pith. machine review for the scientific record. sign in

arxiv: 2605.13627 · v1 · submitted 2026-05-13 · ⚛️ physics.ins-det · physics.data-an

Recognition: unknown

SINAPSE: A lightweight deep learning framework for accurate and explainable neutron-γ discrimination

Thomas Carreau , Adrien Matta , Owen Syrett , Beno\^it Mauss , David Etasse , Audrey Chatillon , Cyril Lenain , Pierre Morfouace
show 6 more authors
Julien Taieb David Regnier Patrick Copp Matthew Devlin Charl\`ene Surault Jason Surbrook
Authors on Pith no claims yet

Pith reviewed 2026-05-14 17:47 UTC · model grok-4.3

classification ⚛️ physics.ins-det physics.data-an
keywords neutron-gamma discriminationpulse shape discriminationdeep learningwaveform denoisingautoencoderexplainable AIorganic scintillatorslow-charge regime
0
0 comments X

The pith

A lightweight dual-branch neural network denoises low-charge waveforms and classifies neutrons versus gammas with calibrated probabilities.

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

The paper introduces SINAPSE to solve neutron-gamma discrimination in organic scintillators when signals are too weak for reliable traditional pulse-shape cuts. It trains a 1D convolutional autoencoder to clean waveforms and a paired classifier to label particles by first applying random augmentations to high-quality data so the model learns to handle simulated low-charge noise. If the approach works, it extends accurate identification into regimes where conventional methods break down, improving measurements in nuclear physics and radiation detection. SHAP explanations confirm the model focuses on established pulse-shape features rather than artifacts. This matters because many experiments operate near the low-charge limit where signal-to-noise ratios drop sharply.

Core claim

SINAPSE uses a dual-branch architecture with a 1-dimensional convolutional autoencoder for waveform denoising and a classifier for particle identification. Random augmentations applied to high-SNR waveforms simulate low-charge conditions, enabling training where conventional PSD labels are unreliable. The framework shows superior denoising performance over conventional digital signal processing, outputs well-calibrated probabilities consistent with traditional graphical cuts, and relies on physically meaningful pulse-shape features according to SHAP analysis.

What carries the argument

Dual-branch architecture of a 1D convolutional autoencoder for denoising paired with a particle classifier, trained via random augmentations on high-SNR waveforms and interpreted with SHAP values.

Load-bearing premise

Random augmentations applied to high-SNR waveforms must faithfully reproduce the noise statistics and pulse-shape distortions present in real low-charge experimental data.

What would settle it

Direct test of classification accuracy and probability calibration on real low-charge experimental waveforms labeled independently by time-of-flight or equivalent ground truth, compared against conventional DSP performance; significant underperformance would falsify the superiority claim.

Figures

Figures reproduced from arXiv: 2605.13627 by Adrien Matta, Audrey Chatillon, Beno\^it Mauss, Charl\`ene Surault, Cyril Lenain, David Etasse, David Regnier, Jason Surbrook, Julien Taieb, Matthew Devlin, Owen Syrett, Patrick Copp, Pierre Morfouace, Thomas Carreau.

Figure 1
Figure 1. Figure 1: FIG. 1. Density matrix of the pulse shape discrimination [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Pulse shape discrimination factor distributions for the train set and validation set before and after augmentation, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Overview of the SINAPSE framework. An input waveform is first augmented using random transformations (temporal [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Distribution of per-signal mean squared error between [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Left: Example signals for which model predictions are in agreement with VENDETA PSD results. Middle: Example [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Examples of signal denoising using four selected ap [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. PSD as a function of the light output for low-energy prompt [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Density matrix of the PSD factor as a function of [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. PSD as a function of light output in MeVee with prob [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 13
Figure 13. Figure 13: FIG. 13. Mean SHAP value as a function of time for predicted [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12. SHAP values for three signals. Left: Signal with [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

Traditionally, neutron-$\gamma$ discrimination in organic scintillators relies on techniques such as time-of-flight (ToF) selection and pulse-shape discrimination (PSD). However, particle identification through graphical cuts remains challenging in the low-charge regime due to poor signal-to-noise ratios (SNR). In this work, we propose SINAPSE, a lightweight deep learning framework for accurate and explainable neutron-$\gamma$ discrimination in the low-charge regime. The framework employs a dual-branch architecture that combines a 1-dimensional convolutional autoencoder for waveform denoising with a classifier for particle identification. Random augmentations are applied to high-SNR waveforms to simulate low-charge conditions, enabling robust extrapolation into regimes where conventional PSD labels are unreliable. We show that SINAPSE achieves superior denoising performance compared to conventional digital signal processing techniques, and outputs well-calibrated probabilities, consistent with traditional graphical cuts. Finally, we apply SHAP (SHapley Additive exPlanations) values to show that model decisions are driven by physically meaningful pulse-shape features, confirming consistency with established PSD principles.

