arxiv: 2605.06779 · v1 · submitted 2026-05-07 · ⚛️ physics.ins-det · hep-ex· hep-ph

Recognition: 2 theorem links

· Lean Theorem

Exploring the Boundaries of Differentiable Radiation Transport and Detector Simulation

Jeffrey Krupa , Yiyang Zhao , Mihaly Novak , Max Aehle , Max Sagebaum , Long Chen , Nicolas Gauger , Miaoyuan Liu

show 2 more authors

Lukas Heinrich Michael Kagan

Authors on Pith no claims yet

Pith reviewed 2026-05-11 00:46 UTC · model grok-4.3

classification ⚛️ physics.ins-det hep-exhep-ph

keywords differentiable simulationradiation transportautomatic differentiationdetector optimizationmaterial boundariesgradient explosionGeant4

0 comments

The pith

Stopping gradient flow at unstable material boundaries yields usable derivatives for optimizing radiation detector designs.

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

The paper applies automatic differentiation to a full electromagnetic particle transport simulation modeled after Geant4. It finds that rare but extreme sensitivities at material boundaries produce exploding gradients that render the derivatives unusable for optimization. The authors introduce a fix that detects those unstable boundary crossings and stops gradient propagation through them, leaving the forward simulation entirely unchanged. This produces stable gradients that support gradient-based optimization in a concrete detector-design task. A sympathetic reader would care because it opens the door to treating radiation transport as a differentiable module inside larger design loops.

Core claim

When step-wise radiation transport is differentiated, rare extreme sensitivities at material boundaries drive exploding gradients that propagate through subsequent shower development. Detecting these identifiable unstable conditions and stopping gradient flow through the associated boundary-crossing operations removes the explosion while preserving the original forward simulation. The resulting derivatives are stable and directly usable for optimization, as shown in a detector-design problem.

What carries the argument

Targeted stopping of gradient propagation through boundary-crossing operations under identifiable unstable conditions.

If this is right

Stable, optimization-ready gradients become available for detector design tasks involving electromagnetic showers.
The forward particle transport and shower development remain identical to the non-differentiated case.
Gradient explosions that previously made differentiation impractical are prevented from propagating through multiple transport steps.
The method applies to Geant4-like simulations with full electromagnetic physics without requiring changes to the underlying physics models.

Where Pith is reading between the lines

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

The same boundary-stopping idea could be tested in other Monte Carlo codes that contain discontinuities at material interfaces.
End-to-end differentiable pipelines that combine transport with downstream reconstruction or machine-learning layers become feasible.
The technique may generalize to any simulation whose discontinuities are localized and detectable at runtime.

Load-bearing premise

Unstable boundary conditions can be detected reliably enough that halting gradients there preserves useful optimization information without bias or missed sensitivities.

What would settle it

Run the mitigated differentiable simulator to optimize a detector geometry parameter such as layer thickness, then compare the obtained optimum against the true minimum found by a grid search or derivative-free optimizer on the same forward model; mismatch beyond numerical tolerance would falsify the claim.

Figures

Figures reproduced from arXiv: 2605.06779 by Jeffrey Krupa, Long Chen, Lukas Heinrich, Max Aehle, Max Sagebaum, Miaoyuan Liu, Michael Kagan, Mihaly Novak, Nicolas Gauger, Yiyang Zhao.

**Figure 2.** Figure 2: Left: a track of step length L incident on a boundary with angle β between the track direction vˆ and the surface normal nˆ. Shifting the boundary by ∆b along nˆ changes the step length by ∆L ∝ ∆b/ cos β. Right: the mean derivative of the step length with respect to boundary position (upper) and the variance of the per-step energy-deposit derivative (lower), both as a function of incidence angle β, confirm… view at source ↗

