We describe a preliminary implementation of a data analysis tool that can characterize features in large scientific datasets. There are two primary challenges in making such a tool both general and practical: first, the definition of an interesting feature changes from domain to domain; second, scientific data varies greatly in format and structure. Our solution uses a hierarchical feature ontology that contains a base layer of objects that violate basic continuity and smoothness assumptions, and layers of higher-order objects that violate the physical laws of specific domains. Our implementation exploits the metadata facilities of the SAF data access libraries in order to combine basic mathematics subroutines smoothly and handle data format translation problems automatically. We demonstrate the results on real-world data from deployed simulators.
Full paper in gzipped postscript (409KB) and pdf (216KB).