explainable artificial intelligence, scalable surrogate modeling, and integration with simulation-based optimization and parametric design tools.

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, 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.

This review focuses specifically on ML applications for OTC assessment at pedestrian and urban design scales. The research object is limited to studies that explicitly link environmental conditions with human thermal perception, either through bioclimatic indices (e.g., PET, UTCI, SET*, PMV) or subjective responses (e.g., TSV, TCV). Input domains considered include meteorological variables, urban morphological parameters, physiological attributes, and behavioral factors. The spatial scope spans point-level measurements, street segments, neighborhood blocks, and district-scale analyses. Broader studies on urban heat island intensity or land surface temperature (LST) are included only when directly connected to human thermal comfort evaluation, and not as standalone climate modeling exercises.

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 literature search was performed using structured keyword combinations applied consistently across bibliometric databases. The core search string combined terms related to OTC (e.g., “outdoor thermal comfort”, “thermal sensation vote”, PET, UTCI) with machine learning concepts (e.g., “machine learning”, “artificial intelligence”, “random forest”, “support vector machine”, “deep learning”, ANN, CNN, LSTM) and urban context descriptors (e.g., urban, microclimate, urban canyon). The initial search results were screened in three stages: (i) duplicate removal, (ii) title and abstract screening to exclude irrelevant or indoor-only studies, and (iii) full-text eligibility assessment based on methodological clarity and validation rigor. Studies were excluded if they did not explicitly address outdoor environments, did not implement ML methodologies, or lacked quantitative validation.

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 its advantages and limitations when applied to environmental datasets.

2.1.1. Regression Algorithms

Regression algorithms in SL are designed to predict continuous outputs by modeling relationships between environmental predictors and thermal comfort indicators. In OTC research, they are primarily used to estimate indices such as PET and UTCI from meteorological and morphological variables. Due to their scalability and interpretability, regression models form a core component of environmental predictive analytics [61].

Representative techniques include linear regression (LR), decision tree regression (DTR), support vector regression (SVR), and gradient boosting regression (GBR). Ensemble and neural-based methods, such as random forest (RF), ANNs, and extreme learning machines (ELMs), are also widely applied when nonlinear interactions dominate.

LR models the relationship between a dependent variable and one or more predictors under assumptions of linearity and normally distributed residuals. For a single predictor, LR 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 OTC 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 rule-based algorithm that predicts continuous outcomes by recursively partitioning data into homogeneous subsets, enabling the modeling of nonlinear relationships and feature interactions without requiring input scaling [61,64]. Its transparent tree structure enhances interpretability and supports diagnostic analysis, while its robustness to outliers and multicollinearity makes it suitable for heterogeneous environmental datasets. At each node, the algorithm selects the optimal split that minimizes prediction error, typically measured using the weighted mean squared error (MSE) [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.

In OTC research, DTR has been applied to predict indices such as PET and UTCI using meteorological variables, including air temperature, relative humidity, wind speed, and solar radiation [46,58]. Its interpretable decision rules make it a practical modeling choice in 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 OTC 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 OTC 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 ${\left\{\left(x_i,y_i\right)\right\}}_{i=1}^N$, RFR constructs $T$ trees ${\left\{h_t\left(x\right)\right\}}_{i=1}^T$, and the predicted output $\widehat{y}$ for a new input $x$ is [17]:

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

(6)

where $h_t\left(x\right)$ 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 SL 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]:

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

(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 OTC 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 fundamental SL task that assigns input observations to predefined categories based on patterns learned from labeled data. In OTC research, classification models are widely used to predict thermal sensation categories (e.g., comfortable vs. uncomfortable), heat stress levels, or subjective perception classes derived from meteorological measurements and field surveys [52,60].

Several established algorithms are employed for such tasks. Decision tree classification (DTC) partitions the feature space through recursive rule-based splits, offering high interpretability and flexibility for both categorical and continuous inputs [88]. Support vector machines (SVM) identify optimal separating hyperplanes by maximizing class margins and can handle nonlinear boundaries through kernel functions [64,67]. K-nearest neighbors (KNN) assigns labels based on proximity in feature space; although simple, it is sensitive to noise and scaling effects [89]. Naive Bayes (NB) applies probabilistic inference under conditional independence assumptions and remains computationally efficient in high-dimensional settings [90].

Ensemble approaches such as the random forest classifier (RFC) enhance robustness by aggregating multiple decision trees through majority voting, reducing variance and overfitting compared to single-tree models [73]. ANNs can also be adapted for classification, learning complex nonlinear decision boundaries in high-dimensional environmental datasets [53,81].

Overall, classification methods in OTC studies present trade-offs between interpretability and predictive performance. While tree-based models provide transparent decision rules suitable for urban planning applications, ensemble and neural-based approaches often yield higher accuracy when modeling nonlinear relationships between environmental variables and subjective thermal perception.

2.2. Unsupervised Learning

UL comprises ML methods that identify latent structures or meaningful groupings in datasets without labeled outputs. In OTC and environmental research, UL is primarily used for exploratory analysis, including the identification of perceptual clusters in comfort surveys, classification of occupant behavior patterns, and detection of microclimatic regimes based on sensor observations [53,75,91,92,93,94,95,96]. Compared to supervised approaches, UL is less frequently applied in OTC studies, reflecting the field’s predominant focus on predictive modeling rather than pattern discovery.

K-means clustering partitions data into k groups by minimizing intra-cluster variance [93]. It is computationally efficient and suitable for large meteorological datasets; however, it requires predefining the number of clusters and may be sensitive to initialization and distance metrics. In OTC contexts, k-means has been used to group similar thermal environments or behavioral responses.

Hierarchical clustering constructs nested clusters without requiring a predefined number of groups, producing a dendrogram that reveals multi-scale structure [97]. This approach is particularly useful in urban climate studies for identifying spatial or temporal comfort typologies across different scales.

Principal Component Analysis (PCA) is a linear dimensionality reduction method that transforms correlated variables into orthogonal components ranked by explained variance [95]. In OTC research, PCA is commonly employed to reduce complexity in high-dimensional meteorological datasets while retaining dominant variability patterns [98]. However, its linear formulation may limit its ability to capture nonlinear interactions among climatic variables.

t-distributed stochastic neighbor embedding (t-SNE) provides nonlinear dimensionality reduction optimized for visualization, preserving local data structure [53,94]. It has been applied to reveal latent perception clusters or spatial microclimate patterns but is primarily exploratory rather than predictive.

Autoencoders, trained in an unsupervised manner, compress input data into lower-dimensional latent representations and reconstruct them through neural encoding–decoding architectures [92,96]. In environmental modeling, they support feature extraction and noise reduction in complex sensor datasets, facilitating subsequent clustering or supervised prediction tasks [99].

Overall, UL methods in OTC research serve mainly as exploratory tools for identifying hidden structures or reducing data dimensionality. While they offer valuable insights into microclimatic variability and perceptual heterogeneity, their standalone use remains limited compared to supervised predictive approaches.

2.3. Semi-Supervised Learning

SSL bridges supervised and unsupervised paradigms by combining a limited labeled dataset with a larger pool of unlabeled data. This approach is particularly relevant in environmental sensing and OTC research, where collecting meteorological measurements is relatively straightforward, but obtaining labeled outcomes (e.g., thermal sensation votes) is resource-intensive. By exploiting the intrinsic structure of unlabeled data, SSL can improve generalization performance compared to purely supervised models trained on small labeled samples [54,100].

