Pith. sign in

REVIEW 3 major objections 5 minor 15 references

CNN cuts charged-particle false hits fourfold in space X-ray detector

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-05 00:51 UTC pith:EBRPFTE6

load-bearing objection Honest, careful measurement of CNN vs grade-based event rejection on a pnCCD, but the headline 3.1% vs 10–12% advantage sits above 4 keV; inside the measured 1.4–4 keV sub-band the CNN is not better, which undercuts the mission-facing claim. the 3 major comments →

arxiv 2608.00476 v1 pith:EBRPFTE6 submitted 2026-08-01 astro-ph.IM

Development and Evaluation of a CNN-Based Charged-Particle Event Rejection Algorithm for Soft X-ray Detection in a pnCCD-Based Satellite System

classification astro-ph.IM
keywords Gamma-ray burstsCCDX-rayCharged particleMachine learningpnCCDonboard event selectionsoft X-ray detector
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that a compact convolutional neural network can replace the threshold-based grade event-selection rules used on X-ray CCDs, specifically for the pnCCD detector planned for the EAGLE wide-field monitor on the HiZ-GUNDAM satellite. The authors built a pnCCD readout system, irradiated a small prototype with Fe-55 X-rays and Sr-90 beta electrons, and measured how often each selection method mistakes one kind of event for the other. Their central numbers: the CNN misclassifies 3.1% of beta events as X-rays, while the conventional 3x3 and 5x5 grade methods misclassify 11.9% and 10.2%, respectively, and the CNN keeps X-ray acceptance at 98.1% over 1.4-23 keV. The advantage comes from the CNN reading the full 5x5 charge pattern, including spatial correlations and relative pixel amplitudes, rather than applying fixed windows and thresholds. If the result holds, onboard event selection for the mission can remove roughly four times more charged-particle background without losing soft X-ray photons, which matters because full-frame images cannot be downlinked.

Core claim

The paper's central claim is that, on this pnCCD prototype, a two-layer CNN with a 5x5 input window, 9,411 trainable parameters, and rotation/reflection augmentation achieves a beta-ray misclassification rate of 3.1 +/- 0.4% while accepting 98.1 +/- 0.2% of Fe-55 X-ray events, compared with 11.9 +/- 0.7% and 10.2 +/- 0.7% misclassification for the 3x3 and 5x5 grade methods, whose X-ray acceptance is 97.8% and 91.9%. The improvement is concentrated above about 4 keV, where grade methods misclassify compact, high-amplitude deposits that can come from charged particles; the CNN rejects these by learning the spatial morphology of charge spread. The paper further claims that the 450 micrometer fu

What carries the argument

The load-bearing object is a two-layer convolutional network that classifies 5x5-pixel charge-distribution patches centered on candidate events, using two 3x3 convolutional layers followed by dropout and two fully connected layers. It replaces the rule-based grade method, which uses event, split, and particle thresholds plus fixed pixel-pattern windows. The CNN's work is to learn the difference between the compact, radially spread charge cloud of an X-ray and the track-like or multi-pixel-heavy pattern of a charged particle, using both spatial correlations and relative signal amplitudes across pixels; that learned morphology is what lets it outperform the fixed windows, especially for compac

Load-bearing premise

Events are labeled by irradiation source: everything in the Fe-55 frames is called an X-ray and everything in the Sr-90 frames is called a charged particle, so if either dataset contains events of the other kind, the reported misclassification and acceptance rates are biased.

