Differentiable nuclear deexcitation simulation for low energy neutrino physics (2404.00180v1)

Abstract: Neutrino-nucleus interactions play an important role in present and future neutrino experiments. The accurate simulation of these interactions at low energies ($<$100 MeV) is crucial for the detection and study of supernova, solar and atmospheric neutrinos. In particular, the reconstruction of the incoming neutrino properties depends on the ability to measure the products from the deexcitation of the final state nucleus after the initial neutrino-nucleus scattering reaction. A realistic nuclear deexcitation model that can correctly manage the theoretical uncertainties in the process is key to determine the response of a detector to low energy neutrinos and estimate the overall systematic uncertainties that affect it. Automatic differentiation frameworks like PyTorch or JAX, widely used in the field of Machine Learning, provide us the tools to compute exact gradients of the simulation outputs with respect to the model parameters. Such differentiable simulators can be applied in simulator tuning to match observed data and forward modeling to efficiently infer the impact of parameter distributions to the physics output, or for parameter inference acting as a likelihood estimator. This paradigm can pave the way to a fully differentiable analysis of the whole simulation chain or direct integration with preexisting Machine Learning tools. As a proof of concept, in this work we implement a fully differentiable nuclear deexcitation simulation based on a simplified version of the Hauser-Feshbach statistical emission model.