Common SSL strategies include self-training, co-training, graph-based label propagation, consistency regularization, and generative models such as semi-supervised variational autoencoders [101,102,103,104,105]. These methods differ in how they incorporate unlabeled data—either through pseudo-labeling, multi-view learning, relational inference, robustness constraints, or latent representation learning.

To date, no OTC-specific studies have implemented SSL frameworks. However, the increasing availability of dense environmental sensor networks and wearable physiological monitoring suggests strong potential for SSL in future thermal comfort modeling. Its capacity to reduce labeling demands while maintaining predictive accuracy makes it particularly promising for scalable, data-efficient, and context-aware OTC applications.

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 OTC 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]:

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

(8)

where $F_{i,j}$ is the output at the 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 ${\mathrm{\Sigma }}_m{\mathrm{\Sigma }}_n$ 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]:

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

(9)

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 SL, 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 $a_{\mathrm{t}}$, moving to a subsequent state $s_{\mathrm{t}+1}$, and receiving a reward signal $r_{\mathrm{t}}$. 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, RL’s ability to improve decision-making in complex, real-world systems underscores its promise as a foundation for intelligent outdoor comfort optimization.

Table 1. Summary of the main ML algorithms with their respective advantages and disadvantages.

Learning Paradigm	Algorithm	Main Advantages	Main Disadvantages
Supervised learning (regression)	LR [56,61,62,97]	Simplicity; interpretability; speed of training and evaluation; and usefulness as a baseline	Linearity assumption; sensitivity to multicollinearity; influence of outliers on performance; homoscedasticity; and normally distributed and uncorrelated errors
	DTR [56,61,64,88]	Non-parametric nature; interpretability; capability for modeling nonlinear relationships; and suitability for both numerical and categorical data	Tendency for overfitting; sensitivity to data noise and small changes
	SVR [64,67,68]	Effectiveness in high-dimensional spaces; robustness against overfitting; use of the kernel trick for modeling nonlinearity; and flexibility through hyperparameter tuning	Computational intensity; requirement for careful kernel and parameter tuning; and low interpretability
	RFR [73,74,75,76]	Reduction of overfitting; modeling of nonlinear relationships and high-dimensional data; and robustness and accuracy	Lower interpretability than single trees; high computational intensity and resource demand compared to simpler models
	GBR [70,71]	High prediction accuracy; ability to handle complex nonlinear patterns; customizability of loss functions; robustness to overfitting; and capacity to process both numerical and categorical data	Tendency for overfitting without regularization; sensitivity to hyperparameters; low interpretability; and high memory requirement
	ANNs (regression) [52,53,75,81,82,83]	Capability to capture complex nonlinear relationships; and scalability to large datasets	Large data requirements; high computational demand; and risk of overfitting
Supervised learning (classification)	DTC [88]	Transparency; ease of visualization; capability to handle both numerical and categorical data as well as nonlinear relationships; and suitability for small- and medium-sized datasets	Overfitting on noisy data; instability with small changes
	SVM [64,67]	Effectiveness in high-dimensional spaces; flexibility through the use of kernels; and robustness to overfitting	Increasing computational cost with dataset size; limited interpretability
	KNN [89]	Simplicity; absence of a training phase; and intuitiveness	Computationally costly during inference; sensitivity to data scaling and noise; and high storage requirements due to keeping all training data in memory
	NB [90]	Speed; efficiency; effectiveness in high-dimensional spaces; and usefulness as a baseline	Strong independence assumption rarely holding; limitation to linear boundaries
	RFC [73]	Reduction of overfitting; robustness to noise; and effectiveness in handling imbalanced data	Low interpretability; high resource demand; and high storage requirements due to keeping multiple trees in memory
	ANNs (classification) [53,81]	Capability to capture complex decision boundaries; adaptability to various data types	Significant data and tuning requirements; low transparency
Unsupervised learning	k-means clustering [93]	Simplicity; efficiency; wide usage; and scalability	Requirement for predefined k; sensitivity to initialization and outliers; and influence of distance metric choice on results
	Hierarchical clustering [97]	Absence of a predefined number of clusters; provision of multi-scale insights; and generation of a dendrogram offering a visual representation of data groupings	High computational cost; sensitivity to noise, outliers, and linkage criteria
	PCA [95]	Reduction in dimensionality; enhancement of interpretability; improvement of performance; and production of uncorrelated components	Linearity of the method; difficulty in interpreting components
	t-SNE [94]	Excellence for visualization; preservation of local structure	Limited scalability to very large datasets; variability of results between runs
	Autoencoders [53,92,96]	Effectiveness for dimensionality reduction; denoising; and unsupervised feature learning	Requirement for careful tuning; low interpretability; and sensitivity to noise or small datasets
Semi-supervised learning	Self-training [101]	Simplicity and ease of implementation; extension of supervised models; and leveraging of unlabeled data	Sensitivity to confidence threshold; variability in performance
	Co-training [102]	Effective exploitation of feature redundancy through multiple, conditionally independent views of the data; and reduction of overfitting	Requirement for input features to be naturally split into distinct and sufficient subsets (views); and increased computational complexity
	Graph-based methods [103]	Efficient utilization of both labeled and unlabeled data through modeling of relationships as graph structures; and capability to model complex topologies and nonlinear manifolds in the input space	High computational intensity for large datasets due to construction and label propagation; and susceptibility to noise in graph connectivity
	Consistency regularization [104]	Encouragement of model robustness through enforcement of stable predictions under input perturbations such as noise or augmentation; and compatibility with modern neural network architectures and training pipelines	Sensitivity to the type and magnitude of perturbations; negative impact of poorly chosen augmentations on performance
	Generative models [105]	Simultaneous modeling of data distribution and classification; and capture of complex latent structures	High computational intensity and implementation complexity compared to simpler SSL methods; and a requirement for careful balancing of supervised and unsupervised loss components
Deep learning	CNNs [106,107,108,109,130]	High effectiveness for spatially structured data, such as images or grids	Limited capability for long-term temporal dependencies; and a requirement for significant training data and computational resources
	RNNs [53,106,107,109,110,111,112,113,114]	Tailoring for sequential/time-series data; capture of temporal dependencies in input signals; and usefulness for microclimate prediction, human behavior modeling, and control systems	Proneness to vanishing or exploding gradients during training; slow training due to sequential computation; and limited ability to retain long-term dependencies
	DBNs [116,117]	A combination of unsupervised pretraining with supervised fine-tuning; and the capability to learn hierarchical feature representations	Difficulty in training and tuning; lower prevalence compared to CNNs and RNNs; and high computational cost
	GANs [118]	Generation of realistic synthetic data, such as for environmental data augmentation; and usefulness in missing data reconstruction and simulation	Difficulty and instability in training (mode collapse and convergence issues)
Reinforcement learning	Value-based methods [55,119]	Well-established theoretical foundation; suitability for discrete action spaces; and effectiveness in tabular and low-dimensional settings	Scalability issues in large or continuous action spaces; and the requirement for full value function estimation
	Policy-based methods [55,119]	Direct optimization of the policy; effectiveness in continuous or stochastic environments; and capability to handle high-dimensional spaces	High variance in gradient estimates; sample inefficiency and slow convergence

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. SL applications in OTC studies.

