SLAC AI Compression Preserves Fine Scientific Data at 10–100× Smaller Size

SLAC targets the data bottleneck facing next-generation experiments

Researchers at the U.S. Department of Energy's SLAC National Accelerator Laboratory published a new AI-based compression method on August 24, 2026 for scientific datasets that can become too large to store, move and analyze efficiently.

SLAC says the method can typically reduce file sizes by roughly 10× to 100×, depending on the underlying data and required fidelity, while preserving fine-scale signals that conventional compression can erase.

The work was published in Nature Machine Intelligence and was tested across multiple data types, including materials and molecular measurements, solar magnetic-field data and photographs.

The method separates information by scale before neural compression

Instead of compressing an entire dataset uniformly, the approach first uses wavelet analysis to separate features by spatial scale. A neural network then learns compact representations of those scale-separated features.

That design matters because subtle structures in scientific measurements can carry the information researchers care about most. In X-ray experiments, for example, tiny speckle patterns can encode details about material structure and dynamics.

The system also supports region-of-interest decompression. Researchers can recover only a selected part of a compressed dataset, at an appropriate scale and resolution, instead of decompressing the entire file.

Why selective decompression could matter for large facilities

SLAC points to the Linac Coherent Light Source as one motivating use case. Future high-rate instruments can approach data-generation rates on the order of one terabyte per second, making conventional storage and retrieval pipelines increasingly expensive.

A compression system that preserves fine detail while allowing targeted retrieval could reduce storage pressure and speed up downstream analysis, particularly when researchers need to revisit only a small subset of a very large experiment.

What the results do and do not show

The reported 10–100× reduction is not a universal compression guarantee. SLAC states that the achieved reduction depends on the dataset and the fidelity researchers need to preserve.

The work is also presented as an additional AI-based approach that can operate alongside broader data-reduction techniques, not as a wholesale replacement for established scientific compression systems.

The important advance is the combination of multi-scale representation, learned compression and selective retrieval—features designed around scientific workflows where losing small but meaningful signals can undermine the experiment itself.