**Figure 3.** Figure 3: The fraction of track energy deposited as a function of the track direction projected [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: An example of a low-energy track exhibiting repeated boundary-limited steps near [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Mean longitudinal energy-deposit derivative per calorimeter layer comparing finite [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Mean longitudinal energy-deposit derivative per calorimeter layer comparing finite [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Trajectories of 50 randomly-seeded optimization runs minimizing the longitudinal-profile [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Mean longitudinal energy-deposit derivative per calorimeter layer comparing finite [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

**Figure 9.** Figure 9: Trajectories of 50 randomly-seeded optimization runs minimizing the longitudinal-profile [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

read the original abstract

We present an application of automatic differentiation for particle transport through matter using a Geant4-like radiation transport simulation with a full electromagnetic physics model. When differentiating this step-based transport, we observe exploding gradients driven by rare but extreme sensitivities at material boundaries, which propagate through subsequent transport and shower development. To obtain usable derivatives for optimization, we introduce a targeted mitigation strategy that stops gradient propagation through boundary-crossing operations under identifiable unstable conditions while leaving the forward (primal) simulation unchanged. We demonstrate that this enables stable, optimization-ready gradients in a detector-design problem.

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 gives a practical engineering fix for exploding gradients in differentiable Geant4-style EM transport by halting flow at detected unstable boundary crossings.

read the letter

The main takeaway is that they apply automatic differentiation to a full electromagnetic physics model in a step-based transport code and handle the resulting gradient explosions at material boundaries by stopping propagation under identifiable unstable conditions. The forward simulation stays untouched, which matters for keeping physics fidelity in existing codes like Geant4. They show this produces usable gradients for a detector-design optimization task. That targeted mitigation is the concrete new piece, and it addresses a real numerical problem that arises because small parameter shifts can dramatically change crossing probabilities or angles in shower development. The approach is straightforward and leaves the primal path intact, which is a plus for reproducibility with standard tools. The soft spot is the risk that stopping gradients discards sensitivities that actually matter for the objective. Material interfaces drive containment, energy deposition, and detector response, and those are exactly where the instabilities occur. If the detection rule misses cases or if the stopped contributions are not negligible in the loss landscape, the derivatives become a biased estimator that could push designs away from better interface-aware solutions. The abstract gives no numbers on trigger frequency, validation metrics, or comparisons to non-differentiated baselines, so it is not yet clear how well the approximation holds in practice. This is for physicists working on gradient-based detector optimization or autodiff in Monte Carlo transport. A reader who needs stable derivatives through complex geometry would find the idea useful if the full paper includes the implementation details and checks against the bias concern. It deserves peer review because the problem is relevant to instrumentation work and the fix is specific, even though the evidence on whether the gradients remain informative needs strengthening.

Referee Report

2 major / 2 minor

Summary. The manuscript applies automatic differentiation to a Geant4-like radiation transport simulation with a full electromagnetic physics model. It identifies exploding gradients caused by rare but extreme sensitivities at material boundaries during step-based transport and shower development. The authors introduce a mitigation that stops gradient propagation through boundary-crossing operations under identifiable unstable conditions while leaving the forward simulation unchanged, and they claim this produces stable, optimization-ready gradients for a detector-design problem.

Significance. If the mitigation can be shown to preserve physically relevant sensitivities without bias, the work would be significant for enabling gradient-based optimization in full-physics detector simulations, a longstanding challenge in the field. The targeted, non-invasive handling of numerical instability is a pragmatic contribution that could influence differentiable physics approaches more broadly. The current lack of quantitative validation metrics, however, prevents a firm assessment of practical utility.

major comments (2)

[Abstract] Abstract: the central claim that the mitigation 'enables stable, optimization-ready gradients in a detector-design problem' is presented without any quantitative results, validation metrics, implementation details on unstable-condition detection, or comparison to unmitigated gradients. This absence makes it impossible to determine whether the reported stability is usable for optimization or merely numerical.
[mitigation strategy] Description of the mitigation strategy: stopping gradients at detected unstable boundary crossings risks omitting dominant sensitivities at material interfaces, which control shower containment, energy deposition, and detector response. The manuscript provides no evidence or test that the heuristic reliably separates unstable from stable contributions without introducing bias into the loss landscape or missing critical parameter sensitivities.

