pith. sign in

arxiv: 2606.12567 · v1 · pith:NHGDKGL3new · submitted 2026-06-10 · ⚛️ physics.plasm-ph

Hierarchical Framework of Runaway Electrons using Deep Learning

Tyler Mark , Christopher McDevitt This is my paper

Pith reviewed 2026-06-27 07:46 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph
keywords runaway electronsphysics-informed neural networksadjoint methodplasma kineticssurrogate modelingdeep learningfusion plasmas
0
0 comments X

The pith

Adjoint formulation with physics-informed neural networks enables orders of magnitude faster predictions of runaway electron kinetics for arbitrary initial distributions.

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

The paper develops an adjoint deep learning framework to track the time evolution of runaway electron fluid moments and energy distributions in plasmas. A careful adjoint setup combined with physics-informed neural networks produces surrogates that work from any starting electron distribution and across a wide span of plasma conditions. This matters because conventional kinetic solvers are computationally heavy, limiting their use in large-scale or repeated modeling tasks. The resulting networks recover runaway current, average energy, and full energy distribution while matching a traditional solver.

Core claim

The authors formulate an adjoint problem for the runaway electron kinetic equations that permits computing the evolution of fluid moments and the energy distribution from arbitrary initial conditions. They then train three physics-informed neural networks to act as fast surrogates for the runaway current, average energy, and full energy distribution. These networks are shown to reproduce results from a conventional runaway electron solver across a broad parameter space.

What carries the argument

Adjoint formulation of the runaway electron kinetic equations paired with physics-informed neural networks trained to evolve the population moments and distribution.

If this is right

  • The surrogates resolve a broad range of plasma parameters.
  • Predictions of RE current, average energy, and energy distribution agree with traditional solvers.
  • The approach supports arbitrary initial electron distributions.
  • Computation is orders of magnitude faster than traditional methods.

Where Pith is reading between the lines

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

  • Such fast surrogates could enable coupling RE models into larger device-scale simulations that were previously intractable.
  • The hierarchical structure might allow consistent switching between fluid and kinetic descriptions within the same framework.
  • Similar adjoint-PINN techniques could be applied to other kinetic problems in plasmas or related fields.

Load-bearing premise

The neural network surrogates accurately represent the underlying kinetic equations even when initial distributions and plasma parameters differ from those used in training.

What would settle it

Running the PINN surrogates on an initial distribution far from the training data and finding that the predicted energy distribution deviates significantly from a high-fidelity traditional solver at later times.

Figures

Figures reproduced from arXiv: 2606.12567 by Christopher McDevitt, Tyler Mark.

Figure 1
Figure 1. Figure 1: Example of a JONTA RE Monte Carlo simulation where [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Panels (a–e) evaluate the CPF for E∥ = 2.5, Zef f = 3, α = 0.1. Panels (a) and (d) show the CPF and residual evaluated at the terminal condition; panels (b) and (e) show the CPF evaluated at τ /τc = 20; panel (c) shows the predicted runaway current for various initial pitch orientations f(ξ0) and an energy distribution as discussed in Sec. III D—PINN [solid], Monte Carlo with frozen particles [dashed], and… view at source ↗
Figure 3
Figure 3. Figure 3: Runaway current versus time for variations in [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Panels (a–e) evaluate the EPF for E∥ = 2.5, Zef f = 3, α = 0.1. Panels (a) and (d) show the EPF and residual at the terminal condition (τ /τc = 0); panels (b) and (e) show them at (τ /τc = 20). Panel (c) shows the evolution of the average energy for various initial pitch orientations f(ξ0) and an energy distribution as discussed in Sec. III D—PINN [solid], Monte Carlo with frozen particles [dashed], and Mo… view at source ↗
Figure 5
Figure 5. Figure 5: Runaway average energy versus time for variations in [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Panels (a) and (d) show a sharp Heaviside terminal condition at [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Panels (a)–(c): Runaway energy distributions for a gaussian and isotropic initialization discussed [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
read the original abstract

