Physics-informed neural networks for predicting the asymptotic outcome of fast neutrino flavor conversions

Abstract

In the most extreme astrophysical environments, such as core-collapse supernovae (CCSNe) and neutron star mergers (NSMs), neutrinos can undergo fast flavor conversions (FFCs) on exceedingly short scales. Intensive simulations have demonstrated that FFCs can attain equilibrium states in certain models. In this study, we utilize physics-informed neural networks (PINNs) to predict the asymptotic outcomes of FFCs, by specifically targeting the first two moments of neutrino angular distributions. This makes our approach suitable for state-of-the-art CCSN and NSM simulations. Through effective feature engineering and the incorporation of customized loss functions that penalize discrepancies in the predicted total number of νe and ν¯e, our PINNs demonstrate remarkable accuracies, with an error margin of ≲3%. Our study represents a substantial leap forward in the potential incorporation of FFCs into simulations of CCSNe and NSMs, thereby enhancing our understanding of these extraordinary astrophysical events.