{"id":"ca439aff-c886-4450-900a-7ae0fd59a278","arxiv_id":"2608.05555","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A Fourier Neural Operator predicts double-null free-boundary Grad-Shafranov equilibria in about 2.8 ms with 0.05% mean relative L2 error over a fixed machine geometry and topology.","lead":"A machine-learning model was trained to replace slow iterative tokamak equilibrium calculations with a single fast network evaluation, producing flux maps in milliseconds instead of seconds. The model is accurate for one double-null machine geometry and could make equilibrium calls cheap enough for control loops and design scans.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GS residual diagnostic in Eq. 11 uses the ground-truth RHS rather than the RHS evaluated on the predicted field, so it does not actually measure whether the FNO field satisfies the Grad-Shafranov equation.","rationale":"The reader correctly identified the residual diagnostic as a weak point but did not pinpoint the deeper issue: Eq. 11 is not the GS residual of the predicted field because it uses RHS_true rather than RHS_pred. This is a concrete, verifiable flaw that affects one of the paper's highlighted validation results. However, the central claim of the paper—that the FNO accurately reproduces FreeGS solutions and provides millisecond-scale speedups—rests on the L2 error, geometry metrics, and latency measurements, which are not undermined by this residual issue. The paper also repeatedly and appropriately scopes the surrogate to a single machine geometry and topology, and it explicitly acknowledges the residual baseline is comparative rather than absolute. Therefore, the reader's CONDITIONAL verdict remains appropriate: the paper should be accepted conditional on correcting or clarifying the residual diagnostic (and ideally on releasing code/data), but the main surrogate claim survives. I chose UNCHANGED because my concern does not move the verdict; it strengthens the rationale for the existing conditional acceptance.","tokens_in":17491,"tokens_out":10977,"duration_ms":88771,"concrete_test":"Recompute the normalized residual for all 500 test predictions using RHS_pred, obtained by evaluating the same profile functions p(ψ) and F(ψ) from the FreeGS setup at ψ_pred (e.g., by interpolating the stored dpdpsi and FdFdpsi data onto ψ_pred values, or by re-running FreeGS's profile evaluation with the predicted field). Compare the corrected residual distribution to the FreeGS baseline computed with RHS_true. If the corrected FNO mean residual exceeds the baseline by more than the reported 0.06 spread, the physics-consistency claim in the abstract and Section III D is not supported and would require revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II D defines the residual as R_i = ||Δ*ψ_pred − RHS_true|| / ||RHS_true||, where RHS_true is loaded from the dataset and evaluated on the ground-truth equilibrium. Because the GS equation is nonlinear, RHS depends on ψ itself; the correct residual for a predicted field should use RHS_pred = −μ0 R^2 (dp/dψ)(ψ_pred) − (1/2)(dF^2/dψ)(ψ_pred). The paper never computes this. What Eq. 11 actually measures is how close Δ*ψ_pred is to Δ*ψ_true (plus truncation error), not whether ψ_pred satisfies the GS equation with its own source term. The reported mean residual of 2.29, identical for FNO and FreeGS, is therefore dominated by the finite-difference truncation error and the mismatch between FreeGS's internal discretization and the external stencil, as the paper itself notes. Because the baseline is nonzero and large relative to expected physical residuals, the test cannot detect, for example, a systematic error in the FNO's boundary-layer scaling or a subtle misalignment of the separatrix that shifts the RHS. The claim in the abstract that 'the predicted fields satisfy an external finite-difference GS residual evaluation' overstates what is verified: the diagnostic only shows that the Laplacian of the predicted field is close to the Laplacian of the ground truth, which is a weaker statement than GS consistency. This is an internal inconsistency in the validation, not merely a disagreement with external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a geometry-conditioned Fourier Neural Operator (FNO) to approximate the forward map from spatial coordinates, scalar operating parameters (P_axis, I_p, f_vac), and prescribed X-point coordinates to the poloidal-flux field for constrained double-null free-boundary Grad-Shafranov equilibria generated by FreeGS on a fixed test-tokamak geometry. On a held-out test set, the best model achieves a mean relative L2 error of 0.052%, sub-millimeter X-point and O-point localization, millisecond-scale inference on GPU