What exactly did the Transformer learn from our physics data?

Dominik Wirtz; Josina Schulte; Martin Erdmann; Niklas Langner

arxiv: 2505.21042 · v1 · submitted 2025-05-27 · 🌌 astro-ph.IM · hep-ex

What exactly did the Transformer learn from our physics data?

Martin Erdmann , Niklas Langner , Josina Schulte , Dominik Wirtz This is my paper

Pith reviewed 2026-05-19 13:27 UTC · model grok-4.3

classification 🌌 astro-ph.IM hep-ex

keywords transformercosmic rayspositional encodingattention mechanismultra-high-energyair showerinterpretabilitygalaxy catalog

0 comments

The pith

Transformers applied to cosmic ray simulations learn physically meaningful features like symmetry and source associations.

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

The paper examines what Transformer networks actually extract when trained on ultra-high-energy cosmic ray data. It tests this in two concrete settings: air-shower simulations that are azimuthally symmetric and events drawn from a galaxy catalog. By inspecting the trained positional encodings and the attention weights, the authors show that the networks recover physically sensible patterns without being explicitly told to look for them. A sympathetic reader would care because these results indicate that the models are not merely fitting statistical correlations but are picking up real structure in the physics data, which could make their outputs more trustworthy for scientific use.

Core claim

In ultra-high-energy cosmic ray simulations, Transformer networks learn plausible, physically meaningful features. Trained positional encodings in azimuthally symmetric air showers respect rotational symmetry around the shower axis. Attention values assigned to cosmic particles from a galaxy catalog highlight plausible source associations.

What carries the argument

Visualization of trained positional encodings and attention values, which reveal that the network has internalized azimuthal symmetry and source catalog information.

If this is right

The networks can be used for physics analyses where explicit symmetry enforcement is not required.
Attention maps may serve as a diagnostic for identifying which input particles carry the most information about origin.
Similar inspection methods could be applied to other symmetry-rich simulation datasets in high-energy physics.
The findings support using Transformers for tasks that combine detector data with astrophysical catalogs.

Where Pith is reading between the lines

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

If the visualized features prove robust, they could guide the design of lighter models that hard-code only the symmetries the network has already discovered.
This interpretability approach might transfer to other domains where data have rotational or catalog-based structure, such as particle collider events or gravitational wave signals.
Quantitative metrics comparing learned encodings to analytic symmetry transformations would strengthen the claim that the features are genuinely physical.

Load-bearing premise

Visualizations of positional encodings and attention maps accurately reflect the model's learned physical understanding without needing extra quantitative checks.

What would settle it

A controlled test that measures how much performance drops when the learned positional encodings or attention patterns are deliberately scrambled or replaced with random equivalents.

read the original abstract

Transformer networks excel in scientific applications. We explore two scenarios in ultra-high-energy cosmic ray simulations to examine what these network architectures learn. First, we investigate the trained positional encodings in air showers which are azimuthally symmetric. Second, we visualize the attention values assigned to cosmic particles originating from a galaxy catalog. In both cases, the Transformers learn plausible, physically meaningful features.

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

The paper shows visualizations of positional encodings and attention in Transformers on cosmic ray simulations that match physical expectations, but the support is only qualitative.

read the letter

The main takeaway is that the authors trained Transformers on ultra-high-energy cosmic ray simulations and then looked at the positional encodings for azimuthally symmetric air showers and the attention maps for particles drawn from a galaxy catalog. In both cases the patterns that emerge line up with what a physicist would expect, such as symmetry around the shower axis or focus on plausible source directions. That is the core observation they report. What is new here is the specific choice of these two scenarios in the cosmic ray domain and the direct application of standard interpretability tools to them. Prior work on Transformers in physics has mostly stayed at the level of performance numbers, so this targeted look at what the model actually encodes is a modest but concrete step. The visualizations themselves are presented clearly and the abstract does not overstate the result. The soft spot is the lack of any quantitative backing. There are no reported correlations between the visualized features and physical observables, no ablation studies with random weights or non-physics inputs, and no test showing that altering the highlighted encodings or attention values changes the model output in a controlled way. Without those checks it remains possible that similar-looking patterns would appear even in models that have not learned the underlying physics. This kind of paper is mainly useful to researchers already working at the intersection of machine learning and cosmic-ray or particle-physics simulations who want to see an interpretability example in their domain. A reader who is comfortable with both air-shower modeling and basic Transformer mechanics will get the most from it. I would send it to peer review. The idea is reasonable and the execution appears careful on the evidence given, even though additional validation would make the central claim more convincing.

