arxiv: 2604.09402 · v1 · submitted 2026-04-10 · 🌌 astro-ph.IM

Recognition: unknown

Enhancing event reconstruction for γ-ray particle detector arrays using transformers

Markus Pirke , Youngwan Son , Jonas Glombitza , Martin Schneider , Ian James Watson , Christopher van Eldik

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:24 UTC · model grok-4.3

classification 🌌 astro-ph.IM

keywords gamma-ray astronomyair-shower reconstructiontransformer networksgamma-hadron separationwater-Cherenkov detectorsdeep learningevent reconstructionparticle detector arrays

0 comments

The pith

Transformer models improve gamma-ray event reconstruction and hadron separation across the energy spectrum using raw detector data.

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

Ground-based gamma-ray observations rely on detecting sparse samples of air showers, which introduces large uncertainties in recovering the primary particle's energy, direction, and location. This paper tests two transformer networks of different sizes that ingest calibrated raw signals from a proposed water-Cherenkov array and directly output both reconstructed parameters and a gamma-hadron classification score. The networks are benchmarked against conventional reconstruction and separation techniques on simulated events spanning hundreds of GeV to PeV. Clear gains appear throughout the range, with the largest advances at low and intermediate energies, and the same model improves both tasks at once. If the gains transfer to real data, future arrays could reach lower energies and achieve higher sensitivity with existing hardware.

Core claim

Using end-to-end transformer architectures on calibrated raw data from simulated future water-Cherenkov arrays yields better angular, core, and energy reconstruction together with stronger gamma-hadron separation than established methods, with the largest improvements at low and intermediate energies. This is the first demonstration that a single architecture can advance both reconstruction quality and background rejection consistently.

What carries the argument

The transformer architecture that ingests sequences of calibrated detector signals and directly produces event parameters plus a classification score.

If this is right

Lower reconstruction errors at low energies extend the reliable energy threshold downward for source detection.
Stronger gamma-hadron separation increases signal purity and thereby improves detection significance for faint sources.
A single end-to-end model replaces separate algorithms for direction, energy, core, and classification tasks.
The method handles variable numbers of active detectors, making it suitable for scaling to larger arrays.

Where Pith is reading between the lines

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

The same sequence-processing approach could be tested on other sparse-sampling experiments such as cosmic-ray or neutrino arrays.
If the gains hold on real data, tighter spectral measurements become possible in the energy band where traditional methods lose precision.
Additional observables such as timing or multi-component signals could be added directly to the input sequence for further gains.

Load-bearing premise

Monte Carlo simulations of the detector response, shower fluctuations, and background faithfully represent real conditions so that models trained on them generalize to actual observations.

What would settle it

Running the trained models on real data from an existing water-Cherenkov array and checking whether the reported improvements in resolution and separation power appear relative to traditional methods.

read the original abstract

Gamma-ray astronomy from hundreds of GeV to PeV is confined to ground-based experiments that detect air showers induced by $\gamma$-rays entering Earth's atmosphere. While particle detector arrays feature huge detection areas, accurately reconstructing the primary particle properties is difficult due to the sparse sampling of the air shower and its intrinsic fluctuations. In this work, using simulations of a future water-Cherenkov array, we investigate two end-to-end deep learning approaches based on the transformer architecture with different computational complexities that utilize calibrated raw data. We benchmark both methods against well-established methods in the field in terms of $\gamma$-hadron separation, angular, core, and energy reconstruction. Our results show significant improvements across the whole energy range, particularly at low and intermediate energies. This work is the first to consistently demonstrate improved performance in both event reconstruction and $\gamma$-hadron separation using a single architecture.

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.

Desk Editor's Note private letter to a colleague

A single transformer beats standard methods on simulated gamma-ray array data for reconstruction and separation, but the gains sit entirely on Monte Carlo for a future detector with no real-data check.

read the letter

The main point is that the authors train two transformer variants on raw calibrated signals from a simulated water-Cherenkov array and show better angular resolution, core location, energy reconstruction, and gamma-hadron separation than the usual reconstruction pipelines, with the largest gains at lower and intermediate energies. They present this as the first consistent single-architecture demonstration for both tasks together.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes two transformer-based end-to-end deep learning models that process calibrated raw data from Monte Carlo simulations of a future water-Cherenkov gamma-ray detector array. The models are benchmarked against established reconstruction and gamma-hadron separation techniques for angular, core, and energy resolution across the energy range, with reported gains especially at low and intermediate energies. The work claims to be the first single-architecture demonstration of simultaneous improvements in both reconstruction and separation tasks.

