This paper uses a fast simulation method to ask whether two Fermilab experiments, MicroBooNE and MiniBooNE, can be described by the same simple extension of the Standard Model that adds one “sterile” neutrino. The authors find that MiniBooNE data strongly prefer the 3+1 sterile‑neutrino hypothesis (3.6σ), MicroBooNE shows a weaker preference (1.8σ), and the tension between the two results is significant (3.3σ). When the authors apply a correction for a normalization difference in MicroBooNE’s muon‑neutrino samples, the tension falls to 2.2σ, which lowers but does not eliminate disagreement.

The 3+1 model adds a single sterile neutrino that does not interact in the usual way but can change identity (oscillate) with the three known neutrinos. Experiments look for three kinds of effects: muon‑to‑electron appearance, electron disappearance, and muon disappearance. Past global analyses have shown that appearance and disappearance datasets pull the model parameters in very different directions, creating internal conflict even when the overall fit looks much better than the Standard Model.

The technical advance in this work is the use of Simulation‑Based Inference (SBI). SBI is a machine‑learning approach that builds fast, simulation‑driven surrogates for statistical tests. It lets the authors run many realistic fake experiments, including the experiments’ published uncertainty information, to map out how likely different data and parameter combinations are. This avoids relying on mathematical approximations that can fail in these cases, and it is much faster than the traditional trial‑based Feldman–Cousins method.