Referee Report

2 major / 2 minor

Summary. The paper examines what Transformer networks learn from physics data in two ultra-high-energy cosmic ray simulation scenarios. It visualizes trained positional encodings for azimuthally symmetric air showers and attention values assigned to particles from a galaxy catalog, claiming that the models acquire plausible, physically meaningful features in both cases.

Significance. If the visualizations are shown to reflect genuine learned representations rather than post-hoc interpretations, the work could advance interpretability studies of Transformers in astrophysical applications. At present the evidence is purely qualitative, so the significance remains exploratory and does not yet establish causal links between visualized patterns and model performance or physical understanding.

major comments (2)

Abstract: the central claim that the Transformers 'learn plausible, physically meaningful features' rests entirely on qualitative visualizations. No quantitative metrics (e.g., correlation coefficients with energy, direction, or shower symmetry), ablation studies, or control models (random weights, non-physics data) are reported, leaving open whether the observed patterns are non-accidental or tied to task performance.
Visualization sections: the absence of baselines (untrained networks or shuffled inputs) and perturbation tests means it is not demonstrated that altering the visualized positional encodings or attention values changes the model's outputs in a physically interpretable way.

minor comments (2)

Add a dedicated methods subsection detailing architecture, training procedure, dataset sizes, and hyper-parameters to allow reproducibility.
Ensure figure captions explicitly link visualized patterns to specific physical symmetries or observables rather than leaving interpretation to the reader.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript exploring what Transformer networks learn from ultra-high-energy cosmic ray simulation data. We address the major comments point by point below.

read point-by-point responses

Referee: Abstract: the central claim that the Transformers 'learn plausible, physically meaningful features' rests entirely on qualitative visualizations. No quantitative metrics (e.g., correlation coefficients with energy, direction, or shower symmetry), ablation studies, or control models (random weights, non-physics data) are reported, leaving open whether the observed patterns are non-accidental or tied to task performance.

Authors: We agree that the interpretations are based on qualitative visualizations of the positional encodings and attention weights. This approach is common in initial interpretability studies to identify potentially meaningful patterns before developing quantitative measures. The observed features, such as symmetry in encodings for air showers and attention to galaxy-origin particles, align closely with physical expectations from the simulation setup. To address this, we will revise the abstract to better reflect the qualitative and exploratory nature of the claims, and we will consider adding simple quantitative correlations where feasible in the revised manuscript. revision: partial
Referee: Visualization sections: the absence of baselines (untrained networks or shuffled inputs) and perturbation tests means it is not demonstrated that altering the visualized positional encodings or attention values changes the model's outputs in a physically interpretable way.

Authors: The referee raises a valid point regarding the need for baselines and perturbation tests to establish that the visualized features are indeed learned and impactful. Our study primarily presents the trained model's internal representations to provide insight into what the network captures from the physics data. We will incorporate comparisons with randomly initialized (untrained) networks in the revised version to demonstrate that the physically plausible patterns arise from training on the cosmic ray data rather than being artifacts of the architecture. revision: yes

