Research On Key Information Extraction And Visualization For Flow Field Data Using Deep Learning

Flow visualization is an indispensable part of computational fluid dynamics(CFD).It provides an important way for domain experts to gain insight into the physical phenomenon and acquire flow field cognition,and plays a significant role in scientific research and engineering practice.With the rapid development of high-performance computer and the continuous progress of computing technology,more and more complex CFD flow data has been generated.How to extract accurately and efficiently the key information from those data has become a research hostpot in the flow visualization field,and there are lots of research works have been done to solve this problem.However,the existing methods usually have the shortcomings of low accuracy,poor generality and high cost.Nowadays,artificial intelligence,especially deep learning,has achieved great success in many scientific and engineering problems and developed into an effective tool for information understanding and learning from large-scale and complex data,which can copy with many challenges and needs in flow feature extraction and visualization.To overcome these disadvantages,this dissertation proposes a series of novel and efficient vortex and key time-steps extraction methods using deep learning.The major work and innovation of this thesis are listed as follows:(1)Three efficient vortex extraction methods based on supervised deep learning are proposed.To deal with the generality and scaliablity deficiency of the existing methods on extracting vortices in flows,this thesis develops multiple vortex extraction models with different architectures including a convolutional neural network(CNN),a fully convolutional segmentation network,and a U-Net.Moverover,critical issues such as data sampling of different grid sizes,irregular grid topology feature fusion,and physicsinformed vortex extraction model design are addressed.The experiments show that compared with the traditional methods,these proposed approaches can detect the vortex structures more quickly and efficiently and achieve a better balance between the accuracy and generalization.(2)Two novel and efficient unsupervised vortex identification methods are proposed.Existing vortex extraction algorithms based on supervised deep learning require a large hand-labeled training set and its accuracy is only just approaching that of the labeled dataset,which limit the applied range.To address this issue,the dissertation first develops an automatic clustering approach to encode vortex-like behavior as the basis for programmatically generating large-scale,highly reliable training labels.Moreover,to speed up the clustering method,a multi-view U-Net(MVU-Net)model is proposed to approximate the clustering results using the knowledge distillation technique.A multi-view learning strategy is further applied to integrate the information across multiple variables.In addition,this thesis proposes a physics-informed loss function,which enables the model to explicitly consider the characteristics of flow fields.The proposed MVU-Net model is subsequently evaluated against several widely used vortex detection methods on both numerically-simulated and analytical flows.Results reveal that the proposed model can achieve both high accuracy and performance.(3)A deep learning framework for key time-steps selection of unsteady CFD datasets is proposed.To overcome the disadvantages in the existing key time-steps selection approaches,such as low accuracy and high cost,this thesis presents KTSS-Net,a deep learning framework for selection of key time-steps.Specifically,KTSS-Net first employs a physics-informed deep convolutional auto-encoder to learn the compact and representative latent feature descriptors of each time-step flow field.These feature descriptors can correctly represent the flow field information,and accurately reconstruct the original flow field.A manifold learning approach is then applied to reduce the noise contained in the feature descriptors and provide a 2D visual representation of the unsteady flow fields.After that,KTSS-Net performs the clustering in the 2D projected space to automatically select key time-steps.Qualitative and quantitative experiments demonstrate the effectiveness and efficiency of this proposed method.