Enhancing Event Reconstruction in Hyper-Kamiokande with Machine Learning: A ResNet Implementation
The forthcoming Hyper-Kamiokande experiment is expected to achieve unprecedented sensitivity to neutrino oscillations. Meeting these precision goals requires substantially larger Monte Carlo datasets than previous experiments to satisfy stringent systematic-uncertainty requirements. While traditional maximum-likelihood reconstruction provides high-quality event reconstruction, its per-event computational cost makes reconstruction of large, systematically varied Monte Carlo samples increasingly impractical. We demonstrate a neural-network-based reconstruction approach for the Hyper-Kamiokande far detector using simulated data with the experiment's detector geometry. Single-particle events are generated with kinetic energies from the Cherenkov threshold up to 2 GeV above threshold. The resulting PMT charge and timing information are mapped to $274\times144$ five-channel images that serve as inputs to ResNet models in the WatChMaL framework. These models (i) classify events into four particle hypotheses ($e$, $μ$, $γ$, $π^{0}$) and (ii) regress the interaction vertex, direction, and momentum of electrons and muons. The regression models achieve momentum resolutions of $1.27\%$ for muons and $2.21\%$ for electrons, angular resolutions of $0.91^\circ$ and $1.60^\circ$, and vertex resolutions of $13.6\,\mathrm{cm}$ and $17.0\,\mathrm{cm}$, better than those reported for traditional methods. The classifier improves $e$-$μ$, $e$-$γ$, and $e$-$π^{0}$ separation relative to the traditional baseline, with ROC AUC values of $0.9999995$, $0.642$, and $0.9659$. Our networks achieve inference times of 1-2 ms per event on a single GPU, yielding speed-ups of $3.2\times10^{4}$-$5.2\times10^{4}$ relative to likelihood-based reconstruction.