Pith

open record

sign in

arxiv: 2605.28332 · v1 · pith:4VIUD3JS · submitted 2026-05-27 · astro-ph.IM · hep-ex

Hybrid neural denoising for resource-efficient near- and sub-threshold radio triggering of extensive air showers

Alperen Aksoy , Ilja Bekman , Markus Cristinziani , Eric-Teunis de Boone , Vesselin Dimitrov , Qader Dorosti , Chimezie Eguzo , Stefan Heidbrink
show 2 more authors
Stefan van Waasen Andre Zambanini
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-29 10:11 UTCgrok-4.3pith:4VIUD3JSrecord.jsonopen to challenge →

classification astro-ph.IM hep-ex
keywords radio detectionair showersneural denoisingtriggeringFPGA implementationsub-threshold signalscosmic ray detectionwaveform classification
0
0 comments X

The pith

A compact convolutional denoiser followed by a classifier retains 41 percent of weak air-shower radio traces at a false-positive rate of 10 to the minus 4, where the classical peak-envelope trigger retains none.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that treating waveform recovery and classification as a single deployment-constrained task yields a hybrid neural trigger that improves signal-background separation for near- and sub-threshold pulses. It tests the approach on measured high-interference background traces paired with detector-folded simulated air-shower pulses and shows the cleaned waveforms carry usable trigger information beyond a final score. The full chain is shown to meet FPGA timing and resource limits after quantisation-aware training and high-level synthesis. A sympathetic reader would care because the result points to a practical way to run sensitive radio self-triggering on edge hardware in noisy environments without raising false-alarm rates.

Core claim

The hybrid neural trigger, built from a compact convolutional denoiser that maps a noisy single-channel trace to a cleaned estimate of the air-shower pulse and a compact classifier that then evaluates it, improves efficiency at fixed false-positive rates. At a false-positive rate of 10^-4 the chain retains about 41 percent of held-out weak-signal traces while the classical peak-envelope trigger retains none. The denoiser alone already converts a simple peak-envelope decision into an efficient weak-pulse trigger, and the cleaned waveform preserves timing and peak-amplitude structure for diagnostics and selective readout.

What carries the argument

The hybrid neural denoiser-classifier chain in which a compact convolutional denoiser produces a cleaned waveform that is then passed to a compact classifier, with the entire pipeline linked through hyperparameter optimisation, quantisation-aware training, fixed-point quantisation, and HLS firmware export.

If this is right

  • The cleaned waveform preserves timing and peak-amplitude structure usable for station-level diagnostics, feature extraction, and selective readout.
  • The firmware implementation meets timing on representative FPGA targets with microsecond-scale latency and compact arithmetic-resource demand.
  • The method enables radio-only triggering for weak and inclined air-shower signals in noisy environments.
  • Model selection and deployment constraints are linked through a single optimisation and export pipeline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same denoiser-first structure could be tested on other radio-detector arrays that already record single-channel traces, to check whether the efficiency gain transfers without retraining.
  • If background statistics drift seasonally, periodic retraining of only the classifier while keeping the denoiser fixed might maintain performance with lower computational cost.
  • The preserved timing information in the cleaned waveform might allow tighter coincidence windows across stations, reducing the overall trigger rate further.

Load-bearing premise

The detector-folded simulated air-shower pulses accurately represent real sub-threshold signals and the measured high-interference background traces represent operational conditions at target sites.

What would settle it

Running the trained firmware on a set of real measured air-shower events recorded at an operational station and comparing the retained fraction at 10^-4 false-positive rate against the 41 percent benchmark obtained on simulated traces.