Circularity Check

0 steps flagged

No derivation chain or load-bearing reductions present

full rationale

The paper contains no equations, derivations, fitted parameters, or predictive claims that reduce to inputs by construction. Its central claim rests entirely on qualitative visualizations of trained positional encodings and attention maps from empirical training on cosmic-ray simulation data. These observations are direct empirical outputs rather than self-definitional, self-cited, or renamed results. No uniqueness theorems, ansatzes, or self-citation chains are invoked as load-bearing justification. The analysis is therefore self-contained and scores at the lowest end of the scale.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility into assumptions; no explicit free parameters or invented entities are stated.

axioms (1)

domain assumption Transformer models trained on simulation data will produce interpretable positional encodings and attention maps that correspond to physical symmetries and source properties.
Central to the two scenarios explored in the abstract.

pith-pipeline@v0.9.0 · 5581 in / 990 out tokens · 24573 ms · 2026-05-19T13:27:22.667760+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Dissecting Jet-Tagger Through Mechanistic Interpretability
hep-ph 2026-05 accept novelty 8.0

A Particle Transformer jet tagger contains a sparse six-head circuit whose source-relay-readout structure recovers most performance and whose residual stream preferentially encodes 2-prong energy correlators.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · cited by 1 Pith paper · 9 internal anchors

[1]

Vaswani, A., Shazeer, N., Parmar, N., Uszko- reit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention Is All You Need (2017) arXiv:1706.03762 [cs.CL] 7

work page internal anchor Pith review Pith/arXiv arXiv 2017
[2]

Dosovitskiy, A., et al.: An Image is Worth 16x16 Words: Transformers for Image Recog- nition at Scale (2020) arXiv:2010.11929 [cs.CV]

work page internal anchor Pith review Pith/arXiv arXiv 2020
[3]

https://openai.com/ blog/chatgpt/

OpenAI: ChatGPT: Optimizing Language Models for Dialogue. https://openai.com/ blog/chatgpt/. Online; Zugriff am 30. Novem- ber 2022 (2022)

work page 2022
[4]

Qu, H., Li, C., Qian, S.: Particle Transformer for Jet Tagging (2022) arXiv:2202.03772 [hep-ph]

work page arXiv 2022
[5]

JHEP 06, 184 (2023) https://doi.org/10.1007/JHEP06(2023)184 arXiv:2303.07364 [hep-ph]

Finke, T., Kr¨ amer, M., M¨ uck, A., T¨ onshoff, J.: Learning the language of QCD jets with transformers. JHEP 06, 184 (2023) https://doi.org/10.1007/JHEP06(2023)184 arXiv:2303.07364 [hep-ph]

work page doi:10.1007/jhep06(2023)184 2023
[6]

PoS ICRC2023, 371 (2023) https://doi.org/10.22323/1.444.0371

Abdul Halim, A., et al.: Deep-Learning-Based Cosmic-Ray Mass Reconstruction Using the Water-Cherenkov and Scintillation Detectors of AugerPrime. PoS ICRC2023, 371 (2023) https://doi.org/10.22323/1.444.0371

work page doi:10.22323/1.444.0371 2023
[7]

Wu, Y., Wang, K., Li, C., Qu, H., Zhu, J.: Jet tagging with more-interaction particle trans- former. Chin. Phys. C 49(1), 013110 (2025) https://doi.org/10.1088/1674-1137/ad7f3d arXiv:2407.08682 [hep-ph]

work page doi:10.1088/1674-1137/ad7f3d 2025
[8]

Brehmer, J., Bres´ o, V., Haan, P., Plehn, T., Qu, H., Spinner, J., Thaler, J.: A Lorentz- Equivariant Transformer for All of the LHC (2024) arXiv:2411.00446 [hep-ph]

work page arXiv 2024
[9]

Feickert, M., Nachman, B.: A Living Review of Machine Learning for Particle Physics (2021) arXiv:2102.02770 [hep-ph]

work page arXiv 2021
[10]

Wang, A., Gandrakota, A., Ngadiuba, J., Sahu, V., Bhatnagar, P., Khoda, E.E., Duarte, J.: Interpreting Transformers for Jet Tagging (2024) arXiv:2412.03673 [hep-ph]

work page arXiv 2024
[11]

