Kinematic reconstruction and particle identification in water-Cherenkov detectors are typically performed with maximum-likelihood methods such as fiTQun. We assess deep neural networks as an alternative for e−/mu− separation and kinematic reconstruction. Using particle-gun simulations of electrons and muons generated with WCSim v1.12.19 for the IWCD detector, we trained on fully contained events and validated on independent samples. Among the tested models, ResNet-152 performed best for kinematic reconstruction, while a lighter ResNet-50 was sufficient for e−/mu− classification. We compared resolutions and biases against fiTQun.  For e−events, ResNet-152 improves upon several key quantities, 3D position resolution, momentum resolution, and momentum bias, relative to nominal fiTQun values. For mu− events, fiTQun is slightly better in 3D position and direction reconstruction, whereas ResNet-152 yields a significantly better momentum resolution. Overall, ResNet-based reconstruction substantially reduces momentum uncertainties—especially for mu−events—while remaining competitive in angle and position; fiTQun retains a small advantage in mu− spatial localization.

gilbertordzp@gmail.com

Gilberto Rodriguez Prado