What would settle it

Use a beam-tagging or fast-coincidence readout that records each event's true source, X-ray line or electron, independently of the irradiation frame, then recompute the CNN's misclassification and acceptance; if the CNN's rates move toward the grade-method rates when labels are corrected, the claimed fourfold improvement is an artifact of source-based labeling.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • In-orbit charged-particle contamination would drop by about a factor of four relative to grade methods while X-ray acceptance stays above 98%.
  • The CNN's advantage is strongest above about 4 keV, so it extends the usable spectral range of the pnCCD beyond the nominal 0.4-4 keV trigger band.
  • Because the network is compact and runs only on threshold-crossing candidates, real-time onboard FPGA implementation is plausible, though processing speed and power remain unmeasured.
  • The classification gain depends on the detector's thick depletion layer; fully depleted sensors with a large sensitive volume will benefit most from this approach.
  • The method still needs validation below 1.4 keV and against proton irradiation before it can be adopted for the mission band.

Where Pith is reading between the lines

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

  • The source-based labeling is the main caveat: if Fe-55 frames contain any non-X-ray background or Sr-90 frames contain X-ray-like compact deposits, both the 3.1% and 98.1% figures shift; the paper says some label ambiguity is unavoidable.
  • Because training used mainly about-6-keV Mn K lines, the classifier's behavior at 0.4-1.4 keV is an open question; a natural test is to retrain on low-energy line sources and measure acceptance as a function of energy.
  • The CMOS comparison is cross-experiment, so a controlled head-to-head with identical source geometry, readout noise, and CNN training would be needed to attribute the difference to depletion depth rather than to other detector characteristics.
  • The pile-up check suggests the CNN accepts many candidate pile-up events; a controlled coincidence experiment with two X-rays or an X-ray plus an electron would show whether pile-up inflates the measured X-ray acceptance.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper reports the development of an FPGA-based readout for a prototype pnCCD and evaluates charged-particle rejection algorithms for the EAGLE wide-field X-ray monitor on HiZ-GUNDAM, whose science band is 0.4–4 keV. Using Fe-55 X-ray and Sr-90 beta-source irradiations, the authors compare conventional 3x3 and 5x5 grade methods with a compact CNN classifier. They report that over the full measured energy range (1.4–23 keV) the CNN reduces the beta misclassification rate from 11.9%/10.2% to 3.1%, while maintaining 98.1% X-ray acceptance compared with 97.8%/91.9% for the grade methods. The paper candidly states that the measured in-band portion (1.4–4 keV) shows comparable performance (0.7% CNN vs 0.4–0.5% grade) and that the main CNN advantage appears above 4 keV. Several limitations are acknowledged: evaluation is restricted to energies above ~1.4 keV, training X-rays are near 6 keV, only electrons are used as charged particles, and source-based labeling involves unavoidable ambiguity.

Significance. If the measured performance is robust, the CNN approach is a promising candidate for onboard event selection in a pnCCD-based X-ray monitor. The paper's strengths include the use of independent training and validation frames, five-fold cross-validation, Monte Carlo style Poisson uncertainties, and a candid discussion of experimental limitations. The comparison with a thin-depletion CMOS sensor is informative even if source conditions differ. However, the headline claim of a four-fold reduction in misclassification is computed over 1.4–23 keV, while the mission's science band is 0.4–4 keV. In the only measured sub-band inside the mission band (1.4–4 keV), the CNN is not better than the grade methods; if anything, it is slightly worse. Thus the mission-relevant benefit of the CNN is not established by the present data, and the abstract overstates the case. The work is still a useful engineering study, but the central claim needs substantial reframing and additional low-energy validation.

