Towards Rapid and Data-Efficient Urban Climate Assessment using Machine Learning Frameworks

Abstract High-resolution physics-based urban climate models provide detailed representations of intra-urban thermal variability, yet their computational and data demands limit their use for rapid city-scale climate assessment and operational planning. This study develops and evaluates two complementary Machine Learning Frameworks (MLFs), FS-MLF (Fine-scale) and CS-MLF (Coarse Scale), that emulate near-surface air temperature from Atmospheric Dispersion Modelling System for urban environments (ADMS-Urban) simulations in Birmingham, UK. The study addresses two questions: (i) to what extent can machine learning support rapid prediction and spatial gap filling of high-resolution urban climate outputs? and (ii) how sensitive are fine- and coarse-scale representations to training-data availability and spatial configuration? In this study, FS-MLF predicts temperature at the Fine-scale using local neighbourhood information, while CS-MLF (U-MOSAIKS) learns compact coarse representations when only coarse-scale predictions are available. Both frameworks reproduced the dominant spatial thermal patterns with high accuracy, with FS-MLF achieving R² = 0.85–0.93 and CS-MLF R² = 0.83–0.95 on held-out test data. Performance remained strong using only 10% of training data, while geographically distributed random sampling (test R² = 0.91–0.93) matched or exceeded concentrated sampling spatially, highlighting the importance of spatial representativeness over sample density alone. Training-data experiments further showed convergence towards test R² ≈ 0.94–0.96 with 50–70% of available samples. The findings demonstrate that machine learning can substantially reduce the computational burden of high-resolution urban climate emulation while retaining spatial fidelity, providing a scalable pathway for citywide climate adaptation assessment, urban digital twins, and climate-resilient planning.