A hybrid AI-physical modelling framework for city-wide heat stress assessment

Urbanisation and climate change are intensifying outdoor heat stress in cities, yet neighbourhood-scale mapping remains constrained by the high computational cost of resolving mean radiant temperature (Tmrt)—a key driver of human heat stress—across heterogeneous urban morphologies. We present a three-phase, modelling framework that emulates high-resolution radiative processes with artificial intelligence to enable rapid, city-wide heat-stress assessment. First, we generated a large training dataset using SOLWEIG simulations over Local Climate Zone (LCZ) archetypes that span realistic ranges of urban form and vegetation, described by building height, building and impervious fractions, canyon aspect ratio, sky view factor, and pervious, tree, grass fractions. Deep neural networks were then trained to predict these Tmrt percentiles from morphology, meteorology, and sun-position predictors. Validation against independent, high-resolution SOLWEIG simulations at two representative sites yielded mean absolute error of 2.59 °C (LCZ06) and 3.28 °C (LCZ02), capturing the daily cycle dynamics. Finally, we deployed the emulator on a 100-m hexagonal grid to produce hourly Tmrt and Universal Thermal Climate Index (UTCI) fields for the 2022 July heatwave in Santander (Spain) using hourly ERA5 as meteorological forcing. The resulting maps reveal strong intra-urban contrasts, with daytime hotspots in more open/low-rise zones and higher nocturnal stress in compact mid/high-rise areas, underscoring the role of shading and longwave trapping. This framework bridges physical realism and operational scalability, supporting rapid identification of heat-risk hotspots and scenario testing of urban adaptation measures.