Speaker
Description
Project 8 measures the neutrino mass using Cyclotron Radiation Emission Spectroscopy (CRES), which detects the cyclotron radiation of single tritium-decay electrons, digitized as a continuous voltage time series. The standard analysis chains independently tuned stages (short-time Fourier transform, background removal, track finding, event building, classification, and conversion to energy) with hand-tuned power and frequency cuts. We present an alternative ML approach using a State-Space Model (SSM) that works directly on the I/Q time series. We present the performance of our best model (S4D architecture, ~300k parameters), which has two jointly trained denoiser and regressor stages and outputs electron observables with a calibrated per-event uncertainty. In high-fidelity simulation of the Cavity CRES Apparatus currently being commissioned at UW, the model reaches 0.168 eV energy resolution, compared with 0.296 eV for the standard algorithm at the same 17.1% efficiency. At the same 0.3 eV target resolution it recovers about 50% more events, with a roughly 1.9 times lower input signal-to-noise ratio. Offline reconstruction of 10,000 events with the SSM needs about 2.5 minutes on one GPU versus O(1) hour on a CPU for the standard algorithm. The same model generalizes across the DAQ chain: As a trigger, it reaches 92.4% signal efficiency at 99.9% background rejection, even though it was never trained on noise-only data. Task-specific heads trained on top of the foundation model allow reconstruction of electron signals affected by scattering with the gas molecules in our apparatus, and improves the detection of early scattering in the signal compared to the standard algorithm, even though the foundation model was trained on scatter-free events. For real-time triggering, the SSM algorithm can reach the 20 µs latency target with O(10) A100 GPUs. We discuss paths to online operation, from combining it with the standard trigger to dedicated hardware.