Reference	Location/Investigation Method	SL Method	Input Parameters	Output Parameters	Main Results
[144]	An open urban space at Naghsh-e Jahan Square, located in Isfahan, Iran. Measurements and survey data.	ELM (single-layer feedforward network with random hidden layer parameters and analytically computed output weights)	Microclimatic and demographic parameters: air temperature, relative humidity, solar radiation, wind speed, clothing insulation, gender, reason for visiting, frequency of attendance	Thermal sensation vote (TSV), MRT, PMV, PET	ELM outperformed ANN and Genetic Programming (GP) in predicting TSV, PMV, PET, and MRT, achieving its highest accuracy for MRT (R2 = 0.99) and TSV (R2 = 0.94), thereby demonstrating superior predictive performance and computational efficiency
[145]	Open park in Tianjin, China. Field measurements and questionnaire surveys conducted.	LR, multinomial logit model, and ordered probability model (supervised statistical models used for TSV prediction)	Air temperature, water vapor pressure, metabolic heat generation, and clothing insulation	TSV	Ordered probability and multinomial logit models outperformed LR in predicting TSVs, achieving 10–30% higher accuracy; multinomial logit was slightly better for individual TSVs, whereas ordered probability better captured TSV distributions under specific conditions
[139]	Tianjin, China, and West Lafayette, USA/Outdoor human subject experiments with 26 participants exposed to varying microclimates. Environmental conditions, skin temperatures, and thermal sensation votes were monitored through instrumentation and questionnaires.	SVM	Environmental parameters (air temperature, relative humidity, global solar radiation, wind speed), Physiological variables (skin temperatures, mean skin temperature, thermal load) and clothing	Cool discomfort, comfort, and warm discomfort, derived from the TSV scale	An SVM model using local skin temperature and thermal load predicted thermal sensation categories (cool discomfort, comfort, warm discomfort), with a highest 91.5% prediction accuracy using hand and head skin temperature points; the model improved up to 6.9% when combining multiple inputs and demonstrated strong generalization across different climates and seasons
[146]	Tehran, Iran, specifically along green sidewalks. Data were collected through field investigations, including on-site microclimatic measurements and questionnaire-based surveys to capture subjective thermal comfort responses.	Supervised ML techniques were applied to predict outdoor thermal comfort, although the specific algorithm was not disclosed.	Air temperature, relative humidity, wind speed, global solar radiation, and PMV values calculated using both ENVI-met simulations and a mathematical formula.	PMV	The supervised ML model effectively predicted thermal comfort from measured environmental conditions, offering a human-centered approach for green sidewalk design optimization, and enhancing urban outdoor quality of life
[137]	A case study in Zürich, Switzerland, crowd-sensing data collected via smartphones to map urban air temperature; validation with fixed weather stations.	Quantile Regression Forest (QRF), an extension of RF used for predicting conditional quantiles of urban temperature distributions.	Smartphone temperature readings, location coordinates, time of day, land use data, surface cover, elevation, and meteorological parameters.	Urban air temperature maps at high spatial resolution	Crowd-sourced sensor data enabled high-resolution urban temperature mapping despite uncertainties, with RF models accurately estimating spatial temperature using few reference stations for validation, offering a cost-efficient and scalable alternative to CFD models for urban heat mitigation and planning
[69]	Southwestern Spain (Mérida, Córdoba, Seville); field monitoring of 22 courtyards during the warm season using on-site environmental measurements and morphological analysis.	SVR	Geometric and environmental features such as courtyard AR, location coordinates (longitude, latitude), height above sea level, and climatic classification of the site	Predicted indoor air temperature within the courtyard environments	The SVM model predicted courtyard microclimates ~1 °C RMSE and <0.05% relative error during peak overheating, matching CFD accuracy at lower computational cost and highlighting ML’s potential for energy-efficient courtyard design
[47]	The study developed a data-driven thermal comfort prediction model using optimized tree-type ML algorithms, including Extreme Gradient Boosting (XGBoost), RF, and DT.	5 ML algorithms: DT, RF, XGBoost, AdaBoost, Bayesian Ridge Regression	Air temperature, relative humidity, MRT, wind speed, solar radiation, and PET calculated using Rayman	PET	A user-friendly data-driven model predicted OTC using optimized tree-based ML algorithms, with XGBoost achieving the highest accuracy (95.21%), validated via five-fold cross-validation, enabling accurate PET prediction from publicly accessible input data, making it suitable for non-experts and practical applications like construction site management
[44]	Guangzhou, China. High-resolution air temperature mapping using sparse weather station data (321 stations), remote sensing (land cover), and socio-economic variables.	Eight ML models, regression kriging, multiple linear regression (MLR), RF, XGBoost, ANN, etc.	Air temperature data from stations, Land cover (1 m resolution, from DL), Socio-economic variables, Geographic coordinates: latitude, longitude, altitude, time variables (hour, day)	Hourly air temperature distribution at neighborhood/building scale (∼200 m resolution)	Among eight spatial prediction models for urban air temperature, the integration of Regression Kriging (RK) combined with MLR achieved the highest predictive accuracy (R2 = 0.959, RMSE = 0.92 °C, Mean Absolute Error or MAE = 0.70 °C) over one month of hourly data, with high-resolution land cover and observed spatial-temporal variability strongly enhancing performance, although residuals increased under high variability
[45]	Hong Kong, a high-density city used as a case study for fine-scale, hourly thermal environment mapping during the summer season. RF algorithm was applied to generate 100 m resolution maps of air temperature, relative humidity, and net effective temperature, incorporating meteorological variables, topography, and local climate zone-based landscape features.	RF	Meteorological drivers, Topographic variables, Local Climate Zone (LCZ)-based landscape drivers:	Hourly air temperature (Ta) and relative humidity (RH) at 100 m spatial resolution; and Net Effective Temperature (NET), a thermal comfort index that combines Ta and RH to estimate perceived heat stress	A high-resolution (100 m) hourly thermal comfort dataset for Hong Kong (2008–2018) analyzed with RF achieved high accuracy for air temperature (R2 = 0.87, RMSE = 1.12 °C) and RH (R2 = 0.80, RMSE = 5.38%), while incorporating NET revealed intensified nighttime heat stress in dense urban areas, underscoring the relevance of fine-scale modeling
[140]	Hyderabad, India, a major metropolitan area with a tropical wet and dry climate. The study utilized Landsat 8 imagery integrated with meteorological data to estimate OTC indices.	SVM	Inputs were derived from Landsat 8 imagery and used to assess spatial variations in thermal comfort across Hyderabad (Normalized Difference Vegetation Index or NDVI, Normalized Difference Water Index or NDWI, New Built-up Index or NBI, last surface temperature or LST, etc.)	Temperature–Humidity Index (THI), which is used as an indicator of outdoor thermal comfort, quantifying the combined effect of temperature and humidity on human well-being.	Distinct spatial variations in THI across Hyderabad showed highest mean values over barren lands (27.3), followed by built-up (26.9), vegetation (24.1), and water (20.7), with LST ranging from 27.39 °C (water) to 31.94 °C (vegetation); mean THI exceeded 25 °C in most land cover classes during the summer, and SVM effectively identified high thermal stress areas for targeted urban cooling strategies
