Towards predicting urban flows with GAVI and SHRED

<title>Abstract</title> <p>Urban wind conditions at street level govern pedestrian comfort, pollutant dispersion and the loads carried by structures, and planning and emergency response increasingly call for estimates delivered faster than high-fidelity simulation can supply them. We reconstruct the wind field over a district of Barcelona at pedestrian height from sparse point measurements, combining a geometry-agnostic variational autoencoder that compresses a large-eddy simulation database of sixteen incident wind directions with a shallow recurrent decoder driven by sensor histories. The reconstruction separates the latent state into a part that depends smoothly on the wind direction, represented by a truncated Fourier series in the angle, and a turbulent fluctuation predicted from the sensors by a model that receives no directional information. Each direction is withheld in turn from both stages. The direction-dependent part carries 71\% of the variance of the latent state and interpolates to an unsampled direction to within 0.6 of the turbulent scatter, recovering 68\% of the streamwise and 31\% of the spanwise field energy against a compression floor of 89\% and 78\%. Eight directions at a spacing of $45^\circ$ recover most of what sixteen provide. The sensor stage transfers to a withheld direction at half the correlation it attains where it was fitted, yet contributes little, because the amplitude at which its prediction should be added varies between directions and cannot be inferred from those available for calibration.</p>