We present an adjoint deep learning framework describing the evolution of fluid moments and the energy distribution of the runaway electron (RE) population. We demonstrate that a careful formulation of the adjoint problem allows for the temporal evolution of these quantities for arbitrary initial electron distributions, and in combination with a physics-informed neural network (PINN), we show that the resulting surrogates can resolve a broad range of plasma parameters. This combination of the adjoint formulation and rapid inference of neural networks enables orders of magnitude faster predictions of RE kinetics than traditional methods. Here, we detail the mathematical formulation and the design of three PINNs which recover the temporal evolution of the RE current, average energy and energy distribution. Predictions are validated against a traditional RE solver, with good agreement across a broad range of scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents an adjoint deep learning framework for modeling the temporal evolution of runaway electron (RE) fluid moments and energy distributions in plasmas. It formulates three physics-informed neural networks (PINNs) to recover the RE current, average energy, and energy distribution, asserting that the adjoint problem enables predictions for arbitrary initial electron distributions across a broad range of plasma parameters. The approach is claimed to yield orders of magnitude faster inference than traditional solvers, with validation showing good agreement against a conventional RE code.

Significance. If the surrogates prove faithful to the underlying kinetic equations, the method could accelerate RE studies in fusion and astrophysical plasmas by enabling rapid parameter sweeps and real-time predictions. The adjoint-PINN combination addresses a genuine computational bottleneck, and the hierarchical structure (moments plus distribution) is a reasonable design choice. However, the absence of quantitative validation metrics means the practical significance cannot yet be assessed.

major comments (2)
  1. [Abstract] Abstract: the assertion of 'good agreement' with a traditional RE solver is unsupported by any error metrics (L2 norms, relative errors, or maximum deviations), training-loss curves, or explicit ranges of initial distributions and plasma parameters tested. This directly undermines the central claim that the three PINNs remain faithful for arbitrary initial conditions.
  2. [Abstract] Abstract: no details are supplied on the adjoint formulation's implementation (e.g., how the adjoint source terms are constructed or how the PINN loss enforces the adjoint equations), making it impossible to verify that the speedup does not come at the cost of hidden parameter tuning or restricted applicability.
minor comments (1)
  1. [Abstract] The abstract refers to 'a broad range of scenarios' without defining the parameter space (density, temperature, electric field, etc.), which should be quantified in the validation section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive suggestions. We address each major comment below and will revise the abstract accordingly to strengthen the presentation of our results.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion of 'good agreement' with a traditional RE solver is unsupported by any error metrics (L2 norms, relative errors, or maximum deviations), training-loss curves, or explicit ranges of initial distributions and plasma parameters tested. This directly undermines the central claim that the three PINNs remain faithful for arbitrary initial conditions.

    Authors: We agree that the abstract would benefit from explicit quantitative support. The full manuscript (Sections 4 and 5, Figures 4–7, and Table 2) reports L2 norms, relative errors (typically <3–5% across cases), maximum deviations, training-loss curves, and the tested ranges of initial distributions and plasma parameters. We will revise the abstract to include representative error metrics and parameter ranges so that the validation claim is self-contained. revision: yes

  2. Referee: [Abstract] Abstract: no details are supplied on the adjoint formulation's implementation (e.g., how the adjoint source terms are constructed or how the PINN loss enforces the adjoint equations), making it impossible to verify that the speedup does not come at the cost of hidden parameter tuning or restricted applicability.

    Authors: The adjoint formulation, source-term construction, and enforcement within the PINN loss are derived and explained in Sections 2.2–2.3 and 3.1–3.2 of the manuscript. We acknowledge that the abstract is too terse on this point and will add a brief clause describing the adjoint approach and its role in enabling arbitrary initial conditions without additional tuning. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on external validation against traditional solver

full rationale

The paper describes an adjoint formulation combined with PINNs to generate surrogates for RE current, average energy, and distribution, with validation performed against a separate traditional RE solver showing good agreement. No equations, fitted parameters, self-citations, or ansatzes are presented that reduce any claimed prediction or result to the inputs by construction. The speedup is asserted relative to external traditional methods rather than through internal redefinition or self-referential fitting, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the framework implicitly assumes the underlying kinetic model is correctly encoded in the PINN loss and that adjoint propagation preserves accuracy for arbitrary initials.

