{"id":"ab681dd1-38bb-45a6-83c5-c062bda6b08d","arxiv_id":"2509.01789","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Neural-network emulators trained on roughly 100,000 to 1,000,000 Grad-Shafranov equilibria can generate virtual circuits with 5-10% displacement accuracy, and shaping currents can be inferred online with few-Ampere residuals.","lead":"This paper tests whether neural-network emulators of tokamak plasma shape can produce virtual circuits accurate enough for real-time control, finding 5-10% displacement accuracy on MAST-U simulations. It also shows that unmeasured vessel currents can be replaced by inferred 'shaping currents' from a short window of coil measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Online shaping-current inference (Eq. 4) assumes a single time-invariant matrix B across all shots; validation on 21 flat-top MAST-U shots is too narrow to support the real-time applicability claim.","rationale":"I agree with the reader's identification of the weakest assumption. The paper provides two internally consistent and held-out-validated results: emulator VCs achieve 5–10% displacement accuracy on 380 GS re-solved equilibria, and online shaping-current inference achieves √(1−R²)=0.015 on a held-out set of MAST-U shots. Both are real evidence and the Section II validation is fairly strong, since it re-solves the GS equation after applying VCs. However, the online-inference result is the load-bearing part for real-time applicability, and it rests on a time-invariant B that is asserted rather than derived from a physical model. The validation is limited to 21 flat-top shots from one device; this is a narrow empirical basis for claiming B is an inherent tokamak property. The lack of an explicitly quantified connection between the reported R² and the shaping-current vector error further weakens the link to VC accuracy. The concrete test proposed—grouped leave-one-out cross-validation plus within-shot stability—would directly establish whether B generalizes across operating conditions or is a curve-fit artifact. If B fails the test, the central real-time claim is not supported; if it passes, the paper's conclusions are reasonable. Therefore the reader's CONDITIONAL verdict is appropriate: the paper should add this cross-validation evidence or soften its generalization claims.","tokens_in":6983,"tokens_out":10389,"duration_ms":131669,"concrete_test":"Using the 21 MAST-U shots, perform leave-one-group-out cross-validation where shots are grouped by flat-top plasma current, plasma shape (e.g., elongation, upper vs lower single null), and line-averaged density. Train Eq. (4) on 20 shots and evaluate on the held-out group; if the validation √(1−R²) degrades by more than a factor of two relative to the random-split value of 0.015, B is not shot-independent. Additionally, fit B separately to the first and second halves of each flat-top phase and test whether the coefficient matrices differ significantly; such drift would falsify time-invariance within a shot.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The real-time control claim depends on Eq. (4), where a single matrix B maps a trailing window of active coil currents and plasma current to shaping-current corrections, and is \"posited to be the same across all shots, i.e. an inherent property of the tokamak.\" This is a strong structural assumption. Vessel currents are driven by inductive coupling whose effective gains depend on the plasma current profile and boundary shape, which affect mutual inductances, and on vessel resistivity that can vary with temperature. B is fitted jointly over 11 flat-top MAST-U shots and validated on 10; all 21 are MAST-U flat-top phases, so the fit could absorb shot-to-shot variation into a common B only if those shots are sufficiently similar. The reported √(1−R²)=0.015 is a scalar goodness-of-fit, and the paper does not directly quantify the resulting error in the shaping-current vector that would enter a VC calculation. If B is not actually an inherent property, the few-Ampere residuals on the 10 validation shots do not generalize, and the emulator VCs—trained without vessel currents as inputs—cannot be supplied with correct shaping currents in real time.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates two aspects of using neural-network emulators for real-time tokamak plasma shape control. First, it validates virtual circuits (VCs) computed from Jacobians of FNN shape emulators against VCs from finite-difference Jacobians of exact Grad-Shafranov solutions, reporting typical realized-displacement errors of 5–10% on a holdout set of 380 equilibria. Second, it addresses the effect of unmeasured vessel currents by introducing an offline least-squares reconstruction of 'shaping currents' and an online linear surrogate (Eq. 4) that maps a trailing window of measured active-coil currents and plasma current to shaping currents, with reported validation on flat-top phases of 21 MAST-U shots.","tokens_in":7268,"tokens_out":5381,"duration_ms":58308,"significance":"If the results hold, the paper offers a useful, quantitative step toward making ML-based sensitivity emulators usable for classical plasma shape control: it checks not just target prediction but the derivatives that enter VC design, and it explicitly tackles the vessel-current/shaping-current problem that is often glossed over. The use of a large synthetic GS library, a separate holdout set for VC evaluation, and the comparison against finite-difference GS solutions are strengths. The online shaping-current regression on real MAST-U data with held-out shots is also a concrete, falsifiable contribution. However, several load-bearing claims are either unsupported or under-validated, and the paper would need substantive revision before publication.","major_comments":[{"comment":"The abstract and §II.A state that a sample of ≈10^5–10^6 synthetic equilibria is 'essential' for training emulators that are not over-regularised or overfitting. The only supporting statement is 'We run training and tuning experiments for different choices of dataset size, up to 10^6,' but no dataset-size ablation is reported. This is load-bearing because it sets the data-generation cost and is part of the abstract's central claim. Please either report the ablation (e.g., validation error vs. dataset size) or soften 'essential' to something like 'we found that our best results used...'