read the original abstract

Autonomous radio self-triggering for extensive air showers must reject variable radio-frequency interference while preserving sensitivity to weak pulses and remaining compatible with station-level edge hardware. This work presents a hybrid neural trigger in which waveform recovery and signal classification are treated as a single deployment-constrained problem. A compact convolutional denoiser maps a noisy single-channel trace to a cleaned estimate of the air-shower pulse, which is then evaluated by a compact classifier. The method is tested with measured high-interference background traces and detector-folded air-shower pulses from the Pierre Auger Offline simulation chain, with signals concentrated in the near- and sub-threshold regime. Model selection and deployment are linked through hyperparameter optimisation, quantisation-aware training, fixed-point quantisation, hls4ml firmware export, high-level synthesis, and register-transfer-level validation. The denoiser alone turns a simple peak-envelope decision into an efficient weak-pulse trigger, showing that the cleaned waveform carries trigger-relevant information beyond a final classifier score. In the full denoiser-classifier chain, the hybrid trigger improves signal-background separation and efficiency at fixed false-positive rates: at a false-positive rate of 10^-4, it retains about 41% of held-out signal traces in the weak-signal benchmark, while the classical peak-envelope trigger retains none. The cleaned waveform preserves timing and peak-amplitude structure for station-level diagnostics, feature extraction, and selective readout. The firmware meets timing on representative FPGA targets with microsecond-scale latency and compact arithmetic-resource demand. These results establish hybrid neural denoising as a practical route toward radio-only triggering for weak and inclined air-shower signals in noisy environments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes a hybrid neural trigger for radio detection of extensive air showers consisting of a compact convolutional denoiser followed by a classifier. Trained and tested on measured high-interference background traces combined with detector-folded air-shower pulses from the Pierre Auger Offline simulation chain (concentrated in the near- and sub-threshold regime), the method reports improved signal-background separation: at a false-positive rate of 10^{-4} the hybrid chain retains ~41% of held-out signal traces while a classical peak-envelope trigger retains none. The work further links model selection to quantisation-aware training, hls4ml export, and FPGA firmware validation, claiming microsecond latency and low resource use while preserving waveform timing and amplitude for diagnostics.

Significance. If the reported separation holds under real operational conditions, the approach would provide a practical, hardware-compatible route to radio-only triggering of weak and inclined showers in noisy environments, addressing a recognised limitation of current autonomous radio arrays. The explicit coupling of denoising to trigger-relevant information and the end-to-end firmware path are concrete strengths; the absence of free parameters in the final architecture and the use of measured backgrounds further support reproducibility.

major comments (3)
  1. [§4] §4 (Testing and results): the headline efficiency numbers (41% signal retention at FPR=10^{-4}) are obtained exclusively on held-out traces drawn from the same Pierre Auger Offline simulation chain used for training and signal folding. No independent test set of real recorded sub-threshold air-shower events is reported, so the central claim that the hybrid trigger improves separation under operational conditions rests on the unverified assumption that the folded simulated pulses accurately reproduce real near-threshold pulse shapes, polarisation statistics and noise correlations after detector response.
  2. [§3.2 and §4] §3.2 (Data partitioning) and §4: the manuscript provides no description of the train/validation/test split ratios, no cross-validation across different simulation runs or background epochs, and no statistical significance tests or error bars on the ROC or efficiency curves. These omissions make it impossible to assess whether the reported gain over the peak-envelope baseline is robust to data-partitioning choices.
  3. [§5] §5 (Firmware implementation): while latency and resource figures are stated, the manuscript does not quantify how quantisation and HLS synthesis affect the denoiser output distribution or the downstream classifier score; without this, it is unclear whether the FPGA implementation preserves the separation performance measured in floating-point simulation.
minor comments (3)
  1. [Figure 3] Figure 3 (or equivalent ROC plot): axis labels and legend should explicitly state whether the curves are computed on the held-out simulation set or include any real-data component.
  2. [§1] The abstract and §1 cite 'measured high-interference background traces' but do not specify the station, frequency band, or duration of the background dataset; a brief table or reference would improve reproducibility.
  3. [§2] Notation for the false-positive rate (10^{-4}) is used without defining the exact definition of 'positive' (trigger decision per trace or per time window); a short clarification in §2 would remove ambiguity.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed report. We respond point-by-point to the major comments, indicating where the manuscript will be revised. The work relies on measured backgrounds and detector-folded simulations because real sub-threshold events cannot be collected without an operational trigger; we address this limitation explicitly while strengthening the other sections.

read point-by-point responses

  1. Referee: [§4] §4 (Testing and results): the headline efficiency numbers (41% signal retention at FPR=10^{-4}) are obtained exclusively on held-out traces drawn from the same Pierre Auger Offline simulation chain used for training and signal folding. No independent test set of real recorded sub-threshold air-shower events is reported, so the central claim that the hybrid trigger improves separation under operational conditions rests on the unverified assumption that the folded simulated pulses accurately reproduce real near-threshold pulse shapes, polarisation statistics and noise correlations after detector response.

