Pith. sign in

REVIEW 2 major objections 4 minor 32 references

Physics-Informed Neural Networks for High-Precision Grad-Shafranov Equilibrium Reconstruction

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read By training a second network on the first network's residual, a physics-informed neural network reproduces analytical Grad-Shafranov equilibria to errors of order $\mathcal{O}(10^{-8})$, three to four orders of magnitude below earlier…

desk verdict The two-stage PINN result looks real on two of three benchmarks, but the triangularity benchmark is internally inconsistent and the paper oversells its scope. read the letter →

arxiv 2507.16636 v1 pith:3IXHRCLP submitted 2025-07-22 physics.plasm-ph

Cuizhi Zhou , Kaien Zhu This is my paper · ORCID
classification physics.plasm-ph
keywords Physics-informedneuralnetworksGrad-Shafranovequationmulti-stagenetworkplasmaequilibriumreconstructiontokamakhigh-precisionPDEsolverLatinhypercubesampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a two-stage physics-informed neural network (PINN) can solve the Grad-Shafranov equation, the central PDE describing magnetically confined fusion plasma equilibrium, with errors of order $\mathcal{O}(10^{-8})$ against known analytical solutions. On three benchmark equilibria, the first network stage stalls at $\mathcal{O}(10^{-4})$ to $\mathcal{O}(10^{-5})$, and the second stage, trained on the residual of the first, lowers the error by three to four orders of magnitude. The authors present this as evidence that multi-stage PINNs can serve as a reliable and fast forward solver for plasma equilibrium reconstruction, with the post-training evaluation cost low enough to be relevant for real-time tokamak diagnostics. Earlier PINN-based Grad-Shafranov solvers plateaued near $\mathcal{O}(10^{-3})$ to $\mathcal{O}(10^{-4})$.

What carries the argument

The load-bearing object is the staged network sum $\psi \approx \sum_{i=0}^{n-1}\epsilon_i u_i$, in which the first network approximates the flux and each subsequent network is trained to fit the residual of the previous stages, with the normalization coefficient $\epsilon_i$ set by the scale of that residual. Training alternates the Adam and L-BFGS optimizers and regenerates the Latin-hypercube collocation points when switching optimizers, which the authors say prevents overfitting to a fixed set of sampling points. This decomposition targets the plateau problem: once the first network's loss stagnates, the next network still has a well-scaled target to learn, so the combined solution continues to improve.

What would settle it

Retrain the three benchmarks with, say, ten random seeds using the same stated architecture, optimizers, and collocation-count schedule, and record the maximum pointwise error on the paper's test grids; if the second-stage maximum error fails to reach the $10^{-8}$ to $10^{-7}$ range in most seeds, the central claim is not stable. A cheaper check is to evaluate the trained network on a finer or shifted test grid and see whether the $\mathcal{O}(10^{-8})$ error persists away from the training collocation points.

Watch

Extended reading notes

Core claim

The central discovery is that the multi-stage training strategy, in which the network output is $\psi=\sum_{i=0}^{n-1}\epsilon_i u_i$ and each later network $u_i$ learns the residual left by the previous stages, lifts PINN solutions of the Grad-Shafranov equation from $\mathcal{O}(10^{-4})$--$\mathcal{O}(10^{-5})$ accuracy to $\mathcal{O}(10^{-8})$ on analytical benchmark problems. The paper reports second-stage test errors of order $\mathcal{O}(10^{-7})$ for the rectilinear Solovév case, $\mathcal{O}(10^{-8})$ for the droplet-shaped case with an X-point, and $\mathcal{O}(10^{-8})$ for the Solovév case with triangularity, with the PDE residual loss driven to $\mathcal{O}(10^{-11})$ or below. Because the boundary values in all three tests are prescribed from the exact analytical solution, what is established is the forward-solving capability of the multi-stage PINN, not yet inverse reconstruction from measurements.

Load-bearing premise

The accuracy claim rests on the assumption that the single training run reported for each benchmark is representative, since no random seeds, run-to-run statistics, or explicit definitions of the reported error are given, and all boundary values come from the exact analytical solution.