.","section":"Abstract / §II.A"},{"comment":"There is an internal contradiction about the holdout set. §II.A says the holdout set is 'never used to train or evaluate the shape emulators,' but §II.B states that emulator Jacobians and VCs are evaluated on a random subset of 380 equilibria in the holdout set. If the holdout was used only for the Jacobian/VC comparison, the text should say 'never used for training or hyperparameter selection'; if it was also used to choose among models, the validation is not truly held-out. Please resolve this, since the validity of the reported 5–10% accuracy depends on it.","section":"§II.A vs §II.B"},{"comment":"Equation (4) posits a single time-invariant matrix B that is 'an inherent property of the tokamak,' but the validation is limited to flat-top phases of 21 MAST-U shots (11 training, 10 validation). This is a strong structural assumption: vessel-current coupling depends on plasma configuration, boundary shape, and resistive state, and the flat-top restriction explicitly excludes ramp-up/down phases. To support the real-time applicability claim, please (a) quantify the residual error in the shaping-current vector in Amperes (not only √(1−R²)), (b) show how the residuals propagate into the VCs of Section II, and (c) either extend validation to ramp-up/down or clearly restrict the claim to flat-top feedback control.","section":"§III.B (Eq. 4)"},{"comment":"The abstract and discussion claim that emulators with 10^5–10^6 parameters can provide 'few-millisecond latency,' but no inference latency measurement is reported; this is a real-time applicability claim that should be benchmarked or explicitly stated as an expectation rather than a measurement. Additionally, the reference finite-difference VCs are computed with a fixed 0.002 relative variation in the plasma-current density distribution, and no sensitivity analysis for this step is given, even though the text acknowledges that the finite-difference step must be 'carefully sized.' Please add a sensitivity check or justify the chosen step.","section":"§IV / §II.B"}],"minor_comments":[{"comment":"The notation b(c,τ) is introduced in §I.A but Eq. (4) uses b(a′,τ) and b(p,τ) without explicitly connecting them to b(c,τ). Please define the index set for c and the ranges of a′, p, and τ.","section":"§I.A / Eq. (4)"},{"comment":"The histograms are hard to compare quantitatively; consider adding a table with mean, median, and standard deviation of realized displacements for each VC and each target.","section":"Fig. 2"},{"comment":"The top/middle/bottom panels would benefit from explicit axis labels and units; in particular, clarify whether √(1−R²) is computed for ψtok residuals or for shaping-current residuals in the online inference.","section":"Fig. 4"},{"comment":"The concept of 'shaping currents' is central to the paper but no reference to the classical control literature where these currents and their use in VC design are defined is given. Please add a citation.","section":"§III.A"},{"comment":"The sentence 'Since PF coil currents can be measured with sub-millisecond cadence, the shaping currents can also be computed and adjusted every millisecond' conflates measurement cadence with control-loop computation. Clarify the intended control-loop timing.","section":"§IV"},{"comment":"There are small typos: 'Toakamak' should be 'Tokamak' in the author affiliation, and 'Amp `eres' has an extra space. Please proofread.","section":"Abstract / metadata"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the core idea is valuable, but the abstract overclaims on two fronts: the necessity of 10^5–10^6 training equilibria is not demonstrated, and the online shaping-current regression is validated on a narrow set of flat-top MAST-U shots. These are fixable with additional experiments or more cautious wording, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things are genuinely new here, and both are done with proper care. First, the paper quantifies how well Jacobians of neural-network shape emulators reproduce finite-difference Jacobians from exact GS solutions, and then checks actual target displacements after applying the VCs and re-solving GS. That is the right way to test an emulator for control, and the 5-10% agreement on 380 held-out equilibria is a useful number for the community. Second, the online shaping-current inference via trailing-window linear regression is simple, physically motivated, and validated on held-out shots; few-Ampere residuals on MAST-U flat-top data are encouraging.\n\nThe soft spots are real but not fatal. The claim that 10^5-10^6 training equilibria are 'essential' is not backed by any dataset-size ablation; the paper mentions running such experiments but doesn't show the curve. The holdout-set description is self-contradictory: one sentence says the holdout is never used to evaluate the shape emulators, then the Jacobian test is explicitly run on 380 equilibria from that holdout. Minor wording fix, but embarrassing. More substantively, the online inference assumes a single time-invariant matrix B, and the validation is on 21 flat-top MAST-U shots. That is good evidence for exactly that regime, and the paper is careful to say so. But the abstract's 'existing and upcoming tokamaks' is an overreach. Vessel current dynamics are nonlinear and machine-specific; a linear B fit on one device's flat-top doesn't establish portability.