Speaker
Description
Radio detection of extensive air showers induced by ultra-high-energy cosmic rays provides crucial information on their origin, composition and energy. Radio arrays detect these events, but cosmic-ray signals are exceedingly rare compared to the overwhelming radio noise and RFI. Since storing all data is not feasible, a trigger system must decide in real time which data to record. FPGAs are a fitting option for this requirement because they provide deterministic low latency and low energy consumption. In this work, we train two AI models on measured noise traces injected with simulated cosmic-ray pulses. The trained models have been quantized with hls4ml and synthesized with Vitis HLS for multiple FPGAs. Both models achieve a latency $\lesssim 8$ $\mu s$ and can be synthesized on medium to small-sized FPGAs. The performance of the proposed neural-network trigger is compared to a threshold trigger used as a baseline. When applied to experimentally measured noisy traces, the threshold trigger fails to detect signals at a false-positive rate of $10^{-4}$. In contrast, the neural-network classifier achieves a detection efficiency of about 0.8 at the same false-positive rate. When combined with an upstream denoising network, the detection efficiency increases to about 0.9 at a false-positive rate of $10^{-4}$.
These results demonstrate that neural-network-based triggers can substantially improve detection performance for radio air-shower signals and represent a viable approach for future FPGA-based station-level triggers. We aim to deploy the models on an FPGA and validate the AI trigger with real antenna data, supported by a data-emulation framework currently under development, enabling end-to-end hardware validation.