Editorial extensions

If this is right

  • If the $\mathcal{O}(10^{-8})$ accuracy is reproducible, the trained two-stage network can act as a near-instant forward solver, replacing repeated expensive Grad-Shafranov solves inside a real-time equilibrium-reconstruction loop.
  • The droplet benchmark suggests the scheme works in geometries with an X-point, since the second stage clearly resolves the X-point after the first stage has already produced a recognizable equilibrium.
  • The construction does not limit the number of stages, so further stages should continue pushing the error toward machine precision, consistent with the multi-stage method on which the paper builds.
  • Because the network requires no mesh generation and evaluates solutions quickly, it is a candidate surrogate for generating many equilibria during discharge design or scenario optimization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not pursue is retraining each benchmark across several random seeds and reporting the spread of the maximum error; that spread is what would let a tokamak control group rely on the claimed accuracy.
  • Since all boundaries in Eqs. (13), (16), and (19) are taken from the analytical solution, extending the same two-stage scheme to inverse reconstruction from synthetic magnetic probe signals is the next test, and nothing in the reported losses guarantees the inversion would inherit the $\mathcal{O}(10^{-8})$ accuracy.
  • The two-stage idea could be combined with parameterized inputs such as inverse aspect ratio, elongation, and triangularity, producing a family of equilibrium surrogates useful for scenario scanning rather than only single-shot solving.
  • The values of $\epsilon_2$, $\lambda_1$, and $\lambda_2$ are not reported; a sensitivity scan of these coefficients would show whether the accuracy gain is robust or depends on a narrow hyperparameter window.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper applies a two-stage physics-informed neural network (PINN) framework, following Wang and Lai's multi-stage training scheme, to solve the Grad-Shafranov equation in three benchmark configurations that have known analytical solutions: the Solovév equilibrium, a droplet-shaped domain, and a triangularity-parameterized Solovév equilibrium. For each case the first-stage network is reported to reach errors of order 10^-4 to 10^-5, and the second-stage network is reported to reduce test-set errors to order 10^-7 or 10^-8, with PDE losses dropping to 10^-11 to 10^-16. The paper concludes that multi-stage PINNs provide a reliable tool for plasma equilibrium reconstruction.

Significance. If the numerical claims are reproducible, the paper provides a useful demonstration that the multi-stage PINN framework can push Grad-Shafranov solution errors well below the O(10^-4)-O(10^-5) values reported for single-stage PINNs. The choice of standard analytical solutions from the literature as benchmarks is appropriate, and the reported improvements between the first and second stages are internally plausible for the first two examples. However, the third benchmark is internally inconsistent as written, and the absence of error-metric definitions, seeds, run-to-run statistics, and code means the headline O(10^-8) accuracy is not verifiable from the manuscript alone.

major comments (2)
  1. [2.2.3, Eqs. (17)-(18)] The triangularity benchmark is not a valid verification as printed. Substituting Eq. (18) into the operator Delta* = d_xx + d_yy - epsilon/(1+epsilon) d_x stated in Eq. (17), with epsilon=3/10, lambda=0, a=1/2, b=7/10, R0=5/3, gives a y^2-dependent residual: the coefficient of y^2 in -Delta* psi is 2(a/b)^2(1-epsilon^2/4) epsilon^3(1-x)/(1+epsilon), which evaluates to approximately 0.0207(1-x). This cannot be matched by the right-hand side alpha (R0(1+epsilon x))^2 + beta for any constant alpha and beta. Independently, the printed formula for alpha evaluates to approximately 4.316 for the stated parameters, not approximately 0.45 as asserted in the text. The operator should presumably be d_xx + d_yy - epsilon/(1+epsilon x) d_x, and alpha must be recomputed consistently; with the equations as printed, the reported O(10^-8) agreement in Figure 5 cannot be interpreted as a successful verification.
  2. [2.1-2.2, Eqs. (5)-(6) and Figs. 2, 4, 5] The central quantitative claim is a test-set error of order 10^-8, but the manuscript never defines the error metric used in the error plots (e.g., maximum absolute error, L2 error, relative error, or the grid on which it is evaluated), reports no seeds or run-to-run statistics, and provides no code. In addition, the loss weights lambda1 and lambda2 in Eq. (5) and the second-stage normalization coefficient epsilon2 in Eq. (6) are never specified, although the text states that they are adjusted per stage. As a result, the reported PDE loss values (10^-11 to 10^-16) and the claimed test-set accuracy cannot be independently checked, and it is not possible to determine whether the single reported run is representative or whether the loss decrease reflects overfitting to the regenerated collocation points.
minor comments (4)
  1. [1, introductory paragraph] The Grad-Shafranov equation is a second-order quasilinear elliptic PDE, so describing it as a "high-order nonlinear partial differential equation" is inaccurate.
  2. [Abstract and 2.2] The Abstract promises an error magnitude of O(10^-8) for the second-stage output, while Section 2.2 states that the test-set errors reach orders of 10^-7 and 10^-8; the Solovév example specifically reports O(10^-7), so the abstract should be qualified accordingly.
  3. [Title and Abstract] The phrase "equilibrium reconstruction" normally denotes inferring equilibrium from measured signals, but all three benchmarks are forward solutions with boundary conditions prescribed from the exact analytical solution and known right-hand sides. The demonstrated capability is high-precision forward solving of the Grad-Shafranov equation, and the title and abstract should be adjusted to avoid overclaiming inverse reconstruction.
  4. [5, Data availability] For a computational methods paper, a statement that data are available "upon reasonable request" is insufficient; providing the exact network configuration, training data, seeds, and code, or a permanent repository link, would be necessary for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported accuracies are benchmarked against independent analytical solutions using an external multi-stage PINN method.

