Intelligent CFD Simulation Based On Deep Learning

With the rapid development of science and technology,traditional industries are transforming into intelligent industries based on internet,big data,and artificial intelligence.Computational Fluid Dynamics(CFD)can comprehensively and accurately predict the complex flow-reaction characteristics in simulated systems,serving as a crucial tool for optimizing designs and scaling up of chemical reactors.Nevertheless,CFD still consumes a lot of researchers’ time in CFD software setup,working condition debugging,etc.and still needs huge computational resources when optimizing operating conditions in a wide range or simulating large/industrial fluidized bed reactors.Therefore,it is urgent to utilize artificial intelligence methods to realize intelligent acceleration of CFD.In this research,Deep Learning(DL)method is used to complete the intelligent acceleration of CFD,which is mainly divided into two aspects: operational intelligent acceleration and computational intelligent acceleration.More specifically,the research involves three contents:CFD interface recognition,natural language processing of CFD manual,and production yields prediction based on the time-series transient flow rate data from CFD.The intelligent operation process is designed for working conditions that need to be set for CFD simulation.Firstly,a dataset,named CFData Set,was created based on CFD interface content.Then,Convolutional Neural Network(CNN)and Long Short-Term Memory(LSTM)were used to establish recognition model of CFD interface.This model was trained with CFData Set and public datasets,and the final recognition accuracy was as high as 94% after data enhancement and parameter optimization.Compared with recognition models for natural scene,the model has an 8% improvement in prediction accuracy on CFD interface content.Meanwhile,the CFD simulation manual described working-condition setting is effectively refined by natural language processing technology,discarding redundant content and keeping operation setting sentences.Then,the operation objects in the statement were found by matching the operation library.Finally,Open CV,recognition model of CFD interface,and pyautogui were used to judge the state of operation object,and locate and operate it in real-time,so as to complete the intelligent operation process.In the accelerated study of computational intelligence,a production-yields prediction model of fluidized-bed reactors for biomass fast pyrolysis was developed using LSTM neural network based on transient flow rate data from CFD.Subsequently,Attention Mechanism(AM)or CNN network was coupled to LSTM to establish ALSTM or CLSTM model.After parameters optimization,the CLSTM model was found to be the most effective in predicting the instantaneous flow rate of species,the relative errors of production yields for bio-oil,gas,and char were 2.44%,0.33%,and 2.73%,respectively,saving CFD simulation time by 30%.