pith-pipeline@v0.9.1-grok · 5651 in / 1058 out tokens · 17363 ms · 2026-06-27T07:46:13.688558+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 11 canonical work pages

  1. [1]

    Bandaru, M

    V . Bandaru, M. Hoelzl, C. Reux, O. Ficker, S. Silburn, M. Lehnen, N. Eidietis, and J. Team, Plasma Physics and Controlled Fusion63, 035024 (2021), URLhttps://iopscience.iop.org/ article/10.1088/1361-6587/abdbcf

    work page doi:10.1088/1361-6587/abdbcf 2021

  2. [2]

    C. Liu, C. Zhao, S. C. Jardin, N. M. Ferraro, C. Paz-Soldan, Y . Liu, and B. C. Lyons, Plasma Physics and Controlled Fusion63, 125031–125031 (2021), URLhttps://iopscience.iop. org/article/10.1088/1361-6587/ac2af8

    work page doi:10.1088/1361-6587/ac2af8 2021

  3. [3]

    A. P. Sainterme and C. R. Sovinec, Physics of Plasmas31(2024), URL https://pubs.aip.org/aip/pop/article/31/1/010701/2932419/ Resistive-hose-modes-in-tokamak-runaway-electron

    2024

  4. [4]

    R. W. Harvey, V . S. Chan, S. C. Chiu, T. E. Evans, M. N. Rosenbluth, and D. G. Whyte, Physics of Plasmas7, 4590–4599 (2000), URLhttps://pubs.aip.org/aip/pop/article/7/11/ 4590/264474/Runaway-electron-production-in-DIII-D-killer

    2000

  5. [5]

    Nilsson, J

    E. Nilsson, J. Decker, Y . Peysson, R. S. Granetz, F. Saint-Laurent, and M. Vlainic, Plasma Physics and Controlled Fusion57, 095006–095006 (2015), URLhttps://ui.adsabs.harvard.edu/ abs/2015PPCF...57i5006N/abstract

    2015

  6. [6]

    C. J. McDevitt, Z. Guo, and X.-Z. Tang, Plasma Physics and Controlled Fusion61, 054008 (2019), URLhttps://iopscience.iop.org/article/10.1088/1361-6587/ab0d6d

    work page doi:10.1088/1361-6587/ab0d6d 2019

  7. [7]

    Hoppe, O

    M. Hoppe, O. Embreus, and T. Fulop, Computer Physics Communications268, 108098 (2021), URL https://www.sciencedirect.com/science/article/pii/S0010465521002101

    2021

  8. [8]

    Beidler, D

    M. Beidler, D. del Castillo-Negrete, D. Shiraki, L. Baylor, E. Hollmann, and C. Lasnier, Nuclear Fusion64, 076038 (2024), URLhttps://iopscience.iop.org/article/10.1088/ 1741-4326/ad4c77/meta

    2024

  9. [9]

    C. Zhao, C. Liu, S. C. Jardin, and N. M. Ferraro, Nuclear Fusion60, 126017–126017 (2020), URL https://iopscience.iop.org/article/10.1088/1741-4326/ab96f4

    work page doi:10.1088/1741-4326/ab96f4 2020

  10. [10]

    Bandaru, M

    V . Bandaru, M. Hoelzl, F. J. Artola, O. Vallhagen, and M. Lehnen, Physics of Plasmas31 (2024), URLhttps://pubs.aip.org/aip/pop/article/31/8/082503/3306459/ Runaway-electron-fluid-model-extension-in-JOREK

    2024

  11. [11]

    M. B. Giles and N. A. Pierce, Flow Turbulence and Combustion65, 393–415 (2000), URLhttps: 24 //link.springer.com/article/10.1023/A:1011430410075

    work page doi:10.1023/a:1011430410075 2000

  12. [12]

    M. B. Giles and N. A. Pierce, Fluid Dynamics and Co-located Conferences (2024), URLhttps: //arc.aiaa.org/doi/10.2514/6.1997-1850

    work page doi:10.2514/6.1997-1850 2024

  13. [13]

