Water Cherenkov detectors have been widely adopted as a low-cost technique for cosmic rays (CR) studies. Thus, an existing CR readout system has been chosen as the base DAQ (data acquisition) design, which has been paired to a Neural Network (NN) in order to work as a trace/event discrimination block. We present the compression of two NN architectures for particle classification, targeting a low-end System-on-Chip (SoC). The hls4ml package is used to translate the final NN models into a high-level synthesis project. Both NNs were implemented and tested on Xilinx SoC ZC7Z020 Zynq. A comparison of the accuracy of the detection, resource utilization and latency of the two NNs are presented. The results show the benefits of using compression techniques to deploy a reduced model, which provides a good compromise between efficiency, effectiveness, latency, as well as resource utilization.