full rationale

The paper does not derive its central claim from a fitted input or from a self-referential uniqueness argument. The multi-stage PINN framework is attributed to Wang (Ref. [24]), which is not the present authors' own work, so there is no self-citation chain carrying the core method. The three benchmark problems use analytical solutions taken from independent literature (Youjun Hu's notes and Deriaz et al.), and the reported errors are measured directly against those analytical solutions. Boundary conditions are prescribed from the exact solutions, which is a standard verification setup rather than a circular prediction: the network still has to solve the PDE in the interior, and there is no guaranteed O(1e-8) accuracy from that setup. No fitted parameter is renamed as a prediction, and no result is defined in terms of the quantity it claims to produce. The internal inconsistency in the third benchmark (Eq. (18) does not satisfy Eq. (17) as printed) is a correctness or reproducibility concern, not a circularity concern, because it does not reduce the benchmark to the method's inputs. Therefore no circular step is identified and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The core result is an empirical numerical demonstration. It relies on standard PDE and benchmark assumptions, on the transferability of the Wang-Lai multi-stage convergence result, and on unstated training choices. No new entities are introduced and no parameter is fitted to external data; the main gaps are unspecified hyperparameters and the absence of code.

free parameters (3)
  • Loss weights lambda1, lambda2 = Not reported
    Eq. (5) defines L_total = lambda1*L_pde + lambda2*L_bd, with values never given; the reported error and loss curves depend on them.
  • Second-stage normalization coefficient epsilon2 = Not reported
    Eq. (6) sums epsilon_i*u_i; the text says epsilon2 is determined from the first-stage residual but gives no value, so the second-stage training cannot be reproduced as specified.
  • Network and training configuration = 3 hidden layers x 20 neurons, tanh, LR 1e-3, batch 100, LHS sampling, 2001/4002 collocation points
    Chosen by hand; the central numerical results depend on these choices, which are standard but not derived.
assumptions (3)
  • standard math The analytical solutions in Eqs. (10), (15), and (18) are exact solutions of the respective Grad-Shafranov forms.
    Used as ground truth for boundary conditions and for computing reported errors; taken from Refs. [27] and [31].
  • domain assumption The two-stage training protocol from Wang and Lai [24] converges to the reported O(10^-8) accuracy when applied to the GS equation.
    The paper does not analyze convergence guarantees, robustness to initialization, or scaling to harder equilibria; it assumes the cited method transfers.
  • domain assumption The reported loss and error values are measured on test sets independent of the training collocation points.
    Stated for the first example (test set 100x100 differs from training set) and implied for the others, but no error metric or uncertainty is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Informed Neural Networks for High-Precision Grad-Shafranov Equilibrium Reconstruction." pith.science (2026). https://pith.science/paper/3IXHRCLP

@misc{pith2026250716636,
  author       = {Pith},
  title        = {Pith review of: Physics-Informed Neural Networks for High-Precision Grad-Shafranov Equilibrium Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3IXHRCLP}},
  note         = {Machine review of arXiv:2507.16636}
}
abstract

The equilibrium reconstruction of plasma is a core step in real-time diagnostic tasks in fusion research. This paper explores a multi-stage Physics-Informed Neural Networks(PINNs) approach to solve the Grad-Shafranov equation, achieving high-precision solutions with an error magnitude of $O(10^{-8})$ between the output of the second-stage neural network and the analytical solution. Our results demonstrate that the multi-stage PINNs provides a reliable tool for plasma equilibrium reconstruction.

Figures

Figures reproduced from arXiv: 2507.16636 by the authors.