and CPU, and an external finite-difference residual whose mean value matches the FreeGS baseline. The paper explicitly scopes the contribution to one topology and one machine geometry, and presents the N^-0.68 scaling and the residual comparison as comparative diagnostics rather than definitive physical laws. The central claim is that, within this fixed geometry and topology, the learned operator can replace cold-started FreeGS solves in speed-critical loops without meaningful loss of accuracy.","tokens_in":17775,"tokens_out":6221,"duration_ms":74704,"significance":"If the accuracy and latency claims hold, the paper is a solid contribution to the neural-surrogate literature for plasma equilibrium: it addresses a constrained free-boundary setting with explicit geometry conditioning, uses a controlled data-generation pipeline with a 100% acceptance rate, performs held-out evaluation with three initialization seeds for field-level metrics, and documents near-deterministic inference latency. These strengths make the core speedup and field-accuracy results credible within the stated scope. However, the physics-consistency claim in the abstract and Section III D is overstated: the reported residual diagnostic compares the predicted Laplacian with the ground-truth right-hand side, not with the right-hand side evaluated on the predicted field, so it does not actually test whether the predicted field satisfies the Grad-Shafranov equation. This is a load-bearing validation gap that must be addressed before the paper is accepted.","major_comments":[{"comment":"The residual defined in Eq. (11) uses RHS_true, the right-hand side evaluated on the ground-truth equilibrium, rather than RHS(ψ_pred). Because the Grad-Shafranov right-hand side depends on ψ through p(ψ) and F(ψ), this diagnostic only measures how close Δ*ψ_pred is to Δ*ψ_true; it does not test whether the predicted field satisfies the GS equation with its own source term. The abstract's statement that predicted fields 'satisfy an external finite-difference GS residual evaluation' is therefore not supported by the reported quantity. I recommend computing the residual with RHS evaluated on ψ_pred (using the same profile functions) and reporting that value, or explicitly rephrasing the claim as Laplacian agreement with the ground-truth source.","section":"§II D, Eq. (11), and §III D"},{"comment":"Even accepting the diagnostic as defined, the comparison has low power: the FreeGS baseline itself has mean residual 2.29±0.06, and the FNO mean is 2.29. The test therefore cannot resolve discrepancies smaller than roughly ±0.06 normalized units, while the claimed field-level accuracy (relative L2 ≈5×10^-4, physical RMSE ≈1.5×10^-5 Wb) is far below that resolution. The phrase 'indistinguishable' should be accompanied by an explicit statement of the diagnostic resolution, and the physics-consistency conclusion should be correspondingly weakened unless a proper RHS(ψ_pred) residual is reported.","section":"§III D"},{"comment":"All geometry-aware metrics in Table II are reported for the single validation-selected best model (N=5000, seed 3), with no seed-to-seed spread or confidence intervals. Since Table I shows that field-level error varies across the three initializations, the reader cannot assess whether the sub-0.2 cm X-point and 0.031 cm O-point accuracy are stable properties of the trained operator or partly an artifact of seed selection. Reporting the three-seed range for the geometry metrics (or at least for the X/O-point errors) would make the headline geometric claims robust.","section":"§III B, Table II"}],"minor_comments":[{"comment":"Equation (10) writes the right-hand side as -μ0 R^2 dp/dψ - F dF/dψ, while Eq. (1) uses the equivalent form -μ0 R^2 dp/dψ - (1/2) dF^2/dψ; using one notation consistently would reduce the chance of confusion about the profile conventions.","section":"§II A, Eq. (10)"},{"comment":"The plasma mask Ω_i used to restrict the residual norm is never defined; please specify how interior points are selected (for example, ψ_true < ψ_bndry or a fixed spatial region).","section":"§II D, Eq. (11)"},{"comment":"Equation (7) is written with unnormalized norms, but the text states that all reported relative L2 errors are computed in normalized units; making the normalization explicit in the equation would remove an ambiguity.","section":"§II C, Eq. (7)"},{"comment":"The power-law exponent N^-0.68 is fit over only four training sizes with no confidence interval; the paper already calls this suggestive, but adding a bootstrap interval or a direct statement that the exponent is not statistically robust would strengthen the presentation.","section":"§III A, Eq. (12)"},{"comment":"The data and code are listed as available 