A surrogate detector model for MINERvA open data reproduces 2p2h events it never saw in training
Synopsis
A probabilistic surrogate trained on the MINERvA open simulation dataset maps generator-level final-state particles directly to reconstructed observables without relying on generator-specific interaction labels. Its fidelity is assessed through conditional response distributions and classifier-based closure tests, and a model trained with all two-particle–two-hole (2p2h) events removed still reproduces their reconstructed response comparably to the full model. On MINERvA charged-current inclusive muon-neutrino data it reproduces selection efficiencies and purities and forward-folds GENIE, NuWro and GiBUU predictions into reconstructed space for direct comparison with the measurement.
Interpretation
Open neutrino data can now carry a new interaction model's predictions into reconstructed-observable space without the original collaboration's full simulation and reconstruction chain. Mapping theoretical predictions into reconstructed observables previously required computationally intensive methods inaccessible outside the experimental collaborations, and reweighting released simulated events works only where that simulation already covers the new model's phase space, with large weights diluting the sample's statistical power. On the MINERvA open simulation test sample, a gradient-boosted classifier separating original simulation from surrogate output using reconstructed quantities alone reaches an AUC of 0.504, and 0.511 when truth-level final-state summary features are added, against a statistical fluctuation of 0.001 for an uninformative classifier at this sample size.
The model takes the set of final-state particles and the interaction geometry as input and generates reconstructed events in the format of the open simulation dataset. Generator-specific interaction labels and quantities describing the underlying interaction, such as the squared four-momentum transfer, the hadronic invariant mass and the neutrino energy, are deliberately excluded so the model can be applied to any generator's final states. Training uses the 15.8 million simulated neutrino interactions of playlist 1A in the MINERvA open simulation dataset, split by subrun into roughly 80% training, 10% validation and 10% test, with all results taken on the test sample.
With all 2p2h events removed from training, the model still gives closure on these unseen events comparable to the full model. 2p2h events populate final-state phase-space regions poorly represented by other interaction modes: 49% of 2p2h events lie where the 2p2h-free sample has lower density, and for 30% a reweighting would need weights above 100. The holdout model trained without all 337,617 2p2h events reaches closure AUCs of 0.505 on reconstructed quantities alone and 0.499 with truth features on 2p2h test events, comparable to the full model's 0.497 and 0.501, while maintaining agreement on the non-2p2h control sample.
Perspective
The demonstration is limited to playlist 1A of the MINERvA open simulation dataset in the medium-energy forward-horn-current configuration with water and helium nuclear targets empty; the end-to-end analysis covers only the charged-current inclusive selection and, as a method demonstration, omits full experimental and theoretical systematic uncertainties. The authors propose extending the approach to other playlists, beam modes and systematic variations, and adapting it to other open simulation releases.
The surrogate aims only to reproduce the open simulation dataset, and MINERvA simulation systematics such as detector-model uncertainties are not evaluated here; model uncertainty is estimated from the spread of four independently trained models and covers only differences between independent training runs. In the end-to-end comparison the surrogate's own uncertainty is estimated below one percent per bin, but full experimental and theoretical systematic uncertainties are not yet incorporated. The classifier AUC deviation from 0.5 is mainly driven by the vertex block, and the authors state that under broad fiducial-volume cuts its impact on efficiency and purity is negligible.
