Overview

Error-controlled lossy compressors are widely used to manage the large amount of data produced by scientific applications. Still, they may produce undesired compression artifacts that distort both raw and post hoc data analytics. This project aims to bridge the gap by developing a novel learning-driven framework to mitigate artifacts produced by scientific lossy compressors. The success of this project is expected to improve the integrity and quality of lossy-compressed scientific data significantly, thus facilitating the use of existing lossy-compression frameworks for efficient data storage, transmission, and analytics in scientific applications. This contributes to scientific discoveries in a broad range of domains, including climatology, cosmology, fusion energy science, and X-ray ptychography, as well as multiple aspects of research and education in advanced cyberinfrastructure.

This project addresses the artifact issue by leveraging recent scientific data compression and deep-learning advancements. In-depth investigations are conducted to generically characterize the compression artifacts produced by scientific compressors on both raw data and post-hoc analysis. This aims to improve the understanding of data quality and establish a benchmark for artifact mitigation. Next, deep learning models are designed to tackle artifact mitigation on both raw data and features of interest, with specifically designed transfer learning to reduce training costs. The quality of the recovered data is improved by fusing model outputs tailored to preserve different features. Finally, the quality of the recovered data is validated through tailored uncertainty quantifications, and the performance of the framework is investigated through careful optimization and parallelization. Integration into state-of-the-art error-controlled lossy compressors and incorporation with real-world scientific applications are expected to advance multiple scientific data management tasks, including data storage, I/O, and transmission.