The objective of the RAPIDS DMU area is to accelerate all the ways applications use data. As DOE simulations grow in scale and adopt new architectures, the bottleneck increasingly lies not in raw compute but in how data is stored, moved, orchestrated, and understood. RAPIDS3 addresses this through a broad portfolio of methods and production software spanning four tightly connected efforts, with AI threaded throughout.

Scientific Data Management optimizes I/O for SciDAC applications across storage, metadata, and compression — for instance, partitioned metadata blocks that speed parallel object creation by up to 196×, and stability-preserving compressors that accelerate checkpointing several-fold while bounding physical error. Data Orchestration enhances workflow management for coupled, near-real-time science; representative advances include command-and-control for in-situ workflows and wide deployment of ADIOS streaming for coupled codes across leadership-class systems. Scientific Visualization and Understanding drives insight at scale through scalable, in-situ, and uncertainty-aware techniques — such as closed-form uncertainty methods reaching hundreds to thousands of times the speed of prior approaches, and production visualization with ParaView/Catalyst. Cutting across all three, AI automates and optimizes data operations, from LLM agents that generate visualization scripts to knowledge-graph-guided reasoning for explainable diagnosis. These are only a sample of a much larger body of work: RAPIDS3 continually develops, hardens, and deploys the I/O, orchestration, visualization, and AI capabilities that DOE science teams depend on.

Poster: DMU-poster.pdf