Physics-informed neural networks for passive scalar emission and transport

Accurate modeling of harmful pollutant concentrations is an important field of interest for protecting public health and the environment. In this study, physics-informed neural networks (PINNs) are applied to a low Reynolds number, time-averaged cylinder wake, which interacts with a variety of different passive scalar regimes. The PINN reconstructs both the time-averaged velocity and scalar concentration from limited measurements. In addition to satisfying the incompressible Reynolds-averaged Navier-Stokes (RANS) equations, the reconstructed fields must also obey the time-averaged advection-diffusion equation. This was done to extend the applicability of mean field reconstruction and to lay the foundations for PINNs to be used in more complex passive scalar modeling in future studies, particularly the prediction of pollutant behavior in urban environments. It was found that the PINN could successfully reconstruct the flow fields on a macro-scale in almost all scenarios, having considerable success with first-order quantities and managing to accurately infer the spatial structure of the unknown closure term of the advection-diffusion equation in all cases. When reconstructing scalar source characteristics, the PINN could identify the source location and size in many cases. The results allowed basic guidelines for optimal sensor placement to be described to inform future studies, and suggests ways in which this technique could be developed to contribute to future air quality forecasting models.