minor comments (2)

[Abstract] The abstract does not specify the detector-design objective function, the parameters being optimized, or the scale of the simulation used in the demonstration.
Notation for 'unstable conditions' and the precise criterion for stopping gradients is not defined in the provided summary, which hinders reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and recognition of the potential significance of applying automatic differentiation to full-physics radiation transport. We address each major comment below and will revise the manuscript to incorporate additional quantitative details and validation as outlined.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that the mitigation 'enables stable, optimization-ready gradients in a detector-design problem' is presented without any quantitative results, validation metrics, implementation details on unstable-condition detection, or comparison to unmitigated gradients. This absence makes it impossible to determine whether the reported stability is usable for optimization or merely numerical.

Authors: We agree that the abstract is concise and omits specific quantitative metrics. The manuscript body presents a detector-design optimization example in which unmitigated gradients explode while the mitigated version yields stable derivatives that enable convergence to a physically reasonable design. We will revise the abstract to include a brief reference to the observed gradient stability improvement and optimization success, and we will expand the main text with explicit implementation details on the unstable-condition detection criteria along with a direct comparison of gradient behavior with and without mitigation. revision: yes
Referee: [mitigation strategy] Description of the mitigation strategy: stopping gradients at detected unstable boundary crossings risks omitting dominant sensitivities at material interfaces, which control shower containment, energy deposition, and detector response. The manuscript provides no evidence or test that the heuristic reliably separates unstable from stable contributions without introducing bias into the loss landscape or missing critical parameter sensitivities.

Authors: This concern is well-founded and we acknowledge that the current manuscript does not provide explicit quantitative tests isolating the effect on interface sensitivities. The mitigation is applied only under identifiable unstable conditions (extreme local sensitivities during boundary crossings in step-based transport) while the forward simulation remains completely unchanged, preserving all physical quantities such as energy deposition and shower development. In the detector-design example the resulting gradients produce optimization outcomes consistent with expected physics. To address the gap, we will add in the revision a set of controlled tests comparing the loss landscape, gradient norms, and final optimized parameters with and without the heuristic, including checks that critical material-interface sensitivities are retained. revision: yes

Circularity Check

0 steps flagged

No significant circularity; mitigation is an independent engineering response to observed instability

full rationale

The paper identifies exploding gradients as an observed numerical issue arising at material boundaries during differentiation of step-based EM transport. It then proposes an explicit mitigation—stopping gradient propagation through boundary-crossing operations only under identifiable unstable conditions—while leaving the primal forward simulation unchanged. This is demonstrated on a detector-design optimization task. No equations or claims reduce by construction to fitted inputs, self-definitions, or load-bearing self-citations; the forward pass and the gradient-stopping rule are described as separate interventions. The derivation chain is therefore self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard automatic differentiation rules applied to an existing Geant4-like transport model; no new free parameters or invented entities are introduced.

axioms (2)

standard math Standard automatic differentiation rules apply to step-based particle transport operations
The paper assumes AD frameworks can be applied to the simulation steps.
domain assumption Unstable boundary conditions can be identified without altering the forward physics
The mitigation relies on this identification being possible and safe.

pith-pipeline@v0.9.0 · 5414 in / 1173 out tokens · 73409 ms · 2026-05-11T00:46:03.402225+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

we introduce a targeted mitigation strategy that stops gradient propagation through boundary-crossing operations under identifiable unstable conditions
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the realized step length L = min(Lphys, Lbnd)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 48 canonical work pages · 1 internal anchor

[1]

Agostinelli Set al.2003 Geant4—a simulation toolkitNucl. Instrum. Methods Phys. Res., Sect. A506250–303 doi:10.1016/S0168-9002(03)01368-8

work page doi:10.1016/s0168-9002(03)01368-8 2003
[2]

