Precision-preserving Compression of Scientific Data: Learn Mechanism from Data

Scientific data often exhibits well-defined model mechanisms, which can be characterized by differential equations. However, existing compression algorithms face challenges in distinguishing between information that encapsulates underlying mechanisms and redundant information in the data. Accordingly, these algorithms struggle to accurately assess the true level of meaningful precision 1 . To address this issue, we propose a compression method that incorporates mechanism learning and effective precision identification. By locally characterizing the data's spatial relationship with linear differential equations and minimizing the source term, we learn the mechanisms. This allows us to separate noise from the data while preserving its precision. Thus, detectable patterns are more likely to emerge, enabling effective compression 2 .