[46]	Xi’an, China (cold climate zone; Köppen BSk/Cwa). Field study combining in situ microclimatic measurements and subjective thermal comfort surveys (questionnaires) during the summer. The dataset was divided into two categories (with shading and without shading) and used to train nine ML models.	Nine supervised machine learning models were evaluated: XGBoost, Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost) with Bayesian optimization, RF, SVM, DT, KNN, Logistic Regression, NB	Environmental variables (air temperature, relative humidity, wind speed, MRT, solar radiation) and human-related features (age, gender, clothing, body mass index or BMI, metabolic rate, emotional state)	TSV (scale of how hot/cold a person feels), thermal acceptability (TA, i.e., whether the person finds the thermal condition acceptable), and thermal comfort vote (TCV, i.e., overall comfort level expressed by the participant)	Interpretable ML models predicted TSV, TA, and TCV from microclimate and physiological data, with shading-based dataset division improving accuracy by 9.2%, 9.31%, and 6.16% in unshaded spaces, and Bayesian optimization further enhancing performance by 6.83%, 4.05%, and 2.55%, respectively
[58]	Xi’an, China. Real-time meteorological measurements and questionnaire surveys conducted in three urban park environments: water-side square, pavilion, and tree-shaded space, differing in SVF and landscape features	Classification (KNN, SVM, DT, AdaBoost, RF, MLP) and regression approaches (Ridge Regression, Least Angle Regression, and Bayesian Ridge Regression) were employed to predict children’s outdoor thermal sensation by linking physiological indices and facial expression data.	(a) Environmental parameters (air temperature, relative humidity, SVF, landscape type) (b) Physiological indices (Mean ear skin temperature, Mean cheek temperature, Heart rate) (c) Facial expression metrics (Proportion of negative emotions) (d) Activity intensity level	TSV, TCV	A highly accurate (97.1%) ML model predicted children’s outdoor thermal sensation from facial expressions, ear skin temperature, and heart rate, identifying ear skin temperature and negative emotions (sadness, disgust) as key non-invasive indicators across thermal stress and activity levels
[59]	Shiraz, Iran—a rapidly urbanizing city experiencing UHI effects. Use of Landsat satellite imagery and machine learning to predict LST variations linked to urban land use/land cover changes.	RF, SVM, DT, XGBoost, and DL	(a) NDVI, Normalized Difference Built-up Index (NDBI), NDWI (b) Land use/land cover (LULC) categories (vegetation, soil, built-up) (c) Elevation, slope (d) Air temperature, humidity, wind speed, solar radiation (e) Configuration/urban morphology metrics	LST	Urban landscape configuration strongly influenced LST, with built-up and vegetated areas showing lower LSTs at the urban boundary; among six ML models, XGBoost performed best, followed by DL, while LULC transformation and urban expansion drove UHI intensification over 2006–2021
[84]	Shenyang, Northeast China (Severe Cold Zone). Field measurements and continuous questionnaire surveys during spring (April 2023) in two microenvironments: open grassland and shaded rest area on a university campus.	RF, Gradient Boosting Trees (GBTs), XGBoost, and ANN	(a) Environmental factors: air temperature, wind speed, relative humidity, radiation temperature (b) Individual factors: gender, age, height, weight, time in Shenyang, clothing insulation	TCV (based on questionnaire responses and representing the subjective thermal comfort level reported by participants); treated as the dependent variable in all ML models applied in the study	XGBoost achieved the highest OTC prediction accuracy (r = 0.9313), outperforming GBT (0.7693), RF (0.7291), and ANN (0.5311), with comfort influenced by environmental and individual factors, showing significant differences across climate origins and clothing insulation levels
[77]	Nagpur City, Maharashtra, India—a city with a tropical wet and dry climate. Investigation methods: (a) Field measurements of microclimatic parameters; (b) Surveys of pedestrian thermal sensation using questionnaires; (c) Urban morphological analysis using remote sensing.	Two RF models were developed. Model included all microclimate and street geometry parameters, while model 2 used only the following most important features: air temperature, solar radiation, and selected street geometry parameters	(a) Microclimatic parameters: air temperature, solar radiation, relative humidity, wind speed (b) Street geometry parameters: AR, orientation, SVF	Modified Physiological Equivalent Temperature (mPET)	RFR models accurately predicted pedestrian thermal comfort (mPET) in tropical urban streets (up to 98% accuracy), while a simplified RFR using only air temperature, solar radiation, and street geometry retained high predictive power, supporting practical urban planning applications for thermal comfort
[138]	Freiburg, Germany. Microclimate data from stationary and mobile measurements were used, combined with high-resolution 3D urban morphology data.	Both SL (RF) and DL (MLP with three hidden layers)	Meteorological parameters Land use factors (vegetation, impervious surface, buildings, water, etc.) Surface characteristics (albedo, emissivity, etc.) Building density SVF	Ta, RH, wind speed, MRT, and UTCI at street-level (and the final output)	An ML framework using RF and MLP predicted high-resolution UTCI with MAE of 2.3 °K, enabling fast, city-wide hourly mapping at 1 m resolution and capturing strong spatial-temporal thermal comfort variability
[131]	Fenghuang County, Hunan Province, China—Field measurements and questionnaire surveys.	Eight ML models were tested: SVM, MLP, Logistic Regression, KNN, DT, XGBoost, NB, RF.	(a) Environmental parameters: air temperature, relative humidity, wind speed, global solar radiation, SVF (b) non-meteorological (questionnaire-based): gender, age, clothing, metabolic rate, time of exposure, etc.	TSV, TCV, thermal acceptance vote (TAV)	XGBoost outperformed empirical and deterministic models in predicting TSV, providing high accuracy, stability, and explainability; TAV achieved the highest prediction accuracy; TCV was less reliable due to subjective variability; and tourists adapted better with longer exposure and fulfilled expectations, while residents were more sensitive to metabolic rate
[65]	Guangzhou, China. Field experiment conducted at a nursing home for elderly residents using outdoor thermal monitoring and facial infrared thermography during 6.5 days of summer. Thermal comfort votes and facial skin temperature from 34 participants (avg. age 83) were collected under varied activity levels and conditions.	Six algorithms: RF, SVM, DT, Logistic Regression, AdaBoost (AB), and Gradient Boosting (GB); models trained on labeled facial and environmental data to predict thermal sensation votes (TSV)	Facial skin temperatures at five landmarks (forehead, left/right cheeks, nose, jaw) measured via infrared thermography, environmental parameters (air temperature, relative humidity, wind speed, solar radiation), and individual characteristics (age, gender, clothing insulation, activity level)	TSV recorded on a 7-point ASHRAE scale ranging from −3 (cold) to +3 (hot)	An RF model achieved the highest accuracy (79.6%) and area under the Receiver Operating Characteristic (ROC) curve (0.889) in predicting TSV; facial skin temperatures, especially on the nose and forehead, were the most influential features; the feasibility of using non-contact infrared thermography and ML models to assess elderly thermal sensation in real-time outdoor conditions was demonstrated
[147]	Isfahan, Iran. Field study in two urban public squares under hot-dry summer conditions, 500 visitors surveyed alongside microclimatic data collection at four distinct points using portable weather stations.	Supervised classification using a Binary Linear Classifier trained via Genetic Algorithm (GA); implemented within a soft-computing framework to distinguish between thermally comfortable and uncomfortable conditions based on labeled field data	Environmental parameters (air temperature, relative humidity, wind speed, solar radiation), clothing insulation (in clo), metabolic rate (MET), and Mean Radiant Temperature. All measured or estimated on-site in urban squares.	PET	A GA-trained binary-linear classifier accurately predicted thermal comfort in urban squares using PET-based thresholds, showing high agreement with actual comfort classes and demonstrating the feasibility of soft-computing for climate-responsive outdoor design
