Collaborative Research: OAC Core: Mitigating Artifacts in Scientific Data Compressors with a Learning-Driven Framework

This research proposal aims to address the critical need for suppressing compression artifacts produced by various error-bounded lossy compressors when they are applied to diverse scienti?c datasets across different domains. This is motivated by the fact that while error-bounded lossy compressors have been widely used to address storage and I/O problems in scienti?c applications, they may not provide suf?cient data quality for domain scientists’ data analysis such as visualization and topological feature extraction. Lossy compression artifacts refer to noticeable data distortion in reconstructed data, which may easily distort the outcomes in those analyses. However, automatically detecting and removing the artifacts from the raw data and user’s post-hoc feature analysis is very challenging, because (1) today’s lossy compressors are generally developed with diverse compression design principles, leading to large different artifacts on the data; (2) the properties of artifacts depend on sophisticated factors including compression design principle, error-control levels (such as error bound), the properties of data to compress, and user’s requirement on the reconstructed data or feature preservation. In this proposal, we plan to develop a learning-driven framework to mitigate artifacts produced by scienti?c lossy compressors. We propose to deeply engage deep learning (DL) in the framework due to their recent success in multiple domains. This project has three key thrusts which are described as follows. • Thrust 1 : Characterization. We will perform in-depth investigations to generically model/characterize the artifact effects produced by scienti?c compressors on both raw data and post-hoc analysis such as topological feature extraction. Knowledge learned in this thrust will not only improve the understanding of data quality in scienti?c compressors but also be leveraged to bene?t model design in the next thrust. • Thrust 2 : Model design. We will develop DL models to mitigate compression artifacts in both raw data and topological features. In particular, we will design a framework to streamline advanced DL models for artifact mitigation with integrated knowledge of the underlying scienti?c compressors. The ultimate goal is to provide artifact suppression that can be tailored toward diverse user requirements for any scienti?c compressor. • Thrust 3 : Optimization. We will optimize the proposed framework to achieve high quality and high performance. On the one hand, we will provide uncertainty quanti?cation to measure and understand the quality of the proposed artifact suppression methods. On the other hand, we will optimize the performance with advanced GPUs and parallelize it in distributed systems for exascale readiness. Our framework will signi?cantly impact both scienti?c domains and the compression community be- cause it has the potential to substantially improve the quality of reconstructed data without impeding com- pression ratios or performance. The framework will be extensively evaluated on multiple DOE applications based on collaboration with corresponding domain scientists. Speci?cally, we will evaluate our artifact suppression framework on multiple scienti?c applications, which include both unstructured datasets and structured datasets.