Author Description

In this dissertation, I will present a reconstruction method based on multi-task deep neural networks which can perform reconstruction of both vertex and incoming neutrino direction with reasonable precision. It can achieve a comparable performance compared to a classic interferometric technique in vertex direction reconstruction, but can also reconstruct vertex distance and neutrino direction that are not achievable with an interferometric technique. After training, this solution is capable of rapid reconstructions (e.g. 0.1~ms/event compared to 10000~ms/event in a conventional interferometric routine) useful for trigger and filter decisions and can be easily generalized to different station configurations for both design and analysis purposes. The model has also be tested on the 2018 deep pulser data set and 2018 SpiceCore data set for its applicability to experimental data.