{"id":"7439d3f7-60f2-4773-a71c-571ff9c7c4d7","arxiv_id":"2607.04865","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Dense, PCA, CNN and physics-informed networks trained on a large pyIPREQ synthetic dataset accurately predict ADITYA-U MHD equilibrium scalars and profiles within the sampled circular-limiter flat-top domain.","lead":"Researchers trained deep learning models on 100,760 synthetic ADITYA-U tokamak equilibria to predict magnetic-axis position, safety-factor and poloidal-flux profiles, and coil currents in about a millisecond. The work offers a fast surrogate for conventional Grad–Shafranov solvers that could support real-time control and shot planning on that machine.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged synthetic-to-experiment transfer gap.","rationale":"The reader's weakest_assumption correctly isolates the only material risk: that a synthetic library built from a parametric Jφ, a coarse βp surrogate, and physics filters may not fully represent real ADITYA-U equilibria, so synthetic test errors need not transfer. That risk is already explicit in §2.1–2.2 and §7 and is the reason the verdict is CONDITIONAL rather than ACCEPT. No stronger internal flaw (incorrect Grad-Shafranov residual formulation, data leakage across shots, contradictory PCA/CNN results, etc.) is present. Consequently the stress-test does not alter the reader's CONDITIONAL/HIGH assessment; the concrete experimental cross-check above is simply the natural next step the authors themselves flag as future work.","tokens_in":23193,"tokens_out":526,"duration_ms":4699,"concrete_test":"Take 20–30 real ADITYA-U flat-top time slices that possess both magnetic-probe/loop-voltage data and independent EFIT/pyIPREQ reconstructions; feed the measured inputs through the published forward models and compare predicted Rax, Zax, βp, ℓi, q(ρ) and ψ(R,Z) against the experimental reconstructions. If median absolute errors remain within a factor of ~2–3 of the synthetic test-set figures, the transfer claim holds; larger systematic bias would confirm the domain-gap concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is carefully scoped to accuracy \"within the operational domain represented by the dataset\" (§Abstract, §7). All reported errors (Tables 9, 11, 13, 15; Figs. 9–42) are synthetic held-out-shot splits of the same pyIPREQ library that uses a fixed Jφ form (Eq. 2), a linear βp prior from only 39 discharges (§2.2), and hard filters on ΔR, Zax, ℓi, q0, q1. The authors already state the resulting limitations (circular limiter only, probe dependence, no formal stability analysis, synthetic-only evaluation). Within that stated domain the architectures, hyperparameter search, shot-wise splits, and quantitative error distributions are internally consistent; no hidden derivation failure or contradictory result appears. The load-bearing premise the reader identified is therefore real but already acknowledged; it does not introduce an additional internal inconsistency that would overturn the scoped claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript develops deep-learning surrogates for ADITYA-U free-boundary MHD equilibria. A 100,760-case synthetic library is generated with pyIPREQ from 766 experimental discharges (circular limiter, near flat-top, Ip>100 kA), using a fixed parametric Jφ form (Eq. 2), Bayesian-optimized γ, a linear probabilistic βp prior fitted on 39 diamagnetic shots plus noise, and physics-informed filters on ΔR, Zax, ℓi, q0, q1. Dense networks predict scalars (Rax, Zax, βp, ℓi, q1, ψaxs/ψlim) and an inverse map from desired equilibrium parameters to selected PF coil currents. PCA and 1d-CNN models reconstruct q(ρ) (the latter via softplus-constrained q′); PCA and 2d-CNN PINNs reconstruct ψ(R,Z) with Grad–Shafranov residual losses and simultaneous prediction of α,β,γ. Shot-wise 80/10/10 splits, Hyperband search, and 95th/99th-percentile absolute errors plus median/worst-case profile figures are reported. Inference of the largest model is ~1 ms on CPU. Claims are scoped to accuracy inside the synthetic operational domain.","tokens_in":23499,"tokens_out":1335,"duration_ms":10518,"significance":"If the reported synthetic accuracy transfers, the work supplies the first