Significance. If the reported gains prove robust, the unified transformer approach could meaningfully advance handling of sparse, fluctuating air-shower data in next-generation arrays. The end-to-end use of calibrated raw data and direct benchmarking against field-standard methods are positive elements. However, the complete reliance on simulations of a hypothetical array without real-data validation or cross-simulation checks substantially limits the immediate practical significance and generalizability of the claimed improvements.

major comments (2)

The central performance claims rest entirely on a single Monte Carlo simulation chain for a future array; the manuscript must demonstrate robustness by repeating the benchmarks under at least one alternative hadronic interaction model and one altered detector-response realization, as any mismatch with reality would invalidate the practical value of the reported gains in angular/core/energy reconstruction and gamma-hadron separation.
The assertion that this is 'the first to consistently demonstrate improved performance in both event reconstruction and γ-hadron separation using a single architecture' requires an explicit, section-by-section comparison in the introduction or discussion to prior single-model or multi-task deep-learning efforts on air-shower arrays; without this, the novelty claim cannot be evaluated.

minor comments (2)

The abstract states improvements were observed but supplies no numerical metrics, training hyperparameters, validation strategy, or error analysis; these details should be added to the abstract or a dedicated methods subsection for reproducibility.
Clarify in the results section how the held-out test set was constructed and whether any post-hoc energy or quality cuts were applied after model inference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We have carefully addressed each major point below, providing our responses and indicating the revisions made to the manuscript.

read point-by-point responses

Referee: The central performance claims rest entirely on a single Monte Carlo simulation chain for a future array; the manuscript must demonstrate robustness by repeating the benchmarks under at least one alternative hadronic interaction model and one altered detector-response realization, as any mismatch with reality would invalidate the practical value of the reported gains in angular/core/energy reconstruction and gamma-hadron separation.

Authors: We acknowledge the referee's concern regarding the reliance on a single simulation chain. Our benchmarks are based on the standard CORSIKA simulation with QGSJet-II-04 hadronic interactions and the nominal detector response for the proposed water-Cherenkov array. While we agree that additional robustness tests would be valuable, generating full alternative Monte Carlo datasets (e.g., using EPOS-LHC or modified detector calibrations) requires substantial computational resources that exceed the scope of this initial study. We have revised the manuscript to include an expanded discussion in Section 4 on the choice of hadronic model, citing relevant literature on model-induced variations in air-shower observables, and noting that the relative gains of the transformer models arise from their handling of sparse, fluctuating data rather than absolute normalization. This provides context for the results while highlighting the limitation. revision: partial
Referee: The assertion that this is 'the first to consistently demonstrate improved performance in both event reconstruction and γ-hadron separation using a single architecture' requires an explicit, section-by-section comparison in the introduction or discussion to prior single-model or multi-task deep-learning efforts on air-shower arrays; without this, the novelty claim cannot be evaluated.

Authors: We thank the referee for this suggestion to strengthen the novelty discussion. The original manuscript referenced several prior deep-learning studies on air-shower arrays (e.g., CNN-based approaches for HAWC and LHAASO data). To explicitly address the request, we have added a new subsection (Section 1.2) in the revised Introduction that provides a systematic, section-by-section comparison to previous single-model and multi-task DL efforts. This includes analysis of works using CNNs, RNNs, and graph networks for gamma-hadron separation or reconstruction, emphasizing the distinctions of our end-to-end transformer approach on calibrated raw data for simultaneous multi-task performance. revision: yes

Circularity Check

0 steps flagged

No circularity: performance claims rest on independent held-out simulation benchmarks against established methods.

full rationale

The paper trains transformer models on Monte Carlo simulations of a future water-Cherenkov array and evaluates reconstruction and separation metrics on held-out simulated events, comparing directly to standard reconstruction algorithms. No derivation chain reduces to self-definition, fitted inputs renamed as predictions, or load-bearing self-citations; the reported improvements are empirical test-set results, not tautological outputs of the training procedure itself. The 'first demonstration' claim is a comparative statement, not a mathematical reduction. The work is self-contained against its stated external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that detailed Monte Carlo simulations of air showers and detector response are sufficiently realistic to train and validate the models for future real-world use. No additional free parameters or invented entities are introduced beyond standard neural-network training.

axioms (1)

domain assumption Simulated air-shower and detector data are representative of real observations
All training, validation, and benchmarking are performed exclusively on simulations of a future water-Cherenkov array.