[85]	Gwalior, India. Field measurements in four urban street canyons during the winter, spring, and summer were used to train and validate ANN models for predicting thermal comfort indices in a composite climate.	Supervised regression using FFNN with a single hidden layer. Model parameters optimized using backpropagation with a 70-15-15 train validation test split.	Meteorological data (air temperature, wind speed, relative humidity, solar radiation), urban canyon geometry parameters (AR, orientation, canyon surface materials), temporal features (hour of day, month)	PET, UTCI	ANN models accurately predicted hourly PET and UTCI with high correlation and low RMSE; the best used full meteorological inputs, while a reduced-input model with only air temperature remained acceptably accurate; supporting ANN use for comfort forecasting with limited microclimatic data
[148]	Chengdu, China. Field surveys conducted in People’s Park during the hottest (July) and coldest (January) months involved on-site meteorological measurements and thermal comfort questionnaires at four sites with different landscape typologies, and with 419 valid responses collected.	RF	Meteorological variables (air temperature, relative humidity, wind speed, globe temperature, solar radiation), landscape type, personal attributes (age, gender, clothing insulation, activity level), seasonal context (summer vs. winter)	TSV (subjective thermal comfort reported by participants) on a 7-point ASHRAE scale from −3 (cold) to +3 (hot)	RF identified RH and Ta as dominant TSV predictors, varying seasonally (Ta in summer, RH in winter), while landscape type significantly influenced thermal comfort, with woods preferred in summer and lakeside/lawn in winter
[66]	Tainan City, Taiwan. ENVI-met simulations conducted for the courtyard of a senior residence under construction to assess OTC under hot and humid summer conditions.	DT algorithm to identify and rank the influence of 11 design parameters on thermal comfort.	Eleven courtyard design parameters including green coverage ratio, tree planting configuration, courtyard width-to-height (W/H) ratio, orientation, pavement material, floor reflectance, tree height, tree spacing, grass planting, and building layout.	Binary thermal comfort classification: comfortable or uncomfortable based on PET thresholds for senior courtyard users under hot-humid summer conditions simulated by ENVI-met	DT classification identified key courtyard design parameters affecting elderly thermal comfort; simulations reduced areas with PET > 38 °C by 14.3% and increased thermally acceptable areas (PET 30–34 °C) to over 54% by late afternoon, showing that vegetation and courtyard geometry modifications significantly enhance outdoor comfort in hot-humid conditions
[49]	Singapore. Year-long, city-scale field study leveraging smartwatches to collect thermal comfort perception and physiological data	Supervised classification using XGBoost optimized via Bayesian methods. Shapley Additive Explanations (SHAP) analysis was applied to interpret feature contributions	Physiological data (heart rate, skin temperature) from smartwatches, geolocation (GPS), time of day, activity level, environmental variables (air temperature, relative humidity, wind speed, solar radiation), urban characteristics (green view index or GVI, land use type, building density)	Thermal comfort levels	XGBoost with Bayesian optimization achieved 66% accuracy in predicting real-time thermal comfort preferences, with heart rate, GPS location, and solar variables as key features, and SHAP analysis confirming nonlinear physiological and environmental impacts, supporting user-aware thermal planning
[135]	Chongqing, China. Field measurements and ENVI-met simulations conducted for sunken urban squares during the summer.	Supervised regression using XGBoost, trained to predict PET from simulation-derived microclimate and geometric features. SHAP analysis was applied to interpret the relative influence of environmental and design variables.	Microclimatic parameters: global horizontal radiation, solar altitude, air temperature, wind speed, relative humidity; Geometric/design parameters: SVF, slope angle, orientation, width-to-height ratio (W/H), square depth	PET	XGBoost predicted PET in sunken squares with high accuracy (R2 = 0.991); SHAP analysis identified SVF, slope angle, and orientation as key nonlinear design parameters, with PET positively correlated with SVF and AR, negatively with slope angle, W/H ratio, and depth, while greening (especially shrubs) unenhanced thermal comfort more than material choices
[86]	Nis, Serbia. Field data collected at the Observatory Nis during the summer months.	Three soft computing models: ANN, ELM, and GP; trained to predict thermal sensation votes based on meteorological inputs.	Meteorological parameters, including air temperature, relative humidity, wind speed, globe temperature, and solar radiation	PET	Among ANN, ELM, and GP, ELM achieved the highest PET prediction accuracy, providing robust results despite environmental complexity, and supporting its use for assessing OTC in urban spaces
[48]	Republic of Cyprus (Nicosia, Limassol, Larnaca, Pafos, Ammochostos). Field surveys conducted in public squares, pedestrian streets, and promenades between July 2019 and February 2020. Simultaneous environmental monitoring and questionnaires were used to collect meteorological and physiological data and thermal sensation votes.	RF, SVM, MLP-ANN with 2 hidden layers, and Linear Discriminant Analysis (LDA)	Environmental variables: air temperature, mean radiant temperature, relative humidity, wind speed, globe temperature, solar radiation and PET. Personal parameters: gender, age, clothing insulation (in clo), metabolic rate (activity level), location and time of day.	Thermal sensation	Τhe RF classifier outperformed MLP, SVM, and LDA in predicting TSV, providing more accurate classifications than PET-based methods, especially for extreme sensations, highlighting PET’s value as input while demonstrating ML’s superior ability to capture complex comfort patterns across urban sites and seasons
[136]	Sendai, Japan. MRT Data at pedestrian height (1.5 m) were obtained from radiative heat balance simulations over the summer of 2019. These were combined with historical weather data (2014–2018) from the Japan Meteorological Agency.	Supervised regression using three ANN approaches: (a) FFNN, (b) Backpropagation Neural Networks or BPNN, and (c) GA–optimized BPNN (GA-BPNN); trained to predict MRT based on weather and solar geometry inputs. PCA and K-means clustering were applied to reduce the training dataset size while maintaining prediction accuracy.	Meteorological and solar parameters: solar altitude, solar azimuth, air temperature at true solar time, minimum daily air temperature, and global solar radiation	MRT at pedestrian height (1.5 m), predicted as a continuous thermal comfort index using supervised neural network models	The GA-optimized BPNN outperformed FFNN and BPNN, achieving the highest MRT prediction accuracy (Mean Absolute Percentage Error or MAPE < 1%), while PCA and k-means clustering reduced the training dataset by ~70% while preserving accuracy
[149]	Shenzhen, China. Field measurements + on-site thermal comfort surveys + machine learning analysis (XGBoost + SHAP).	Extreme Gradient Boosting (XGBoost) with SHAP analysis for model interpretability	(a) Climatic factors (air temperature, wind speed, solar radiation, black globe temperature, relative humidity; (b) physiological factors (respiratory rate, BMI); and (c) environmental and morphological factors (GVI, SVF, vegetation density)	Thermal comfort	Green exposure significantly improved OTC in subtropical environments, with optimal comfort at GVI 20–56%; low BMI and respiratory rate further improve comfort, amplified by moderate air velocity (0.5–1 m/s) and SVF (12–20%)