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

2 major / 1 minor

Summary. The manuscript proposes SINAPSE, a lightweight dual-branch deep learning framework for neutron-gamma discrimination in organic scintillators. It combines a 1D convolutional autoencoder for denoising waveforms with a classifier for particle identification, trained via random augmentations applied to high-SNR data to extrapolate into the low-charge regime where conventional PSD labels are unreliable. The paper claims superior denoising performance relative to digital signal processing techniques, well-calibrated output probabilities consistent with traditional graphical cuts, and explainability via SHAP values that align model decisions with established pulse-shape features.

Significance. If the central claims hold after addressing validation gaps, the work could offer a practical, lightweight, and interpretable tool for improving particle identification in low-SNR scintillator data, where traditional methods struggle. The emphasis on physical consistency via SHAP and the dual-branch design for denoising-plus-classification represent potential strengths for applications in nuclear instrumentation and radiation detection experiments.

major comments (2)
  1. [Methods] Methods (augmentation procedure): The central extrapolation claim relies on random augmentations of high-SNR waveforms to simulate low-charge conditions, yet no specific augmentation parameters, noise power spectra, or quantitative fidelity metrics (e.g., Kolmogorov-Smirnov tests on baseline fluctuations or pulse-shape distortions) are provided to demonstrate that the simulated data match real low-SNR experimental statistics. This directly affects both the denoising superiority and calibration claims.
  2. [Results] Results (performance evaluation): The abstract asserts superior denoising and well-calibrated probabilities relative to DSP and graphical cuts, but the manuscript supplies no quantitative metrics, baseline comparisons, error bars, or held-out real low-SNR validation set details. Without these, the load-bearing performance claims cannot be assessed for statistical significance or robustness.
minor comments (1)
  1. [Abstract] Abstract: Key quantitative results (e.g., denoising MSE or AUC values) should be included to allow readers to immediately gauge the magnitude of the reported improvements.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help strengthen the manuscript. We address each major point below and will incorporate revisions to improve methodological transparency and quantitative rigor.

read point-by-point responses

  1. Referee: [Methods] Methods (augmentation procedure): The central extrapolation claim relies on random augmentations of high-SNR waveforms to simulate low-charge conditions, yet no specific augmentation parameters, noise power spectra, or quantitative fidelity metrics (e.g., Kolmogorov-Smirnov tests on baseline fluctuations or pulse-shape distortions) are provided to demonstrate that the simulated data match real low-SNR experimental statistics. This directly affects both the denoising superiority and calibration claims.

    Authors: We agree that the augmentation procedure requires fuller specification to support the extrapolation claims. In the revised manuscript we will add the exact augmentation parameters (including noise amplitude ranges, frequency content, and application order), describe the noise power spectra employed, and report quantitative fidelity metrics such as Kolmogorov-Smirnov tests comparing baseline fluctuations and pulse-shape statistics between simulated and real low-SNR data. These additions will allow readers to evaluate the realism of the simulated low-charge regime. revision: yes

  2. Referee: [Results] Results (performance evaluation): The abstract asserts superior denoising and well-calibrated probabilities relative to DSP and graphical cuts, but the manuscript supplies no quantitative metrics, baseline comparisons, error bars, or held-out real low-SNR validation set details. Without these, the load-bearing performance claims cannot be assessed for statistical significance or robustness.

    Authors: We acknowledge that the current results section would benefit from expanded quantitative reporting. The revised manuscript will include explicit denoising metrics (e.g., SNR gain, MSE, and waveform fidelity scores) with error bars from multiple runs, direct numerical comparisons against DSP baselines, and a clearer description of the validation strategy. Because conventional PSD labels become unreliable at low charge, our primary validation relies on consistency with graphical cuts on higher-SNR data plus SHAP-based physical interpretability; we will expand this discussion to address robustness. revision: yes

