Machine Learning Neutrino-Nucleus Cross Sections

TL;DR

The paper tests a data-driven approach to neutrino-nucleus cross sections by learning inclusive structure functions from near-detector data in a toy DUNE setup. A neural network models the structure functions $W_i$ and, when used in a far-detector oscillation analysis, yields parameter constraints close to the ideal case of known cross sections. Systematic uncertainties—finite ND statistics, detector smearing, flux shape, and ML initialization—are quantified and incorporated, showing a conservative, calibrated uncertainty framework. The results suggest that data-driven cross-section modeling can complement traditional generator-based methods and enable robust oscillation measurements with near-term near-detector datasets. The work highlights future avenues, including semi-inclusive channels and multi-experiment synergy, to maximize the discovery potential of upcoming neutrino experiments.

Abstract

Neutrino-nucleus scattering cross sections are critical theoretical inputs for long-baseline neutrino oscillation experiments. However, robust modeling of these cross sections remains challenging. For a simple but physically motivated toy model of the DUNE experiment, we demonstrate that an accurate neural-network model of the cross section -- leveraging only Standard-Model symmetries -- can be learned from near-detector data. We perform a neutrino oscillation analysis with simulated far-detector events, finding that oscillation analysis results enabled by our data-driven cross-section model approach the theoretical limit achievable with perfect prior knowledge of the cross section. We further quantify the effects of flux shape and detector resolution uncertainties as well as systematics from cross-section mismodeling. This proof-of-principle study highlights the potential of future neutrino near-detector datasets and data-driven cross-section models.

Figures (19)

• Figure 1: Illustration of our analysis procedure as applied in the context of our closure test. In a real application, the mock ND and FD data would be replaced by experimental data.
• Figure 2: Event distributions at the ND (left) and FD (right) as predicted using either the true (top) or learned model (middle) cross sections and the true ND and FD fluxes. Differences between true and learned cross sections rescaled for visibility are also shown (bottom).
• Figure 3: One-dimensional marginal event densities at the ND (top) and FD (bottom) as predicted using either the true (dashed black) or model (blue lines) cross sections and the true ND and FD fluxes. Computed from discretized integrals on a $256^3$ grids, specifically over $(E_\nu, E_{\ell}, \cos\theta)$ for $E_{\ell}$ and $\cos\theta$, and over $(E_\nu, v_1, v_2)$ for $v_1$ and $v_2$. The marginals for $E_\nu$ can be computed equivalently using either grid.
• Figure 4: Comparisons of true and model cross sections along slices of fixed $E_\nu$ interpolating the full range of $0 \leq E_\nu \leq 10~\mathrm{GeV}$. Each cross section is first normalized as described in the text to remove an overall scale, then within each row the maximum value over either true or model is divided out of both. Not visible given this normalization convention is that $\frac{d^2\sigma}{dv_1 dv_2}$ increases as a function of $E_\nu$, as shown in Fig. \ref{['fig:int-xsec-comp']}.
• Figure 5: Comparison of true and model normalized (total) cross sections.