A 3D discrepancy modeling framework for urban pollution prediction in accelerated time

Computational Fluid Dynamics provides reliable high-resolution simulations of atmospheric transport and dispersion. However, its high computational cost limits applicability in time-sensitive scenarios such as emergency response to toxic releases in urban areas. As a faster alternative for decision-making, we previously proposed MCxM, a cost-effective surrogate learning framework for pollutant exposure prediction. A key limitation was its restriction to a single horizontal layer near ground level, suitable for efficiency, but insufficient to capture the full 3D behavior of pollutant dispersion. In this work, we extend the discrepancy modeling to 3D with MCxM-3D, which refines a 3D simplified physical prior using neural operators. The model is trained on realistic pollutant distributions in built-up areas, generated by a Lagrangian particle dispersion model under varying meteorological conditions. Evaluation on unseen urban configurations shows an average 20% (up to 51%) reduction in prediction error over the 2D approach, with millisecond-scale inference enabling real-time deployment.