Author(s)

REUBEN NODDER, MOHAMMAD JADIDI, MAHDI AZARPEYVAND

Abstract

Accurate and computationally efficient prediction of turbulent flow in urban street canyons is needed for rapid assessment of ventilation and pollutant dispersion in decision-making. High-fidelity computational fluid dynamics (CFD) is accurate but too costly for real-time use and parametric studies, while field measurements are sparse. Purely data-driven AI surrogates can be fast, but often require large, labelled datasets, generalise poorly across changing geometry and meteorology, and may violate conservation constraints. Physics-informed neural networks (PINNs) address these issues by embedding governing equations and boundary conditions in the loss, enabling data-efficient, physics-consistent prediction. For turbulent street canyons, however, physics-only PINNs can struggle to couple the outer flow above roof level with the in-canyon recirculation across the roof-level shear layer that controls momentum exchange. We study data-augmented PINNs for an idealised street canyon with aspect ratio one, modelled as a cavity-type configuration driven by the over-roof flow. The reference case has Reynolds number 9000 (based on H and mean inlet velocity). We train a fully connected PINN for incompressible flow with interior collocation points enforcing PDE residuals and edge geometry points enforcing boundary constraints. To strengthen learning where gradients are sharp, we inject sparse CFD derived velocity samples and systematically vary (i) data placement (windward/leeward walls, multiple elevations) and (ii) sampling density. Performance is quantified with field-wise error metrics, focusing on the mean recirculation structure. Results show that targeted CFD augmentation improves robustness relative to a physics-only baseline; beyond a case-dependent threshold, additional samples yield diminishing returns. It is observed that increasing the number of supervised points per line from 20 to 334, a 16 fold increase, yields only a 9.5% decrease in mean pointwise velocity magnitude error. In contrast, reducing from 20 to 5 points per line increases the same metric by 71%, indicating that the model is markedly more sensitive to data sparsity than further data enrichment.

Keywords

data-augmented PINNs, urban street canyons, roof-level shear layer, hybrid physics-data learning, turbulent flow