Analytical and Machine Learning Methods for Model Discernment at CE$\nu$NS Experiments

TL;DR

This paper explores how correlations in CEνNS data can improve discrimination among BSM neutrino models, using likelihood and neural network methods to analyze shape information beyond total event rates.

Contribution

It demonstrates that multidimensional shape analysis and machine learning can significantly enhance model discrimination and localization in neutrino experiments.

Findings

Likelihood analysis distinguishes BSM scenarios using shape information.

Neural networks retain discrimination power even without total rate.

Shape-based observables enable approximate localization of sterile-neutrino parameters.

Abstract

Neutrino experiments are often limited by low statistics, sizable systematic uncertainties, and coarse observable binning, which can hinder discrimination among competing beyond-the-Standard-Model (BSM) explanations of anomalous signals. In particular, analyses based primarily on total event-rate differences are vulnerable to source-normalization uncertainties and to degeneracies among models that induce similar inclusive yields. Using stopped-pion coherent elastic neutrino-nucleus scattering (CE$\nu$NS) as a benchmark environment, we study how much model-discrimination power can be obtained from correlations in baseline, recoil energy, and timing that are less sensitive to the total rate. As benchmark BSM scenarios, we consider a $3+1$ sterile-neutrino framework and neutral-current non-standard neutrino interactions (NSI). We show with a likelihood-based analysis that these scenarios can be distinguished in nontrivial regions of parameter space once multidimensional shape information is retained. We further demonstrate with convolutional neural networks that substantial discrimination remains possible even after the total event rate is explicitly removed from the input, indicating that the relevant information is genuinely encoded in the shape of the CE$\nu$NS distribution. Finally, through multi-class classification within the sterile parameter space, we show that in favorable regions the same observables can support approximate localization of the underlying sterile-neutrino benchmark point. Our results highlight the complementary roles of conventional and machine-learning-based inference in moving neutrino new-physics searches from anomaly detection to physics interpretation.