major comments (3)
  1. [§4.2, Table 4; Abstract] The headline comparison (3.1% vs 10.2–11.9%) is over 1.4–23 keV. Table 4 also gives the 1.4–4.0 keV misclassification rates: 0.4±0.1% (3x3 grade), 0.5±0.2% (5x5 grade), and 0.7±0.2% (CNN). The CNN is therefore not better than the grade methods in the measured portion of EAGLE's 0.4–4 keV band. The Abstract's statement that the CNN 'substantially reduces charged-particle contamination' is not justified for the mission band. The conclusion in §6.1 does acknowledge comparable low-energy performance, but the abstract and the introduction frame the improvement as broadly applicable. Separate the full-range and in-band results clearly in the abstract and avoid claiming mission-relevant improvement without this qualification.
  2. [§5.4] The evaluation is limited to deposited energies above ~1.4 keV, and the CNN was trained on Fe-55 X-rays near 6 keV. EAGLE's science band is 0.4–4 keV, so the entire 0.4–1.4 keV interval is untested. As the paper itself notes, lower-energy events produce simpler charge distributions and potentially reduce the CNN's advantage. Because the mission's primary science band is exactly the energy range in which the CNN has not been shown to help, the central mission-relevance claim is unsupported by the present measurements. This is not a minor caveat: the abstract's 'promising approach for onboard event selection in future pnCCD-based wide-field X-ray missions' depends on performance in the unmeasured low-energy band.
  3. [§5.3, §4.1] Event labels are assigned by irradiation source: all Fe-55 events are treated as X-rays and all Sr-90 events as charged particles after simple selection cuts. If the Fe-55 dataset contains non-X-ray background or the Sr-90 dataset contains compact X-ray-like deposits, both the 3.1% misclassification rate and the 98.1% acceptance rate are biased. The paper acknowledges this ambiguity, and the use of separate validation frames reduces circularity, but the reported uncertainties (Poisson only) do not include this systematic. The numerical values should be presented as conditional on the source-based labeling assumption, and a mixed-event or simulated-background validation is needed before the in-orbit performance can be trusted.
minor comments (5)
  1. [Table 4] The column header formatting is garbled: 'Misclassification rate X-ray acceptance rate1.4 − 23 keV 1.4 − 4.0 keV' is difficult to parse. Please restructure the table so that the energy range for each metric is explicitly clear.
  2. [§5.2] The CMOS comparison uses a different beta source (Bi-210, max 1.16 MeV) and a different energy range (0.4–7.5 keV) than the pnCCD measurement. This is noted in the text, but the table should carry a footnote so the comparison is not read as a direct apples-to-apples measurement.
  3. [§4.1] The particle threshold is set to the 14-bit ADC maximum (16383 ADU), which means no real rejection is performed for most sub-MeV electrons. The text explains that this is intended to catch muons, but it should be stated more explicitly that the grade methods effectively operate without a particle threshold for the beta source used here.
  4. [§5.4] The paper repeatedly says the CNN advantage is 'particularly pronounced at higher deposited energies' but does not quantify the threshold above which the advantage becomes statistically significant. Consider adding a cumulative or energy-binned comparison with confidence bounds.
  5. [§6.2] The phrase 'charge-transferefficiency' in the future-work paragraph is missing a space. Minor typographical issue.

Circularity Check

0 steps flagged

No significant circularity; the central result is a measured performance comparison on held-out data, not a derivation from fitted inputs.

full rationale

The central claim—CNN misclassification of beta-like events at 3.1% versus 10.2–11.9% for grade methods—is an experimental measurement, not a derivation. The CNN was trained on an 80% subset of labeled events and evaluated on a held-out 20% test subset plus an independently acquired 100-frame validation dataset for each source. Misclassification and acceptance rates are computed directly from counts of events that pass each classifier; no equation reduces one metric to another, and no fitted parameter is renamed as a prediction. The only self-citation, Ref. [11], is used to introduce the machine-learning approach and to provide a comparative CMOS measurement; it does not supply the pnCCD results, which are measured in this work. The paper also explicitly acknowledges limitations—label ambiguity due to source-based labeling, the 1.4 keV lower energy bound, the energy-band mismatch with the mission band, and the untested 0.4–1.4 keV region. These limit generalization and interpretation of the mission-relevant benefit, but they are not circularity: they do not make the reported 3.1% figure equivalent to the training labels or to any prior self-cited result by construction. Accordingly, there is no circular step to flag.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The headline numbers depend on user-set thresholds and on source-based labels. The CNN itself is a fitted model, but that is the contribution, not an unstated input. No new physical entities are introduced.

