Fast And Accurate Flow Simulation Based On Deep Graph Network

In recent years,with the enhancement of computing power,the demand for high-precision simulation of aerodynamic flow fields in aviation,automobiles,and weather forecasting is also increasing.However,fine simulation will lead to an increase in the dimensionality and scale of data generated by calculations,making engineers unable to efficiently conduct design space searches.With the breakthrough progress of deep learning methods in the field of computer vision,researchers have tried to use deep learning techniques such as convolutional neural networks and recurrent neural networks to predict aerodynamic flow fields.However,these work are still still in terms of prediction accuracy and prediction speed.There are huge challenges.This paper attempts to apply the existing graph neural network to the aerodynamic flow field simulation,and effectively integrate it with the physical solver in the traditional numerical simulation,aiming to provide a new solution for the aerodynamic flow field simulation.The main research work and innovations of this article are as follows:(1)The capabilities of three graph neural networks for aerodynamic flow field simulation are tested,compared the error and speed performance differences of different graph neural networks in predicting aerodynamic flow fields,and analyzed the number of network layers,optimizer types and activation functions The influence of the type on the network forecasting effect.(2)Based on the graph neural network aerodynamic flow field acceleration model,an improved model of the coupled physics solver is proposed.In the improved model,two preheating methods are used to quickly obtain better iterative initial solutions,one uses the physical solver in the preheat module,and the other uses neural network pretraining to preheat the backbone network.The physics solver is also used to refine the module to ensure that the model output conforms to the laws of physics.Numerical experiment results show that the improved model can effectively reduce the number of solver iteration steps and speed up the flow field calculation while ensuring the validity of the predicted data.