{"id":"b348462d-2c37-4533-b1ce-295a1a144fd3","arxiv_id":"2507.16636","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Two-stage PINNs solve three analytical Grad-Shafranov benchmarks to O(10^-8) accuracy, far below the 10^-3 to 10^-4 errors of earlier PINN solvers cited in the paper.","lead":"This paper applies a two-stage physics-informed neural network to three Grad-Shafranov plasma equilibrium problems with known answers, reporting errors around 10^-8, much smaller than earlier PINN-based solvers. The result is a benchmark demonstration, not a production tool: no code is released and the tests use ideal boundaries, not experimental data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The third benchmark is internally inconsistent: Eq. (18) does not satisfy Eq. (17) as printed with the stated parameters, so its reported O(1e-8) error is not a valid verification.","rationale":"The central claim is the O(1e-8) accuracy of the two-stage PINN on three analytically solvable Grad-Shafranov benchmarks. The reader's conditional verdict focused on missing seeds, code, and error-metric definitions. Those are valid reproducibility concerns, but the more load-bearing issue is that one of the three benchmarks appears internally inconsistent: the stated analytical solution and the stated PDE do not match. Because the third benchmark is one of only three pieces of evidence for the headline accuracy, a single invalid benchmark materially weakens the claim. The concern is checkable by direct symbolic substitution, so it should be resolved before relying on the paper. If the substitution passes after correcting a typo, for example replacing 1+ε with 1+εx and α≈0.45 with α≈4.316, the benchmark may be valid and the conditional verdict can stand. I therefore keep the reader's verdict unchanged: the final recommendation remains CONDITIONAL, but the condition is now specifically the internal consistency of Eqs. (17) and (18), not merely the availability of code or seeds.","tokens_in":7699,"tokens_out":21208,"duration_ms":199200,"concrete_test":"Evaluate the residual R(x,y)=−Δ*ψ − [α(R0(1+εx))²+β] at (x,y)=(0.5,0.5) using Eqs. (17) and (18) with the stated parameters and the operator exactly as printed in Eq. (17); if |R| exceeds 1e-6 relative to the dominant term, the benchmark is internally inconsistent and the reported O(1e-8) error for this case is not meaningful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the triangularity benchmark, substitute Eq. (18) into Eq. (17) using the operator printed in Eq. (17), Δ* = ∂xx + ∂yy − ε/(1+ε)∂x, with ε=3/10, λ=0, a=1/2, b=7/10, R0=5/3. Writing p=ε, k=1−p²/4, m=(a/b)², and B=m k(1+px)², the y² coefficient of −Δ*ψ is B_xx − p/(1+p) B_x = 2 m k p³(1−x)/(1+p) ≈ 0.0207(1−x), while the proposed RHS α(R0(1+εx))² + β has no y² dependence. Hence Eq. (18) is not a solution of Eq. (17) for any constant α, β. Independently, the printed formula for α evaluates to ≈4.316, not ≈0.45; Eq. (18) is consistent with the exact cylindrical operator only at α≈4.316 (or α≈1.08 under the simplified operator with 1+εx). Thus the third benchmark as written is not a valid test of the claimed O(1e-8) accuracy, and if reproduced literally the reported comparison cannot hold. This is an internal consistency problem, not merely a missing-code or missing-seed issue.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7992,"tokens_out":15032,"duration_ms":143776,"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":[{"comment":"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.","section":"2.2.3, Eqs. (17)-(18)"},{"comment":"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.","section":"2.1-2.2, Eqs. (5)-(6) and Figs. 2, 4, 5"}],"minor_comments":[{"comment":"The Grad-Shafranov equation is a second-order quasilinear elliptic PDE, so describing it as a \"high-order nonlinear partial differential equation\" is inaccurate.","section":"1, introductory paragraph"},{"comment":"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.","section":"Abstract and 2.2"},{"comment":"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.","section":"Title and Abstract"},{"comment":"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.","section":"5, Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal, but the review process should require a corrected and rerun third benchmark, explicit error-metric definitions, and full configuration details before publication. The missing x in the denominator of Eq. (17) may be a typographical error, but the discrepancy between the printed alpha formula and the stated alpha value is an independent error that must be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a legitimate extension of Wang and Lai's multi-stage PINN trick to the Grad-Shafranov equation, and the two solid benchmarks suggest the O(10^-8) claim is real. But the third benchmark is wrong as written, and the paper's conclusions reach beyond what it actually demonstrates.\n\nThe good parts: the method is applied cleanly, the Solov'ev and droplet benchmarks are standard closed-form solutions, and the reported loss curves drop to 10^-11 or lower, which is consistent with the claimed accuracy. Prior PINN-GS results were at 10^-3 to 10^-4, so the improvement is genuine and new for this equation. The authors are transparent about borrowing the multi-stage scheme from [24]; no attempt to pass it off as their own.