Denoising radio pulses from air showers using machine-learning methods

Ars\`ene Ferri\`ere (for the GRAND Collaboration); Aur\'elien Benoit-L\'evy; Oscar Macias; Zhisen Lai

REVIEW 4 major objections 6 minor 12 references

Denoising radio pulses from air showers using machine-learning methods

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper reports that an encoder-decoder neural network can denoise GRANDProto300 radio traces well enough to recover air-shower pulses at signal-to-noise ratios around 4 with over 95% efficiency and a false-positive rate near 1%.

desk verdict A solid ML denoising feasibility study for GRANDProto300 with real noise in the test set, but the claimed sensitivity enhancement is not benchmarked against a raw threshold trigger or a conventional filter. read the letter →

arxiv 2507.06688 v1 pith:N7M3UZZN submitted 2025-07-09 astro-ph.IM astro-ph.HE

Aurélien Benoit-Lévy , Zhisen Lai , Oscar Macias , Arsène Ferrière (for the GRAND Collaboration) This is my paper · ORCID

classification astro-ph.IMastro-ph.HE

keywords denoisingencoder-decoderconvolutionalneuralnetworkGRANDProto300air-showerradiodetectionsignal-to-noiseratiocosmicraysZHAireS

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether machine learning can make the GRANDProto300 radio array detect fainter cosmic-ray air showers by cleaning voltage traces before they are used for triggering and reconstruction. The authors build an encoder-decoder convolutional network, train it on simulated ZHAireS air-shower signals mixed with Gaussian noise whose power spectra are taken from real detector noise recordings, and test it on real noise traces. They report that for signals with signal-to-noise ratio above about 4, more than 95% of traces come out with a detectable peak, while pure-noise traces cross the detection threshold only about 1.4% of the time in the south-north channel and 1.1% in the east-west channel. They also report that denoising keeps the recovered peak close to the true peak position and amplitude, where the raw noisy traces fail. If these numbers survive on real events, the method lowers the effective detection threshold of the array without a large false-trigger penalty.

What carries the argument

The load-bearing mechanism is an encoder-decoder convolutional network with two parallel encoder branches: one operates in the time domain with a convolutional layer and two residual blocks; the other applies a fast Fourier transform to the input, stacks the real and imaginary parts, and feeds them through the same residual-block structure. The decoder reconstructs the trace from the latent representation, residual connections allow gradient flow in the deep network, and full convolutions make the model independent of trace length. The training set is built from 8000 ZHAireS air-shower simulations processed through GRANDLib to mimic the GRANDProto300 electronics, with only traces whose peak exceeds 15 ADC retained; the pulse position is decorrelated from the trace by training on random 768-bin windows taken from 1024-bin traces. Noise is synthesized as Gaussian traces matching the average power spectra of real ADC-noise recordings, and the network minimizes the L1 distance between the denoised and noiseless trace, a choice that suppresses false reconstructions relative to L2 losses.

What would settle it

Use the trained model on real GRANDProto300 traces into which calibration pulses of known amplitude and arrival time have been injected at signal-to-noise ratio about 4; if fewer than about 95% of them come out above the 15-ADC threshold, or if the recovered peak is off by more than 10 ns in a non-negligible fraction, the central claim fails. On the noise side, run the model on a long stream of pure recorded noise and count threshold crossings; a false-positive fraction above about 1.4% at 15 ADC would contradict the reported control.

Watch

Extended reading notes

Core claim

The central claim is that a fully convolutional encoder-decoder, trained with an L1 reconstruction loss on 138,769 simulated signal-plus-noise trace pairs, can denoise GRANDProto300 radio data: for clean peak amplitudes above 15 ADC, the denoised trace exceeds the same threshold in over 95% of cases once the signal-to-noise ratio reaches about 4, on both polarizations. The recovered peak amplitude stays close to the true amplitude as the signal-to-noise ratio drops, whereas the raw noisy traces' peak amplitudes diverge; the fraction of badly timed peaks (offset by more than 10 or 20 ns) is sharply reduced. On pure noise, the false-positive fraction at the 15-ADC threshold is 1.4% for the south-north channel and 1.1% for the east-west channel, and it vanishes when the threshold is raised to 40 ADC. The authors conclude that the denoiser's success on real noise traces, despite being trained on Gaussian noise, supports its use on real detector data.

