Fourier neural operator for real-time simulation of 3D dynamic urban microclimate

Global urbanization has underscored the significance of urban microclimates for human comfort, health, and building/urban energy efficiency. However, analyzing urban microclimates requires considering a complex array of outdoor parameters within computational domains at the city scale over a longer period than indoors. As a result, numerical methods like Computational Fluid Dynamics (CFD) become computationally expensive when evaluating the impact of urban microclimates. The rise of deep learning techniques has opened new opportunities for accelerating the modeling of complex nonlinear interactions and system dynamics. Recently, the Fourier Neural Operator (FNO) has been shown to be very promising in accelerating solving the Partial Differential Equations (PDEs) and modeling fluid dynamic systems. In this work, we apply the FNO network for real-time three-dimensional (3D) urban microclimate simulation. For modeling large-scale urban microclimate problems, CityFFD simulates urban microclimate features based on the semi-Lagrangian approach and fractional stepping method with the Smagorinsky large eddy simulation model. In our simulation, the 1200 sequential time steps are used as training data. We retain and analyze the data from all stages, including the spin-up period, because we wish to understand how the flow develops transiently from initial conditions, and both one-step and sequential timestep predictions are analyzed. When applied to unseen data with different wind directions, the FNO model has a 0.3% one-step prediction error and a maximum error of 5%. A real-time simulation of urban microclimates in 3D is possible with the FNO approach, which is 25 times faster than the traditional numerical solver. •Fourier neural operator (FNO) was applied to simulate 3D urban wind field.•FNO is demonstrated to learn 3-D N–S equations for reconstruction of urban microclimate.•Different instantaneous wind direction results obtained when trained only for one wind direction.•Real-time predictions achieved within 0.08 s: 25 times faster than physics-based GPU solver.•Limitations identified and FNO codes and training data available for public sharing.