OmniLearned Jumps From Collider Jets to Neutrino Physics 50% Faster
A foundation model trained on collider jets transfers to few-GeV neutrino scattering, beating from-scratch baselines on MINERvA data.
Takeaways
- OmniLearned, a jet-physics foundation model, transfers effectively to few-GeV neutrino scattering on MINERvA data.
- Pre-trained 3M-parameter model beats identically-sized from-scratch transformer at the same FLOP budget on all tasks.
- Pre-training reaches best from-scratch validation loss in about 45 to 50 percent fewer steps.
- Largest gains appear at low pion energies and low three-momentum transfer, where nuclear effects dominate.
- Backbone was frozen for the 53M-parameter variant; multi-pion high-W tagging still trails the cut-based baseline.
- Code released at github.com/gregorkrz/minerva-ml; training fits on a single A100 in under 8 hours.
A foundation model built for high-energy collider jets just proved it can moonlight as a neutrino physicist. In a new study, researchers took OmniLearned, a transformer pre-trained on particle-level representations of proton-proton and electron-proton collisions, and dropped it into a completely different corner of physics: few-GeV neutrino-nucleus scattering events from the MINERvA experiment at Fermilab. Pre-trained OmniLearned models consistently outperform similarly sized models trained from scratch, hitting better performance at the same compute budget and at the same number of training steps.
That result matters because the two domains barely look like the same field of science. Collider jets involve hundreds of particles slamming into a hermetic detector at TeV energies, while neutrino scattering produces a handful of reconstructed objects at GeV energies inside an asymmetric segmented calorimeter. If a single pre-trained backbone can generalize across that gap, foundation models for physics start to look a lot more like foundation models for language.
Why neutrino experiments needed help
The motivation is not academic. A multi-decade experimental campaign is trying to pin down neutrino oscillation parameters using few-GeV accelerator beams, and the biggest limiting factor is the messy nuclear physics of how neutrinos interact with the target material. Few-GeV neutrino-nucleus scattering uncertainties are dominated by a large number of complex, relatively low-energy nuclear effects, and those uncertainties are expected to be limiting for next-generation flagship experiments like DUNE and Hyper-Kamiokande.
To attack this, the MINERvA collaboration built a finely segmented scintillator tracking calorimeter at Fermilab and recently released their simulated data and analysis tools publicly. That release is what made the current work possible.
Bridging a medium-transfer gap
OmniLearned had already been stress-tested on what its authors call near transfers (between collider experiments, across pp and ep collisions) and far transfers (to cosmological structure formation and molecular dynamics). What was missing was a medium transfer: still recognizably particle physics, but with different detector technology and underlying interaction physics. Neutrino scattering fits that slot exactly.