Aab, A., et al.: The Pierre Auger Cosmic Ray Observatory. Nucl. Instrum. Meth. A 798, 172–213 (2015) https://doi.org/10.1016/j. nima.2015.06.058 arXiv:1502.01323 [astro- ph.IM]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j 2015
[12]

: Deep-learning based reconstruction of the shower maxi- mum Xmax using the water-Cherenkov detectors of the Pierre Auger Observa- tory

Aab, A., et al. : Deep-learning based reconstruction of the shower maxi- mum Xmax using the water-Cherenkov detectors of the Pierre Auger Observa- tory. JINST 16(07), 07019 (2021) https: //doi.org/10.1088/1748-0221/16/07/P07019 arXiv:2101.02946 [astro-ph.IM]

work page doi:10.1088/1748-0221/16/07/p07019 2021
[13]

: Extraction of the muon signals recorded with the surface detector of the Pierre Auger Observa- tory using recurrent neural networks

Aab, A., et al. : Extraction of the muon signals recorded with the surface detector of the Pierre Auger Observa- tory using recurrent neural networks. JINST 16(07), 07016 (2021) https: //doi.org/10.1088/1748-0221/16/07/P07016 arXiv:2103.11983 [hep-ex]

work page doi:10.1088/1748-0221/16/07/p07016 2021
[14]

: Measurement of the depth of maximum of air-shower profiles with energies between 10 18.5 and 10 20 eV using the surface detector of the Pierre Auger Observatory and deep learning

Abdul Halim, A., et al. : Measurement of the depth of maximum of air-shower profiles with energies between 10 18.5 and 10 20 eV using the surface detector of the Pierre Auger Observatory and deep learning. Phys. Rev. D 111(2), 022003 (2025) https: //doi.org/10.1103/PhysRevD.111.022003 arXiv:2406.06319 [astro-ph.HE]

work page doi:10.1103/physrevd.111.022003 2025
[15]

Hoogeboom, E., Peters, J.W.T., Cohen, T.S., Welling, M.: HexaConv (2018) arXiv:1803.02108 [cs.LG]

work page internal anchor Pith review Pith/arXiv arXiv 2018
[16]

Deriving global structure of the Galactic Magnetic Field from Faraday Rotation Measures of extragalactic sources

Pshirkov, M.S., Tinyakov, P.G., Kronberg, P.P., Newton-McGee, K.J.: Deriving global structure of the Galactic Magnetic Field from Faraday Rotation Measures of extra- galactic sources. Astrophys. J. 738, 192 (2011) https://doi.org/10.1088/0004-637X/ 738/2/192 arXiv:1103.0814 [astro-ph.GA]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/0004-637x/ 2011
[17]

The Galactic Magnetic Field

Jansson, R., Farrar, G.R.: The Galactic Magnetic Field. Astrophys. J. Lett. 761, 11 (2012) https://doi.org/10.1088/2041-8205/ 761/1/L11 arXiv:1210.7820 [astro-ph.GA]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/2041-8205/ 2012
[18]

Unger and G

Unger, M., Farrar, G.R.: The Coherent Mag- netic Field of the Milky Way. Astrophys. J. 970(1), 95 (2024) https://doi.org/10.3847/ 1538-4357/ad4a54 arXiv:2311.12120 [astro- ph.GA]

work page arXiv 2024
[19]

Unger, M., Farrar, G.R.: Uncertainties in the Magnetic Field of the Milky Way (2017) arXiv:1707.02339 [astro-ph.GA] 8

work page internal anchor Pith review Pith/arXiv arXiv 2017
[20]

PoS ICRC2023, 198 (2023) https://doi.org/10.22323/1.444

Schulte, J., Bister, T., Erdmann, M.: An all- sky search method for coherent magnetic field deflections of ultra-high-energy cosmic rays based on Deep Learning. PoS ICRC2023, 198 (2023) https://doi.org/10.22323/1.444. 0198

work page doi:10.22323/1.444 2023
[21]

