Hardware-efficient neural networks for FPGA-based radio triggering of extensive air showers

Abstract

We present a hardware-efficient hybrid trigger for FPGA-based radio detection of extensive air showers. The hybrid design consists of a lightweight denoiser that cleans raw ADC traces and a compact classifier that operates on the denoised output, enabling robust near-threshold pulse detection in high-interference environments. Both neural networks are trained quantization-aware. Signals are generated from detector-folded CoREAS/CORSIKA simulations and embedded into measured noise to form a realistic benchmark. The trigger reaches an AUC of 0.992 while fitting comfortably within the resource budget of a Zynq-7000 Z-7020, with microsecond-scale latency and sub-watt power consumption. RTL validation confirms agreement between the fixed-point hardware and the quantized software model, demonstrating that neural denoising combined with classification provides reliable, low-cost radio triggering in noisy environments.