'upon reasonable request'; for a neural-operator paper whose reproducibility depends on exact data splits, training configuration, and evaluation code, a public repository would be preferable.","section":"Data and code availability"}],"recommendation":"major_revision","confidential_remarks":"I am broadly supportive of the paper: the core surrogate-accuracy and latency results appear well supported by the held-out evaluation, and the paper is appropriately cautious about its fixed-geometry, fixed-topology scope. The one point that blocks acceptance is the residual diagnostic, which is presented as physics consistency but in fact only compares the predicted Laplacian to the ground-truth source. If the authors compute a proper RHS(ψ_pred) residual or remove the physics-consistency claim from the abstract and conclusions, the paper would be acceptable after a minor revision; as written, the overstatement warrants a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good, scoped empirical demonstration of a geometry-conditioned Fourier Neural Operator as a millisecond-scale surrogate for constrained double-null free-boundary Grad-Shafranov equilibria. The central accuracy and speed claims hold up: 0.05% relative L2 error, sub-mm X-point and O-point localization, 2.77 ms GPU / 25.6 ms CPU inference, and near-deterministic latency. The error analysis is unusually careful for this genre — the Tukey-outlier analysis and the demonstration that the Hausdorff tail is a geometric amplification of small field errors rather than a separate failure mode are genuinely informative. The scaling exponent N^-0.68 is appropriately hedged as suggestive, and the framing as a design map conditioned on prescribed X-points, distinct from reconstruction or actuator-level forward models, is clear and honest.\n\nNow the soft spots, in proportion. The biggest one is the GS residual diagnostic in Eq. 11. It uses RHS_true loaded from the dataset evaluated on the ground-truth equilibrium, not RHS computed from the predicted field. So the diagnostic actually measures how close the Laplacian of the prediction is to the Laplacian of the ground truth, not whether the predicted field satisfies Grad-Shafranov with its own source term. The paper does call this a comparative diagnostic and notes the 2.29 baseline is dominated by truncation error, but the abstract's phrasing that predicted fields 'satisfy an external finite-difference GS residual evaluation' overstates what is verified. This is fixable in revision: compute RHS from the predicted psi and report that residual separately. It would also help to report error bars for Table II geometry metrics; they exist in the distributions in Fig. 6 but the table only gives means and p95s. No code or data is released, which limits reproducibility of the exact numbers, though the training setup is described in enough detail to reimplement. There is no ablation of the X-point conditioning or comparison to a simpler surrogate like a plain CNN or a learned low-dimensional fit; those would strengthen the claim that the FNO architecture and conditioning are doing useful work.\n\nNone of these issues sinks the paper. The central empirical claim — that within one fixed machine geometry and topology, an FNO can replace cold-started FreeGS solves in speed-critical loops at negligible accuracy loss — is well supported. The paper is a useful contribution to the fusion-ML surrogate literature, and it deserves serious peer review. My recommendation: send it out, with a request to fix the residual diagnostic and add a brief comparison or ablation, plus a reproducibility statement on code/data.","headline":"A solid, well-scoped FNO surrogate for FreeGS double-null equilibria with strong accuracy and latency numbers, but the 'GS residual' diagnostic is weaker than the abstract suggests because it compares the predicted Laplacian to the ground-truth RHS, not to the RHS evaluated on the predicted field.","tokens_in":18344,"tokens_out":1160,"would_cite":true,"duration_ms":40880,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.55.Fa","52.65.-y","52.30.