Muon–Electron Pulse Shape Discrimination for Water Cherenkov Detectors Based on FPGA/SoC

Abstract

The distinction of secondary particles in extensive air showers, specifically muons and electrons, is one of the requirements to perform a good measurement of the composition of primary cosmic rays. We describe two methods for pulse shape detection and discrimination of muons and electrons implemented on FPGA. One uses an artificial neural network (ANN) algorithm; the other exploits a correlation approach based on finite impulse response (FIR) filters. The novel hls4ml package is used to build the ANN inference model. Both methods were implemented and tested on Xilinx FPGA System on Chip (SoC) devices: ZU9EG Zynq UltraScale+ and ZC7Z020 Zynq. The data set used for the analysis was captured with a data acquisition system on an experimental site based on a water Cherenkov detector. A comparison of the accuracy of the detection, resources utilization and power consumption of both methods is presented. The results show an overall accuracy on particle discrimination of 96.62% for the ANN and 92.50% for the FIR-based correlation, with execution times of 848 ns and 752 ns, respectively.

A Simplified Correlation Index for Fast Real-Time Pulse Shape Recognition

“…An important compression ratio was achieved by quantizing the data to a 14-bit fixed-point representation (as done by [ 45 , 46 ]), rather than using the double-precision floating-point numeric resolution of the original Python simulation. Such optimization methods have been proven to reduce the required hardware resources in PSD and machine learning applications without significantly affecting the accuracy [ 47 , 48 , 49 ].…”

Pulse Shape Discrimination for Online Data Acquisition in Water Cherenkov Detectors Based on FPGA/SoC

“…Neural networks have also been used due to their potential as classifiers because of the use of experimental data for training [6]. We presented a PSD study with Artificial Neural Newtorks (ANN) and FIR-Based Correlation Digital Signal Processing (DSP) [7] capable of discriminating between two different type of particles. Thus, following this study, an upgraded version from the acquisition system is presented heading towards higher speed data sampling and processing capabilities.…”

“…Such classification was done with a centroid-based clustering algorithm. Due to the fact that measurement of similarity is based on shape, the computation of the distance was performed using Pearson correlation function, instead of Euclidean distance (such as in traditional K-means algorithm) [8] allowing the system to distinguish among pulse waveforms [7]. A set of ∼3 million triggered pulses have been acquired and classified using this method, a subset of raw pulses within each cluster of interest is shown in Fig.…”