\n\nNow the soft spots. The third benchmark, Eq. (18) vs Eq. (17), does not hold up. Substituting the stated parameters (epsilon=0.3, lambda=0, a=0.5, b=0.7) into Eq. (18) and applying the operator in Eq. (17) gives a y^2-dependent term proportional to (1-x), while the RHS of Eq. (17) is only a function of x. So Eq. (18) is not a solution of Eq. (17) for any constant alpha, beta. Also, the printed formula for alpha evaluates to about 4.3, not 0.45. This is not a missing-code issue; the example as written is internally inconsistent. The authors need to fix the formula or the displayed solution, otherwise that part of the verification is invalid.\n\nBeyond that, the paper lacks reproducibility details: no code, no seeds, no run-to-run statistics, and the loss weights lambda1/lambda2 and epsilon2 are not specified. For a paper claiming 10^-8 accuracy, one run per benchmark is thin. The conclusion also overstates the scope: these are forward solves with boundaries taken from the analytical solution, not inverse reconstruction from measurements. Calling the method a 'reliable tool for plasma equilibrium reconstruction' is premature.\n\nVerdict: the first two examples are promising, but the third example is a load-bearing flaw. I'd send this to peer review, but the authors need to correct the triangularity case, provide more details, and tone down the conclusions. A good referee will catch the inconsistency quickly.","headline":"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.","tokens_in":8509,"tokens_out":5226,"would_cite":false,"duration_ms":47741,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["Physics-informed neural networks","Grad-Shafranov equation","multi-stage neural network","plasma equilibrium reconstruction","tokamak equilibrium","high-precision PDE solver","Latin hypercube sampling"],"falsifier":"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.","tokens_in":7499,"feed_emoji":"⚛️","tokens_out":8910,"duration_ms":86418,"temperature":0.7,"pith_summary":"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 $\\mathcal{O}(10^{-4})$.","feed_headline":"Physics-informed net pushes plasma equilibrium to 1e-8","feed_subtitle":"Sequential residual training lifts physics-informed solvers from 1e-4 to 1e-8 on tokamak equilibrium benchmarks.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"This citation supplies the multi-stage neural network scheme, including the residual-targeted training and normalization coefficients that the paper applies to the Grad-Shafranov equation.","marker":"[24]"},{"why":"This citation provides the physics-informed neural network framework used to define the PDE and boundary losses in Eqs. (3)-(5).","marker":"[15]"},{"why":"This citation gives the basic-PINN baseline accuracy of order $10^{-3}$ that the multi-stage result improves upon.","marker":"[21]"},{"why":"This citation gives the data-free PINN baseline accuracy of order $10^{-4}$ and the parameterized Grad-Shafranov setup that the paper's two-stage approach surpasses in precision.","marker":"[22]"},{"why":"This citation supplies the Solovév equilibrium parameters and analytical solution used in the first benchmark.","marker":"[27]"},{"why":"This citation supplies the droplet-shaped and triangularity benchmark cases with their analytical solutions and boundary expressions used in the second and third examples.","marker":"[31]"},{"why":"This citation defines the Latin hypercube sampling used to generate collocation points in the solution domain.","marker":"[28]"}],"fun_headline_variants":["Multi-stage PINN lifts Grad-Shafranov accuracy to 1e-8","PINN residual training hits 1e-8 on tokamak equilibrium","Sequential PINN stages push plasma equilibrium error to 1e-8","From 1e-4 to 1e-8: multi-stage PINN for plasma equilibrium","Residual-learning PINN achieves 1e-8 in Grad-Shafranov solvers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Multi-stage PINN lifts Grad-Shafranov accuracy to 1e-8","PINN residual training hits 1e-8 on tokamak equilibrium","Sequential PINN stages push plasma equilibrium error to 1e-8","From 1e-4 to 1e-8: multi-stage PINN for plasma equilibrium","Residual-learning PINN achieves 1e-8 in Grad-Shafranov solvers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1556,"prompt_tokens":829,"completion_tokens":727,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":615}},"tokens_in":445,"tokens_out":727,"duration_ms":6762,"temperature":1.0,"reasoning_tokens":615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:04:50.287466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation gives the basic-PINN baseline accuracy of order $10^{-3}$ that the multi-stage result improves upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation gives the data-free PINN baseline accuracy of order $10^{-4}$ and the parameterized Grad-Shafranov setup that the paper's two-stage approach surpasses in precision."},{"cited_title":"Hu, Notes on tokamak equilibrium*","cited_arxiv_id":null,"evidence_quote":"This citation supplies the Solovév equilibrium parameters and analytical solution used in the first benchmark."},{"cited_title":"Deriaz, B","cited_arxiv_id":null,"evidence_quote":"This citation supplies the droplet-shaped and triangularity benchmark cases with their analytical solutions and boundary expressions used in the second and third examples."}],"review_version":1}