Bayesian Deep Learning for Shower Parameter Reconstruction in Water Cherenkov Detectors

Deep Learning methods are among the state-of-art of several computer vision tasks, intelligent control systems, fast and reliable signal processing and inference in big data regimes. It is also a promising tool for scientific analysis, such as gamma/hadron discrimination. We present an approach based on Deep Learning for the regression of shower parameters, namely the its core position and ground energy, using water Cherenkov detectors. We design our method using simulations. We evaluate the limits of such estimation near the borders of the arrays, including when the center is outside the detector's range. We used Bayesian Neural Networks and derived and quantified systematic errors arising from Deep Learning models and in an EfficientNet model design. The method could be easily adapted to estimate other parameters.