(for SL),

Table 3. UL applications in OTC studies.

Reference	Location/Investigation Method	UL Method	Input Parameters	Output Parameters	Main Results
[98]	Tehran, Iran, where OTC was assessed using satellite imagery and ground-based climatic data	PCA was applied to reduce dimensionality and inter-correlation among input features derived from surface biophysical indices and meteorological variables	Landsat-derived LST, NDVI, NDWI, land-cover, DEM, in situ meteorological data (air temperature, relative humidity, wind speed)	Principal components forming a composite comfort index calibrated to the Discomfort Index (DI)	The PCA model demonstrated strong predictive performance (correlation up to 0.89, RMSE = 1.15 °C), revealing that DI was on average 8.5 °C higher in the warm season, with bare land being the most uncomfortable surface and water bodies the most comfortable, while the approach proved efficient for large-scale thermal comfort modeling
[150]	Taipei, Taiwan (subtropical dense urban area) with meteorological big data from the Taipei Weather Station covering June–August (2011–2020).	K-means clustering was applied to categorize rainfall weather patterns and assess their impact on apparent temperature (AT) and physiological equivalent temperature (PET) under different rainfall scenarios	Hourly meteorological data (temperature, RH, wind speed, precipitation, cloud cover)	Apparent temperature (AT) and PET under different rainfall scenarios	In subtropical oceanic climates, summer rainfall significantly influences urban OTC, with 37.7% of rainfall events improving comfort and 62.3% leading to its deterioration

(for UL), and

Table 4. DL applications in OTC studies.

Reference	Location/Investigation Method	DL Method	Input Parameters	Output Parameters	Main Results
[143]	Simulation-based study using 2400 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 3848 samples collected from eight sites on 19 December 2021	Heat Island Intensity (HII), Coupling Effect Intensity (CEI), and UTCI	The ANN model showed high predictive performance, with a relative area under the ROC curve of 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 OTC 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 the 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 conditional GAN	The conditional GAN model accurately predicted UTCI maps from urban configuration images with a Structural Similarity Index (SSIM) of 96%, generating results in ~3 s compared to 30 min 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

(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 the 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 intercorrelation 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.

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 conditional GANs) 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 conditional GANs 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, 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.

Across the reviewed literature, several consistent performance patterns emerge. Tree-based ensemble models (RF, XGBoost, LightGBM) consistently produced the most accurate and stable predictions in structured environmental datasets, especially when predicting PET, UTCI, or TSV using meteorological and urban morphology variables [46,47,84,131,135,148]. Their dominance in field-based OTC applications is explained by their robustness to multicollinearity, ability to capture nonlinear interactions, and minimal preprocessing requirements.

In contrast, linear and classical statistical models (LR, ordered probability, multinomial logit) performed adequately when the relationships were approximately linear or when sample sizes were limited [145], but they generally underperformed relative to ensemble ML methods under complex, highly variable outdoor conditions.

DL approaches demonstrated clear advantages in high-dimensional and spatially structured problems, such as image-based UTCI mapping, CFD emulation, and facial-expression-based comfort prediction [58,138,143,152]. CNNs and generative models were particularly effective when spatial invariance and feature abstraction were required. However, in smaller field datasets, DL did not consistently outperform optimized tree-based models and typically required larger training samples and higher computational cost.

Dataset structure proved decisive. Studies incorporating physiological inputs (skin temperature, heart rate, facial features) reported substantial gains in predicting subjective comfort responses [49,58,65], indicating that purely meteorological inputs may be insufficient to capture individual variability. Conversely, simplified models using reduced feature sets retained acceptable predictive performance in several cases [77,85], highlighting opportunities for scalable and cost-efficient urban applications.

Transferability remains a critical limitation. Most models were trained and validated within single climatic or cultural contexts, and few studies assessed cross-city or cross-climate robustness [151]. Macroclimatic conditions, adaptation levels, and behavioral expectations significantly influenced prediction reliability, particularly for subjective indicators such as TCV and TAV.

Across the reviewed studies, datasets varied substantially in structure and scale. Most field-based investigations relied on in situ microclimatic monitoring combined with thermal perception surveys, typically involving sample sizes ranging from a few hundred to a few thousand observations. Target variables were predominantly thermal indices (UTCI, PET) or subjective responses such as TSV and TCV. Model evaluation was generally performed using hold-out validation or k-fold cross-validation schemes, with performance reported through metrics such as R², RMSE, MAE, classification accuracy, or F1-score. External validation across different.

Overall, the evidence indicates that ensemble SL models offer the most reliable balance between accuracy, interpretability, and operational feasibility for OTC applications. Meanwhile, DL methods offer distinct advantages for high-resolution spatial modeling and surrogate simulation frameworks.

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 in 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:

$$\mathrm{Min}/\max \left\{\begin{array}{c}{f}_1\left(x\right)=c_1^T·x\\ f_2\left(x\right)=c_2^T·x\\ ⋮ ⋮ ⋮ \\ f_k\left(x\right)=c_k^T·x\end{array}\right\},$$

$$\mathrm{subject} \mathrm{to} x\in S, \mathrm{where} S=\left\{x\in \left.{\mathbb{R}}^n\right|A·x\begin{array}{c}\le \\ =\\ \ge \end{array}b, x\ge 0, b\in {\mathbb{R}}^m\right\},$$

(10)

where k is the number of objective functions, n is the number of decision variables, m is the number of constraints, f_k(x) is the k-st objective function, x is n-dimensional vector of the decision variables, c_k 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 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 SBO 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. Case studies applying optimization methods for OTC enhancement across diverse urban and climatic contexts.

