Three topologies of deep neural networks for pulse height extraction

Pulse shaping is a common technique for optimizing signal-to-noise ratio (SNR) in particle detectors. Although analog or digital linear shapers are typically used for this purpose, there are nonlinear approaches, such as neural networks (NN), which have demonstrated their potential to outperform linear ones. Their nonlinear nature makes it possible to optimize the SNR of incoming pulses and extract diverse information, such as particle type and energy, with extremely short shaping time to avoid crowding. This paper shows three different NNs for shaping pulses: (a) convolutional NN (CNN); (b) recurrent NN (RNN); (c) self-attenuating NN. These NNs shape the pulses and return them unfolded avoiding stacking, and even estimate the height of the pulses when there has been saturation in the preamplifier. In this work we show the architectures of the NNs and their results using CR-RC pulses with Brownian and white noise, but they could be extrapolated to any shape and type of noise. The results obtained show that when the noise level is low and the frequency is low, all the topologies presented are a valid solution, but with white noise and high pulse arrival frequency, CNN is a better solution than the others. In the case of Brownian noise, the three topologies presented give similar results.