Event Generator Tuning as a Robustness Test

TL;DR

This study investigates how tuning GENIE's interaction-model configurations to external cross-section data affects predictions across multiple neutrino experiments. By applying both full-covariance and diagonal chi-square tuning to G18 and AR23 using five cross-section datasets, it tests robustness and universality of the tunings. Key findings show AR23 better describes T2K 2020 and MicroBooNE but cannot fit MINERvA, while G18 struggles similarly; MicroBooNE's tuning results depend heavily on the chosen dataset, underscoring the risk of tuning for SBL searches. The authors recommend SBL analyses avoid external tuning and instead rely on a base model with sufficient uncertainty, constrained by SBND data in joint fits.

Abstract

Neutrino oscillation experiments use Monte Carlo event generators to predict neutrino-nucleus interactions. Cross section uncertainties are typically implemented by varying the parameters of the model(s) used in the generator. We study the performance of two commonly-used model configurations of the GENIE generator (G18_10a_02_11a and AR23_0i_00_000) and their uncertainties by tuning parameters to cross section data, and then comparing the resulting tuned prediction to a suite of other measurements from T2K, MicroBooNE, and MINERvA. This reveals whether the model can simultaneously describe several datasets, as well as whether the uncertainties are adequately robust. We find that G18 and especially AR23 are reasonable in predicting lower-energy measurements from T2K and MicroBooNE, but unable to describe MINERvA data, and discuss the implications for short-baseline oscillation searches. We attempt to replicate a tuning procedure developed by MicroBooNE using several different measurements, and find substantially different results depending on which measurement is used, and that the MicroBooNE tune does not agree with other measurements. We conclude that the SBN experiment should not tune its generator to external data.

Figures (11)

• Figure 1: The sensitivity ranges of the accelerator-based neutrino experiments from Table \ref{['tab:experiments']} to $L/E_\nu$ (left) and $\Delta m^2$ (right). The LBL far detectors from Table \ref{['tab:experiments']} are excluded, as they all have sensitivity in the few $\times 10^{-3}$ eV$^2$ region.
• Figure 2: MicroBooNE CC1$\mu$1p single-differential cross section data in muon momentum (top left), muon angle (top right), proton momentum (bottom left), and proton angle (bottom right) plotted with the nominal predictions of the data made by the G18 and AR23 model configurations.
• Figure 3: The T2K 2016 data is compared to predictions from the nominal G18 model configuration, as well as five different tunes obtained by fitting G18 to the data given in the legend using the "diagonal" method.
• Figure 4: MiniBooNE data in the $-0.1<\cos\theta_\mu<0.0$ bin (top) and $0.9<\cos\theta_\mu<1.0$ bin (bottom) is compared to predictions from the nominal G18 and AR23 models and their tunes to the data in the legend using the method in the plot title
• Figure 5: The $\chi^2_{diag}$ from Eqn. \ref{['eq:chi2diag']} obtained by comparing the G18 model configuration to five different measurements, indicated by each column. The top row is the nominal prediction, subsequent rows are the best-fit tune to the measurement indicated by the row, using the diagonal method. The cell color is a $p$-value assuming $\chi^2_{diag}$ follows a $\chi^2$ distribution.