Speaker
Description
A machine learning approach for background suppression and noise hit identification in the Baikal-GVD neutrino telescope is presented, addressing signal-to-background ratios of approximately 10^-6 and noise contamination of 85-90% of hits from water luminescence. Two complementary Transformer-based neural network models are developed: a fast background filter that retains 99% of neutrino events while suppressing atmospheric muon background by a factor of 20, and a noise hit classifier achieving ~90% precision and recall on Monte Carlo simulations. To address discrepancies between simulations and experimental data, domain adaptation (DA) using gradient reversal layers is applied. It is shown that DA reduces the Kullback-Leibler divergence between models outputs from 0.267 to 0.049 for background suppression and from 0.014 to 0.008 for noise classification.