Reference	Location	Optimization Method	Design Variables	Objectives	Main Results
[136]	Sendai, Japan	GA was applied to optimize the training of a backpropagation-based ANN, adjusting weights and thresholds to predict MRT	Meteorological and solar parameters: solar altitude; solar azimuth; air temperature at true solar time; minimum daily air temperature; and global solar radiation	MRT	The GA-optimized ANN significantly improved the prediction accuracy of MRT under future climate conditions in Sendai, Japan, while the hybrid optimized model more reliably captured long-term seasonal variations and microclimatic dynamics than the standard ANN, with reduced MAE
[185]	Tianjin Tuanbo tennis center, Tianjin, China (semi-outdoor space)	ANN and GA	Stadium geometric and morphological parameters	Maximize thermal comfort, expressed in ${PC}_{ave}$ (average percentage of comfortable seats with UTCI between 9 °C and 26 °C)	The study proposed a surrogate-based optimization method combining ANN and GA to improve semi-outdoor stadium design, introduced ${PC}_{ave}$ as a new comfort evaluation metric, and achieved an 8.96% improvement in thermal comfort, demonstrating the feasibility of AI-driven morphological design
[187]	Shenyang, China	Two-stage SBO was performed using a hierarchical GA: Stage 1 optimized network morphology, while Stage 2 refined patch-level configuration, without the use of surrogate models	Stage 1: network-level morphology (street orientation, road network density, intersection spacing, block size and layout); Stage 2: patch-level morphology (vegetation distribution, green space configuration, patch shape and size)	Urban thermal field variance index (UTFVI)	UTFVI was reduced by up to 35% in key areas, with optimization identifying 20 TC sources, 20 thermal discomfort (TD) sources, and 78 thermal corridors; barrier points and core areas were spatially mapped; and network-level design was found to be the most effective
[171]	Beijing, China	Simulation-based multi-objective optimization was conducted using NSGA-II, integrated with parametric modeling	Tree parameters: height, crown diameter, spacing, layout pattern, species selection, and greenbelt width	Minimize UCLI for thermal comfort; maximize daylight factor; maximize GVI	UCLI was reduced by 12.7%, daylight factor improved by 26.3%, GVI increased by 35.8%, and trade-offs between shading and daylight were effectively balanced through tree layout optimization
[188]	Guangzhou, China (hot and humid climate)	Deterministic SBO was conducted using a dynamic local energy balance model	Land-use configuration: building layout and spacing; vegetation coverage; and road orientation	Minimize cooling energy demand and improve OTC (PET, MRT)	A parametric optimization strategy using a dynamic energy balance model identified optimal ranges for building density (0.45–0.5), building height (10–20 floors), and greenland/woodland ratios (0.25–0.4); building density and floor number were the most influential factors on a comprehensive indicator (CI) combining cooling demand and PET; and the study demonstrated that spatial configuration significantly affects energy and thermal performance, offering guidance for high-density residential design in hot-humid regions
[189]	Xi’an, China	Parametric spatial optimization was performed based on TSV and UTCI evaluations	Vegetation arrangement, surface materials, user type (patients vs. healthy), and microclimate characteristics	Maximize OTC for different user groups in hospital open spaces	Patients have higher neutral UTCI (18.8 °C) than healthy individuals (16.9 °C); thermal comfort design should be tailored by user type; and the key drivers are air temperature and black globe temperature
[172]	District 12, Tehran, Iran	Multi-Objective Genetic Algorithm (MOGA) optimization was performed using NSGA-II implemented via the Wallacei plugin in Grasshopper	Block orientation, height, spacing, and vegetation percentage	Minimize MRT and UTCI; and maintain RH within a comfortable range	Optimized block orientation (+35°), spacing (6 m alleys), and H/W ratio reduced MRT by 3.34 °C, UTCI by 2.91 °C, and PET by 0.52 °C; peak thermal stress shifted earlier in the day, enhancing OTC
[173]	Various kindergartens located in high-density residential areas, China	GA optimization was implemented using the Galapagos plugin in Grasshopper	Building shape, façade orientation, and location	Minimize outdoor thermal stress, expressed by UTCI	Optimization reduced thermal stress from 6.53 °C to 5.37 °C in Tianjin and from 3.57 °C to 2.87 °C in Shanghai; ladder-shaped buildings with context-sensitive orientations were most effective; solar radiation was more influential than wind in improving thermal comfort; and the framework offers actionable guidance for early-stage building design focused on OTC
[190]	Nanjing, China (subtropical monsoon climate)	Two-level parametric and enumeration-based optimization	Building orientation, height, spacing, layout type, and site coverage	Improve UTCI-based thermal comfort across seasons: reduce summer heat stress and mitigate winter cold stress	The two-level optimization framework improved OTC by modifying both the layout and form of urban building clusters; compared to the base case, the optimized plan reduced mean UTCI by 0.73 °C at noon in the summer and increased it by 1.91 °C at night in the winter, indicating better thermal comfort in both cooling and heating seasons; layout had a stronger effect in the winter, while lower-level refinements (e.g., wind corridors, solar access) further improved local microclimate conditions
[191]	Shenyang City, China (cold climate)	GA via Galapagos in Grasshopper, coupled with UTCI-based simulation using Ladybug	Urban block type (e.g., multistory vs. high-rise); street width; building placement layout; and north-side building configuration	Maximize OTC by increasing the percentage of time during which the UTCI remains within the acceptable range (−17 °C to 20 °C) during daytime hours on a typical winter day	Optimization achieved 87.7% UTCI compliance at the block level and 90.3% at the street level; multistory blocks outperformed high-rise and open layouts in cold climates; and placing dense block enclosures on the north side was effective for wind protection and thermal comfort enhancement
[192]	Kashgar, China (hot and dry climate)	SBO using GA via Galapagos in Grasshopper, coupled with Ladybug, Butterfly (Computational Fluid Dynamics), and OpenFOAM for microclimate simulation	Building height, orientation, length and width; open space layout; and street orientation	Minimize outdoor thermal stress by reducing UTCI and MRT through the optimization of urban block form and spatial layout	Average UTCI decreased from 31.17 °C to 27.43 °C (−3.74 °C) and mean MRT reduced from 43.94 °C to 41.29 °C (−2.65 °C); the optimization framework effectively improved OTC in a hot-dry urban setting by adjusting urban form parameters
[186]	Beijing, China (cold climate)	Multi-objective optimization (MOO) integrating daylight, sunlight hours, SVF, and OTC metrics for high-rise residential layouts in cold climates; a parametric model-controlled building layout; and simulation outputs were accelerated using a trained ANN as a surrogate model	Building layout parameters, including position; spacing; orientation; and rotation of high-rise residential buildings on a large urban site	Multi-objective optimization: maximize daylight; maximize annual sunlight hours; maximize SVF; and improve OTC	The MOO framework identified the top 10 building layout solutions (out of 150) that improved combined performance metrics by approximately 21% compared to the baseline; ANN surrogate modeling accelerated performance evaluation with an average prediction accuracy of 89.9%; and the optimized layouts offered balanced improvements across daylight, sunlight hours, SVF, and UTCI, demonstrating the potential of integrated optimization for sustainable high-rise residential design
[193]	Greater Cairo region, Egypt (hot-arid climate)	Parametric SBO using Ladybug tools in Grasshopper 3D; variations in building height, street width, and orientation were simulated to assess and optimize UTCI across different urban canyon configurations	Building height; street width; street orientation; symmetrical and asymmetrical height-to-width (H/W) ratios	Maximize OTC by minimizing diurnal average UTCI through the optimization of canyon geometry in hot-arid urban environments	UTCI reductions of up to ~6 °C were achieved through optimized H/W ratio and orientation; a strong negative correlation was found between UTCI and H/W ratio (R2 = 0.71 and R2 = 0.91, excluding E–W orientations); and optimized asymmetrical canyons outperformed regulatory baselines, supporting climate-adaptive planning in hot-arid cities
[194]	Cairo, Egypt (hot-arid climate)	Parametric SBO, a Design of Experiments approach, evaluated 3430 hypothetical urban configurations based on three typologies (scattered, linear, and courtyard) in a hot-arid climate with Ladybug Tools integrated in Grasshopper, aiming to identify optimal trade-offs between OTC (UTCI and PET) and energy consumption	Urban typology (scattered, linear, courtyard); building height; street width; plot ratio (density); block orientation; building spacing; footprint dimensions; and arrangement and distribution of open spaces	Simultaneously enhance OTC and reduce building energy consumption across diverse urban forms in hot-arid climates	(a) Strong correlation found between design parameters and combined performance (UTCI + EUI), with R2 = 0.84. (b) Urban density had the highest impact on both thermal comfort (R2 = 0.70) and energy use (R2 = 0.95).
[195]	Cairo, Egypt	MOGA implemented via Octopus in Grasshopper, using Pareto-based evolutionary optimization	Courtyard block orientation; building height; courtyard proportions (W/L); and interspace dimensions	Minimize outdoor thermal stress (UTCI) and reduce summer weekly cooling loads	Optimal layouts reduced UTCI by up to 1.6 °C and cooling loads by 31.7% compared to the baseline; preferred orientations were 45° for square blocks and 135° for elongated blocks; minimum interspaces were consistently favored; and solar exposure increased by ~4%, offering winter heating potential
[196]	Western Sydney, Australia	A MOO parametric framework was employed, integrating urban- and building-scale design variables within Grasshopper, with environmental performance metrics—OTC (PET) and energy demand for cooling and heating—simulated using EnergyPlus Version 8.8 via Honeybee and Ladybug plugins	Urban grid rotation (street layout angle); AR; building typology; building use; and window-to-wall ratio (WWR)	Minimize simulated annual building cooling and heating energy loads, and maximize OTC (evaluated using PET)	The optimization framework achieved up to 25.85% improvement in OTC and reductions of 72.76% in cooling loads and 93.67% in heating loads compared to the baseline; optimal solutions typically featured compact urban forms with lower AR, north–south grid orientation, and moderate window-to-wall ratios
[197]	Madinaty, New Cairo, Egypt	A generative design tool using a mutation-based evolutionary algorithm via Dynamo in the Revit/Autodesk environment was coupled with Ladybug and Grasshopper for simulation-informed iteration to evaluate OTC via UTCI	Tree count and spatial distribution, and neighborhood morphology types (clustered, semi-clustered, fully open)	Minimize UTCI in outdoor communal areas by optimizing vegetation layout (tree distribution)	Optimized tree distributions improved UTCI across all neighborhood types, with clustered layouts reducing UTCI from 38.0 °C to 37.55 °C using fewer trees, decreasing from 43 to 33; semi-clustered layouts improving UTCI from 39.40 °C to 38.01 °C with more trees, increasing from 27 to 45; and fully open layouts showing a slight UTCI reduction from 39.60 °C to 39.55 °C using fewer trees, reduced from 31 to 25

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.

