Developing a GNN-based large-time-step surrogate model towards real-time prediction of urban wind environment for air mobility

The rapid emerging of urban air mobility (UAM) in recent years necessities real-time and high-fidelity prediction of low-altitude urban wind environments, mainly to support the operational safety regulation and path planning. State-of-the-art approaches including computational fluid dynamics (CFD) cannot meet the requirement of instant prediction. This study proposes a graph neural network (GNN)-based large-time-step surrogate model to provide real-time and high-fidelity flow field prediction to support UAM operation safety. High-precision urban wind environment datasets, generated via CFD, were first validated against wind tunnel experiments and then expanded to train the GNN model. The prediction performance of the surrogate model was comprehensively evaluated in both idealized and real-world urban building settings. The results demonstrate that the proposed model overcomes the constraints of the Courant-Friedrichs-Lewy (CFL) condition, and the time step is 3 orders of magnitude larger than that of CFD. The single-step (0.5 s) inference of flow field consumes millisecond level computational time, achieving 3 to 4 orders of magnitude speedup and faster-than-real-time high-fidelity prediction. However, its accuracy declines as the iterative steps increase due to the accumulation of recursive errors, resulting in insufficient precision for long-term prediction. Spatial analysis indicates that prediction errors are primarily localized in the leeward regions of buildings, manifesting as the underprediction or smoothing of velocity gradients within complex vortex structures. This research demonstrates the potential of proposed model to realize faster-than-real-time urban wind environment prediction to support UAM safety.