\n\nAlso, no code or data is released. The FreeGSNKE solver is referenced, but the trained emulators and datasets are not available, which makes independent reproduction harder. The citation pattern is fine—they build on their own FreeGSNKE/Agnello work, which is appropriate.\n\nOverall, the central claims are plausible and internally consistent. The stress-test worry about Eq. (4) doesn't sink the paper because the held-out shot validation is genuine evidence for the stated regime; it just means the generalizability claim should be toned down. I'd send this to a plasma control journal after asking for an ablation of dataset size, a correction of the holdout phrasing, and a more measured abstract. It deserves a serious referee.","headline":"A useful, honestly-hedged validation of emulator-based virtual circuits for MAST-U shape control; the broad real-time applicability claim overreaches the evidence, but the core results are likely sound.","tokens_in":7798,"tokens_out":2625,"would_cite":true,"duration_ms":28141,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.55.Fa"],"model":"deepseek-v4-flash","headline":"Neural-network emulators of Grad-Shafranov equilibria can supply virtual circuits that move tokamak shape targets with 5-10% accuracy, and shaping currents can be inferred online from a trailing window of coil-current measurements.","keywords":["tokamak plasma shape control","virtual circuits","neural network emulation","Grad-Shafranov equilibria","sensitivity Jacobians","shaping currents","real-time control","MAST-U"],"falsifier":"Run the published shaping-current linear regression on a shot with strong ramp-up or disruption transients, compare the predicted shaping currents with those obtained by solving the full circuit equations including vessel currents, and check whether the residual scatter stays at the few-Ampere level; if it does not, the single-matrix assumption fails. Equivalently, request a 5 mm displacement via the emulator virtual circuit for an equilibrium far outside the training distribution and re-solve the Grad-Shafranov equation to see whether the realized displacement error exceeds the observed 5-10%","tokens_in":6906,"feed_emoji":"⚡","tokens_out":5857,"duration_ms":60942,"temperature":0.7,"pith_summary":"The paper tests whether machine-learning emulators of tokamak plasma shape can be trusted for real-time control. It reports that virtual circuits, built from the Jacobian of a neural-network emulator trained on roughly 1e5-1e6 synthetic Grad-Shafranov equilibria, produce actual target displacements within about 5-10% of the requested 5 mm shift, comparable to virtual circuits from finite-difference Jacobians of an exact solver. It also shows that the unmeasured effect of vessel currents can be absorbed into effective \"shaping currents\" inferred from a trailing window of measured active-coil and plasma currents, with a residual scatter of about 1.5%. If these results hold on other machines and across all shot phases, emulated virtual circuits offer an explainable and fast alternative to reinforcement-learning controllers that can generalize to configurations not yet seen.","feed_headline":"Emulated virtual circuits steer tokamak shape within 5-10%","feed_subtitle":"Neural-net Jacobians plus online shaping currents enable millisecond-scale plasma control on MAST-U","key_machinery":"Virtual circuits: linear combinations of poloidal-field coil current changes, obtained from the pseudo-inverse of the Jacobian of shape targets with respect to coil currents, designed to move one shape target at a time; the paper uses Jacobians of neural-network emulators trained on shape targets rather than on the Jacobians themselves. Shaping currents: effective currents in the active PF coils that alone would reproduce the flux produced by both active coils and passive vessel structures; inferred offline via linear regression of the Green-function flux (Eq. 3) and online via a linear map from a trailing window of measured currents (Eq. 4).","core_discovery":"The central claim is that differentiable emulators of plasma shape, trained on large synthetic libraries of Grad-Shafranov equilibria, can provide the Jacobian needed to construct virtual circuits at every time step, and that these emulator-derived virtual circuits are accurate enough for feedback control. Across 380 held-out equilibria, requesting a 5 mm displacement in inner or outer radius, X-point position, or strike-point and re-solving the Grad-Shafranov equation gives realized displacements within 5-10% of the request when the virtual circuit comes from emulator Jacobians, often matching the performance of finite-difference Jacobians on the same equilibria. The paper further claims th","pith_inferences":["Because the single-matrix shaping-current map is fit only on flat-top MAST-U phases, its most natural test is on ramp-up, ramp-down, and disruption-recovery phases of the same machine; if the map holds there, it is closer to a genuine machine transfer function.","The emulators are trained on shape targets, not their derivatives, so Jacobian error may be only loosely correlated with target error; a controller that also tracks a cheap finite-difference Jacobian estimate during operation would catch excursions outside