Rapid prediction of human perceived microclimate through multiphysics-informed reduced order models

High resolution modeling of the urban microclimate is crucial for evaluating thermal exposure and mitigation strategies, yet remains computationally prohibitive when carried out with full Multiphysics simulations. This study presents a Reduced Order Modeling (ROM) framework that combines transient multiphysics simulations with machine learning to reproduce key microclimate variables needed for the assessment of physiological equivalent temperature, including mean radiant temperature (MRT), air temperature, and wind speed magnitude, at sub meter resolution. Twelve representative days in spring and summer 2024 in a university campus in Barcelona Spain were simulated, with detailed geometry and materials including trees, to capture multi seasonal and diurnal variability and to generate training data. After dimensionality reduction with Principal Component Analysis and hyperparameter optimization, Random Forest and Gradient Boosting regressors achieved high predictive accuracy, with mean absolute errors of 0.84 °C for MRT, 0.34 °C for air temperature, and 0.17 m/s for wind speed, and coefficients of determination (R2) exceeding 0.89 across all variables. Model inference reduced computational time by more than four orders of magnitude, enabling near instantaneous prediction of high-resolution microclimate fields from five meteorological inputs. The proposed approach provides a scalable pathway for real time forecasting, mesoscale to microscale downscaling, and rapid scenario evaluation in urban thermal exposure modeling.