Skip to content
Open access

A physics-informed surrogate model for rapid prediction and source inference of pollutant dispersion in complex urban environments

Oct 2026 · Environmental Fluid Mechanics · Vol 26 · 0 citations · 53 references

Abstract

Pollutant transport in complex urban environments is strongly shaped by building-induced flow distortion, channeling, recirculation, and geometrical blockage, which together produce highly heterogeneous concentration fields. Although high-fidelity numerical simulations can capture these effects, their computational cost limits rapid prediction and source inference. This study proposes a physics-informed surrogate modeling framework for rapid prediction and source inference of near-ground pollutant dispersion in complex urban environments under quasi-steady wind conditions. Four source-dependent features are introduced to encode transport direction and geometry-constrained accessibility, namely along-wind distance, crosswind distance, line-of-sight indicator, and shortest-path distance. These features are used as inputs to deep-learning surrogate models, with DeepLabV3 + adopted as the primary architecture and compared against U-Net, Fourier neural operator (FNO), and Pix2Pix. Experiments based on simulation data from eight realistic urban blocks in Paris and Grenoble show that the proposed feature encoding consistently improves cross-layout prediction within the dataset, yielding R2(log) values of 0.72–0.85 across the tested architectures. An eight-fold single-block-to-others transfer experiment yields an average R2(log) of 0.80. When coupled with Bayesian source-location inference based on a precomputed surrogate response library, the framework achieves a finding rate of approximately 96%, compared with about 46% for a Gaussian-plume baseline. Response library construction for a new block requires approximately 1 min, while each online inference step takes about 1 ms. A preliminary evaluation on two ANSYS Fluent cases, neither used for training nor fine-tuning, yielded R2(log) values of 0.667 and 0.572, and the inference procedure reached the source from all 20 tested downstream starting locations, providing limited but encouraging evidence of cross-simulator applicability.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.