Physics-Informed Neural Networks for Predicting Steady Incompressible Flow Around Obstacles in Urban and Aerodynamic Settings

This study presents the development and application of Physics-Informed Neural Networks (PINNs) framework for solving steady-state incompressible fluid flow around canonical 2D obstacles, such as block structures, Quonset huts, and cylindrical bodies. The method integrates the governing equations and boundary conditions directly through the neural network loss function, enabling mesh-free prediction, and data-efficient learning of velocity and pressure fields. Boundary condition strategies such as slip, no-slip, and pressure outlet constraints are implemented. Collocation points are generated using randomized and ground-biased distributions, with ongoing investigations into adaptive sampling methods for enhanced resolution in critical flow regions. Additionally, a direct neural network (NN) model trained purely on data is developed and compared against the PINN results. Predictions from both approaches are benchmarked against high-fidelity CFD simulations using STAR-CCM+, showing agreement in vortex structure and residual convergence. This PINN approach demonstrates potential for rapid and flexible flow prediction in scenarios where traditional CFD may be computationally expensive or constrained by geometry—such as urban wind modeling, environmental hazard prediction, and early-stage aerodynamic design.