Learning Astrophysical Uncertainties in Dark Matter Direct Detection with Simulation-Based Inference

Abstract

We present a proof-of-concept study of neural ratio estimation (NRE) - a simulation-based inference algorithm - applied to WIMP direct detection using the XENONnT experiment as an example. A binary classifier is trained on simulated parameter-observation pairs to approximate the likelihood ratio for the WIMP mass $m_χ$ and spin-independent proton coupling $c_p$, using realistic scintillation and ionization signals produced by the fuse detector simulation. We demonstrate that astrophysical uncertainties from halo model variations are marginalized over implicitly and at no additional inference cost by including multiple halo models in the training data. Using extreme $\pm5σ$ Standard Halo Model variations, we show that training on a combined dataset recovers well-calibrated posteriors even under severe misspecification, while more realistic halo models produce nearly indistinguishable results. The signal-plus-background classifier yields an exclusion limit that is both qualitatively and (after accounting for differences between frequentist and Bayesian approaches) quantitatively consistent with the XENONnT expected sensitivity. These results establish NRE as a scalable complement to the profile likelihood ratio approach, with particular advantages as experiments probe deeper into the neutrino fog.