Load-bearing premise

The reported performance numbers assume that simulated ZHAireS air-shower pulses, after GRANDLib's model of the detector electronics, have the same shape and polarization behavior as real pulses recorded by GRANDProto300 in the field.

Editorial extensions

If this is right

The GRANDProto300 detection threshold can be pushed down to pulses with signal-to-noise ratio around 4 without paying more than about one percent false triggers.
Peak-time recovery becomes reliable at the level of the detector's GPS timing precision (10–20 ns) for pulses that would otherwise be lost in noise.
Because the model is fully convolutional, the same trained network can denoise traces of arbitrary length, including continuous readout windows.
A denoiser trained only on simulated Gaussian noise demonstrably works when applied to real noise traces, so the pipeline does not require contaminating training data with hidden real signals.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper stops at simulated signal pulses, so the open question is whether ZHAireS and GRANDLib trace shapes match real air-shower pulses; if they do, the same recipe should transfer to other radio arrays by swapping the RF-chain simulation.
An ablation removing the frequency-domain branch would isolate how much of the gain comes from the FFT path, a test the paper does not run.
Because false positives vanish at threshold 40 ADC, roughly SNR 3, the denoiser could plausibly serve as an online filter on the trigger path rather than only as an offline analysis step.
Training on power-spectrum-matched Gaussian noise avoids the hidden-signal contamination that raw noise traces would introduce, a subtlety that could be measured by comparing denoisers trained on the two noise types.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The denoiser does what it claims on simulated signals buried in real GRAND noise, and the use of real AN noise power spectra to train on matched noise is a genuinely nice choice. The secondary metrics—peak-time error and amplitude-ratio recovery—show clear improvements over the noisy traces, and the false-positive characterization on pure noise is a step often skipped in this literature. As an ICRC proceedings contribution, this is a competent status report from the GRAND collaboration.

What is actually new here is modest but real: the frequency-domain encoder branch, the residual-block architecture, and the specific application to GRANDProto300 with its noise environment. The training/test separation is clean, with simulated Gaussian noise for training and real measured noise for testing, and the authors avoided letting the network learn the pulse position by using random temporal windows. Those are good methodological instincts.

The soft spots are in the headline claim. The abstract and conclusion say the method produces a 'sensitivity enhancement,' and the 95% efficiency at SNR>4 is the supporting number. But that efficiency is computed only on the denoised output, with no comparison to a raw threshold trigger or a matched/Wiener filter. At SNR>4, a simple amplitude threshold on the raw trace may already pass for most events; the false-positive rate of that raw threshold is never reported. Without a ROC curve, or at least efficiency at fixed false-positive rate for both raw and denoised traces, the sensitivity-enhancement claim is unestablished, even under the paper's own simulation assumptions. The stress-test note gets this right.

Two smaller issues: there are no error bars on the efficiency curves, and the test set uses simulated signals with real noise, so the 'reality gap' is bridged only on the noise side, not the signal side. The conclusion overstates things slightly when it claims the method 'successfully meets' the reality-gap challenge. The peak-time and amplitude-ratio results do show real reconstruction gains, so I do not think the core method is flawed—just the sensitivity framing.

For an ICRC proceedings, I would accept this as-is with minor comments. If the authors want to claim sensitivity enhancement in a journal version, they need to add the baseline and a proper ROC comparison. I would not desk-reject it; a referee can push them to quantify the claimed gain and soften the conclusion. If I were in the radio-detection subfield, I would want this comparison in the paper before citing the sensitivity number.

Referee Report

4 major / 6 minor

Summary. The paper presents a supervised encoder-decoder convolutional network for denoising radio pulses from extensive air showers in the GRANDProto300 experiment. Signals are simulated with ZHAireS and processed through GRANDLib; noise is generated as Gaussian noise matching the power spectra of measured ADC noise traces, while real noise traces are used for testing. The model is trained with an L1 loss on 768-bin windows. The authors report a denoising efficiency above 95% for signals with SNR above about 4, improved peak-timing accuracy relative to noisy traces, amplitude recovery close to unity, and false-positive fractions of 1.4% (S-N) and 1.1% (E-W) at a 15-ADC threshold on pure-noise traces.

