AI-Enhanced Self-Triggering for Extensive Air Showers: Performance and FPGA Feasibility

TL;DR

This work tackles autonomous self-triggering for radio-based extensive air shower detection in environments with strong RF interference by developing a deep‑learning 1‑D convolutional trigger trained on real noise and MGMR3D pulses. The approach is demonstrated end‑to‑end: GPU training with robust data splits, fixed‑point quantisation via HLS4ML, and FPGA synthesis with Vitis HLS across multiple platforms, achieving sub‑microsecond inference and preserving near‑perfect physics performance (AUC ≈ $0.997$ float, $0.996$ quantised). Compared to a baseline SNR threshold, the AI trigger delivers markedly higher efficiency at a fixed false‑positive rate (≈ $68\%$ vs. ≈ $17\%$ at FPR $=10^{-4}$), enabling radio‑only triggering that can access weaker and highly inclined showers. The results imply a substantial physics impact: lowering effective trigger thresholds broadens energy reach, expands sky coverage, and paves the way for ultra‑high‑energy neutrino detection with ground‑based radio arrays, all while maintaining practical hardware latency and power budgets. This work establishes a complete AI-to-FPGA workflow for real‑time cosmic‑ray radio detection and outlines clear paths for future hardware and data‑set enhancements.

Abstract

Autonomous self-triggering for radio detection of extensive air showers remains a long-standing challenge, particularly in environments dominated by strong and variable radio-frequency interference. Current radio arrays usually rely on external particle-detector triggers: while this lowers thresholds for vertical showers, it excludes very inclined events, where radio detection is uniquely powerful because the particle cascade is absorbed while the radio pulse remains measurable. In this work, I present a proof-of-principle study showing that deep-learning-based triggering can overcome these limitations, operating robustly under realistic high-interference conditions and within the strict latency constraints of large-scale observatories. Using measured noise traces combined with simulated cosmic-ray pulses, a fully convolutional network was trained and optimised for MHz-scale trigger-trial rates, and its performance was compared to a simple threshold-based trigger, which proved markedly less efficient at the same false-positive rate. The trained model was then quantised with HLS4ML and synthesised with Vitis HLS for multiple FPGA targets. The quantised implementation preserves the floating-point model performance (AUC_float = 0.997, AUC_quant = 0.996) while achieving microsecond-scale inference latencies.These results show that modern FPGA-based AI inference can provide low-latency, radio-only triggering for next-generation cosmic-ray observatories, particularly in regions with strong and variable radio backgrounds. This capability unlocks access to weak and highly inclined air-shower signals--even in noisy environments--thereby broadening the energy range, extending sky coverage, and opening the path toward ultra-high-energy neutrino detection with ground-based radio arrays.

Figures (7)

• Figure 1: Power spectrum of the recorded signal after applying a bandpass filter in the range of 30–80. Two prominent narrowband RFI sources are observed at approximately 45 MHz and 71 MHz.
• Figure 2: SNR distributions for background and signal traces in Validation Dataset 1 (left) and Validation Dataset 2 (right). Validation Dataset 1 follows the distribution of the training dataset, while Validation Dataset 2 contains low-amplitude signals.
• Figure 3: ROC curve for the classifier evaluated on signal from Validation 1 and background from Validation 3. The solid line shows the model response; the area under the curve (AUC) is 0.969. The vertical red dashed line indicates the chosen operating point at a false positive rate of $10^{-4}$.
• Figure 4: Representative trace with a faint signal pulse correctly classified by the model despite transient noise, demonstrating its ability to recover weak air shower signals undetectable by threshold-based methods.
• Figure 5: ROC curve for the SNR-threshold baseline (Hilbert-envelope peak squared divided by noise variance). The area under the curve is $0.529$. The vertical red dashed line indicates the operating point at a false positive rate of $10^{-4}$; the corresponding signal efficiency is $\approx 16.8\%$.