the training distribution.","The same pipeline—synthetic equilibria, emulator Jacobians, linear shaping-current surrogates—could in principle be run for a not-yet-built tokamak using design equilibria, producing a first set of virtual circuits before any experimental data exists.","Since the shaping-current regression is linear and trained on data, its residuals embed machine-specific passive-structure time constants; transferring the same linear map to another machine without retraining is unlikely to keep the 1.5% accuracy, which hints that each tokamak needs its own calibration."],"forward_implications":["Feedback controllers can query fresh virtual circuits every few milliseconds instead of interpolating lookup tables separated by hundreds of milliseconds.","Shaping-current estimation no longer requires real-time equilibrium reconstruction, only sub-millisecond coil-current and plasma-current measurements.","For core shape targets, emulator virtual circuits are accurate at the 5-10% level, comparable to finite-difference virtual circuits, making classical control more robust to non-negligible displacements.","If higher accuracy is needed, emulators can be fine-tuned on finite-difference Jacobians, and aggregating the five best emulators reduces target prediction scatter to about 1%.","The same approach can be applied to new or upcoming tokamaks using synthetic equilibria seeded from design or recent campaign data, without requiring a long experimental history."],"supporting_citations":[{"why":"Supplies the MCMC-based synthetic equilibrium library generation method and the fully-connected neural-network emulator architecture that this paper builds on.","marker":"[11]"},{"why":"Provide the free-boundary Grad-Shafranov equilibrium solver and its validation, used to produce the exact equilibria and finite-difference Jacobians that serve as the accuracy reference.","marker":"[14, 15]"},{"why":"Defines the flux function and plasma physics notation that underpin the shaping-current regression and target definitions.","marker":"[13]"},{"why":"Describes the MAST-U tokamak, the machine whose shots, geometry, and coil layouts are used in both the emulator validation and the shaping-current study.","marker":"[12]"},{"why":"Describes the MAST Upgrade plasma control system, establishing the real-time control context and latency requirements the emulators must satisfy.","marker":"[5]"},{"why":"Earlier work on neural-network modeling of equilibria in another tokamak, providing context for the emulation approach and the state of the art it extends.","marker":"[10]"}],"fun_headline_variants":["Emulated virtual circuits hit 5-10% accuracy on tokamaks","Neural-net emulators enable millisecond tokamak shape control","5-10% shape accuracy from emulated Jacobians","Online shaping currents improve tokamak control","Real-time plasma shaping with emulator-trained Jacobians"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The claim rests on the assumption that one fixed linear map from a trailing window of measured coil and plasma currents to the shaping currents is valid for every shot and every operating condition; the paper only tests this on the flat-top phases of 21 MAST-U shots.","fun_headline_variants_meta":{"raw":{"variants":["Emulated virtual circuits hit 5-10% accuracy on tokamaks","Neural-net emulators enable millisecond tokamak shape control","5-10% shape accuracy from emulated Jacobians","Online shaping currents improve tokamak control","Real-time plasma shaping with emulator-trained Jacobians"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001011,"raw_usage":{"total_tokens":4138,"prompt_tokens":802,"completion_tokens":3336,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":3253}},"tokens_in":546,"tokens_out":3336,"duration_ms":22167,"temperature":1.0,"reasoning_tokens":3253,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:10:18.899687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published shaping-current linear regression on a shot with strong ramp-up or disruption transients, compare the predicted shaping currents with those obtained by solving the full circuit equations including vessel currents, and check whether the residual scatter stays at the few-Ampere level; if it does not, the single-matrix assumption fails. Equivalently, request a 5 mm displacement via the emulator virtual circuit for an equilibrium far outside the training distribution and re-solve the Grad-Shafranov equation to see whether the realized displacement error exceeds the observed 5-10%","supporting_citations":[{"cited_title":"P., Plasma Physics and Fusion Energy , Cambridge, UK: Cambridge University Press, 2008","cited_arxiv_id":null,"evidence_quote":"Defines the flux function and plasma physics notation that underpin the shaping-current regression and target definitions."},{"cited_title":"MAST-upgrade divertor facility and assessing performance of long-legged divertors","cited_arxiv_id":null,"evidence_quote":"Describes the MAST-U tokamak, the machine whose shots, geometry, and coil layouts are used in both the emulator validation and the shaping-current study."},{"cited_title":"The MAST Upgrade plasma control system","cited_arxiv_id":null,"evidence_quote":"Describes the MAST Upgrade plasma control system, establishing the real-time control context and latency requirements the emulators must satisfy."},{"cited_title":"Emulation techniques for sce- nario and classical control design of tokamak plasmas","cited_arxiv_id":null,"evidence_quote":"Supplies the MCMC-based synthetic equilibrium library generation method and the fully-connected neural-network emulator architecture that this paper builds on."}],"review_version":1}