Unfolding using deep learning and its application on pulse height analysis and pile-up management

Abstract

Traditionally, electronics for pulse processing can be modeled as linear transfer functions. In contrast, due to the fact that artificial Neural Networks (NNs) are generally non-linear systems, their behavior against noise is significantly different as in linear systems. We take advantage of this non-linearity to achieve acceptable Signal-to-Noise Ratios (SNR) with a extremely short shaping time. This article shows an approach to a concrete NN named U-net as pulse shaper. It filters the pulses and return them unfolded solving the pile-up problem, and even estimates the height of the pulses when there has been saturation in the detector. In this article, the NN architecture and results using simulated pulses and real pulses from scintillators are shown. The results clearly show the effectiveness of the approach.