    Bui-Thanh,Adjoint and its roles in sciences, engineering, and mathematics: A tutorial(2023), URL https://arxiv.org/abs/2306.09917

    T. Bui-Thanh,Adjoint and its roles in sciences, engineering, and mathematics: A tutorial(2023), URL https://arxiv.org/abs/2306.09917

    arXiv 2023

  14. [14]

    C. J. McDevitt, J. S. Arnaud, and X.-Z. Tang, Physics of Plasmas32(2025), URLhttps://pubs.aip.org/aip/pop/article/32/4/042503/3342014/ A-physics-constrained-deep-learning-treatment-of

    2025

  15. [15]

    C. J. McDevitt, J. S. Arnaud, and X.-Z. Tang, Physics of Plasmas32(2025), URLhttps://pubs.aip.org/aip/pop/article/32/4/042504/3342096/ An-efficient-surrogate-model-of-secondary-electron

    2025

  16. [16]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. Karniadakis, Journal of Computational Physics378, 686–707 (2019), URLhttps://www.sciencedirect.com/science/article/abs/ pii/S0021999118307125

    2019

  17. [17]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Nature Reviews Physics3, 422 (2021)

    2021

  18. [18]

    J. S. Arnaud, X.-Z. Tang, and C. J. McDevitt, Nuclear Fusion65, 106013 (2025)

    2025

  19. [19]

    Risken, SpringerLink (2019), URLhttps://link.springer.com/book/10.1007/ 978-3-642-61544-3

    H. Risken, SpringerLink (2019), URLhttps://link.springer.com/book/10.1007/ 978-3-642-61544-3

    2019

  20. [20]

    Z. Guo, C. J. McDevitt, and X.-Z. Tang, Plasma Physics and Controlled Fusion59, 044003 (2017), URLhttps://iopscience.iop.org/article/10.1088/1361-6587/aa5952

    work page doi:10.1088/1361-6587/aa5952 2017

  21. [21]

    Connor and R

    J. Connor and R. Hastie, Nuclear Fusion15, 415–424 (1975), URLhttps://iopscience.iop. org/article/10.1088/0029-5515/15/3/007

    work page doi:10.1088/0029-5515/15/3/007 1975

  22. [22]

    C. F. Karney and N. J. Fisch, The Physics of Fluids29, 180–192 (1986), URLhttps://pubs.aip.org/aip/pfl/article/29/1/180/944104/ Current-in-wave-driven-plasmas

    1986

  23. [23]

    T. M. Antonsen and K. R. Chu, The Physics of Fluids25, 1295–1296 (1982), URLhttps://pubs.aip.org/aip/pfl/article/25/8/1295/841467/ Radio-frequency-current-generation-by-waves-in

    1982

  24. [24]

    Taguchi, Journal of the Physical Society of Japan52, 2035–2040 (1983), URLhttps:// journals.jps.jp/doi/10.1143/JPSJ.52.2035?mobileUi=0

    M. Taguchi, Journal of the Physical Society of Japan52, 2035–2040 (1983), URLhttps:// journals.jps.jp/doi/10.1143/JPSJ.52.2035?mobileUi=0. 25

    work page doi:10.1143/jpsj.52.2035 2035

  25. [25]

    C. Liu, D. P. Brennan, A. H. Boozer, and A. Bhattacharjee, Plasma Physics and Controlled Fusion59, 024003 (2016)

    2016

  26. [26]

    J. S. Arnaud, T. Mark, and C. J. McDevitt,A physics-constrained deep learning surrogate model of the runaway electron avalanche growth rate(2024), URLhttps://arxiv.org/abs/2403. 04948

    2024

  27. [27]

    Zhang and D

    G. Zhang and D. del Castillo-Negrete, Physics of Plasmas24(2017), URL https://pubs.aip.org/aip/pop/article/24/9/092511/107983/ A-backward-Monte-Carlo-method-for-time-dependent

    2017

  28. [28]

