Estimating pollutant dispersion based on reconstructed mean flow field in a two-dimensional street canyon using two-step physics-informed neural networks

Source abstract

Urban environments exhibit complex physical phenomena involving multiple interacting and imbalanced variables, making high-resolution measurements challenging and expensive. This study aimed to reduce the measurement demand while accurately estimating the mean flow and pollutant concentration fields using physics-informed neural networks (PINNs). To improve the robustness of and mitigate the difficulties faced by PINNs when handling imbalanced outputs, this article proposes a two-step framework in which the airflow and pollutant dispersion within a two-dimensional street canyon are trained separately. In the first step, a PINN reconstructs the mean flow field from sparse velocity data derived from a large eddy simulation and simultaneously estimates the eddy viscosity. In the second step, these results are imposed as fixed fields in the advection–diffusion equation, enabling estimation of the pollutant dispersion from a ground-level line source. By leveraging the accurately reconstructed flow field, the second-step PINN achieves reliable reconstruction of the mean concentration field and turbulent diffusivity. Incorporating the eddy viscosity as additional physical information not only reduces the training cost and improves the accuracy, but also substantially reduces the data requirements. The eddy viscosity data ensure system closure, enabling physics-driven concentration field estimation even in the absence of concentration measurements. Furthermore, a comparison of the optimization strategies shows that the L-BFGS is more effective than Adam because of its second-order characteristics, facilitating convergence in ill-conditioned residual loss landscapes. Overall, the proposed method demonstrates the strong capability of PINNs to handle multiple outputs with imbalanced scaling in urban environments, highlighting their potential for broader applications.