Allison Jet al.2006 Geant4 developments and applicationsIEEE Trans. Nucl. Sci.53270–278 doi:10.1109/TNS.2006.869826

work page doi:10.1109/tns.2006.869826 2006
[3]

Allison Jet al.2016 Recent developments in Geant4Nucl. Instrum. Methods Phys. Res., Sect. A835186–225 doi:10.1016/j.nima.2016.06.125

work page doi:10.1016/j.nima.2016.06.125 2016
[4]

Zein S A, Bordage M C, Francis Z, Macetti G, Genoni A, Cappello C D, Shin W G and Incerti S 2021 Electron transport in DNA bases: An extension of the Geant4-DNA Monte Carlo toolkit Nucl. Instrum. Methods Phys. Res., Sect. B48870–82 doi:10.1016/j.nimb.2020.11.021

work page doi:10.1016/j.nimb.2020.11.021 2021
[5]

Dom´ ınguez-Mu˜ noz A D, Gallardo M I, Bordage M C, Francis Z, Incerti S and Cort´ es- Giraldo M A 2022 A model for Geant4-DNA to simulate ionization and excitation of liquid water by protons travelling above 100 MeVRadiat. Phys. Chem.199110363 doi:10.1016/j.radphyschem.2022.110363

work page doi:10.1016/j.radphyschem.2022.110363 2022
[6]

Chattopadhyay, M

Ganjgah A A and Taherparvar P 2024 The effects of cell displacement on DNA damages in targeted radiation therapy using Geant4-DNASci. Rep.1481863 doi:10.1038/s41598-024- 81863-4

work page doi:10.1038/s41598-024- 2024
[7]

Tabbakh F 2024 Significance of the proton energy loss mechanism to gold nanoparticles in proton therapy: a Geant4 simulationScientific Reports1476244 doi:10.1038/s41598-024-76244-w

work page doi:10.1038/s41598-024-76244-w 2024
[8]

Shin W Get al.2021 A Geant4-DNA Evaluation of Radiation-Induced DNA Damage on a Human FibroblastCancers134940 doi:10.3390/cancers13194940

work page doi:10.3390/cancers13194940 2021
[9]

Chappuis F, Tran H N, Zein S A, Bailat C, Incerti S, Bochud F and Desorgher L 2023 The general-purpose Geant4 Monte Carlo toolkit and its Geant4-DNA extension to investigate mechanisms underlying the FLASH effect in radiotherapy: Current status and challenges Physica Med.110102601 doi:10.1016/j.ejmp.2023.102601

work page doi:10.1016/j.ejmp.2023.102601 2023
[10]

Simhony Y, Segal A, Amrani O and Etzion E 2025 High-end space electronics: Active shielding to mitigate catastrophic single-event effectsAerosp. Sci. Technol.169111344 doi:10.1016/j.ast.2025.111344

work page doi:10.1016/j.ast.2025.111344 2025
[11]

Negm H and Ohgaki H 2024 Validation of GEANT4 simulation for active materials in- terrogation using nuclear resonance fluorescenceJ. Radiat. Res. Appl. Sci.17101209 doi:10.1016/j.jrras.2024.101209

work page doi:10.1016/j.jrras.2024.101209 2024
[12]

MacFadden N, Peggs S and Gulliford C 2018 Development and Validation of a Geant4 Radiation Shielding Simulation Framework doi:10.2172/1515417

work page doi:10.2172/1515417 2018
[13]

Baydin A G, Pearlmutter B A, Radul A A and Siskind J M 2018 Automatic differentiation in machine learning: a surveyJ. Mach. Learn. Res.181–43 doi:10.48550/arXiv.1502.05767 15

work page Pith review doi:10.48550/arxiv.1502.05767 2018
[14]

Graphics371–12 doi:10.1145/3272127.3275109

Li T M, Aittala M, Durand F and Lehtinen J 2018 Differentiable Monte Carlo ray tracing through edge samplingACM Trans. Graphics371–12 doi:10.1145/3272127.3275109

work page doi:10.1145/3272127.3275109 2018
[15]