large-scale, multi-architecture ML equilibrium framework for ADITYA-U, with practical utility for real-time control, rapid discharge analysis, and actuator planning. Strengths include the experimentally motivated input ranges, shot-wise splits that avoid temporal leakage, systematic PCA-vs-CNN comparison, physics-informed constraints (monotonicity of q, progressive GS residual with uncertainty weighting), and explicit quantitative error distributions rather than only mean metrics. The inverse coil-current model and cascade-ready scalar predictors are useful engineering contributions within the stated domain. The principal limitation is that all validation remains synthetic; experimental transfer is left for future work and is already acknowledged.","major_comments":[{"comment":"The central claim is carefully scoped to the synthetic domain, yet the abstract, introduction and conclusion repeatedly advertise utility for real-time control and experimental planning. No experimental reconstruction or even a single pyIPREQ-vs-diagnostic comparison on held-out ADITYA-U shots is shown. A minimal experimental sanity check (or a clearly labeled “synthetic-only” caveat in the abstract) is needed so that the transfer premise of §2.1–2.2 does not over-extend the reported numbers.","section":null},{"comment":"§2.2: the βp prior is a linear model trained on only 39 diamagnetic discharges; Fig. 3 shows clear bias at high/low βp and the residual noise N(0,0.04) is then injected into all 100k cases. Because βp directly sets the Jφ parameter β (Eq. 2) and therefore shapes ℓi and q, the sensitivity of the downstream scalar and profile errors to this prior should be quantified (e.g., by re-training with a wider or alternative βp distribution).","section":null},{"comment":"§4.2 and cascade discussion: Rax/Zax (and later q1, ψaxs/ψlim) are used as inputs to subsequent models, yet training uses ground-truth values; cascading error is never measured. Because the intended real-time pipeline is cascaded, the reported test-set errors for βp/ℓi, q-profile and ψ-profile are optimistic. A short end-to-end cascade evaluation on the test shots is required.","section":null}],"minor_comments":[{"comment":"Table 1 lists coil geometry but omits the actual current ranges present in the dataset; a short summary row would help readers judge the inverse-model domain.","section":null},{"comment":"Figs. 9–22 report absolute errors with medians and 95th percentiles in the captions; adding the same numbers to the main text or a summary table would improve readability.","section":null},{"comment":"Eq. (3) and the progressive factor f are introduced without a short derivation or reference to the uncertainty-weighting paper beyond [21]; a one-sentence justification of the numerical constants (−50, 0.15) would help reproducibility.","section":null},{"comment":"The 1d-CNN predicts 100 values of q′ and reconstructs q by integrating from the separately predicted q1; the accumulation of integration error should be stated explicitly when comparing core-region accuracy with the PCA model.","section":null},{"comment":"Minor typographical issues: “two-dimensionalpoloidalfluxprofiles” (abstract), “physics-informedneuralnetworks” (abstract), and occasional missing spaces after commas in the introduction.","section":null},{"comment":"Acknowledgment notes LLM rephrasing; a brief statement that scientific content and all numerical results were verified by the authors would be appropriate.","section":null}],"recommendation":"major_revision","confidential_remarks":"The work is solid engineering for a specific device and is appropriately scoped once the synthetic-to-experiment gap is made more prominent. It is a good fit for a plasma-physics or nuclear-fusion journal that accepts ML-for-control papers; I would not recommend a pure ML venue. The three major points are fixable without new experimental campaigns if the authors either add a cascade study on the existing synthetic test set or clearly demote the real-time-control language."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a careful engineering paper that does exactly what it claims: builds a large, experimentally motivated free-boundary library for ADITYA-U and trains dense/PCA/CNN/PINN surrogates (plus an inverse coil model) that run in ~1 ms inside that domain.