Investigation of BGO Coincidence Time Resolution with Deep Learning
Francis Loignon-Houle, N. Kratochwil, Maxime Toussaint, C. Lowis, Gerard Ariño‐Estrada, Antonio J. González, E. Auffray, Roger Lecomte · 2024
Significant advancements in coincidence time resolution (CTR) of BGO scintillators have been made in recent years, driven by the enhanced capability of new SiPMs to detect Cherenkov photons. However, BGO also emits slower scintillation light, causing time walk and degrading CTR when measured with leading edge discrimination (LED). This can be partially counteracted by using a second, higher threshold to perform time walk correction (TWC). Convolutional neural networks (CNNs) were also shown to provide enhanced timing estimation in other scintillators by training with digitized waveforms. The present study compares the CTR performance of LED, TWC, and CNN approaches in BGO scintillators read out by NUV-HD-MT SiPMs and high-frequency electronics. For BGO 2×2×3 mm3utilizing TWC produces a CTR of 129 ± 2 ps FWHM, whereas the CNN achieves 115 ± 2 ps, representing 18% and 26% improvements over LED, respectively. For BGO 2×2×20 mm3, both approaches produce comparable CTR (around 240 ps FWHM, ~15% improvement over LED), but the CNN shows superior tail suppression in the coincidence time distribution. The increased complexity associated with waveform digitization required for CNNs might be alleviated by adopting a simpler dual-threshold approach, which so far appears to recover the most crucial features of the signal for enhancing CTR in longer BGO crystals.