Figure 1
Figure 1. Multi-stage PINNs architecture for Grad-Shafranov equation. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Solution results of the Solovév equilibrium via Physics-Informed Neural Networks. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of point distributions: Latin hypercube method vs pseudo-random [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Solution results of the Grad-Shafranov equation in drop domain. (a) Solution [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Solution results of the Solovév equilibrium with triangularity parameter via [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 19 canonical work pages

  1. [24]

    Wang, C.-Y

    Y. Wang, C.-Y. Lai, Multi-stage neural networks: Function approxima- tor of machine precision, Journal of Computational Physics 504 (2024) 112865.doi:https://doi.org/10.1016/j.jcp.2024.112865

    arXiv 2024

  2. [1]

    C. L. Smith, S. Cowley, The path to fusion power, Phil. Trans. R. Soc. A 368 (2010) 1091–1108.doi:10.1098/rsta.2009.0216

    work page arXiv 2010

  3. [2]

    H. Grad, H. Rubin, Hydromagnetic equilibria and force-free fields, Jour- nal of Nuclear Energy 7 (1958) 284–285, (1954)

    work page 1958

  4. [3]

    Shafranov, On magnetohydrodynamical equilibrium configurations, Soviet Physics JETP 6 (1958) 1013

    V. Shafranov, On magnetohydrodynamical equilibrium configurations, Soviet Physics JETP 6 (1958) 1013

    work page 1958

  5. [4]

    L. Lao, H. S. John, Q. Peng, J. Ferron, E. Strait, T. Taylor, W. Meyer, C.Zhang, K.You, Mhdequilibriumreconstructioninthediii-dtokamak, Fusion science and technology 48 (2005) 968–977. 11

    work page 2005

  6. [5]

    Albanese, A

    R. Albanese, A. Iaiunese, P. Zumbolo, Solution of grad-shafranov equa- tion with linear triangular finite elements providing magnetic field con- tinuity with a second order accuracy, Computer Physics Communica- tions291(2023)108804.doi:https://doi.org/10.1016/j.cpc.2023. 108804

    doi:10.1016/j.cpc.2023 2023

  7. [6]

    Ricketson, A

    L. Ricketson, A. Cerfon, M. Rachh, J. Freidberg, Accurate derivative evaluation for any grad–shafranov solver, Journal of Computational Physics 305 (2016) 744–757.doi:https://doi.org/10.1016/j.jcp. 2015.11.015

    doi:10.1016/j.jcp 2016

  8. [7]

    Heumann, F

    H. Heumann, F. Rapetti, A finite element method with overlapping meshes for free-boundary axisymmetric plasma equilibria in realistic geometries, Journal of Computational Physics 334 (2017) 522–540. doi:https://doi.org/10.1016/j.jcp.2017.01.006

    doi:10.1016/j.jcp.2017.01.006 2017

Show all 32 references
  1. [8]

    H. Li, P. Zhu, Solving the grad–shafranov equation using spectral ele- mentsfor tokamak equilibriumwith toroidalrotation, ComputerPhysics Communications 260 (2021) 107264.doi:https://doi.org/10.1016/ j.cpc.2020.107264

    2021

  2. [9]

    Palha, B

    A. Palha, B. Koren, F. Felici, A mimetic spectral element solver for the grad–shafranov equation, Journal of Computational Physics 316 (2016) 63–93.doi:https://doi.org/10.1016/j.jcp.2016.04.002

    2016 doi

  3. [10]

    doi:https://doi.org/10.1016/j.cpc.2014.02.008

    E.Howell, C.Sovinec, Solvingthegrad–shafranovequationwithspectral elements, Computer Physics Communications 185 (5) (2014) 1415–1421. doi:https://doi.org/10.1016/j.cpc.2014.02.008

    2014 doi

  4. [11]

    Fambri, E

    F. Fambri, E. Zampa, S. Busto, L. Río-Martín, F. Hindenlang, E. Son- nendrücker, M. Dumbser, A well-balanced and exactly divergence- free staggered semi-implicit hybrid finite volume / finite element scheme for the incompressible mhd equations, Journal of Computa- tional Physi...

    2023

  5. [12]

    Zampa, S

    E. Zampa, S. Busto, M. Dumbser, A divergence-free hybrid finite vol- ume / finite element scheme for the incompressible mhd equations based on compatible finite element spaces with a posteriori limiting, Applied 12 Numerical Mathematics 198 (2024) 346–374.doi:https://doi.org/ ...

    2024 doi

  6. [13]

    G.Karniadakis, I.Kevrekidis, L.Lu, Physics-informedmachinelearning, Nat Rev Phys 3 (2021) 422–440.doi:10.1038/s42254-021-00314-5

    2021 doi

  7. [14]

    M.Raissi, P.Perdikaris, G.E.Karniadakis, Physicsinformeddeeplearn- ing (part i): Data-driven solutions of nonlinear partial differential equa- tions, arXiv preprint arXiv:1711.10561 (2017)

    2017 arXiv

  8. [15]

    Raissi, P

    M. Raissi, P. Perdikaris, G. Karniadakis, Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics 378 (2019) 686–707.doi:https://doi.org/ 10.1016...

