The proposed trigger accepted 97% of benchmark signal events in an independent validation test, with a frame-equivalent candidate rate of 3.39 kHz. The rate is a per-second estimate based on accepted fixed-length data frames; 95% confidence intervals were 0.96–0.98 for efficiency and 2.68–4.23 kHz for the rate.
But this is a constructed benchmark, not a live station-wide performance figure. It used simulated, detector-folded air-shower pulses injected into measured backgrounds from one station in the Auger Engineering Radio Array, and it was not a full end-to-end deployment.
A trigger shaped around the background
At its core is one 16-sample template, a short signal pattern learned from detector-folded cosmic-ray pulses. The template is whitened using measured background covariance, an adjustment for noise correlations, and applied to both high-gain polarization channels. Their responses are combined when they agree within a radius of eight samples before a fixed threshold is applied.
The underlying traces contained two channels and 2,016 samples per event. Measured backgrounds were combined with pulse traces at random allowed times; the pulse library came from 844 high-quality AERA events, with one proton and one iron simulation for each event.
Three disjoint sets held 7,000 training, 2,000 validation set 1 and 2,000 validation set 2 events per class. Training learned the template and covariance, validation set 1 searched architectures and calibrated thresholds, and validation set 2 supplied the independent evaluation.
Validation set 1 set the operating point at a 3.57 kHz background-candidate budget, defined as 5% occupancy of a downstream module with 71.4 kHz service capacity, before the frozen trigger was applied to validation set 2.
Strong results within a narrow test
In the architecture search, candidates were compared at 1% background-frame acceptance. The selected one-template sum reached 0.96 efficiency and was the smallest candidate to meet the 0.90 requirement; the corresponding coincidence rule reached 0.94.
The comparison with simpler alternatives was stark: at their calibration-set operating points, the raw dual-polarization template sum reached 0.41 efficiency, the absolute-amplitude trigger 0.16, and the whitened coincidence rule 0.96. The authors attribute the difference to pulse morphology after whitening plus local timing agreement, but present that as a benchmark interpretation rather than causal evidence.
Across eight equal-population signal-strength bins of 250 validation events each, measured efficiency ranged from 0.94 to 1.00. Those bins describe the sampled benchmark range and do not establish performance on independently observed cosmic-ray events.
Continuous replay raises the harder question
In two later recording periods, without period-specific signal rescaling or threshold adjustment, efficiency was 0.92, with a 95% confidence interval of 0.91–0.93, and 0.91, with an interval of 0.90–0.93. Each period accepted 15 of 2,000 background frames, corresponding to 0.67 kHz with a 95% interval of 0.38–1.10 kHz.
A separate continuous replay of 14 chronological native background records produced 1,057 above-threshold score clusters, with contiguous runs counted together. The mean rate was 6.64 kHz, with a 95% interval of 5.19–8.08 kHz, while individual records ranged from 2.37 to 11.87 kHz. The mean stayed below the downstream capacity of 71.4 kHz but exceeded the 3.57 kHz design target in most records.
The two headline rates are not interchangeable: 3.39 kHz came from independent frame acceptance, while 6.64 kHz came from counting above-threshold clusters in continuous data. The replay remained below downstream service capacity, but its average did not meet the design target.
The hardware passed a basic check
The design targeted 200 MHz, sustained one new dual-polarization sample per clock and reported 0.100 ns of setup slack and 0.018 ns of hold slack after routing. It used 2,377 lookup tables, 5,330 registers, 32 digital signal-processing blocks and no block RAM.
Power was a vectorless estimate of 0.310 W, not a board measurement. In a separate hardware-level validation, all 400 tested decisions matched the reference bit-for-bit after a one-sample channel skew in the testbench was corrected.
Deployment is still untested
The central limitation is the evidence base: results from one station, simulated pulses injected into measured backgrounds and technical replay do not establish final accidental-trigger rates or efficiency across stations and operating conditions. They also do not validate live buffering, deadtime, trigger menus or array-level event building, or measure efficiency for independently observed cosmic-ray events.
Next steps are to test longer data across multiple stations, adjust thresholds as noise changes, and complete board integration with activity-based power measurements. The report is an arXiv version 1 preprint dated 20 August 2026.
Paper data and sources
Original title: A dual-polarization whitened-template trigger for real-time radio detection of extensive air showers
Authors: Qader Dorosti, Alperen Aksoy, Ilja Bekman et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text