Long Time Simulation of Flow Dynamics Based on Transfer Learning and Fourier Neural Operator

The use of deep learning for fluid dynamics simulation has become a hot topic. However, whether physics informed neural network or neural operator are used to solve the long time-dependent partial differential equation, the accuracy of the predicted solutions inevitably decreases gradually with time. We propose a transfer learning enhanced Fourier neural operator to improving accuracy and stability of Fourier neural operator for the long term prediction of partial differential equations. In the experiments of 2D cavity flow, our method reduces the error at the 1000th time step by 87.83% compared to the traditional Fourier neural operator in autoregressive prediction.