Abstract
Neural network models in deep learning have represented expectable predictive performance as surrogate models for accelerating numerical fluid dynamics applications. However, existing neural network models still face challenges in accurately capturing the details of unsteady velocity fields, particularly when dealing with high Reynolds number airflow. Therefore, we address this problem to predict Large-Eddy Simulation (LES) results for the three-dimensional velocity field around a building using the emerging Fourier Neural Operator (FNO). The FNO can effectively learn the solutions to partial differential equations and predict them in a purely data-driven manner with high generalization ability. In task predicting the three-dimensional unsteady velocity magnitude over ten consecutive time steps, we found that FNO significantly outperformed the traditional Deep Neural Network (DNN), and the more training data volume used, the higher the FNO accuracy. We compared key FNO parameters such as training data volume, and the numbers of Fourier modes and Fourier layers to summarize their influence on model performance. Results show that the default Fourier mode number 12 can be enlarged to 18 as more training data is applied, thus increasing FNO accuracy. In addition, the default
Fourier layer number four can be decreased to one, avoiding overfitting.