free parameters (5)
  • Event threshold = 5 sigma / 1000 ADU (~1.4 keV)
    Chosen to reject noise; sets the lower energy bound of all metrics.
  • Split threshold (grade method) = 3 sigma / 600 ADU (~0.8 keV)
    Used to include neighboring pixels in grade classification; varied from 400-800 ADU in the trade-off curve.
  • Particle threshold (grade method) = 16383 ADU (14-bit maximum)
    Rejects muons; effectively excludes only saturation-level events.
  • CNN decision threshold = not stated for headline result (likely 0.5 softmax)
    Figure 9 shows misclassification rate varies with this threshold, so the 3.1% number is operating-point dependent.
  • CNN trained weights and hyperparameters = 9,411 parameters; 2 conv layers, dropout, 2 FC; Adam, batch 32, 250 epochs
    These are fitted to the training data; the architecture was selected by cross-validation.
axioms (4)
  • domain assumption Events in Fe-55 frames are X-rays and events in Sr-90 frames are charged particles
    Used as ground-truth labels for training and evaluation; acknowledged in Sec 5.3 that label ambiguity is unavoidable.
  • domain assumption Pixel noise is sufficiently Gaussian that 5-sigma and 3-sigma thresholds provide clean event selection
    Thresholds chosen from dark-frame RMS; used to define candidates and split pixels (Sec 4.1).
  • ad hoc to paper Fe-55 X-rays near 6 keV are representative of X-ray charge morphologies in the 0.4-4 keV mission band
    The classifier was trained only on ~6 keV photons; the paper states applicability to lower energies is not established (Sec 5.4).
  • domain assumption The AE8/SPENVIS model describes the orbital electron environment for HiZ-GUNDAM
    Used to argue the Sr-90 electron energy range overlaps with the trapped electron population (Sec 5.4).

pith-pipeline@v1.3.0-alltime-deepseek · 12921 in / 12584 out tokens · 117695 ms · 2026-08-05T00:51:45.069181+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Development and Evaluation of a CNN-Based Charged-Particle Event Rejection Algorithm for Soft X-ray Detection in a pnCCD-Based Satellite System." pith.science (2026). https://pith.science/paper/EBRPFTE6

@misc{pith2026260800476,
  author       = {Pith},
  title        = {Pith review of: Development and Evaluation of a CNN-Based Charged-Particle Event Rejection Algorithm for Soft X-ray Detection in a pnCCD-Based Satellite System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EBRPFTE6}},
  note         = {Machine review of arXiv:2608.00476}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

All-sky surveys in the soft X-ray band are essential for detecting transient objects such as high-redshift gamma-ray bursts (GRBs), which provide key insights into the early universe. HiZ-GUNDAM is a future satellite mission designed to detect and localize high-redshift GRBs. Its wide-field X-ray monitor, EAGLE, combines Lobster Eye Optics with a pnCCD imaging detector operating in the 0.4-4 keV band. Because of limited satellite telemetry, full-frame pnCCD images cannot be downlinked, requiring onboard event selection. Charged particles in the space environment produce background events that can be misidentified as X-ray photons, degrading detection sensitivity and potentially triggering false alerts. In this study, we developed a pnCCD readout system and evaluated charged-particle rejection using conventional grade methods and a convolutional neural network (CNN). Performance was evaluated using X-ray events from an Fe-55 source and electron events from a Sr-90 beta source. The CNN reduced the misclassification rate from 10.2-11.9% for conventional grade methods to 3.1% while maintaining a high acceptance rate for X-ray events. The improvement is particularly pronounced at higher deposited energies, reflecting the CNN's ability to distinguish track-like particle events from X-ray events by capturing detailed spatial features of charge distributions. Comparison with a thin-depletion-layer CMOS sensor further indicates that the thicker depletion layer of the pnCCD enhances discrimination performance. These results demonstrate that CNN-based event classification can substantially reduce charged-particle contamination while maintaining high X-ray acceptance, making it a promising approach for onboard event selection in future pnCCD-based wide-field X-ray missions.