Circularity Check

0 steps flagged

No circularity; claims rest on independent validation against traditional PSD cuts

full rationale

The paper trains a dual-branch autoencoder+classifier on randomly augmented high-SNR waveforms and evaluates denoising and probability calibration against separate traditional graphical PSD cuts. No equations, fitted parameters, or self-citations are shown that reduce the performance claims to tautological inputs by construction. Augmentation is presented as an empirical modeling choice whose fidelity is an external assumption, not a definitional loop. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that synthetic low-charge waveforms generated by random augmentation preserve the essential statistical and shape properties of real experimental data; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption Random augmentations on high-SNR waveforms accurately simulate the noise and pulse-shape distortions of real low-charge data
    This assumption enables training the model for regimes where conventional labels are unavailable.

pith-pipeline@v0.9.0 · 5540 in / 1223 out tokens · 58563 ms · 2026-05-14T17:47:40.108538+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

34 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    R. C. Haight, H. Y. Lee, T. N. Taddeucci, J. M. O’Donnell, B. A. Perdue, N. Fotiades, M. Devlin, J. L. Ullmann, A. Laptev, T. Bredeweg,et al., Two detector arrays for fast neutrons at lansce, Journal of Instrumen- tation7(03), C03028

  2. [2]

    Mart´ ınez, D

    T. Mart´ ınez, D. Cano-Ott, J. Castilla, A. Garcia, J. Marin, G. Martinez, E. Mendoza, C. Santos, F. Tera, D. Villamarin,et al., Monster: a tof spectrometer forβ- delayed neutron spectroscopy, Nuclear Data Sheets120, 78 (2014)

    2014

  3. [3]

    Talou and R

    P. Talou and R. Vogt,Nuclear fission: theories, experi- ments and applications(Springer Nature, 2023)

    2023

  4. [4]

    Brooks, Development of organic scintillators, Nuclear Instruments and Methods162, 477 (1979)

    F. Brooks, Development of organic scintillators, Nuclear Instruments and Methods162, 477 (1979)

    1979

  5. [5]

    Fabian, G

    X. Fabian, G. Baulieu, L. Ducroux, O. St´ ezowski, A. Boujrad, E. Cl´ ement, S. Coudert, G. de France, N. Er- duran, S. Ert¨ urk,et al., Artificial neural networks for neutron/γdiscrimination in the neutron detectors of neda, Nuclear Instruments and Methods in Physics Re- search Section A: Accelerators, Spectrometers, Detectors and Associated Equipment986...

    2021

  6. [6]

    R. L. Garnett and S. H. Byun, Neutralnet: Develop- ment and testing of a machine learning solution for pulse shape discrimination, Applied Radiation and Isotopes 211, 111384 (2024)

    2024

  7. [7]

    Panda, P

    S. Panda, P. Netrakanti, S. Behera, R. Sahu, K. Ku- mar, R. Sehgal, D. Mishra, and V. Jha, Discrimination of neutron-γin the low energy regime using machine learning for an ej-276d plastic scintillator, arXiv preprint arXiv:2506.13802 (2025)

    work page arXiv 2025

  8. [8]

    Syrett, C

    O. Syrett, C. Lenain, B. Mauss, J. Taieb, P. Morfouace, A. Chatillon, and D. Etasse, Vendeta: Versatile neu- tron detector array, a new high-resolution neutron time- of-flight measurement array, Nuclear Instruments and 13 Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment , 170824 (2025)

    2025

  9. [9]

    Laurent, J

    B. Laurent, J. Taieb, G. B´ elier, P. Marini, and P. Mor- fouace, New developments of a fission chamber for very high radioactivity samples, Nuclear Instruments and Methods in Physics Research Section A: Acceler- ators, Spectrometers, Detectors and Associated Equip- ment990, 164966 (2021)

    2021

  10. [10]

    FASTER, Faster,https://faster.in2p3.fr/

  11. [11]

    G. F. Knoll,Radiation detection and measurement(John Wiley & Sons, 2010)

    2010

  12. [12]

    Harrison, J

    J. Harrison, J. Willes, and J. Snoek, Variational bayesian last layers, inInternational Conference on Learning Rep- resentations (ICLR)(2024)

    2024

  13. [13]

    Watson, J

    J. Watson, J. A. Lin, P. Klink, J. Pajarinen, and J. Pe- ters, Latent derivative bayesian last layer networks, in International Conference on Artificial Intelligence and Statistics(PMLR, 2021) pp. 1198–1206

    2021

  14. [14]

    GRIT, Grit,https://grit.in2p3.fr/

  15. [15]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, Advances in neural information processing systems32(2019)

    2019

  16. [16]