-q"],"model":"deepseek-v4-flash","headline":"A geometry-conditioned Fourier Neural Operator can reproduce constrained double-null tokamak equilibria to 0.05% field error and sub-centimeter critical-point accuracy in milliseconds.","keywords":["Grad-Shafranov equation","free-boundary equilibrium","Fourier Neural Operator","double-null divertor","poloidal flux","tokamak equilibrium","surrogate modeling"],"falsifier":"Generate the same operating points with an independent, finer-grid free-boundary Grad-Shafranov solver and compare its flux fields and critical points with the network's predictions; if agreement is materially worse than the 0.052% L2 error and sub-0.2 cm X-point errors reported against the training solver, the physical-accuracy claim fails. Alternatively, run the finite-difference residual diagnostic on the same fields with a stricter solver-internal residual norm or a 129 by 129 grid; a predicted-versus-truth gap much larger than the current 0.998 ratio would falsify the physics-consistency claim.","tokens_in":17248,"feed_emoji":"⚡","tokens_out":9815,"duration_ms":74179,"temperature":0.7,"pith_summary":"The paper aims to show that a constrained free-boundary Grad-Shafranov equilibrium, normally found by iterating a nonlinear solver, can be replaced by a single forward pass of a learned neural operator for one fixed machine geometry and double-null topology. Training on examples from an open-source free-boundary solver, the map takes spatial coordinates, three scalar operating parameters, and prescribed X-point locations and outputs the full poloidal-flux field. The best model reaches 0.052% mean relative L2 error, localizes both X-points to within about 0.2 cm and the O-point to 0.031 cm, and evaluates one equilibrium in 2.77 ms on GPU and 25.6 ms on CPU. If the claim holds, equilibrium evaluations inside control, optimization, and integrated-modeling loops could drop from seconds to milliseconds with near-deterministic latency.","feed_headline":"Neural operator returns tokamak equilibria in 2.8 ms at 0.05% error","feed_subtitle":"It recovers X-points to 0.2 cm and the O-point to 0.03 cm, so speed-critical loops can afford full equilibrium calls.","key_machinery":"The central object is the operator G mapping nine gridded input channels -- spatial coordinates (R,Z), broadcast scalars (P_axis, I_p, f_vac), and broadcast lower and upper X-point coordinates -- to the poloidal-flux field psi(R,Z). The architecture is a Fourier Neural Operator with four spectral-convolution layers, 16 by 16 retained modes per dimension, width 64, and 4.77 million parameters; each layer performs a truncated spectral convolution K v = $F^{{-1}}$(R(k) F[v]), whose global mixing matches the elliptic, globally coupled nature of the Grad-Shafranov equation. The explicit X-point conditioning is what anchors the double-null topology and makes the map a constrained design surrogate rather than a diagnostic reconstruction from measurements.","core_discovery":"The central claim is that, within a single fixed machine geometry and prescribed double-null topology, a geometry-conditioned Fourier Neural Operator learns the constrained forward map from spatial coordinates, scalar operating parameters, and X-point locations to the poloidal-flux field, and that the result is accurate enough to serve as a surrogate for iterative equilibrium solves. The evidence is a 500-sample test set on which the best model attains mean relative L2 error 0.052%, physical RMSE 1.54e-5 Wb, separatrix mean deviation 0.072 cm, upper and lower X-point errors 0.112 and 0.161 cm, and O-point error 0.031 cm. An external finite-difference Grad-Shafranov residual gives mean 2.29 on predicted fields, indistinguishable from the 2.29 plus or minus 0.06 baseline on the ground-truth fields. The paper also claims test error follows an approximate N to the minus 0.68 power law over training sizes 500 to 5000, GPU and CPU speedups of about 640 times and 69 times, and a p95/median latency ratio of 1.01. The intended reading is that the learned surrogate can replace cold-started iterative solves in speed-critical loops, not that it generalizes across devices or topologies.","pith_inferences":["The paper leaves implicit that its X-point conditioning doubles as an ablation-ready control handle: setting those coordinates to measured values would let the same network annotate reconstructed equilibria, not only prescribe them.","A natural next test is to train with random vertical flips of every sample; the reported 0.161 cm versus 0.112 cm upper/lower X-point asymmetry would be expected to vanish, since the data-generation pipeline is exactly symmetric.","Because the 0.05% field error concentrates near the X-point pinch, replacing the pure L2 loss with a physics-residual or shape-aware term could shrink the heavy Hausdorff tail without changing the architecture.","If the N to the minus 0.68 trend continues, training on tens of thousands of equilibria could plausibly push field error toward 0.01%, but the four-point fit is not a reliable law; the same pipeline could be tested on single-null and snowflake topologies."],"forward_implications":["In the tested operating range, equilibrium evaluation drops from a median of roughly 1.77 s with the iterative solver to 2.77 ms on GPU and 25.6 ms on CPU, with p95/median 1.01, making equilibrium calls affordable inside control and optimization