Figures

Figures reproduced from arXiv: 2608.00476 by Akihiro Doi, Daisuke Yonetoku, Hiro Otsuka, Hiroshi Nakajima, Hiroshi Tomida, Hsien-chieh Shen, Junko Hiraga, Lothar Str\"uder, Makoto Arimoto, Robert Hartmann, Ryuji Kondo, Shutaro Ueda, Takaaki Tanaka, Takanori Sakamoto, Tatsuro Kanenaga, Tatsuya Sawano.

Figure 1
Figure 1. Figure 1: Photograph of the small pnCCD used in this study. The pnCCD sensor is mounted at the center of the chip, with two CAMEXs mounted at each end [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Block diagram of the developed pnCCD readout system [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Grade method. Values in the pixels show the analog￾to-digital units (ADU). The event pixel exceeds the event threshold, while surrounding pixels exceeding the split thresh￾old are included for energy reconstruction. Representative pixel patterns used to classify X-ray and non–X-ray events are shown in the lower panels. Shen et al.: Preprint submitted to Elsevier Page 9 of 9 [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 6
Figure 6. Figure 6: Representative X-ray (left) and 𝛽-ray (right) pnCCD images. X-ray events appear as compact, dot-like patterns, whereas 𝛽-ray events form extended, track-like structures. Insets show magnified examples. Pixel values are expressed in analog-to-digital units (ADU) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Workflow of the CNN-based X-ray event selection method. Candidate center pixels are first identified using a threshold. A 5 × 5 pixel patch around each candidate is extracted and classified by a CNN event classifier. Events identified as X-rays are recorded, while charged-particle events are rejected. Values in the pixels show the analog-to-digital units (ADU) [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The experimental setup. The pnCCD is mounted on a socket and connected to FPGA and ADC boards, with power supplies. The system was placed in a thermostatic chamber. The radioactive sources were positioned at a distance of 4 mm from the pnCCD [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 9
Figure 9. Figure 9: Trade-off between the X-ray acceptance rate and the 𝛽-ray misclassification rate for the 3 × 3 grade method, the 5 × 5 grade method, and the CNN method. For the grade methods, the split threshold was varied from 400 to 800 ADU. For the CNN method, the decision threshold was varied using the softmax probability of the X-ray class from 0.1 to 0.9 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Representative 𝛽-ray events misclassified as X-ray events by (a) the 3 × 3 grade method, (b) the 5 × 5 grade method, and (c) the CNN method. The orange boxes indicate the pixel regions used for discrimination. The grade methods misidentify extended or track-like charge distributions, whereas the CNN method suppresses most such events, with only a few compact or edge events remaining misclassified. Shen et… view at source ↗
Figure 10
Figure 10. Figure 10: Energy spectra of 𝛽-ray events misclassified as X￾ray events (false X-ray events) obtained with the 3 × 3 grade method, the 5 × 5 grade method, and the CNN method on pnCCD image data irradiated with 𝛽-rays. Error bars represent statistical uncertainties assuming Poisson statistics, given by the square root of the counts in each bin [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗

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

15 extracted references · 15 canonical work pages

  1. [1]

    D.Yonetokuetal.Conceptofhigh-zgamma-rayburstsunravelingthe dark ages and extreme space-time mission—hiz-gundam.Journal of Astronomical Telescopes, Instruments, and Systems, 11(4):044002, 2025

  2. [2]

    Arimoto et al

    M. Arimoto et al. Conceptual study of the eagle wide-field x-ray monitor onboard hiz-gundam.Journal of Astronomical Telescopes, Instruments, and Systems, 12(2):024006, 2026

  3. [3]

