Unsteady flow prediction based on a hybrid network-operator learning model

Fast and accurate forecasting of unsteady flow is a challenge for both fluid dynamics and machine learning study. In this paper, a hybrid network-operator learning model is proposed to predict the spatiotemporal dynamics of unsteady flow. The hybrid model first integrates the Fourier neural operator (FNO) and convolutional long short-term memory (ConvLSTM) network in parallel. The model is systematically evaluated in three typical unsteady problems, including the two-dimensional turbulence, flow around a stationary cylinder, and flow around an oscillated airfoil. Numerical experiments indicate that the proposed hybrid network-operator model exhibits higher accuracy than the vanilla FNO and ConvLSTM model. Additionally, the hybrid model shows excellent performance on long-period prediction and can be well generalized to cases with different dimensionless parameters (such as Reynolds number and Mach number).