    L. Sun, H. Gao, S. Pan, and J.-X. Wang, Computer Methods in Applied Mechanics and Engineering 361, 112732 (2020)

    2020

  29. [29]

    McDevitt, E

    C. McDevitt, E. Fowler, and S. Roy, inAIAA Scitech 2024 F orum(2024), p. 1692

    2024

  30. [30]

    Lagaris, A

    I. Lagaris, A. Likas, and D. Fotiadis, IEEE Transactions on Neural Networks9, 987–1000 (1998), URLhttps://arxiv.org/abs/physics/9705023

    Pith/arXiv arXiv 1998

  31. [31]

    Karpatne, G

    A. Karpatne, G. Atluri, J. H. Faghmous, M. Steinbach, A. Banerjee, A. Ganguly, S. Shekhar, N. Sam- atova, and V . Kumar, IEEE Transactions on Knowledge and Data Engineering29, 2318–2331 (2017), URLhttps://ieeexplore.ieee.org/document/7959606

    arXiv 2017

  32. [32]

    Lusch, J

    B. Lusch, J. N. Kutz, and S. L. Brunton, Nature Communications9(2018), URLhttps://www. nature.com/articles/s41467-018-07210-0

    2018

  33. [33]

    R. Wang, K. Kashinath, M. Mustafa, A. Albert, and R. Yu, Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining p. 1457–1466 (2020), URL https://dl.acm.org/doi/10.1145/3394486.3403198

    work page doi:10.1145/3394486.3403198 2020

  34. [34]

    N. Vyas, D. Morwani, R. Zhao, M. Kwun, I. Shapira, D. Brandfonbrener, L. Janson, and S. Kakade, Soap: Improving and stabilizing shampoo using adam(2024), URLhttps://arxiv.org/abs/ 2409.11321

    Pith/arXiv arXiv 2024

  35. [35]

    C. Wu, M. Zhu, Q. Tan, Y . Kartha, and L. Lu, Computer Methods in Applied Mechanics and Engi- neering403, 115671–115671 (2022), URLhttps://arxiv.org/abs/2207.10289

    arXiv 2022

  36. [36]

    C. J. McDevitt and J. S. Arnaud, Journal of Plasma Physics92(2026), URLhttps: //www.cambridge.org/core/journals/journal-of-plasma-physics/ article/an-adjoint-formulation-of-energetic-particle-confinement/ B8D26545ABD691BFC5C01F9F26038404

    2026

  37. [37]

    A. H. Boozer and G. Kuo-Petravic, The Physics of Fluids24, 851 (1981). 26

    1981

  38. [38]

    Decker, E

    J. Decker, E. Hirvijoki, O. Embreus, Y . Peysson, A. Stahl, I. Pusztai, and T. Fulop, Plasma Physics and Controlled Fusion58, 025016 (2016), URLhttps://iopscience.iop.org/article/ 10.1088/0741-3335/58/2/025016

    work page doi:10.1088/0741-3335/58/2/025016 2016

  39. [39]

    Hesslow,Kinetic modeling of runaway-electron dynamics in partially ionized plas- mas(2000), URLhttps://research.chalmers.se/publication/518256/file/ 518256_Fulltext.pdf

    L. Hesslow,Kinetic modeling of runaway-electron dynamics in partially ionized plas- mas(2000), URLhttps://research.chalmers.se/publication/518256/file/ 518256_Fulltext.pdf

    2000

  40. [40]

    Hesslow, O

    L. Hesslow, O. Embr ´eus, A. Stahl, T. DuBois, G. Papp, S. Newton, and T. F ¨ul¨op, Physical Review Letters118(2017)

    2017

  41. [41]

    Hesslow, O

    L. Hesslow, O. Embr ´eus, M. Hoppe, T. C. DuBois, G. Papp, M. Rahm, and T. F ¨ul¨op, Journal of Plasma Physics84(2018), URLhttps://www. cambridge.org/core/journals/journal-of-plasma-physics/article/ generalized-collision-operator-for-fast-electrons-interacting-with-partially-ionized-impurities/ 085B7218B631AA7056B79C1CF66D79F5. 27

    2018