pith-pipeline@v0.9.0 · 5462 in / 1320 out tokens · 41438 ms · 2026-05-10T16:24:49.728878+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

46 extracted references · 31 canonical work pages · 7 internal anchors

[1]

Albert et al.,Science Case for a Wide Field-of-View Very-High-Energy Gamma-Ray Observatory in the Southern Hemisphere,1902.08429

A. Albert et al.,Science Case for a Wide Field-of-View Very-High-Energy Gamma-Ray Observatory in the Southern Hemisphere,1902.08429. [2]CTA Consortiumcollaboration,Introducing the CTA concept,Astropart. Phys.43(2013) 3

work page arXiv 1902
[2]

Abreu et al.,The Southern Wide-Field Gamma-Ray Observatory (SWGO): A Next-Generation Ground-Based Survey Instrument for VHE Gamma-Ray Astronomy, 1907.07737

P. Abreu et al.,The Southern Wide-Field Gamma-Ray Observatory (SWGO): A Next-Generation Ground-Based Survey Instrument for VHE Gamma-Ray Astronomy, 1907.07737. [4]Milagrocollaboration,Observation of TeV gamma-rays from the Crab nebula with MILAGRO using a new background rejection technique,Astrophys. J.595(2003) 803 [astro-ph/0305308]. [5]Milagrocollabora...

work page arXiv 1907
[3]

Parsons and J.A

R.D. Parsons and J.A. Hinton,A Monte Carlo Template based analysis for Air-Cherenkov Arrays,Astropart. Phys.56(2014) 26 [1403.2993]. – 17 –

work page arXiv 2014
[4]

Joshi et al.,A template-basedγ-ray reconstruction method for air shower arrays,JCAP 2019(2019) 012–012

V. Joshi et al.,A template-basedγ-ray reconstruction method for air shower arrays,JCAP 2019(2019) 012–012

2019
[5]

S. Ohm, C. van Eldik and K. Egberts,Gamma-Hadron Separation in Very-High-Energy gamma-ray astronomy using a multivariate analysis method,Astropart. Phys.31(2009) 383 [0904.1136]

work page arXiv 2009
[6]

Krause, E

M. Krause, E. Pueschel and G. Maier,Improvedγ/hadron separation for the detection of faint γ-ray sources using boosted decision trees,Astropart. Phys.89(2017) 1 [1701.06928]

work page arXiv 2017
[7]

Albert et al.,Implementation of the Random Forest Method for the Imaging Atmospheric Cherenkov Telescope MAGIC,Nucl

J. Albert et al.,Implementation of the Random Forest Method for the Imaging Atmospheric Cherenkov Telescope MAGIC,Nucl. Instrum. Meth. A588(2008) 424 [0709.3719]

work page arXiv 2008
[8]

LeCun, Y

Y. LeCun, Y. Bengio and G. Hinton,Deep learning,Nature521(2015) 436

2015
[9]

Erdmann, J

M. Erdmann, J. Glombitza, G. Kasieczka and U. Klemradt,Deep Learning for Physics Research, WORLD SCIENTIFIC (2, 2021), 10.1142/12294

work page doi:10.1142/12294 2021
[10]

Westerhoff, B

S. Westerhoff, B. Funk, N. Magnussen, H. Meyer, H. Moeller, W. Rhode et al.,Separating gamma and hadron induced cosmic ray air showers with feed forward neural networks using the charged particle information,Astropart. Phys.4(1995) 119

1995
[11]

Boinee, F

P. Boinee, F. Barbarino, A. De Angelis, A. Saggion and M. Zacchello,Neural networks for gamma-hadron separation in MAGIC, in6th International Symposium on Frontiers of Fundamental Physics, pp. 297–302, 3, 2005, DOI [astro-ph/0503539]. [19]HA WCcollaboration,Gamma/hadron separation with the HAWC observatory,Nucl. Instrum. Meth. A1039(2022) 166984 [2205.121...

work page arXiv 2005
[12]

Erdmann, J

M. Erdmann, J. Glombitza and D. Walz,A deep learning-based reconstruction of cosmic ray-induced air showers,Astropart. Phys.97(2018) 46 [1708.00647]. [22]Pierre Augercollaboration,Inference of the Mass Composition of Cosmic Rays with Energies from 1018.5 to 1020 eV Using the Pierre Auger Observatory and Deep Learning, Phys. Rev. Lett.134(2025) 021001 [240...

work page arXiv 2018
[13]

