Performance Assessment of CFD and Physics-Informed Neural Network Approaches in Modeling Pollutant Dispersion within Complex Urban Geometries

This research investigates the potential of Physics-Informed Neural Networks (PINNs) as an alternative to Computational Fluid Dynamics (CFD) for simulating urban air pollutant dispersion within a representative street canyon geometry in the Bangladesh University of Engineering and Technology (BUET) area, Dhaka. While CFD methods provide high-fidelity spatial resolution, they demand extensive computational resources and rely on well-defined boundary conditions. Conversely, real-world urban sensing predominantly yields sparse scalar concentration measurements from low-cost sensors, rendering the inference of underlying velocity fields a severely ill-posed inverse problem.

This extended research monograph presents two complementary physics-informed deep learning frameworks: (1) a forward reconstruction PINN under the SST k-ω closure trained on 20,000 ANSYS Fluent CFD nodes, and (2) an inverse PINN that reconstructs in-canyon velocity, kinematic pressure, and concentration fields exclusively from 14 spanwise-averaged wall concentration measurements from the CODASC wind-tunnel benchmark without velocity supervision. Key innovations include Fourier feature input encoding, regularized tanh lid-driven cavity vortex seeding, and a two-stage curriculum learning strategy.

Note: A condensed 6-page version of this research appeared in the Proceedings of the 2026 IEEE 2nd International Conference on Quantum Photonics, AI & Networking (QPAIN) [DOI: 10.1109/qpain69676.2026.11546537]. This extended research monograph presents the complete mathematical formulations, comprehensive CFD grid sensitivity studies, full inverse PINN implementation, and hyperparameter sensitivity analyses.