loops.","The separatrix is reproduced with 0.072 cm mean closest-point deviation, both X-points within 0.2 cm, and the O-point at 0.031 cm, so divertor geometry and critical-point structure survive at sub-millimeter scale.","Predicted fields pass the same finite-difference Grad-Shafranov residual test as the ground-truth fields, with mean 2.29 versus baseline 2.29 plus or minus 0.06, so the surrogate's agreement is not merely label memorization.","Test error scales as roughly N to the minus 0.68 with training data over 500 to 5000 samples, so the controlled equilibrium family is learnable with systematic, better-than-random improvement.","The surrogate is scoped to one machine geometry and one topology; it does not claim cross-device transfer, and it is an X-point-conditioned design map rather than an actuator-to-equilibrium model."],"supporting_citations":[{"why":"Defines the ideal-MHD equilibrium equation that the surrogate approximates.","marker":"1–3"},{"why":"Establishes the constrained free-boundary solver paradigm with isoflux and X-point constraints, the data-generation setting.","marker":"5–7"},{"why":"Makes the case for fast real-time equilibrium reconstruction, the practical motivation for millisecond latency.","marker":"9"},{"why":"Prior neural forward model predicting flux surfaces from coil and plasma parameters; the comparison that defines this paper's constrained design map.","marker":"30"},{"why":"Gives the function-space operator-learning framing that justifies training one map over a family of equilibria.","marker":"40"},{"why":"Introduces the spectral-convolution architecture whose global mixing carries the approximation.","marker":"42"},{"why":"Open-source free-boundary solver that produced every training, validation, and test field used as ground truth.","marker":"45"}],"fun_headline_variants":["2.77 ms tokamak equilibria with 0.05% error via neural operator","Millisecond surrogate: 0.05% error, X-points to 0.2 cm","640x faster: FNO predicts tokamak equilibrium in 2.77 ms","Neural surrogate matches FreeGS residual while running 640x faster","0.05% L2 error and X-points within 0.2 cm: millisecond equilibrium"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the solver-generated equilibria used as training labels are the correct ground truth; every reported accuracy number is an agreement with that solver, and if the solver's constrained solutions are wrong or grid-dependent, the surrogate inherits those errors.","fun_headline_variants_meta":{"raw":{"variants":["2.77 ms tokamak equilibria with 0.05% error via neural operator","Millisecond surrogate: 0.05% error, X-points to 0.2 cm","640x faster: FNO predicts tokamak equilibrium in 2.77 ms","Neural surrogate matches FreeGS residual while running 640x faster","0.05% L2 error and X-points within 0.2 cm: millisecond equilibrium"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000947,"raw_usage":{"total_tokens":4195,"prompt_tokens":1249,"completion_tokens":2946,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":865,"completion_tokens_details":{"reasoning_tokens":2830}},"tokens_in":865,"tokens_out":2946,"duration_ms":18107,"temperature":1.0,"reasoning_tokens":2830,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:58:42.112788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate the same operating points with an independent, finer-grid free-boundary Grad-Shafranov solver and compare its flux fields and critical points with the network's predictions; if agreement is materially worse than the 0.052% L2 error and sub-0.2 cm X-point errors reported against the training solver, the physical-accuracy claim fails. Alternatively, run the finite-difference residual diagnostic on the same fields with a stricter solver-internal residual norm or a 129 by 129 grid; a predicted-versus-truth gap much larger than the current 0.998 ratio would falsify the physics-consistency claim.","supporting_citations":[{"cited_title":"Journal of Fusion Energy , volume =","cited_arxiv_id":null,"evidence_quote":"Prior neural forward model predicting flux surfaces from coil and plasma parameters; the comparison that defines this paper's constrained design map."},{"cited_title":"Nuclear Fusion , volume =","cited_arxiv_id":null,"evidence_quote":"Gives the function-space operator-learning framing that justifies training one map over a family of equilibria."},{"cited_title":"and Hutter, F","cited_arxiv_id":null,"evidence_quote":"Introduces the spectral-convolution architecture whose global mixing carries the approximation."},{"cited_title":"and Adebayo-Ige, P.O","cited_arxiv_id":null,"evidence_quote":"Open-source free-boundary solver that produced every training, validation, and test field used as ground truth."}],"review_version":1}