Enhancing Event Reconstruction in Hyper-Kamiokande with Machine Learning: A ResNet Implementation

TL;DR

This paper presents a ResNet-based neural network approach for event reconstruction in Hyper-Kamiokande, achieving comparable accuracy to traditional methods but with vastly improved computational efficiency, enabling large-scale data processing.

Contribution

It introduces a deep learning framework that classifies particle types and regresses event parameters with high accuracy and significantly faster inference times.

Findings

Achieves momentum resolutions of 1.35% for muons and 2.39% for electrons.

Attains angular resolutions of 1.25° for muons and 1.94° for electrons.

Provides inference speed-ups of over 30,000 times compared to traditional likelihood methods.

Abstract

The forthcoming Hyper-Kamiokande experiment requires substantially larger Monte Carlo datasets than previous experiments to satisfy stringent systematic-uncertainty requirements. While traditional maximum-likelihood reconstruction provides high-quality results, its per-event computational cost makes processing these large samples increasingly impractical. We demonstrate a neural-network-based reconstruction approach for the Hyper-Kamiokande far detector using simulated data. Single-particle events with kinetic energies from the Cherenkov threshold up to 2 GeV are propagated through the detector, with PMT charge and timing information mapped to $190\times189$ two-channel images serving as inputs to ResNet models in the WatChMaL framework. These models (i) classify events into four particle hypotheses ($e$, $\mu$, $\gamma$, $\pi^{0}$) and (ii) regress the vertex, direction, and momentum of electrons and muons. Averaged over the full kinematic range, the regression models achieve momentum resolutions of $1.35\%$ and $2.39\%$, angular resolutions of $1.25^\circ$ and $1.94^\circ$, and vertex resolutions of $28.2$ cm and $25.4$ cm, for muons and electrons respectively, broadly consistent with traditional methods. The classifier improves $e$-$\mu$, $e$-$\gamma$, and $e$-$\pi^{0}$ separation, with ROC curve areas of $0.9999992$, $0.633$, and $0.9526$. Crucially, 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, highlighting deep learning as a scalable alternative for Hyper-Kamiokande event reconstruction.