\n\nWhat is actually new is the device-specific package, not a new algorithm. They generate ~1e5 pyIPREQ cases from 766 shots, filter to circular limiter flat-top, and systematically compare PCA vs CNN under Grad–Shafranov residual losses, with a softplus monotonicity constraint on q′. Shot-wise 80/10/10 splits, Hyperband search, repeated retrainings, and 95th/99th-percentile absolute errors plus median/worst-case profile figures are all there. Scalar models (Rax, Zax, βp, ℓi, q1, ψaxs/ψlim) look tight; q and ψ reconstructions are usable for control-oriented work. The inverse PF-current model is a practical bonus for planning, with the non-uniqueness caveat stated clearly.\n\nThe soft spot is the one they already flag: everything is synthetic held-out shots of the same library (fixed Jφ form, linear βp prior from 39 diamagnetic shots plus noise, hard filters on ΔR, Zax, ℓi, q0, q1). Transfer to real ADITYA-U discharges is unproven, and the models need the full magnetic-probe + Vloop suite. That is a real limitation for operational claims, not a hidden flaw in the math or the reporting. Circularity is mild and intentional—they are surrogates of pyIPREQ, not a new equilibrium theory. Citations cover the usual EFIT/ML-equilibrium literature; no red flags.\n\nThis is for people who run or plan ADITYA-U, or who want a reproducible recipe for device-scoped equilibrium surrogates. It deserves a serious referee, not a desk reject. I would engage if I needed fast ADITYA-U equilibria or a clean PCA-vs-CNN PINN comparison; I would not treat the error bars as experimental until they validate on real reconstructions.","headline":"Solid first-for-ADITYA-U equilibrium surrogate suite: careful methods and honest scoping, limited mainly by synthetic-only evaluation.","tokens_in":24090,"tokens_out":545,"would_cite":false,"duration_ms":9942,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Deep-learning surrogates predict ADITYA-U MHD equilibrium parameters and profiles from a large synthetic free-boundary dataset within the machine’s circular-limiter flat-top domain.","keywords":["MHD equilibrium","Deep learning","Real-time control","ADITYA-U Tokamak","Grad-Shafranov","Physics-informed neural networks","Safety factor profile","Poloidal flux"],"falsifier":"Apply the trained forward models to a set of real ADITYA-U flat-top discharges that have independent equilibrium reconstructions (or diamagnetic βp and magnetic-axis measurements) never used in the synthetic generation pipeline; systematic errors larger than the reported synthetic percentiles would falsify the claimed transferability.","tokens_in":24078,"feed_emoji":"⚡","tokens_out":1125,"duration_ms":9587,"temperature":0.7,"pith_summary":"Tokamak control and analysis need the Grad-Shafranov equilibrium solution (plasma boundary, magnetic axis, safety factor, flux, inductance, and related scalars), but classical free-boundary solvers are too slow for real-time use. This paper builds a 100,760-case synthetic library of free-boundary equilibria for the ADITYA-U tokamak, generated from experimentally motivated inputs drawn from 766 discharges and filtered to circular limiter plasmas near flat-top. Dense networks, PCA reduced-order models, 1-D and 2-D CNNs, and physics-informed networks that penalize Grad-Shafranov residuals are trained to map readily available inputs (coil currents, Ip, Bφ, magnetic probes, loop voltage, and cascaded scalars) onto magnetic-axis location, βp, internal inductance, edge safety factor, full q(ρ), full ψ(R,Z), and selected poloidal-field coil currents. Within the domain spanned by the library the surrogates recover the target quantities accurately, with the largest model evaluating in roughly one millisecond. The practical claim is that these models can serve as fast substitutes for conventional equilibrium estimation in ADITYA-U real-time control, rapid discharge analysis, and experiment planning.","feed_headline":"Deep nets predict ADITYA-U plasma equilibria in ~1 ms","feed_subtitle":"Trained on 100k synthetic free-boundary cases, they recover axis, q, flux and coil currents inside the flat-top domain","key_machinery":"A large, physics-filtered synthetic