: CRPropa 3.2 — an advanced framework for high-energy particle propagation in extragalactic and galactic spaces

Alves Batista, R., et al. : CRPropa 3.2 — an advanced framework for high-energy particle propagation in extragalactic and galactic spaces. JCAP 09, 035 (2022) https: //doi.org/10.1088/1475-7516/2022/09/035 arXiv:2208.00107 [astro-ph.HE]

work page doi:10.1088/1475-7516/2022/09/035 2022
[22]

Halim, A.A., et al. : Constraining models for the origin of ultra-high-energy cosmic rays with a novel combined analysis of arrival directions, spectrum, and compo- sition data measured at the Pierre Auger Observatory. JCAP 01, 022 (2024) https: //doi.org/10.1088/1475-7516/2024/01/022 arXiv:2305.16693 [astro-ph.HE]

work page doi:10.1088/1475-7516/2024/01/022 2024
[23]

Xiong, Y., Zeng, Z., Chakraborty, R., Tan, M., Fung, G., Li, Y., Singh, V.: Nystr¨ omformer: A nystr¨ om-based algorithm for approximating self-attention (2021) arXiv:2102.03902 [cs.CL]

work page arXiv 2021
[24]

HEALPix -- a Framework for High Resolution Discretization, and Fast Analysis of Data Distributed on the Sphere

G´ orski, K.M., Hivon, E., Banday, A.J., Wan- delt, B.D., Hansen, F.K., Reinecke, M., Bartelman, M.: HEALPix - A Framework for high resolution discretization, and fast anal- ysis of data distributed on the sphere. Astro- phys. J. 622, 759–771 (2005) https://doi.org/ 10.1086/427976 arXiv:astro-ph/0409513

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1086/427976 2005
[25]

Journal of Open Source Software 4(35), 1298 (2019) https://doi.org/10.21105/ joss.01298

Zonca, A., Singer, L., Lenz, D., Reinecke, M., Rosset, C., Hivon, E., Gorski, K.: healpy: equal area pixelization and spherical har- monics transforms for data on the sphere in Python. Journal of Open Source Software 4(35), 1298 (2019) https://doi.org/10.21105/ joss.01298

work page 2019
[26]

Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks (2017) arXiv:1703.01365 [cs.LG] 9

work page internal anchor Pith review Pith/arXiv arXiv 2017

[1] [1]

Vaswani, A., Shazeer, N., Parmar, N., Uszko- reit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention Is All You Need (2017) arXiv:1706.03762 [cs.CL] 7

work page internal anchor Pith review Pith/arXiv arXiv 2017

[2] [2]

Dosovitskiy, A., et al.: An Image is Worth 16x16 Words: Transformers for Image Recog- nition at Scale (2020) arXiv:2010.11929 [cs.CV]

work page internal anchor Pith review Pith/arXiv arXiv 2020

[3] [3]

https://openai.com/ blog/chatgpt/

OpenAI: ChatGPT: Optimizing Language Models for Dialogue. https://openai.com/ blog/chatgpt/. Online; Zugriff am 30. Novem- ber 2022 (2022)

work page 2022

[4] [4]

Qu, H., Li, C., Qian, S.: Particle Transformer for Jet Tagging (2022) arXiv:2202.03772 [hep-ph]

work page arXiv 2022

[5] [5]

JHEP 06, 184 (2023) https://doi.org/10.1007/JHEP06(2023)184 arXiv:2303.07364 [hep-ph]

Finke, T., Kr¨ amer, M., M¨ uck, A., T¨ onshoff, J.: Learning the language of QCD jets with transformers. JHEP 06, 184 (2023) https://doi.org/10.1007/JHEP06(2023)184 arXiv:2303.07364 [hep-ph]

work page doi:10.1007/jhep06(2023)184 2023

[6] [6]

