Development of charge reset multiplexing for SiPMs using deep learning architecture

초록

Objective. To present a charge-reset multiplexing method for silicon photomultipliers (SiPMs) that reduces 16 readout channels to a single line while preserving per-channel waveform information. Approach. Each of the 16 channels is equipped with a charge-reset preamplifier that encodes channel identity into a unique pulse width; these pulses are then summed into a single output. An autoencoder, a deep learning model, is trained to reconstruct the 16 individual signals from this single summed trace by minimizing the error between the original and reconstructed outputs. Performance was evaluated using a 4 x 4 Ce:GAGG scintillator array coupled to a 4 x 4 SiPM array with 137Cs and 22Na sources. Main results. The reconstructed signals achieved energy resolutions of 10.84% (137Cs) and 16.12% (22Na), comparable to a 1:1 non-multiplexed baseline measured under identical conditions. Recovering per-channel signals further enabled inter-crystal scatter (ICS) identification and removal, yielding clear flood maps. Significance. As SiPM-based systems scale, channel count drives cost and system complexity. Charge-reset multiplexing offers a practical path to drastic channel reduction without significant loss of performance, while retaining per-channel information for advanced processing (e.g. ICS rejection). The approach is amenable to integrated implementations and broadly applicable to high-density detector readout in nuclear medicine and related fields.

키워드