    Authors: We agree that no independent set of real recorded sub-threshold events is available, as such events are by definition below existing trigger thresholds and cannot be identified without a working trigger. The manuscript already uses real measured high-interference background traces; the signal component is obtained by folding validated Pierre Auger Offline simulations through the detector response. In revision we will expand §4 with additional discussion of simulation fidelity (citing prior Auger validation studies), explicit limitations of the folded-pulse approach, and the fact that the method is evaluated under realistic measured noise conditions rather than idealised noise. revision: partial

  2. Referee: [§3.2 and §4] §3.2 (Data partitioning) and §4: the manuscript provides no description of the train/validation/test split ratios, no cross-validation across different simulation runs or background epochs, and no statistical significance tests or error bars on the ROC or efficiency curves. These omissions make it impossible to assess whether the reported gain over the peak-envelope baseline is robust to data-partitioning choices.

    Authors: These details were omitted from the original submission. The revised manuscript will specify the train/validation/test split ratios, describe the partitioning procedure used to prevent leakage between background epochs and simulation runs, report results from cross-validation across independent background epochs, and include error bars on all ROC and efficiency curves obtained via bootstrap resampling. revision: yes

  3. Referee: [§5] §5 (Firmware implementation): while latency and resource figures are stated, the manuscript does not quantify how quantisation and HLS synthesis affect the denoiser output distribution or the downstream classifier score; without this, it is unclear whether the FPGA implementation preserves the separation performance measured in floating-point simulation.

    Authors: We will add to §5 a direct comparison of denoiser output distributions and classifier scores in floating-point versus post-quantisation and post-HLS versions. This will quantify any degradation and confirm that the reported separation performance is preserved within acceptable bounds for the target FPGA. revision: yes

standing simulated objections not resolved
  • Absence of an independent test set of real recorded sub-threshold air-shower events, which cannot be obtained without an existing operational trigger.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper reports an empirical performance result (41% signal retention at FPR=10^-4 on held-out traces vs. 0% for peak-envelope) measured on a test split of detector-folded simulations and measured backgrounds. No equations, fitted parameters renamed as predictions, self-definitional steps, or load-bearing self-citations appear in the abstract or described claims. The evaluation uses separate held-out data, so the reported efficiency gain is a measured outcome rather than a quantity equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; all technical details required for ledger population are absent.

pith-pipeline@v0.9.1-grok · 5878 in / 1163 out tokens · 27409 ms · 2026-06-29T10:11:37.648720+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 7 canonical work pages

  1. [1]

    Schröder,Radio detection of cosmic-ray air showers and high-energy neutrinos,Prog

    F.G. Schröder,Radio detection of cosmic-ray air showers and high-energy neutrinos,Prog. Part. Nucl. Phys.93(2017) 1

    2017

  2. [2]

    Huege,Radio detection of cosmic ray air showers in the digital era,Phys

    T. Huege,Radio detection of cosmic ray air showers in the digital era,Phys. Rep.620(2016) 1

    2016

  3. [3]

    F. D. Kahn and I. Lerche,Radiation from cosmic ray air showers,Proc. R. Soc. Lond. A289(1966) 206

    1966

  4. [4]

    Conti and G

    E. Conti and G. Sartori,On the coherent emission of radio frequency radiation from high energy particle showers,Int. J. Mod. Phys. D26(2017) 1750083. [7]Pierre Augercollaboration, A. Abdul Halim et al.,Radio measurements of the depth of air-shower maximum at the pierre auger observatory,Phys. Rev. D.109(2024) . [8]Pierre Augercollaboration, A. Aab et al.,...