Representative studies can be used to clarify how ML-assisted optimization is formulated in practice. These studies can be expressed explicitly in terms of decision variables, objectives, constraints, algorithms, and verification procedures.

For example, in study [186] (Beijing, cold climate), the optimization problem can be structured as follows:

➢

Decision variables: building position, spacing, orientation, and rotation within a high-rise residential layout.

➢

Objective functions: maximization of daylight availability, annual sunlight hours, sky view factor (SVF), and improvement of OTC (UTCI-based metric).

➢

Constraints: regulatory spacing requirements between buildings, minimum daylight thresholds, geometric feasibility of high-rise typologies, and climatic boundary conditions defined by local weather data.

➢

Optimization algorithm: multi-objective evolutionary algorithm (NSGA-II).

➢

ML integration: a trained ANN surrogate model was used to approximate simulation outputs and accelerate performance evaluation.

➢

Verification: optimized solutions were re-evaluated through high-fidelity simulations; surrogate model performance was validated using prediction accuracy (89.9%), and Pareto-optimal layouts were compared against a baseline configuration, demonstrating approximately 21% improvement in combined performance metrics.

Similarly, study [185] (semi-outdoor stadium, Tianjin) formulated the problem as:

➢

Decision variables: stadium geometric and morphological parameters.

➢

Objective: maximizing the percentage of thermally comfortable seats (PC_ave, UTCI-based).

➢

Constraints: architectural feasibility of stadium form and predefined spatial envelope.

➢

Algorithm: genetic algorithm coupled with ANN surrogate modeling.

➢

Verification: improvement of 8.96% in comfort metric relative to baseline, with predictive error quantified through MAE reduction.

The reviewed literature typically follows one of three approaches to verify optimized outcomes:

(i)

Re-simulation of Pareto-optimal solutions using high-fidelity microclimate tools (e.g., ENVI-met, EnergyPlus);

(ii)

Statistical validation of surrogate models using RMSE, MAE, or cross-validation techniques;

(iii)

Comparison with baseline or regulatory reference configurations.

Making these elements explicit enhances methodological transparency and shows that ML-assisted optimization in OTC is based on rigorous, validated computational workflows, not just a concept.

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.
• 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.
• 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.
• 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.
• 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.
• 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.
• The reviewed literature demonstrates that optimization applications have matured from simple geometric manipulations into 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 OTC 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 synthesized ML methodologies applied to OTC, examined their integration with SBO and surrogate modeling, and articulated a practical framework to support scalable, intelligent, and thermally adaptive urban design. By consolidating predictive modeling and optimization strategies, the study bridges methodological advances with implementation-oriented decision-making.

The reviewed literature spans diverse climatic and urban contexts—including Mediterranean, humid subtropical, arid, and tropical environments—and demonstrates the growing versatility of ML-based approaches in modeling nonlinear outdoor thermal dynamics. SL dominates the field, reflecting the availability of labeled datasets such as TSV, PET, UTCI, and DI. Tree-based ensemble models (e.g., RF, XGBoost) and neural networks consistently provide high predictive accuracy for both comfort indices and subjective responses. DL methods are increasingly adopted in high-resolution spatial mapping, image-based analysis, and surrogate simulation frameworks, where hierarchical feature extraction enhances performance. In contrast, UL remains limited in OTC applications, primarily serving exploratory clustering and dimensionality reduction tasks. RL and SSL, despite their success in related domains, have yet to be meaningfully explored in outdoor contexts, representing a substantial opportunity for future research.

Optimization studies reveal a clear methodological evolution from geometric parameter tuning to multi-objective, multi-scalar frameworks integrating thermal comfort, daylight access, and energy performance. While SBO remains the dominant paradigm, surrogate-assisted approaches are emerging as powerful alternatives, significantly reducing computational cost while maintaining acceptable predictive fidelity. Reported performance gains of 2–6 °C in PET or UTCI across diverse case studies demonstrate the tangible implications for public health and livability of optimization-driven design. The integration of parametric modeling environments with evolutionary algorithms further enhances practical applicability in early-stage urban planning.

Overall, the findings highlight three central insights. First, ML enhances predictive accuracy and enables the integration of heterogeneous data sources—including meteorological, morphological, physiological, and behavioral variables—thereby overcoming limitations of steady-state indices. Second, optimization transforms predictive outputs into prescriptive design strategies, allowing systematic exploration of trade-offs among competing objectives. Third, the convergence of ML, surrogate modeling, and multi-objective optimization offers a scalable pathway for climate-responsive and human-centered urban environments.

Despite these advances, key challenges remain. Model transferability across climatic zones and socio-cultural contexts is still limited, as most frameworks rely on location-specific training data. The subjective and adaptive nature of thermal perception complicates generalization and necessitates context-sensitive calibration. Furthermore, interpretability and uncertainty quantification remain insufficiently addressed in many applications, particularly in DL workflows.

A critical methodological consideration concerns validation rigor and model generalization. Many reviewed studies rely on random train–test splits or k-fold cross-validation without explicitly accounting for spatial or temporal autocorrelation. Data leakage may occur when observations from the same site, season, or measurement campaign are distributed across training and testing subsets, leading to overly optimistic accuracy estimates. Similarly, strong spatial clustering of microclimatic variables or repeated measurements over time can inflate predictive performance unless it is addressed through spatially or temporally disjoint validation schemes. Few studies implemented leave-one-site-out, cross-city, or cross-season validation protocols to assess true transferability. As a result, reported metrics such as R², RMSE, or classification accuracy may reflect within-context performance rather than genuine generalization across climates, urban morphologies, or populations. Future research should adopt more robust validation strategies, including spatial blocking, temporal holdout testing, and external cross-city validation, to ensure that ML models for OTC remain reliable beyond their original training environments.

Also, future research should prioritize three strategic directions. First, improving cross-climate generalization through multi-site datasets, domain adaptation techniques, and hybrid physics-informed ML frameworks. Second, expanding the use of advanced paradigms such as SSL, RL, and physics-guided neural networks to reduce labeling requirements and support adaptive, feedback-based outdoor systems. Third, embedding ML-driven optimization within real-time sensing networks, digital twins, and participatory planning platforms to enable dynamic and citizen-centered comfort management.

By integrating methodological synthesis with a structured optimization framework, this review shifts OTC modeling from a predominantly diagnostic exercise toward a proactive design instrument. The convergence of predictive ML, surrogate acceleration, and multi-objective optimization provides the foundation for resilient, scalable, and implementation-ready thermal comfort strategies in rapidly urbanizing and climate-vulnerable cities.