2019 , issue_date =

Nimier-David M, Vicini D, Zeltner T and Jakob W 2019 Mitsuba 2: A Retargetable Forward and Inverse RendererACM Trans. Graphics381–17 doi:10.1145/3355089.3356498

work page doi:10.1145/3355089.3356498 2019
[16]

Hu Y, Anderson L, Li T M, Sun Q, Carr N, Ragan-Kelley J and Durand F 2020 DiffTaichi: Differentiable Programming for Physical SimulationICLR 2020doi:10.48550/arXiv.1910.00935

work page doi:10.48550/arxiv.1910.00935 2020
[17]

Newbury R, Collins J, He K, Pan J, Posner I, Howard D and Cosgun A 2024 A Review of Differentiable SimulatorsIEEE Access1297581–97604 doi:10.1109/ACCESS.2024.3425448

work page doi:10.1109/access.2024.3425448 2024
[18]

Chen R T Q, Rubanova Y, Bettencourt J and Duvenaud D K 2018 Neural Ordi- nary Differential EquationsAdvances in Neural Information Processing Systemsvol 31 doi:10.48550/arXiv.1806.07366

work page doi:10.48550/arxiv.1806.07366 2018
[19]

Aehle M, Nov´ ak M, Vassilev V, Gauger N R, Heinrich L, Kagan M and Lange D 2024 Optimization Using Pathwise Algorithmic Derivatives of Electromagnetic Shower Simulations Comput. Phys. Commun.309109491 doi:10.1016/j.cpc.2024.109491

work page doi:10.1016/j.cpc.2024.109491 2024
[20]

Nov´ ak M 2023 HepEmShow: a compact EM shower simulation application documentation, accessed 2026-01-20 URLhttps://hepemshow.readthedocs.io/en/latest/index.html

2023
[21]

Nov´ ak M, Hahnfeld J and Morgan B 2022 g4hepem: The G4HepEm R&D Project gitHub repository, accessed 2026-01-20 URLhttps://github.com/mnovak42/g4hepem

2022
[22]

Fabjan C W and Gianotti F 2003 Calorimetry for particle physicsRev. Mod. Phys.751243–1286 doi:10.1103/RevModPhys.75.1243

work page doi:10.1103/revmodphys.75.1243 2003
[23]

B., Abdo, A

Atwood W Bet al.2009 The Large Area Telescope on the Fermi Gamma-Ray Space Telescope MissionAstrophys. J.6971071–1102 doi:10.1088/0004-637X/697/2/1071

work page doi:10.1088/0004-637x/697/2/1071 2009
[24]

Seuntjens J and Duane S 2009 Photon absorbed dose standardsMetrologia46S39–S58 doi:10.1088/0026-1394/46/2/S04

work page doi:10.1088/0026-1394/46/2/s04 2009
[25]

Aehle M, Nguyen X T, Nov´ ak M, Dorigo T, Gauger N R, Kieseler J, Klute M and Vassilev V 2024 Efficient Forward-Mode Algorithmic Derivatives of Geant4 doi:10.48550/arXiv.2407.02966

work page doi:10.48550/arxiv.2407.02966 2024
[26]

Kagan M and Heinrich L 2023 Branches of a Tree: Taking Derivatives of Programs with Discrete and Branching Randomness in High Energy Physics doi:10.48550/arXiv.2308.16680

work page doi:10.48550/arxiv.2308.16680 2023
[27]

Instrum.17P08021 doi:10.1088/1748-0221/17/08/p08021

Cheong S, Frisch J C, Gasiorowski S, Hogan J M, Kagan M, Safdari M, Schwartzman A and Vandegar M 2022 Novel light field imaging device with enhanced light collection for cold atom cloudsJ. Instrum.17P08021 doi:10.1088/1748-0221/17/08/p08021

work page doi:10.1088/1748-0221/17/08/p08021 2022
[28]

