A Data-driven Based Extraction Method Of Fluid Characteristic Information And Its Applications

With the development of computer graphics,people began to simulate various physical phenomena in nature.The fluid flow is a kind of motion which seems simple but complicated,and the simulation of fluid is widely used in daily life.However,in conventional framework,the iteration method is time-consuming,so it is very important to accelerate the process of fluid simulation.In recent years,fluid feature extraction is a widely studied topic in the field of fluid simulation.The method of extracting fluid features based on vortex filaments can provide a simplified representation of flow field from the real world or simulation.This feature extraction method is to establish a large sparse complex energy matrix based on the data in a given 3D fluid velocity field,and to solve the minimum eigenvalue and its corresponding eigenvector of this matrix.The eigenvector is used to extract the filaments through one-dimensional contour tracking to obtain the characteristic information of the fluid velocity field.The fluid velocity field is reconstructed using the extracted vortex filaments.The key problem of this feature extraction method is to solve the minimum eigenvalue and its corresponding eigenvectors of the large sparse complex energy matrix.This paper aims to study the acceleration method of solving eigenvalue and eigenvectors of the large sparse complex energy matrix,to optimize the process of fluid feature extraction based on the vortex filament,and to achieve the acceleration of the feature extraction.The research of this paper can be divided into the following aspects:Firstly,considering that the previous method software package APPACK in MATLAB is too slow to calculate the minimum eigenvalue and its corresponding eigenvectors of large sparse matrix,in order to optimize the fluid feature extraction process,three traditional numerical iteration methods are tested: inverse power iteration(IPM),implicit restart block Lanczos method(IRBL),Jacobi-Davidson and preconditioned conjugate gradient method(JDCG),which complete the process of fluid feature extraction.By comparing the calculation cost,we not only find that JDCG method can optimize the feature extraction process,but also draw the conclusion that the traditional method is time-consuming.Secondly,in order to accelerate the fluid feature extraction,a datadriven method is proposed.The model includes:(1)Data acquisition module:we obtain a large number of sample data by the parameters adjustment of fluid velocity field and the basic operation of matrix;(2)Convolution neural network(CNN)model building module: we construct a convolution neural network suitable for input and output data;(3)Model Generation module:according to different inputs and outputs in different modes,three modes are constructed in the model with different parameters,which are training mode,evaluation mode and prediction mode.For training mode,the above convolutional neural network is trained by using the sample data to obtain the convolutional neural network model suitable for the calculation of the minimum eigenvalue and its corresponding eigenvectors of the large sparse matrix;(4)Model evaluation module: for the evaluation mode,the availability of our trained model is evaluated;(5)Prediction and application module: for prediction mode,we use the above evaluated convolution neural network model to calculate the minimum eigenvalue and its corresponding eigenvector of the complex energy matrix,in order to achieve the process of extracting feature information from the given three-dimensional fluid velocity field and reconstructing with vortex filament.Thirdly,this paper applies this data-driven feature extraction method to 9 different fluid velocity fields,including 5 simple streamline models and4 complex flow field models.The feature extraction process of the above different models is implemented respectively,and the images of the extracted features are obtained.We compare and analyze each experimental result and create computation overhead tables in order to compare the datadriven method(CNN)and the minimum overhead method(JDCG)among traditional numerical we find,showing the performance merit of our datadriven method,drawing the conclusion that the data-driven method proposed by our paper can be widely used in the feature extractions of different fluid models for acceleration,which is of great significance.