Scalable machine learning for urban thermal morphology classification: clustering high-resolution microclimate and dynamic mapping of radiant temperature

Recent advances in ultra-fine and ultra-fast urban climate modelling have improved the identification and representation of complex urban environments. However, their practical application of Mean Radiant Temperature (MRT) time series generated by these tools remain difficult to interpret for non-expert stakeholders, necessitating a transduction of complex numerical data into actionable spatial information. Leveraging LUCIDiT—an ultra-fast simulation tool enabling high-resolution MRT maps across continuous urban domain—this research proposes a scalable, automated framework for classifying different “thermal morphologies” based on the MRT time series analysis. The capacity of unsupervised machine learning to capture the nonlinear relationships of urban morphology is investigated trough a rigorous comparison between a physics-based heuristic approach, a shape-based algorithm (K-Shape), and K-Means clustering applied to both engineered thermal features and MRT time series. Results demonstrate that applying K-Means directly to the raw time series outperforms the other methods in terms of both computational efficiency and mathematical robustness, as confirmed by the Silhouette Score and Davies-Bouldin Index. Furthermore, the physical accuracy and interpretability of the data-driven clusters are successfully validated by projecting the spatial outputs into a novel 2D Thermal Phase Space. This computationally efficient methodology, applicable at different urban scales, converts the physical complex and synergistic dynamics of the urban microclimate into intuitive and easy-to-use spatial thermal maps. It connects advanced numerical simulations to practical urban design, providing a compact representation of urban thermal behaviour and enables the large-scale identification of recurrent microclimatic patterns fundamental for urban heat mitigation strategies.