Shilon, M

I. Shilon, M. Kraus, M. Büchele, K. Egberts, T. Fischer, T.L. Holch et al.,Application of Deep Learning methods to analysis of Imaging Atmospheric Cherenkov Telescopes data,Astropart. Phys.105(2019) 44 [1803.10698]

work page arXiv 2019
[14]

Nieto, A

D. Nieto, A. Brill, Q. Feng, T.B. Humensky, B. Kim, T. Miener et al.,CTLearn: Deep Learning for Gamma-ray Astronomy,PoSICRC2019(2020) [1912.09877]

work page arXiv 2020
[15]

Brill, Q

A. Brill, Q. Feng, T.B. Humensky, B. Kim, D. Nieto and T. Miener,Investigating a Deep Learning Method to Analyze Images from Multiple Gamma-ray Telescopes, in2019 New York Scientific Data Summit: Data-Driven Discovery in Science and Industry, 6, 2019, DOI [2001.03602]

work page arXiv 2019
[16]

Jacquemont, T

M. Jacquemont, T. Vuillaume, A. Benoit, G. Maurin, P. Lambert and G. Lamanna,First Full-Event Reconstruction from Imaging Atmospheric Cherenkov Telescope Real Data with Deep Learning, inInternational Conference on Content-Based Multimedia Indexing, 5, 2021, DOI [2105.14927]

work page arXiv 2021
[17]

Spencer, T

S. Spencer, T. Armstrong, J. Watson, S. Mangano, Y. Renier and G. Cotter,Deep learning with photosensor timing information as a background rejection method for the Cherenkov Telescope Array,Astropart. Phys.129(2021) 102579 [2103.06054]. – 18 –

work page arXiv 2021
[18]

Glombitza, V

J. Glombitza, V. Joshi, B. Bruno and S. Funk,Application of graph networks to background rejection in Imaging Air Cherenkov Telescopes,JCAP11(2023) 008 [2305.08674]

work page arXiv 2023
[19]

Schwefer, R

G. Schwefer, R. Parsons and J. Hinton,A hybrid approach to event reconstruction for atmospheric Cherenkov Telescopes combining machine learning and likelihood fitting,Astropart. Phys.163(2024) 103008 [2406.17502]. [31]HA WCcollaboration,Deep Learning for the HAWC Observatory,PoSICRC2023(2023) 927

work page arXiv 2024
[20]

Glombitza, M

J. Glombitza, M. Schneider, F. Leitl, S. Funk and C. van Eldik,Application of graph networks to a wide-field water-Cherenkov-based Gamma-ray Observatory,JCAP02(2025) 066 [2411.16565]

work page arXiv 2025
[21]

Conceição, B.S

R. Conceição, B.S. González, A. Guillén, M. Pimenta and B. Tomé,Discriminating sub-TeV gamma and hadron-induced showers through their footprints,Phys. Rev. D111(2025) 043047 [2409.11093]

work page arXiv 2025
[22]

R. Conceição et al.,Muon identification in a compact single-layered water Cherenkov detector and gamma/hadron discrimination using machine learning techniques,EPJ C81(2021) 542 [2101.10109]

work page arXiv 2021
[23]

Attention Is All You Need

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez et al.,Attention Is All You Need, in31st International Conference on Neural Information Processing Systems, 6, 2017 [1706.03762]

work page internal anchor Pith review Pith/arXiv arXiv 2017
[24]

D. Heck, J. Knapp, J.N. Capdevielle, G. Schatz and T. Thouw,Corsika: A monte carlo code to simulate extensive air showers,FZKA6019(1998)

1998
[25]

Microscopic Models for Ultrarelativistic Heavy Ion Collisions

S.A. Bass et al.,Microscopic models for ultrarelativistic heavy ion collisions,Prog. Part. Nucl. Phys.41(1998) 255 [nucl-th/9803035]

work page Pith review arXiv 1998
[26]

Ostapchenko,QGSJET-II: towards reliable description of very high energy hadronic interactions,Nuclear Physics B - Proceedings Supplements151(2006) 143

S. Ostapchenko,QGSJET-II: towards reliable description of very high energy hadronic interactions,Nuclear Physics B - Proceedings Supplements151(2006) 143

2006
[28]

A. U. Abeysekara (HAWC Collaboration) et al.,The High-Altitude Water Cherenkov (HAWC) observatory in México: The primary detector,Nucl. Instrum. Methods Phys. Res. A1052 (2023) 168253

2023
[29]

Watson,Event Reconstruction Performance for SWGO using Attention-based Neural Network,PoSICRC2025(2025) 878

