Author(s)

MOHAMMAD JADIDI, MAHDI AZARPEYVAND

Abstract

Physics-informed neural networks (PINN) offer a route to rapid, physically consistent prediction of urban street canyon flow and thermal fields when full computational fluid dynamics (CFD) is too costly for parametric screening. This study evaluates a coupled momentum-energy PINN for an idealised street canyon with aspect ratio AR =1.0, using k − ω turbulence model as reference data. A single façade is prescribed a temperature increase (Δω) as a surrogate for solar heating, while remaining walls are adiabatic and no-slip. The governing equations are enforced in non-dimensional form with buoyancy represented through the Richardson number Ri. To represent turbulent transport with a compact computational graph, the PINN employs an algebraic eddy-viscosity closure based on wall distance from the signed distance function, and training minimises a composite loss combining PDE residuals, boundary constraints and sparse supervised line data. Results show that the PINN accurately reconstructs the dominant canyon recirculation, with strong pointwise agreement for momentum variables (u*, v*) over the canyon region (R² = 0.96–0.97). The temperature field exhibits lower pointwise parity (R² = 0.56), with errors concentrated in the roof-level exchange band and near-wall thermal layer. The findings indicate that PINN can provide momentum reconstruction for ventilation assessment, while temperature accuracy is governed by resolution of thin thermal boundary layers and interface placement at the opening.

Keywords

physics-informed neural networks (PINN), urban street canyon, thermal instability, mixed convection, turbulent heat transfer