A model pretrained on simulated detector readouts matched a scratch-trained model on one neutrino-flavor classification task while using far fewer labeled examples, according to a paper published in Nature Machine Intelligence on September 30. The comparison—about 1,000 labeled events versus 10,000—uses simulated data for a proposed CERN detector upgrade, not measurements from a working FASERCal detector.
Image: Simulated events and self-supervised reconstruction from the authors’ paper, shared under CC BY 4.0.
The study models FASERCal, a proposed calorimeter upgrade for CERN’s FASER experiment. Its central 3D calorimeter divides detector volume into more than 460,000 small readout cells, or voxels, though only a fraction register activity in any one event. At teraelectronvolt (TeV) energies, particle showers and tracks can overlap, making conventional reconstruction difficult and labeled examples expensive to prepare.
Instead of beginning with examples tagged for each final task, the researchers first trained a sparse vision-transformer model to learn patterns from simulated detector readouts. It hid most active patches and inferred missing activity, then learned related structure: which signals were “ghost” hits with no matching true particle, how to distinguish primary from secondary activity, and which broad particle category contributed to a detector cell. The team then fine-tuned the shared encoder for tasks including neutrino-flavor and charm-quark identification, momentum estimation and interaction-vertex reconstruction.
Neutrino flavor means whether a neutrino is electron-, muon- or tau-type; the paper’s six classification categories also separate tau-decay channels and neutral-current events. On this task, the model using both masked-patch reconstruction and relational training scored 0.818 macro-AUROC with 1,000 labeled events, compared with 0.807 for the same downstream architecture trained from scratch with 10,000. Macro-AUROC averages how well a classifier separates each category. Both models were fine-tuned; this was not a frozen-model comparison.
Pretraining also improved several simulated reconstruction tasks, and the authors report transfer tests on public plastic-scintillator and liquid-argon benchmarks. Those tests do not establish performance on experimental FASERCal data: the FASERCal study itself uses simulation. The authors describe “foundation-style” narrowly and say a general detector model would need broader detector coverage, stronger systematic checks and calibration against experimental data.