    2019 doi

  9. [16]

    B. P. van Milligen, V. Tribaldos, J. A. Jiménez, Neural network differen- tial equation and plasma equilibrium solver, Phys. Rev. Lett. 75 (1995) 3594–3597.doi:10.1103/PhysRevLett.75.3594

    1995 doi

  10. [17]

    Joung, J

    S. Joung, J. Kim, S. Kwak, J. Bak, S. Lee, H. Han, H.-S. Kim, G. Lee, D. Kwon, Y.-C. Ghim, Deep neural network grad–shafranov solver con- strained with measured magnetic signals, Nuclear Fusion 60 (1) (2019) 016034

    2019

  11. [18]

    Merlo, D

    A. Merlo, D. Böckenhoff, J. Schilling, U. Höfel, S. Kwak, J. Svensson, A. Pavone, S. A. Lazerson, T. S. Pedersen, et al., Proof of concept of a fast surrogate model of the vmec code via neural networks in wendelstein 7-x scenarios, Nuclear Fusion 61 (9) (2021) 096039

    2021

  12. [19]

    J. Wai, M. Boyer, E. Kolemen, Neural net modeling of equilibria in nstx-u, Nuclear Fusion 62 (8) (2022) 086042

    2022

  13. [20]

    Y. Liu, C. Akcay, L. L. Lao, X. Sun, Surrogate models for plasma dis- placement and current in 3d perturbed magnetohydrodynamic equilib- ria in tokamaks, Nuclear Fusion 62 (12) (2022) 126067.doi:10.1088/ 1741-4326/ac9d4c

    2022

  14. [21]

    H. Baty, V. Vigon, Modelling solar coronal magnetic fields with physics- informed neural networks, Monthly Notices of the Royal Astronomical Society 527 (2) (2023) 2575–2584.doi:10.1093/mnras/stad3320. 13

    2023 doi

  15. [22]

    B. Jang, A. A. Kaptanoglu, R. Gaur, S. Pan, M. Landreman, W. Dor- land, Grad–shafranov equilibria via data-free physics informed neural networks, Physics of Plasmas 31 (3) (2024)

    2024

  16. [23]

    F. N. Rizqan, M. Hole, C. Gretton, Evaluation and verification of physics-informed neural models of the grad-shafranov equation, arXiv preprint arXiv:2504.21155 (2025)

    2025 arXiv

  17. [25]

    Y. Wu, S. Şahin, M. Wang, Q. Yang, Nuclear energy production: Fusion energy, in: Reference Module in Earth Systems and Envi- ronmental Sciences, Elsevier, 2024.doi:https://doi.org/10.1016/ B978-0-44-313219-3.00019-8

    2024

  18. [26]

    Krishnapriyan, A

    A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. W. Mahoney, Char- acterizing possible failure modes in physics-informed neural networks, in: M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, J. W. Vaughan (Eds.), Advances in Neural Information Processing Systems, Vol. 34, C...

    2021

  19. [27]

    Hu, Notes on tokamak equilibrium*

    Y. Hu, Notes on tokamak equilibrium*. URLhttps://youjunhu.github.io/research_notes/tokamak_ equilibrium.pdf

  20. [28]

    Loh, On latin hypercube sampling, The Annals of Statistics 24 (5) (1996) 2058–2080.doi:10.1214/aos/1069362310

    W.-L. Loh, On latin hypercube sampling, The Annals of Statistics 24 (5) (1996) 2058–2080.doi:10.1214/aos/1069362310

    1996

  21. [29]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, CoRR abs/1412.6980 (2014). URLhttps://api.semanticscholar.org/CorpusID:6628106

    2014 arXiv

  22. [30]

    D. C. Liu, J. Nocedal, On the limited memory BFGS method for large scale optimization, Mathematical Programming 45 (1) (1989) 503–528. doi:10.1007/BF01589116. URLhttps://doi.org/10.1007/BF01589116 14

    1989 doi

  23. [31]

    Deriaz, B

    E. Deriaz, B. Despres, G. Faccanoni, K. P. Gostaf, L.-M. Imbert-Gérard, G. Sadaka, R. Sart, Magnetic equations with FreeFem++: The Grad- Shafranov equation & the current hole, ESAIM: Proceedings 32 (2011) 76–94.doi:https://doi.org/10.1051/proc/2011013

    2011

  24. [32]

    H. Baty, A hands-on introduction to Physics-Informed Neural Net- works for solving partial differential equations with benchmark tests taken from astrophysics and plasma physics, arXiv e-prints (2024) arXiv:2403.00599doi:10.48550/arXiv.2403.00599. 15

    2024 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.