    2017

  5. [5]

    Schmidt,Realization of a self-triggered detector for the radio emission of cosmic rays, Ph.D

    A. Schmidt,Realization of a self-triggered detector for the radio emission of cosmic rays, Ph.D. thesis, Karlsruher Institut für Technologie (KIT), 2011. 10.5445/IR/1000030957. [10]CODALEMAcollaboration, D. Torres Machado et al.,Latest results of the CODALEMA experiment: cosmic rays radio detection in a self trigger mode,J. Phys. Conf. Ser.409(2013) 012074

    work page doi:10.5445/ir/1000030957 2011

  6. [6]

    Kelley,Data acquisition, triggering, and filtering at the auger engineering radio array,Nucl

    J. Kelley,Data acquisition, triggering, and filtering at the auger engineering radio array,Nucl. Instrum. Methods Phys. Res. A725(2013) 133

    2013

  7. [7]

    Dorosti,AI-enhanced self-triggering for extensive air showers: performance and FPGA feasibility, JINST20(2025) P10010 [2502.21198]

    Q. Dorosti,AI-enhanced self-triggering for extensive air showers: performance and FPGA feasibility, JINST20(2025) P10010 [2502.21198]

    work page arXiv 2025

  8. [8]

    Sun et al.,HGQ: High granularity quantization for real-time neural networks on FPGAs,Proc

    C. Sun et al.,HGQ: High granularity quantization for real-time neural networks on FPGAs,Proc. ACM/SIGDA Int. Symp. Field Programmable Gate Arrays(2026) 79. – 30 –

    2026

  9. [9]

    Argirò et al.,The offline software framework of the pierre auger observatory,Nucl

    S. Argirò et al.,The offline software framework of the pierre auger observatory,Nucl. Instrum. Methods Phys. Res. A580(2007) 1485. [15]Pierre Augercollaboration, B. Fuchs,The auger engineering radio array,Nucl. Instrum. Meth. A 692(2012) 93. [16]Pierre Augercollaboration, T. Huege et al.,Determination of the energy scale of cosmic ray measurements using t...

    work page arXiv 2007

  10. [10]

    FastML Team,fastmachinelearning/hls4ml,10.5281/zenodo.1201549(2025)

    work page doi:10.5281/zenodo.1201549(2025 2025

  11. [11]

    Duarte et al.,Fast inference of deep neural networks in FPGAs for particle physics,JINST13 (2018) P07027 [1804.06913]

    J. Duarte et al.,Fast inference of deep neural networks in FPGAs for particle physics,JINST13 (2018) P07027 [1804.06913]

    work page arXiv 2018

  12. [12]

    2101.05108 , archiveprefix =

    T. Aarrestad et al.,Fast convolutional neural networks on FPGAs with hls4ml,Mach. Learn. Sci. Tech.2(2021) 045015 [2101.05108]

    work page arXiv 2021

  13. [13]

    Li et al.,Hyperband: A novel bandit-based approach to hyperparameter optimization,Journal of Machine Learning Research18(2018) 1

    L. Li et al.,Hyperband: A novel bandit-based approach to hyperparameter optimization,Journal of Machine Learning Research18(2018) 1

    2018

  14. [14]

    Coelho, Jr et al.,Automatic heterogeneous quantization of deep neural networks for low-latency inference on the edge for particle detectors,Nat

    C.N. Coelho, Jr et al.,Automatic heterogeneous quantization of deep neural networks for low-latency inference on the edge for particle detectors,Nat. Mach. Intell.3(2021) 675

    2021

  15. [15]

    AMD,Vitis High-Level Synthesis (HLS) User Guide, October, 2023

    2023

  16. [16]

    AMD,Vivado Design Suite User Guide: Design Flows Overview (UG892), 2024

    2024

  17. [17]

    Gadde et al.,Effective design verification – constrained random with python and cocotb, 2407.10312

    D.N. Gadde et al.,Effective design verification – constrained random with python and cocotb, 2407.10312. – 31 –

    work page arXiv