Speaker
Description
A next-generation neutrinoless double beta decay (0νββ) search in ¹³⁶Xe has the potential to uncover lepton number violation, and determine if neutrinos are their own antiparticle. This rare decay, if discovered, would demonstrate Physics beyond the Standard Model and provide key insights into the evolution of the universe. 0νββ searches in ¹³⁶Xe require detector systems capable of reconstructing rare candidate events with high accuracy while supporting sophisticated real time data processing.
The Ba-tagging technique will increase the sensitivity in future searches in a liquid xenon time projection chamber by probing the decay volume for the presence of the ¹³⁶Xe decay daughter, ¹³⁶Ba. Such techniques motivate rapid reconstruction of the event location so that a subsequent Ba-extraction system can be directed toward the candidate decay region. At UC San Diego, our group is developing a low latency machine learning framework capable of identifying candidate 0νββ events using FPGA hardware for real time event reconstruction. The present model is a compact, quantized, dual path convolutional neural network that processes charge readout projections and reconstructs transverse vertex localization of the event from simulated nEXO charge readout data. The architecture was designed with FPGA implementation in mind, using a fully convolutional regression to enable fast triggering and real time decision making, while preserving reconstruction accuracy detector operation.
We evaluated several candidate network architectures for vertex reconstruction accuracy with quantization aware training including FINN, and nn2FPGA, and validated the final implementation on an RFSoC 4x2 platform. Preliminary vertex reconstruction results show that the deployed network maintains millimeter scale vertex reconstruction. Using nn2FPGA, the physical FPGA implementation achieves an accelerator latency of approximately 1.9 ms per event, while preserving agreement with the software and HLS reference models. These results demonstrate the feasibility of performing accurate neural network inference on detector side hardware on millisecond timescales. The design, implementation, and performance of the fast machine learning trigger will be presented, together with planned future developments that will lead us towards a more complete real time trigger for Ba-tagging operations.