Significance. If the central claim of sensitivity enhancement were established, this work would be a useful contribution to the GRAND software pipeline and to radio-detection experiments more generally. The paper has several genuine strengths: it uses a realistic noise model derived from measured ADC noise spectra, it randomizes the pulse position to avoid the network learning a fixed time bin, it evaluates on a held-out test set with real noise traces, and it demonstrates large improvements in peak-timing accuracy and amplitude ratio over the noisy traces. However, the headline quantitative claim of a sensitivity enhancement is not yet supported because the denoising efficiency is not compared with the raw traces or a standard filter baseline; additionally, the test set uses simulated signals, so the real-world deployment claim requires qualification. The approach is promising and the gaps are addressable within the scope of a revision.

major comments (4)

[Section 5, Figures 3-4] The denoising efficiency is defined through Conditions 1 and 2, which compare only the clean-trace peak and the denoised-trace peak against the ADC threshold. No analogous efficiency is reported for the raw noisy traces, nor for a conventional filter such as a Wiener or matched filter. This is load-bearing for the claimed 'sensitivity enhancement': with 768 time bins, the maximum of pure Gaussian noise can reach several sigma, and the paper itself states that 15 ADC corresponds roughly to SNR = 1. Many traces with SNR > 4 may already satisfy the 15-ADC condition on the raw trace, in which case the 95% denoising efficiency does not demonstrate additional sensitivity. Please add the raw-trace efficiency curve, a simple amplitude-threshold baseline, and ideally a standard filter baseline, and recast the sensitivity claim based on the comparison.
[Section 3 and Section 6] The test set uses simulated ZHAireS signals added to real noise traces, so the evaluation validates denoising of simulated pulses under realistic noise, not of real air-shower pulses. The conclusion states that the model was 'successfully applied to real noise traces from the experiment' and 'paves the way for its integration into ongoing and future experiments,' which overstates the evidence. The reality gap for the signal component—pulse shape, polarization, and RF-chain response—remains untested. Please state this limitation explicitly in the conclusion and temper the deployment claim accordingly.
[Section 5, false positive fraction] The text is ambiguous about whether the false-positive test uses real AN noise traces or simulated Gaussian noise generated from AN power spectra. The earlier description says real noise traces are used for the testing set, but the false-positive paragraph says 'pure Gaussian AN noise.' This distinction is important because the model was trained on Gaussian noise, so evaluation on simulated Gaussian noise could underestimate the false-positive rate on real, non-Gaussian noise. Please clarify which dataset was used and, if real noise traces are available, report the false-positive fraction on them.
[Section 5, Figures 3-5] The efficiency, timing, amplitude-ratio, and false-positive results are quoted without statistical uncertainties. The test and validation sets contain 21,145 traces, so binomial confidence intervals on the 95% efficiency and the 1.4% false-positive fraction would be simple to compute and would strengthen the quantitative claims. Please add error bars or report the relevant counts.

minor comments (6)

[Section 3] The sentence 'Out of the many events that are simulated, many antennas produce very weak signals. In order to have a balanced dataset, we only retain traces for which the maximum amplitude in one of the two polarizations is above 15 ADC' describes a selection on the clean signal amplitude. This is consistent with Condition 1 of the efficiency metric, but the resulting bias toward high-amplitude events should be stated explicitly when interpreting the efficiency.
[Equation (1)] The SNR definition uses 'the maximum of the Hilbert envelope of the noiseless trace'; please clarify that this is the maximum over time of the envelope, and specify the units of the standard deviation in the denominator.
[Section 5, Figures 3 and 4] The figure captions refer to 'South-North axis' and 'East-West axis'; please define these axes in the text or caption, as the two polarizations are not otherwise described.
[Section 5, paragraph on efficiency] The statement 'the denoising efficiency is above 95% for signals with SNR≈4' should be accompanied by the exact SNR bin or the functional form of the efficiency curve, so the reader can see how sharply the transition occurs.
[Section 3, 'Noise' paragraph] The phrase 'As we will illustrate in Section 5, our approach successfully meets this challenge' is forward-looking; consider moving the supporting evidence to the results section or rephrasing to avoid a dangling promise.
[References] Reference [16] is cited as 'NUTRIG proceedings'; please provide the full arXiv or journal reference so that the ADC noise dataset can be located.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; held-out evaluation and independent data sources support the denoising claims.