free-boundary equilibrium library (pyIPREQ solutions of the Grad-Shafranov equation with a three-parameter Jφ profile, Bayesian-optimized γ, and a probabilistic βp prior) that supplies training targets for dense, PCA, CNN and Grad-Shafranov-residual PINN models.","core_discovery":"Key ADITYA-U MHD equilibrium parameters and profiles—Rax, Zax, βp, ℓi, q1, ψaxs/ψlim, the full q(ρ) profile, the full ψ(R,Z) map, and selected PF coil currents—can be accurately estimated by deep-learning surrogates trained on a 100,760-case pyIPREQ free-boundary synthetic dataset, with inference times of order 1 ms, inside the circular-limiter flat-top operational domain represented by that dataset.","pith_inferences":["Because the models rely on magnetic-probe and loop-voltage inputs, they cannot yet replace free-boundary solvers that operate from coil currents alone; a pure-actuator forward map would be a natural next library.","The low-dimensional PCA manifolds for both q and ψ suggest that simple parametric families already capture most ADITYA-U flat-top equilibria, so uncertainty-aware or active-learning extensions could focus data collection on the residual high-order modes.","Localized CNN distortions versus globally smooth PCA errors imply that hybrid PCA–CNN or ensemble predictors may be needed before the maps are trusted for stability calculations that depend on local shear or curvature."],"forward_implications":["Magnetic-axis position, βp, ℓi and edge q can be obtained in ~1 ms from magnetic diagnostics and coil currents, enabling real-time equilibrium feedback on ADITYA-U.","Full q(ρ) and ψ(R,Z) maps become available for rapid post-shot analysis without repeated free-boundary solves.","The inverse coil-current model supplies candidate PF actuator settings for desired plasma parameters inside the trained domain, supporting experimental planning.","Physics-informed residual losses keep the predicted flux maps approximately consistent with the Grad-Shafranov equation, reducing the chance of non-physical reconstructions.","The same library-plus-surrogate pattern can be reused for other circular-limiter machines once an analogous synthetic database is generated."],"fun_headline_variants":["Deep nets map ADITYA-U MHD equilibria from 100k cases in ~1 ms","Surrogates recover ADITYA-U axis q flux and coils inside flat-top domain","DL models predict ADITYA-U equilibrium scalars profiles and PF currents","Neural nets estimate ADITYA-U Grad-Shafranov solutions in 1 ms","1 ms deep learning surrogates for ADITYA-U plasma equilibria"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The synthetic library—built with a fixed parametric current-density form, a linear βp model fitted on only a few dozen measured discharges, and hard filters on axis shift, inductance and safety factor—must faithfully cover the equilibria that real ADITYA-U flat-top plasmas actually produce, so that synthetic test errors transfer to experiment.","fun_headline_variants_meta":{"raw":{"variants":["Deep nets map ADITYA-U MHD equilibria from 100k cases in ~1 ms","Surrogates recover ADITYA-U axis q flux and coils inside flat-top domain","DL models predict ADITYA-U equilibrium scalars profiles and PF currents","Neural nets estimate ADITYA-U Grad-Shafranov solutions in 1 ms","1 ms deep learning surrogates for ADITYA-U plasma equilibria"]},"model":"grok-4.5","effort":"low","cost_usd":0.003834,"raw_usage":{"total_tokens":1203,"prompt_tokens":798,"num_sources_used":0,"completion_tokens":94,"cost_in_usd_ticks":38340000,"prompt_tokens_details":{"text_tokens":798,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":311,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":798,"tokens_out":94,"duration_ms":3086,"temperature":1.0,"reasoning_tokens":311,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T12:17:23.605951+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Apply the trained forward models to a set of real ADITYA-U flat-top discharges that have independent equilibrium reconstructions (or diamagnetic βp and magnetic-axis measurements) never used in the synthetic generation pipeline; systematic errors larger than the reported synthetic percentiles would falsify the claimed transferability.","supporting_citations":[],"review_version":1}