Machine Learning-Informed 3+1 Sterile Neutrino Global Fits using Posterior Density Estimation of Electron Disappearance Data

TL;DR

This work addresses the challenge of globally fitting a 3+1 sterile-neutrino model to electron-flavor disappearance data by leveraging simulation-based inference to estimate posteriors without resorting to costly MCMC. It systematically compares several posterior-estimation strategies—most notably flow matching (FMPE), SNPE-C, NPSE, and DNRE—using a large training set of pseudoexperiments across reactor, source, and accelerator datasets. The FMPE approach yields well-calibrated, high-fidelity posterior surfaces that align with traditional Bayesian and frequentist results while offering substantial speedups, and the authors show how these posteriors can augment or replace components of trials-based frequentist fits via fast MAP estimation. They also demonstrate that SNPE-C provides the fastest MAP computations, suggesting a practical path to upgrading existing frequentist pipelines. Overall, the paper argues for flow-based posterior estimation as a robust, scalable complement to established methods, enabling end-to-end SBI 3+1 global fits and guiding inference in complex beyond-Standard-Model scenarios.

Abstract

Global analyses of particle physics data are integral for validating and scrutinizing published results of experiments. Global fits of anomalous oscillation data which search for one or more eV-scale sterile neutrinos are particularly challenging both to evaluate and to reconcile in the global picture. Fits (especially joint ones) to oscillation data suffer from significant computational burdens, such as likelihood intractability, making traditional Markov Chain-Monte Carlo all but impossible. Given evidence both supporting and challenging beyond Standard Model physics across neutrino experiments of various baselines, energies, and detection techniques, the global search for sterile neutrinos requires additional tools in order to determine whether sterile neutrinos remain a viable solution to unexplained anomalies. Furthermore, both a Bayesian and frequentist interpretation of sterile neutrino data is needed for a complete assessment of longstanding tensions in the field. Techniques from the machine learning subfield of simulation-based inference have a natural application to such a problem. In this contribution, we illustrate some of the outstanding questions of the global picture of light sterile neutrinos by focusing on experiments searching with the disappearance of electron (anti)neutrinos, and look to posterior density estimation strategies to craft answers, including comparisons to a machine-learning-based frequentist approach.

Figures (10)

• Figure 1: Comparison of the posterior density estimation methods based on flows and the neural likelihood ratio estimator from Ref. 10.1088/2632-2153/ae040c using (a) log-score distributions, (b) MAP error statistics, and (c) rank score distributions.
• Figure 2: Using simulated data, posterior density estimation performed by FMPE on out-of-sample null-like (left) and signal-like (right) pseudoexperimental throws. The fit to the signal sample includes an inset demonstrating excellent agreement between the sharply peaked posterior estimate and the true injected sterile oscillation parameters. These fits are to all of the electron neutrino disappearance experiments presented in Tab. \ref{['tab:experiments']}.
• Figure 3: Posterior density estimated with FMPE for the published $\nu_e/\overline{\nu}_e$ disappearance data in Tab. \ref{['tab:experiments']}. White contours show highest-posterior-density credibility regions. The MAP estimate lies at $U_{e4}=0.28$ ($\sin^2 2\theta_{ee}=0.29$), $\Delta m_{41}^2 = 48.69\,\text{eV}^2$. Edge effects cause the normalization seen in the source experiments to taper at large $\Delta m_{41}^2$ (see discussion). This posterior approximates the MCMC result in Fig. \ref{['fig:true-fit-sblmc']} and is qualitatively consistent with the frequentist allowed region in Fig. \ref{['fig:freq-fit']}.
• Figure 4: Fits to published data from each of the three experiment classes (reactors, including STEREO, PROSPECT, DANSS, and NEOS; sources, including SAGE/GALLEX and BEST; accelerators, including the Karmen-LSND cross section analysis) using FMPE. These agree qualitatively with the frequentist allowed regions in Fig. \ref{['fig:diff-exps-freq']}.
• Figure 5: Acceptance regions for each experiment class within the frequentist framework, computed using the DNRE posterior evaluation.