Lattice Boltzmann Method Model Optimization Based On Machine Learning

The multi-relaxation lattice Boltzmann method has attracted much attention because of its advantages such as flexible boundary conditions,easy programming and changing working conditions,etc.However,its theoretical convergence criteria are not clear and the calculation efficiency is slow,which leads to the accuracy and time cost problems of calculation fluctuation.In this paper,Machine Learning is introduced to excavate the convergence relationship from high-dimensional data in order to realize the optimization scheme with strong generality.The Lattice Boltzmann method was used as the source of data production,and its convergence was optimized to meet the requirements of data purity.The advanced neural network was used to train the data set produced by the lattice Boltzmann algorithm to predict the flow field information of low-dimensional labels,and the time cost of solving the problem was avoided.In this paper,the research of the relaxation of the Lattice Boltzmann precision convergence and optimization provides a new thought,also realized with a large number of the flow field information of prediction to improve the efficiency of specific conditions.The optimization process of Lattice Boltzmann provides a new way of research and general conclusion,also offers a new network architecture for the prediction of flow information.(1)Verification of multi-relaxation Lattice Boltzmann optimization strategy for machine learning.In order to verify the feasibility of machine learning method for multi-relaxation lattice Boltzmann convergence optimization,a large number of data sets of heat transfer conditions were generated,the Random Forest method,whose degree of freedom is low,was adopted to train and establish the black box relationship between the Boltzmann input characteristics and its output results.Gini coefficient based on Random Forest algorithm identifies the input feature,the relaxation parameter s1 and s2,which has a great influence on the convergence of calculation.After screening and dimensionality reduction,a decision boundary which can express the convergence explicitly is obtained as 0.52s1+0.51s2=1.(2)Selection strategy of multiple relaxation Lattices Boltzmann relaxation parameters for neural networks.In order to fully test the convergence of multirelaxation Lattice Boltzmann,the dataset with initial flow values and convergence labels is generated by the pipeline flow,and the initial value features,s2,s3,s4 and Re which have great influence on the results are screened out.The multilayer perceptron is used to fit the convergent/divergent decision space pointed by the input initial vector,and the high-dimensional decision space is visualized,which can observe the decision region of relaxation parameters s3 and s4 under specific s2 and Re intuitively.This decision region is extended to the square cavity flow problem,and the vortex center position and flow plot are compared in one normal case and two extreme cases,and the characteristics of the residual curve are analyzed.It is proved that machine learning has a high accuracy in the convergence/divergence prediction of input parameters.(3)Velocity prediction of square cavity flow field based on Generative Adversarial Network.After ensuring the data reliability of with multi-relaxation lattice Boltzmann method,the flow velocity information of the top square cavity with labels was produced to provide training data.Adjust the conditions to generate the overall architecture,generator architecture and discriminator architecture of the antagonistic network,select the most suitable flow field information prediction architecture and scheme,and finally realize the low-dimensional vector tag to predict the flow field information.By testing the model on unknown data sets,the rationality of the network architecture and the accuracy of the prediction effect are proved.