Flow field reconstruction and wind pressure estimation from sparse measurements using physics-informed neural networks: Application to two-dimensional street canyon flow

• Physics-informed neural networks (PINNs) are used to reconstruct street canyon flow. • Key airflow characteristics are captured from sparse velocity sensor data. • Fourier feature PINN resolves high-frequency flow and reduces training steps by 30%. • Relative pressure on surfaces is estimated without relying on direct sensor data. Wind tunnel experiments offer precise insights into wind environments while they have notable limitations, particularly sparse data measurements and high time costs. In this study, we employed physics-informed neural networks (PINNs) to reconstruct the mean flow field in a two-dimensional urban street canyon by leveraging sparse sensor data to improve real-world measurement strategies. A large-eddy simulation dataset from the authors’ previous study was used to provide the mean velocity data at selected sensor locations and to evaluate the prediction accuracy for the entire domain. The performances of vanilla PINN (V-PINN) and Fourier feature PINN (FF-PINN) were compared for different sensor configurations. With a baseline sensor configuration of approximately H /6 intervals within the canyon ( H : canyon size), both models successfully captured the key airflow characteristics, including the recirculation vortex. The majority of the errors were concentrated near the building surfaces and shear layer, reflecting a lower accuracy in areas with steep velocity gradients. FF-PINN demonstrated a superior capability in resolving high-frequency phenomena, such as shear flow. It increased the spatial relations between nearby input coordinates and reduced training steps by 30%. With the maximum number of sensors, V-PINN and FF-PINN achieved the RMSE values of 0.008 U r e f and 0.010 U r e f ( U r e f : mean velocity at 2 H ), respectively. Additionally, the relative wind pressure on the building surfaces were estimated through integrated physics constraints, without relying on direct sensor data. V-PINN achieved a high correlation coefficient of 0.96 for the wind pressure on the building surfaces, whereas FF-PINN achieved an even higher value of 0.97.