PoS ICRC2023, 371 (2023) https://doi.org/10.22323/1.444.0371

Abdul Halim, A., et al.: Deep-Learning-Based Cosmic-Ray Mass Reconstruction Using the Water-Cherenkov and Scintillation Detectors of AugerPrime. PoS ICRC2023, 371 (2023) https://doi.org/10.22323/1.444.0371

work page doi:10.22323/1.444.0371 2023

[7] [7]

Wu, Y., Wang, K., Li, C., Qu, H., Zhu, J.: Jet tagging with more-interaction particle trans- former. Chin. Phys. C 49(1), 013110 (2025) https://doi.org/10.1088/1674-1137/ad7f3d arXiv:2407.08682 [hep-ph]

work page doi:10.1088/1674-1137/ad7f3d 2025

[8] [8]

Brehmer, J., Bres´ o, V., Haan, P., Plehn, T., Qu, H., Spinner, J., Thaler, J.: A Lorentz- Equivariant Transformer for All of the LHC (2024) arXiv:2411.00446 [hep-ph]

work page arXiv 2024

[9] [9]

Feickert, M., Nachman, B.: A Living Review of Machine Learning for Particle Physics (2021) arXiv:2102.02770 [hep-ph]

work page arXiv 2021

[10] [10]

Wang, A., Gandrakota, A., Ngadiuba, J., Sahu, V., Bhatnagar, P., Khoda, E.E., Duarte, J.: Interpreting Transformers for Jet Tagging (2024) arXiv:2412.03673 [hep-ph]

work page arXiv 2024

[11] [11]

Aab, A., et al.: The Pierre Auger Cosmic Ray Observatory. Nucl. Instrum. Meth. A 798, 172–213 (2015) https://doi.org/10.1016/j. nima.2015.06.058 arXiv:1502.01323 [astro- ph.IM]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j 2015

[12] [12]

: Deep-learning based reconstruction of the shower maxi- mum Xmax using the water-Cherenkov detectors of the Pierre Auger Observa- tory

Aab, A., et al. : Deep-learning based reconstruction of the shower maxi- mum Xmax using the water-Cherenkov detectors of the Pierre Auger Observa- tory. JINST 16(07), 07019 (2021) https: //doi.org/10.1088/1748-0221/16/07/P07019 arXiv:2101.02946 [astro-ph.IM]

work page doi:10.1088/1748-0221/16/07/p07019 2021

[13] [13]

: Extraction of the muon signals recorded with the surface detector of the Pierre Auger Observa- tory using recurrent neural networks

Aab, A., et al. : Extraction of the muon signals recorded with the surface detector of the Pierre Auger Observa- tory using recurrent neural networks. JINST 16(07), 07016 (2021) https: //doi.org/10.1088/1748-0221/16/07/P07016 arXiv:2103.11983 [hep-ex]

work page doi:10.1088/1748-0221/16/07/p07016 2021

[14] [14]

: Measurement of the depth of maximum of air-shower profiles with energies between 10 18.5 and 10 20 eV using the surface detector of the Pierre Auger Observatory and deep learning

Abdul Halim, A., et al. : Measurement of the depth of maximum of air-shower profiles with energies between 10 18.5 and 10 20 eV using the surface detector of the Pierre Auger Observatory and deep learning. Phys. Rev. D 111(2), 022003 (2025) https: //doi.org/10.1103/PhysRevD.111.022003 arXiv:2406.06319 [astro-ph.HE]

work page doi:10.1103/physrevd.111.022003 2025

[15] [15]

Hoogeboom, E., Peters, J.W.T., Cohen, T.S., Welling, M.: HexaConv (2018) arXiv:1803.02108 [cs.LG]

work page internal anchor Pith review Pith/arXiv arXiv 2018

[16] [16]

Deriving global structure of the Galactic Magnetic Field from Faraday Rotation Measures of extragalactic sources