    Haar, Zur theorie der orthogonalen funktionensys- teme, Mathematische Annalen71, 38 (1911)

    A. Haar, Zur theorie der orthogonalen funktionensys- teme, Mathematische Annalen71, 38 (1911)

    1911

  17. [17]

    L. I. Rudin, S. Osher, and E. Fatemi, Nonlinear total vari- ation based noise removal algorithms, Physica D: nonlin- ear phenomena60, 259 (1992)

    1992

  18. [18]

    M. Pavlovic, Understanding model calibration–a gen- tle introduction and visual exploration of calibration and the expected calibration error (ece), arXiv preprint arXiv:2501.19047 (2025)

    work page arXiv 2025

  19. [19]

    ONNX, Open neural network exchange,https://onnx. ai/

  20. [20]

    J. Y. Araz, A. Buckley, G. Kasieczka, J. Kieseler, S. Kraml, A. Kvellestad, A. Lessa, T. Procter, A. Raklev, H. Reyes-Gonzalez,et al., Les houches guide to reusable ml models in lhc analyses, SciPost Physics Community Reports , 003 (2024)

    2024

  21. [21]

    org/abs/2308.01222

    C. Wang, Calibration in deep learning: A survey of the state-of-the-art, arXiv preprint arXiv:2308.01222 (2023)

    work page arXiv 2023

  22. [22]

    M. Kull, T. M. S. Filho, and P. Flach, Beyond sigmoids: How to obtain well-calibrated probabilities from binary classifiers with beta calibration, Electronic Journal of Statistics11, 5052 (2017)

    2017

  23. [23]

    Longo, M

    L. Longo, M. Brcic, F. Cabitza, J. Choi, R. Con- falonieri, J. Del Ser, R. Guidotti, Y. Hayashi, F. Her- rera, A. Holzinger,et al., Explainable artificial intelli- gence (xai) 2.0: A manifesto of open challenges and inter- disciplinary research directions, Information Fusion106, 102301 (2024)

    2024

  24. [24]

    Towards A Rigorous Science of Interpretable Machine Learning

    F. Doshi-Velez and B. Kim, Towards a rigorous sci- ence of interpretable machine learning, arXiv preprint arXiv:1702.08608 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  25. [25]

    S. M. Lundberg and S.-I. Lee, A unified approach to in- terpreting model predictions, Advances in neural infor- mation processing systems30(2017)

    2017

  26. [26]

    Adebayo, J

    J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, Sanity checks for saliency maps, Advances in neural information processing systems31 (2018)

    2018

  27. [27]

    Kindermans, S

    P.-J. Kindermans, S. Hooker, J. Adebayo, M. Alber, K. T. Sch¨ utt, S. D¨ ahne, D. Erhan, and B. Kim, The (un) reliability of saliency methods, inExplainable AI: Interpreting, explaining and visualizing deep learning (Springer, 2019) pp. 267–280

    2019

  28. [28]

    Shrikumar, P

    A. Shrikumar, P. Greenside, and A. Kundaje, Learning important features through propagating activation dif- ferences, inInternational conference on machine learning (PMlR, 2017) pp. 3145–3153

    2017

  29. [29]

    S. M. Lundberg and S.-I. Lee, A unified approach to interpreting model predictions (Curran Associates, Inc., 2017)

    2017

  30. [30]

    Matta, P

    A. Matta, P. Morfouace, N. d. S´ er´ eville, F. Flavigny, M. Labiche, and R. Shearman, Nptool: a simulation and analysis framework for low-energy nuclear physics ex- periments, Journal of Physics G: Nuclear and Particle Physics43, 045113 (2016)

    2016

  31. [31]

    O. R. developers, Onnx runtime,https://onnxruntime. ai/(2021)

    2021

  32. [32]

    S. Owen, C. Thomas, A. Matta, B. Mauss, D. Etasse, A. Chatillon, C. Lenain, P. Morfouace, J. Taieb, D. Reg- nier, M. Delvin, P. Copp, C. Surault, and J. Surbrook, Sinapse neutron-gamma discrimination, 10.5281/zen- odo.19922303 (2026)

    work page doi:10.5281/zen- 2026

  33. [33]

    NVIDIA, Tensorrt,https://github.com/NVIDIA/ TensorRT

  34. [34]

    Berthelot, N

    D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, and C. A. Raffel, Mixmatch: A holistic ap- proach to semi-supervised learning, Advances in neural information processing systems32(2019)

    2019