Learn.: Sci

Strong G Cet al.2024 TomOpt: differential optimisation for task- and constraint-aware design of particle detectors in the context of muon tomographyMach. Learn.: Sci. Technol.5035002 doi:10.1088/2632-2153/ad52e7

work page doi:10.1088/2632-2153/ad52e7 2024
[29]

Learn.: Sci

Gasiorowski S, Chen Y, Nashed Y, Granger P, Mironov C, Tsang K V, Ratner D and Terao K 2024 Differentiable simulation of a liquid argon time projection chamberMach. Learn.: Sci. Technol.5025012 doi:10.1088/2632-2153/ad2cf0 16

work page doi:10.1088/2632-2153/ad2cf0 2024
[30]

Alterkait O, Jes´ us-Valls C, Matsumoto R, de Perio P and Terao K 2026 End-to-end Differentiable Calibration and Reconstruction for Optical Particle Detectors doi:10.48550/arXiv.2602.24129

work page doi:10.48550/arxiv.2602.24129 2026
[31]

Aehle Met al.2025 Progress in end-to-end optimization of fundamental physics ex- perimental apparata with differentiable programmingReviews in Physics13100120 doi:10.1016/j.revip.2025.100120

work page doi:10.1016/j.revip.2025.100120 2025
[32]

Mokhtar F, Pata J, Garcia D, Wulff E, Zhang M, Kagan M and Duarte J 2025 Fine-tuning machine-learned particle-flow reconstruction for new detector geometries in future colliders Phys. Rev. D111doi:10.1103/physrevd.111.092015

work page doi:10.1103/physrevd.111.092015 2025
[33]

Adelmann Aet al.2022 New directions for surrogate models and differentiable programming for High Energy Physics detector simulationSnowmass 2021doi:10.48550/arXiv.2203.08806

work page doi:10.48550/arxiv.2203.08806 2022
[34]

Shirobokov S, Belavin V, Kagan M, Ustyuzhanin A and Baydin A G 2020 Black-Box Optimization with Local Generative SurrogatesAdvances in Neural Information Pro- cessing SystemsURL https://proceedings.neurips.cc/paper_files/paper/2020/file/ a878dbebc902328b41dbf02aa87abb58-Paper.pdf

2020
[35]

Learn.: Sci

Wozniak K A, Mulligana S, Kieseler J, Klute M, Fleuret F and Golling T 2025 End-to-End Optimal Detector Design with Mutual Information SurrogatesMach. Learn.: Sci. Technol.6 045047 doi:10.1088/2632-2153/ae1acb

work page doi:10.1088/2632-2153/ae1acb 2025
[36]

Schmidt Ket al.2025 End-to-End Detector Optimization with Diffusion Models: A Case Study in Sampling CalorimetersParticles847 doi:10.3390/particles8020047

work page doi:10.3390/particles8020047 2025
[37]

, keywords =

Dorigo T (MODE Collaboration) 2023 Toward the end-to-end optimization of par- ticle physics instruments with differentiable programmingRev. Phys.10100085 doi:10.1016/j.revip.2023.100085

work page doi:10.1016/j.revip.2023.100085 2023
[38]

Dorigo T (MODE Collaboration) 2023 Toward artificial-intelligence assisted design of experi- mentsNucl. Instrum. Methods Phys. Res., Sect. A1047167873 doi:10.1016/j.nima.2022.167873

work page doi:10.1016/j.nima.2022.167873 2023
[39]

Dorigo T (MODE Collaboration) 2026 On the Codesign of Scientific Experiments and Industrial Systems doi:10.48550/arXiv.2603.26613

work page doi:10.48550/arxiv.2603.26613 2026
[40]

Instrum.15P05009 doi:10.1088/1748-0221/15/05/P05009

Cisbani Eet al.2020 AI-optimized detector design for the future Electron-Ion Collider: the dual-radiator RICH caseJ. Instrum.15P05009 doi:10.1088/1748-0221/15/05/P05009

work page doi:10.1088/1748-0221/15/05/p05009 2020
[41]

