High-dimensional flow field prediction method based on cognitive physical information neural network

Abstract

The invention discloses a high-dimensional flow field prediction method based on a cognitive physical information neural network, and the method comprises the following steps: collecting and preprocessing high-dimensional physical field data, and obtaining a training data set and a test data set; constructing a separable physical information neural network model; inputting a training data set with a physical field and a partial differential equation corresponding to the training data set, and training by utilizing cognitive learning to obtain a trained neural network model; and through the trained neural network model, calculating physical quantities in different time-space coordinates in the test data set, and outputting a prediction result. According to the method, independent one-dimensional coordinates are coded by adopting the separable ion network, so that the limitation of a physical information neural network on computing resources and the complexity of flow field prediction are effectively reduced; the difficulty of predicting a sample is dynamically evaluated through the gradient size of a PDE residual error, and the model is adaptively optimized through a cognitive training scheduler, so that the robustness and generalization ability of flow field prediction in a physical boundary region are effectively enhanced.

Images