Research And Application Of Deep Learning Reduced Order Model Coupled With Self-attention

Numerical simulation is an emerging technology for the analysis of fluid mechanics.It uses computers to solve the flow control equation of the fluid to obtain the flow state of the fluid,so as to analyze the flow phenomenon.Numerical simulation method can achieve high accuracy when the algorithm parameters are set reasonably.However,its computational cost is high,and it is difficult to meet the needs of high real-time application fields.Reduced order models are approximation used in place of a complex dynamic governing equation model,which performs simulations within an acceptable time frame and limited memory capacity with sufficiently reliable results.Reduced-order models have been widely used due to their extremely fast simulation speed and relatively accurate results.However,the existing reduced order models still have some shortcomings.The traditional linear system reduced order models represented by proper orthogonal decomposition perform poorly in nonlinear scenarios.Auto-encoders have strong nonlinear mapping ability.But the models have too many parameters,and are difficult to train.So auto-encoders have poor performance.To address these issues,this paper proposes a deep learning reduced order model based on convolutional auto-encoder.Convolutional auto-encoder is used to project in the model,and long short-term memory networks are used for regression fitting.The reduced order model based on the convolutional auto-encoder proposed in this paper is compared and verified with other models in the 2D case of flow around a cylinder.The experimental results show that the long short-term memory network has the best comprehensive performance among several regression methods,and the convolutional auto-encoder also has higher accuracy and fewer parameters than the fully connected auto-encoder.But there is still a certain gap that compared to the accuracy of the proper orthogonal decomposition.To further improve the accuracy of convolutional auto-encoder,self-attention module is inserted into the reduced order model.The experimental results on the 2D case of flow around a cylinder show that the accuracy of reduced order model based on convolutional autoencoder with self-attention has been greatly improved,even surpassing the reduced order model based on proper orthogonal decomposition.The convolutional auto-encoder reduced order model coupled with self-attention is applied to air flow prediction on the London South Bank University.The result shows its performance in practical application.The 3D model of London South Bank University is modeled according to the real building size.There are more than 60 buildings in the 3D model,so the flow field is very complex.The experimental results show that the convolutional auto-encoder reduced order model with self-attention can accurately predict even the flow state with strong nonlinearity in complex practical applications,and has excellent performance.