Pshirkov, M.S., Tinyakov, P.G., Kronberg, P.P., Newton-McGee, K.J.: Deriving global structure of the Galactic Magnetic Field from Faraday Rotation Measures of extra- galactic sources. Astrophys. J. 738, 192 (2011) https://doi.org/10.1088/0004-637X/ 738/2/192 arXiv:1103.0814 [astro-ph.GA]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/0004-637x/ 2011

[17] [17]

The Galactic Magnetic Field

Jansson, R., Farrar, G.R.: The Galactic Magnetic Field. Astrophys. J. Lett. 761, 11 (2012) https://doi.org/10.1088/2041-8205/ 761/1/L11 arXiv:1210.7820 [astro-ph.GA]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/2041-8205/ 2012

[18] [18]

Unger and G

Unger, M., Farrar, G.R.: The Coherent Mag- netic Field of the Milky Way. Astrophys. J. 970(1), 95 (2024) https://doi.org/10.3847/ 1538-4357/ad4a54 arXiv:2311.12120 [astro- ph.GA]

work page arXiv 2024

[19] [19]

Unger, M., Farrar, G.R.: Uncertainties in the Magnetic Field of the Milky Way (2017) arXiv:1707.02339 [astro-ph.GA] 8

work page internal anchor Pith review Pith/arXiv arXiv 2017

[20] [20]

PoS ICRC2023, 198 (2023) https://doi.org/10.22323/1.444

Schulte, J., Bister, T., Erdmann, M.: An all- sky search method for coherent magnetic field deflections of ultra-high-energy cosmic rays based on Deep Learning. PoS ICRC2023, 198 (2023) https://doi.org/10.22323/1.444. 0198

work page doi:10.22323/1.444 2023

[21] [21]

: CRPropa 3.2 — an advanced framework for high-energy particle propagation in extragalactic and galactic spaces

Alves Batista, R., et al. : CRPropa 3.2 — an advanced framework for high-energy particle propagation in extragalactic and galactic spaces. JCAP 09, 035 (2022) https: //doi.org/10.1088/1475-7516/2022/09/035 arXiv:2208.00107 [astro-ph.HE]

work page doi:10.1088/1475-7516/2022/09/035 2022

[22] [22]

Halim, A.A., et al. : Constraining models for the origin of ultra-high-energy cosmic rays with a novel combined analysis of arrival directions, spectrum, and compo- sition data measured at the Pierre Auger Observatory. JCAP 01, 022 (2024) https: //doi.org/10.1088/1475-7516/2024/01/022 arXiv:2305.16693 [astro-ph.HE]

work page doi:10.1088/1475-7516/2024/01/022 2024

[23] [23]

Xiong, Y., Zeng, Z., Chakraborty, R., Tan, M., Fung, G., Li, Y., Singh, V.: Nystr¨ omformer: A nystr¨ om-based algorithm for approximating self-attention (2021) arXiv:2102.03902 [cs.CL]

work page arXiv 2021

[24] [24]

HEALPix -- a Framework for High Resolution Discretization, and Fast Analysis of Data Distributed on the Sphere

G´ orski, K.M., Hivon, E., Banday, A.J., Wan- delt, B.D., Hansen, F.K., Reinecke, M., Bartelman, M.: HEALPix - A Framework for high resolution discretization, and fast anal- ysis of data distributed on the sphere. Astro- phys. J. 622, 759–771 (2005) https://doi.org/ 10.1086/427976 arXiv:astro-ph/0409513

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1086/427976 2005

[25] [25]

Journal of Open Source Software 4(35), 1298 (2019) https://doi.org/10.21105/ joss.01298

Zonca, A., Singer, L., Lenz, D., Reinecke, M., Rosset, C., Hivon, E., Gorski, K.: healpy: equal area pixelization and spherical har- monics transforms for data on the sphere in Python. Journal of Open Source Software 4(35), 1298 (2019) https://doi.org/10.21105/ joss.01298

work page 2019

[26] [26]

Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks (2017) arXiv:1703.01365 [cs.LG] 9

work page internal anchor Pith review Pith/arXiv arXiv 2017