full rationale

The paper's derivation chain is not circular. The denoiser is trained by minimizing the L1 loss between the reconstructed trace and the noiseless simulated trace (Eq. 2), using 138,769 training pairs, and then evaluated on a held-out test set of 21,145 pairs constructed from separate simulated signals and real AN noise traces (Section 5). The denoising efficiency, false-positive fraction, peak-time error, and amplitude ratio are all measured on this held-out test set, and the denoised results are explicitly compared to the noisy traces in Figures 3-5. No fitted parameter is renamed as a prediction, and no conclusion is derived from an equation that already contains the claimed result. The cited simulation packages (ZHAireS, GRANDLib) and the real noise dataset are independent inputs, not conclusions derived from the model. The paper's reliance on collaboration-internal references for noise data and simulation software is a provenance issue, not a circularity issue. The absence of a raw-threshold or classical-filter baseline means the 'sensitivity enhancement' claim is not fully benchmarked, but that is a missing comparison rather than a circular reduction. Therefore, no specific circular step can be identified.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the fidelity of the signal simulation chain and the representativeness of the real-noise dataset, not on any new physical entities. The chosen ADC thresholds are analysis choices that set the scale of the reported metrics.

free parameters (2)

ADC threshold for efficiency definition = 15 ADC counts
Conditions 1 and 2 in Section 5 define successful denoising relative to this threshold; changing it changes the reported efficiency and false positive rates.
ADC threshold for false positive vanishing = 40 ADC counts
In Section 5, the false positive fraction is reported as vanishing above this threshold; it is a chosen analysis threshold, not a fitted model parameter.

assumptions (4)

domain assumption ZHAireS and GRANDLib simulations accurately reproduce GRANDProto300 pulse shapes.
Used throughout to generate the signal training/test set (Section 3). If false, the reported efficiency does not transfer to real signals.
domain assumption The real noise traces from the AN dataset are representative of GRANDProto300's homogeneous noise environment.
Power spectra from these traces are used to generate Gaussian training noise and the real traces form the test set (Sections 3 and 5).
domain assumption Real noise is sufficiently well approximated by Gaussian noise with matched power spectra for training.
Training noise is generated as Gaussian with the measured power spectra (Section 3); non-Gaussian structure, if significant, would cause a training-test mismatch.
standard math Standard neural network training assumptions, e.g., Adam optimizer and cyclic learning rate, are sufficient for convergence.
Training procedure in Section 4 relies on well-established ML practice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Denoising radio pulses from air showers using machine-learning methods." pith.science (2026). https://pith.science/paper/N7M3UZZN

@misc{pith2026250706688,
  author       = {Pith},
  title        = {Pith review of: Denoising radio pulses from air showers using machine-learning methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7M3UZZN}},
  note         = {Machine review of arXiv:2507.06688}
}

read the original abstract

The Giant Radio Array for Neutrino Detection (GRAND) aims to detect radio signals from extensive air showers (EAS) caused by ultra-high-energy (UHE) cosmic particles. Galactic, hardware-like, and anthropogenic noise are expected to contaminate these signals. To address this problem, we propose training a supervised convolutional network known as an encoder-decoder. This network is used to learn a coded representation of the data and remove specific features from it. This denoiser is trained using high-fidelity air shower simulations specifically tailored to replicate the characteristics of signals detected by GRAND. In this contribution, we describe our machine-learning model and report initial results demonstrating the sensitivity enhancement resulting from our denoising algorithm when applied to realistically simulated GRAND signals with varying signal-to-noise ratios.

Figures

Figures reproduced from arXiv: 2507.06688 by the authors.

**Figure 1.** Architecture of our encoder-decoder architecture. The encoder part (yellow blocks) is composed of a time domain and a frequency domain branches, while the decoder part (red blocks) reconstructs the input signal to produce a denoised trace with the same dimension as the noisy input. procedure and/or provide better results. First, for both the encoder and the decoder, we use residual blocks [9], which are a key ingred… view at source ↗

**Figure 2.** Examples of denoised traces for varying signal-to-noise ratios (SNRs). In all panels, the raw or noisy traces are shown as orange lines, the true signal that we aim to recover is depicted in red, and the denoised traces are illustrated in blue. The dashed lines represent the 1𝜎-level of the noise in the corresponding trace. Panels a) to d) present examples where the denoising procedure was successful in terms of bot… view at source ↗

