Reliable and timely estimation of air pollution exposure at high spatial and temporal resolution remains challenging in complex urban environments, where pollutant concentrations vary due to traffic emissions, urban morphology, and meteorological conditions. This study presents a physics-informed machine learning framework for near-real-time estimation of NO2 concentrations at fine spatial scales. The approach combines a limited set of steady-state computational fluid dynamics (CFD) simulations with operational meteorological and air-quality data. CFD simulations under specific wind directions are first used to characterize site-specific dispersion patterns. These outputs are then scaled using hourly meteorological observations to generate physics-based concentration descriptors. A machine learning predictor, implemented using Random Forest and Extreme Gradient Boosting, is trained to refine these estimates by incorporating additional environmental and observational features. The method is applied to a 1 km × 1 km urban district in Antwerp, Belgium, within the FAIRMODE intercomparison framework. Validation against measurements from 105 passive samples collected over one month shows substantial improvement compared to standalone dispersion modeling, with coefficients of determination up to R2 = 0.965 and reduced bias across locations. These findings demonstrate that integrating physical modeling with machine learning enables accurate and computationally efficient high-resolution exposure assessment in urban settings.
Bartzis et al. (Sat,) studied this question.