Natural Language Processor based real-time cyclotron radiation detection in Project 8

Project 8 aims to determine the absolute neutrino mass scale by analyzing the endpoint of tritium β^--decay spectrum using Cyclotron Radiation Emission Spectroscopy (CRES). The Cavity CRES Apparatus (CCA), a key demonstrator for Project 8, employs a resonant cavity for the cyclotron electrons. Within the cavity, β^--decay electrons undergo cyclotron motion in a magnetic trap and emit characteristic CRES signals. Efficient identification of these CRES signals in the data stream is challenging due to low signal-to-noise ratios (SNR) and complex signal structures. Because endpoint electrons are both rare and short-lived, and considering storage limitations, a real-time trigger is essential to record data only when a CRES signal is present.

I will present a novel machine learning trigger system based on stacked Long Short-Term Memory (LSTM) networks. After training on simulated CCA data spanning a wide range of physical parameters, the LSTM-based model effectively and efficiently discerns CRES signal patterns embedded in noisy backgrounds. I will show that the optimized machine learning model significantly outperforms the existing power-based trigger, and rivals a matched filter approach.