Benchmarking Supernova Pointing with Machine Learning

Core-collapse supernovae are some of the most energetic events in the universe, but many aspects of the explosion mechanism and the associated neutrino emission remain poorly understood. Most of a supernova’s energy is carried away by neutrinos, making their detection essential for probing the physics at the heart of the collapse and for constraining neutrino properties themselves. The next-generation Hyper-Kamiokande (HK) experiment in Japan will be the world’s largest water Cherenkov detector and is expected to collect orders of magnitude more neutrinos than the current Super-Kamiokande (SK) detector during the next nearby core collapse supernova. In anticipation of this event, we simulate supernova events in HK and apply machine-learning algorithms to reconstruct the supernova direction and differentiate between different theoretical emission models. In this report, we compare three architectures—-convolutional neural networks (CNNs), graph neural networks (GNNs), and sparse CNNs—-and evaluate their ability to infer the supernova direction from simulated photomultiplier tube hit patterns. Our results show that the sparse CNN significantly outperforms the other architectures: for a supernova at 8 kpc, it reconstructs the direction to within 2.4° on average assuming no contamination from inverse beta decay events.