    AstrophysicalJournal, 233:364–373, 1979

    J.R.P.Angel.Lobstereyesasx-raytelescopes. AstrophysicalJournal, 233:364–373, 1979

  4. [4]

    Strüder, P

    L. Strüder, P. Holl, G. Lutz, and J. Kemmer. Development of fully depletable CCDs for high energy physics applications. Nuclear Instruments and Methods in Physics Research A, 257(3):594–602, July 1987

  5. [5]

    Kissel, B

    H.Nakajima,H.Yamaguchi,H.Matsumoto,T.G.Tsuru,K.Koyama, S. Kissel, B. LaMarr, and M. Bautz. The ground calibration of X- ray CCD cameras (XIS) with front-illuminated chips onboard Astro- E2. NuclearInstruments andMethodsinPhysicsResearchA,541(1- 2):365–371, April 2005

  6. [6]

    Meidinger et al

    N. Meidinger et al. Development of the focal plane PNCCD camera system for the X-ray space telescope eROSITA.NuclearInstruments andMethodsinPhysicsResearchA, 624:321–329, (Dec. 2010)

  7. [7]

    Mercier et al

    K. Mercier et al. MXT instrument on-board the French-Chinese SVOM mission. In Space Telescopes and Instrumentation 2018: Ultraviolet to Gamma Ray, volume 10699 of Proc. SPIE, page 1069921, 2018

  8. [8]

    Camex readout asics for pnccds

    Sven Herrmann, Werner Buttler, Robert Hartmann, Norbert Mei- dinger, Matteo Porro, and Lothar Strueder. Camex readout asics for pnccds. In2008 IEEE Nuclear Science Symposium Conference Record, pages 2952–2957, 2008

  9. [9]

    InAstronomicalTelescopes+ Instrumentation, Proc

    H.C.Shenetal.Thestatusofpnccdwithanfpga-basedelectronicsys- tem for hiz-gundam. InAstronomicalTelescopes+ Instrumentation, Proc. SPIE, 2024. Shen et al.:Preprint submitted to ElsevierPage 8 of 9 Development and Evaluation of a CNN-Based Charged-Particle Event Rejection Algorithm for Soft X-ray Detection in a pnCCD-Based Satellite System

  10. [10]

    Kondo et al

    R. Kondo et al. Design and development of an fpga-based pnccd driver and readout system for future satellite mission hiz-gundam. In AstronomicalTelescopes+Instrumentation, Proc. SPIE, 2024

  11. [11]

    H. C. Shen et al. Application of the grade selection of X-ray events using machine learning for a CubeSat mission.Journal of Instrumentation, 18:C12012, (Dec. 2023)

  12. [12]

    TheX-RayAstronomySatelliteASCA

    Y.Tanakaetal. TheX-RayAstronomySatelliteASCA. Publications ofthe AstronomicalSocietyof Japan, 46:L37–L41, (Jun. 1994)

  13. [13]

    TheX-RayObservatorySuzaku

    K.Mitsudaetal. TheX-RayObservatorySuzaku. Publicationsofthe AstronomicalSocietyof Japan, 59:S1–S7, (Jan. 2007)

  14. [14]

    Ogino et al

    N. Ogino et al. Performance verification of next-generation si cmos soft x-ray detector for space applications.Nuclear Instruments and MethodsinPhysicsResearchSectionA:Accelerators,Spectrometers, DetectorsandAssociated Equipment, 987:164843, 2021

  15. [15]

    Heynderickx, B

    D. Heynderickx, B. Quaghebeur, E. Speelman, and E. Daly. Esa’s spaceenvironmentinformationsystem(spenvis):Awwwinterfaceto modelsofthespaceenvironmentanditseffects. In Proceedingsofthe 2000IEEE RadiationEffectsData Workshop, pages 48–50, 2000. Figure 1:Photograph of the small pnCCD used in this study. The pnCCD sensor is mounted at the center of the chip, ...