I. Watson,Event Reconstruction Performance for SWGO using Attention-based Neural Network,PoSICRC2025(2025) 878. [42]Swgocollaboration,Deep Learning Methods for Gamma/Hadron Separation in SWGO,PoS ICRC2025(2025) 836

2025
[30]

Krizhevsky, I

A. Krizhevsky, I. Sutskever and G.E. Hinton,ImageNet classification with deep convolutional neural networks,Commun. ACM60(2017) 84

2017
[31]

et al.,Comparison of vision transformers and convolutional neural networks in medical image analysis: A systematic review,Journal of Medical Systems48(2024) 84

S.T. et al.,Comparison of vision transformers and convolutional neural networks in medical image analysis: A systematic review,Journal of Medical Systems48(2024) 84

2024
[32]

Kudo and J

T. Kudo and J. Richardson,Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing, 2018

2018
[33]

Jumper et al.,Highly accurate protein structure prediction with alphafold,nature596(2021) 583

J. Jumper et al.,Highly accurate protein structure prediction with alphafold,nature596(2021) 583

2021
[34]

Neural Machine Translation by Jointly Learning to Align and Translate

B. Dzmitry, C. Kyunghyun and B. Yoshua,Neural machine translation by jointly learning to align and translate,arXiv preprint arXiv:1409.0473(2014)

work page internal anchor Pith review arXiv 2014
[35]

Y. Tay, M. Dehghani, D. Bahri and D. Metzler,Efficient transformers: A survey, 2022

2022
[36]

Perceiver: General perception with iterative attention

A. Jaegle et al.,Perceiver: General perception with iterative attention,2103.03206. – 19 –

work page arXiv
[37]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

A. Dosovitskiy et al.,An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,2010.11929

work page internal anchor Pith review Pith/arXiv arXiv 2010
[38]

Vision Transformers Need Registers

T. Darcet, M. Oquab, J. Mairal and P. Bojanowski,Vision transformers need registers, 2309.16588

work page internal anchor Pith review arXiv
[39]

Rosenblatt,The perceptron: a probabilistic model for information storage and organization in the brain.,Psychological review65(1958) 386

F. Rosenblatt,The perceptron: a probabilistic model for information storage and organization in the brain.,Psychological review65(1958) 386

1958
[40]

K. He, X. Zhang, S. Ren and J. Sun,Deep Residual Learning for Image Recognition, 1512.03385

work page internal anchor Pith review arXiv
[41]

J.L. Ba, J.R. Kiros and G.E. Hinton,Layer Normalization,1607.06450. [55]HA WCcollaboration,Deep Learning for the HAWC Observatory,PoSICRC2023(2023) 927

work page internal anchor Pith review Pith/arXiv arXiv 2023
[42]

Rasley,Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters, KDD ’20, p

J.e.a. Rasley,Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters, KDD ’20, p. 3505–3506, 2020, DOI

2020
[43]

Abeysekara et al.,Observation of the Crab Nebula with the HAWC Gamma-Ray Observatory,Astrophys

A.U. Abeysekara et al.,Observation of the Crab Nebula with the HAWC Gamma-Ray Observatory,Astrophys. J.843(2017) 39 [1701.01778]. [58]Swgocollaboration,Status of the SWGO air shower reconstruction using a template-based likelihood method,PoSICRC2023(2023) 593

work page arXiv 2017
[44]

Conceição, L

R. Conceição, L. Gibilisco, M. Pimenta and B. Tomé,Gamma/hadron discrimination at high energies through the azimuthal fluctuations of air shower particle distributions at the ground, JCAP10(2022) 086 [2204.12337]. [60]Swgocollaboration,Deep Learning Methods for Gamma/Hadron Separation in SWGO,PoS ICRC2025(2025) 836. [61]Swgocollaboration,Status of the SWG...

work page arXiv 2022
[45]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter,Decoupled Weight Decay Regularization, 11, 2017 [1711.05101]

work page internal anchor Pith review Pith/arXiv arXiv 2017
[46]

Bukhari, D

H. Bukhari, D. Chakraborty, P. Eller, T. Ito, M.V. Shugaev and R. Ørsøe,IceCube – Neutrinos in Deep Ice: The top 3 solutions from the public Kaggle competition,Eur. Phys. J. C84(2024) 646 [2310.15674]

work page arXiv 2024
[47]

Cipolla, Y

R. Cipolla, Y. Gal and A. Kendall,Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7482–7491, 2018, DOI. – 20 – A Training details A.1 Point Cloud T ransformer We applied the same input normalization procedure as described in [32] for our Poin...

2018