What the study found
The study found that BOA Constrictor, a lossless neural compressor built on the Mamba state space model, achieved competitive compression on several scientific datasets. It sometimes matched or exceeded standard lossless compressors such as LZMA, ZSTD, and ZLIB, and it performed well on some high-energy physics data, computational fluid dynamics data, and cosmology data.
Why the authors say this matters
The authors say this matters because petabyte-scale data from high energy physics experiments creates a storage challenge. They conclude that BOA is a first step toward improving compression for next-generation scientific data.
What the researchers tested
The researchers tested BOA, a pseudo-streaming lossless neural compressor, on structured scientific datasets. These included ATLAS Open Data in HDF5 format, simulated particle collision records in HepMC v3, CMS Open Data in NanoAOD format, as well as datasets from computational fluid dynamics and cosmology.
What worked and what didn't
BOA achieved an effective compression ratio of 7.23× on ATLAS Open Data and 9.13× on simulated particle collision records when model size was included. In those tests, it outperformed the next-best traditional algorithms, and on CMS Open Data it obtained comparable or improved effective compression ratios within 5% of the next-best traditional algorithm. Its throughput was reported as about 3.5 to 45 MB/s for compression and about 1.5 to 25 MB/s for decompression, which the abstract says is not yet competitive with optimized algorithms such as ZSTD or LZMA.
What to keep in mind
The results come from a proof-of-principle implementation, so the abstract does not present them as a finished production system. The abstract also notes that BOA is strongest on high-entropy float32 payloads, that FP16 weights reduce model size without reducing predictive accuracy, and that the available summary does not describe other limitations beyond slower throughput.
Key points
- BOA is a lossless neural compressor based on the Mamba state space model.
- It achieved 7.23× effective compression on ATLAS Open Data and 9.13× on simulated particle collision records.
- It outperformed the next-best traditional algorithm on those two datasets when model size was included.
- On CMS Open Data, it was comparable to or better than the next-best traditional algorithm within 5%.
- The abstract says BOA is slower than optimized compressors such as ZSTD and LZMA.
- FP16 weights reduced the model footprint without degrading predictive accuracy.
Disclosure
- Research title:
- Mamba-based BOA compressor matches or exceeds standard lossless tools
- Authors:
- Akshat Gupta, C. Doglioni, Thomas Joseph Elliott
- Institutions:
- University of Manchester, University of Manchester, University of Manchester
- Publication date:
- 2026-04-24
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.