Instrum.19C07001 doi:10.1088/1748-0221/19/07/C07001

Diefenthaler M (AID(2)E Collaboration) 2024 AI-assisted detector design for the EIC (AID(2)E) J. Instrum.19C07001 doi:10.1088/1748-0221/19/07/C07001

work page doi:10.1088/1748-0221/19/07/c07001 2024
[42]

Roussel Ret al.2024 Bayesian optimization algorithms for accelerator physicsPhys. Rev. Accel. Beams27(8) 084801 doi:10.1103/PhysRevAccelBeams.27.084801

work page doi:10.1103/physrevaccelbeams.27.084801 2024
[43]

Fanelli Cet al.2023 AI-assisted optimization of the ECCE tracking system at the Electron Ion ColliderNucl. Instrum. Methods Phys. Res., Sect. A1047167748 doi:10.1016/j.nima.2022.167748

work page doi:10.1016/j.nima.2022.167748 2023
[44]

Learn.: Sci

Qasim S R, Owen P and Serra N 2025 Physics Instrument Design with Reinforcement Learning Mach. Learn.: Sci. Technol.035033 doi:10.1088/2632-2153/adf7ff 17

work page doi:10.1088/2632-2153/adf7ff 2025
[45]

Learn.: Sci

Kortus T, Keidel R, Gauger N R and Kieseler J (Bergen pCT) 2026 Constrained collaborative optimization of charged particle tracking with multi-agent reinforcement learningMach. Learn.: Sci. Technol.7015021 doi:10.1088/2632-2153/ae352b

work page doi:10.1088/2632-2153/ae352b 2026
[46]

Zoccheddu S, Qasim S R, Owen P and Serra N 2026 Large Language Models for Physics Instrument Design doi:10.48550/arXiv.2601.07580

work page doi:10.48550/arxiv.2601.07580 2026
[47]

Nguyen X T, Chen L, Dorigo T, Gauger N R, Vischia P, Nardi F, Awais M, Hanif H, Abbas S and Kapoor R 2026 Differentiable Surrogate for Detector Simulation and Design with Diffusion Models doi:10.48550/arXiv.2601.07859

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.07859 2026
[48]

Hill J and Ryoo H J 2026 GRACE: an Agentic AI for Particle Physics Experiment Design and Simulation doi:10.48550/arXiv.2602.15039

work page doi:10.48550/arxiv.2602.15039 2026
[49]

SciCompKL 2024 CoDiPack (Code Differentiation Package) gitHub repository, accessed 2026- 01-20 URLhttps://github.com/SciCompKL/CoDiPack

2024
[50]

Methods Software331207–1231 doi:10.1080/10556788.2018.1471140

Sagebaum M, Albring T and Gauger N R 2018 Expression templates for primal value taping in the reverse mode of algorithmic differentiationOptim. Methods Software331207–1231 doi:10.1080/10556788.2018.1471140

work page doi:10.1080/10556788.2018.1471140 2018
[51]

M Sagebaum T Albring N G 2019 High-Performance Derivative Computations using CoDiPack ACM Trans. Math. Software (TOMS)45doi:10.1145/3356900

work page doi:10.1145/3356900 2019
[52]

com/mnovak42/hepemshow

Nov´ ak M 2023 HepEmShow gitHub repository, accessed 2026-01-20 URL https://github. com/mnovak42/hepemshow

2023
[53]

Urb´ an L 2006 A Model for Multiple Scattering in GEANT4 Tech. Rep. CERN-OPEN-2006-077 CERN URLhttps://cds.cern.ch/record/1004190 A Supplementary Results 18 fPropagation Stopgrad Fraction (%) Energy Steps Tracks 0.0 Track-only 9.3 20.2 11.8 Track + desc. 15.1 26.9 32.3 0.05 Track-only 9.6 20.4 11.9 Track + desc. 15.5 27.2 32.8 0.1 Track-only 9.9 20.7 12.2...

work page arXiv 2006