Research On Flow In Situ Visualization Based On Deep Learning

With the improvement of computer performance,scholars of Computational Fluid Dynamics(CFD)can use more and more complex models to simulate the flow of fluid,so the accuracy of numerical simulation is greatly improved.But at the same time,the scale of numerical simulation is also expanding.This is a difficult I/O bottleneck for the traditional post-processing method that needs to save the result data in the hard disk first and then read it out for visualization.In high performance computing,although supercomputers can generate and process a large number of flow field data quickly,problems such as high data storage overhead and long I/O time seriously restrict the efficiency of flow visualization.This I/O bottleneck has been greatly alleviated by the in situ visualization.Scholars can use the computing resources on the same computing node to run numerical simulation and visualization code,which can visualize the generated data while the simulation program is running,avoiding the I/O operation on the hard disk.As one of the feasible methods of large-scale flow visualization,flow in situ visualization has received more and more attention.However,with the increasing data,the resource competition between flow in situ visualization and numerical simulation becomes more and more intense.With limited resources,how to efficiently mine the key data of flow field is a great challenge for flow in situ visualization.In the past few years,deep learning has achieved great success in CFD,showing its potential in processing flow field data and bringing opportunities to solve the problems faced by in situ visualization.Therefore,based on the deep learning technology,this paper conducts researches on the key technical issues of flow in situ visualization,and the main work and innovation achievements are as follows:1)A key time step selection method based on Deep Metric Learning(DML)is proposed.The the key time step selection of flow field is a common method of sampling the data in temporal dimension,which can reduce the frequent occupation of system resources.Aiming at how to find the time step with key information accurately,this paper introduces DML into the key time step selection,and proposes a method of key time step selection based on DML.In this method,we first design a neural network for similarity learning.After using Convolution Neural Network(CNN)to fully mine the data characteristics,we use metric learning to accurately obtain the similarity between time steps.On this basis,this paper proposes a key time step selection algorithm based on time interval,which can analyze the similarity between time steps to select the key time steps.Compared with the existing local methods,this method has advantages in accuracy,precision and recall.2)An in situ compression method for flow field data based on Generative Adversarial Network(GAN)is proposed.In situ compression is an effective method to reduce the network transmission pressure in loosely coupled mode.In order to solve the problem that the existing data compression methods are not efficient enough,this paper introduces GAN into the data compression of flow field,and proposes an in situ compression method for flow field data based on GAN.In this method,a neural network for data compression and reconstruction is designed.Compared with the existing lossy compression method,this method has advantages in compression time.Besides,it can adjust the compression ratio according to the acceptable reconstruction effect.3)A shock detection method based on CNN is proposed.It is the key and timeconsuming part to detect the characteristics of shock wave and locate its position in the flow field.Aiming at the defects of the existing shock detection methods,this paper proposes a CNN-based shock detection method.In this method,combining with the physical characteristics of shock wave,a loss function for shock wave detection is designed.On this basis,this paper designs a neural network for shock feature detection.Compared with the existing shock detection methods which are not based on deep learning,this method has advantages in detection time.Compared with the shock detection method based on deep learning,the effect of the shock detected by this method is better.