**Figure 3.** Denoising efficiency (solid black line) for an ADC threshold of 15 ADC. The gray dashed lines represent the 5% and 95% levels. The colored lines represent the fraction of denoised (solid lines) or noisy (dashed lines) traces that have their peak time off by more than 10 ns (orange lines) or 20 ns (blue lines) compared to the peak time of the clean signal. South-North axis [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Same as [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Left: Ratio of the peak amplitude of successfully denoised (red and blue dots) and noisy (orange and cyan triangles) traces to the peak amplitude of the noise-free signals. Denoising efficiently enables the recovery of the real signal for low SNR traces. Right: Distrib…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[1]

10, (2012) P10011

The Pierre Auger CollaborationJINST7no. 10, (2012) P10011

work page 2012
[2]

Ardouinet al

D. Ardouinet al. NIM-A555no. 1, (2005) 148–163. [3]Tunka-RexCollaboration, D. Kostuninet al. PoSICRC2019(2020) 319. [4]LOPESCollaboration, T. Huegeet al. J. Phys: Conf. Ser.110no. 6, (May, 2008) 062012

work page 2005
[5]

M. P. van Haarlemet al. A&A556(2013) A2

work page 2013
[6]

Schröder and A

F. Schröder and A. RehmanPoSARENA2024(2024) 034

work page 2024
[7]

Erdmannet al

M. Erdmannet al. Journal of Instrumentation14no. 04, (Apr, 2019) P04005

work page 2019
[8]

China- Phys

The GRAND CollaborationSci. China- Phys. Mech. Astron.63no. 1, (2019)

work page 2019
[9]

Heet al.in2016 IEEE CVPR, pp

K. Heet al.in2016 IEEE CVPR, pp. 770–778. 2016

work page 2016
[10]

Shelhameret al

E. Shelhameret al. IEEE TPAMI39no. 4, (2017) 640–651

work page 2017

Show all 12 references

[11]

Alvarez-Muñizet al

J. Alvarez-Muñizet al. Astroparticle Physics35no. 6, (Jan., 2012) 325–341. [12]GRANDCollaboration, R. Alves Batistaet al. Comp. Phys. Comm.308(2025) 109461. [13]GRANDCollaboration, O. MartineauPoSICRC2025(these proceedings) 1114. [14]GRANDCollaboration, P. MaPoSICRC2025(2025) 453

2025
[15]

Le Coz, O

S. Le Coz, O. Martineau-Huynh, and A. Benoit-LevyPoSARENA2022(2023) 041. [16]GRANDCollaboration, P. Correa and J. KöhlerPoSICRC2025(these proceedings) 229

2023
[17]

D. P. Kingma and J. Ba, 2017.https://arxiv.org/abs/1412.6980

2017 arXiv
[18]

Emergences

L. N. Smith, 2017.https://arxiv.org/abs/1506.01186. 8 Denoising radio pulses from air showers using machine-learning methodsAurélien Benoit-Lévy Full Author List: GRAND Collaboration J. Álvarez-Muñiz1, R. Alves Batista2,3, A. Benoit-Lévy4, T. Bister5,6, M. Bohacova7, M. Bustam...

1900 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

10, (2012) P10011

The Pierre Auger CollaborationJINST7no. 10, (2012) P10011

work page 2012

[2] [2]

Ardouinet al

D. Ardouinet al. NIM-A555no. 1, (2005) 148–163. [3]Tunka-RexCollaboration, D. Kostuninet al. PoSICRC2019(2020) 319. [4]LOPESCollaboration, T. Huegeet al. J. Phys: Conf. Ser.110no. 6, (May, 2008) 062012

work page 2005

[3] [5]

M. P. van Haarlemet al. A&A556(2013) A2

work page 2013

[4] [6]

Schröder and A

F. Schröder and A. RehmanPoSARENA2024(2024) 034

work page 2024

[5] [7]

Erdmannet al

M. Erdmannet al. Journal of Instrumentation14no. 04, (Apr, 2019) P04005

work page 2019

[6] [8]

China- Phys

The GRAND CollaborationSci. China- Phys. Mech. Astron.63no. 1, (2019)

work page 2019

[7] [9]

Heet al.in2016 IEEE CVPR, pp

K. Heet al.in2016 IEEE CVPR, pp. 770–778. 2016

work page 2016

[8] [10]

Shelhameret al

E. Shelhameret al. IEEE TPAMI39no. 4, (2017) 640–651

work page 2017