Modern scientific data sets, including both the results of numerical simulations and experimental measurements, are often of a scale and complexity that far exceeds current analysis capabilities. To address this data generation-analysis gap, we will develop effective and trustworthy algorithms and computational tools tailored for scientific data reduction that preserves critical features and quantities of interest.

Our approach consists of four interconnected technical aims, including: (i) graph active learning and contrastive learning to identify a "representative set" – reducing data demands for feature selection and extraction; (ii) symmetry exploitation to reveal inherent structure (e.g. invariant features and equivariant group actions), yielding interpretable data representation and vast reduction without information loss; (iii) multi-scale latent diffusion compression, supporting different trade-offs between computational cost and reconstruction fidelity; and (iv) implementation of an integrated HPC-scalable framework.

Achieving the proposed innovations in data reduction will result in robust, scalable, end-to-end automated scientific workflows, with guarantees of preserving information in features, structures, and physics, actionable uncertainty quantification (UQ), and improved resource management for data storage and movement. All of these are fundamental resources for supporting effective decision making and informing experimental design and tuning, which are essential to accelerate scientific discovery.