{"id":"fac77663-d187-46fa-a867-80691fd84dc3","arxiv_id":"2502.07805","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Bayesian framework simultaneously reconstructs kinetic profiles and magnetic equilibrium from simulated ITER diagnostics, giving MAP results with uncertainties in about three minutes that mostly agree with MCMC verification.","lead":"This paper reports a software framework that reconstructs a fusion plasma's density, temperature, and magnetic field shape in one Bayesian analysis instead of in separate steps. Tested on simulated ITER data, it produced a self-consistent plasma state with uncertainty bars in about three minutes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop validation leaves surrogate bias unquantified; the central claim rests on E-Forward-NN accuracy, which is shown by only one qualitative comparison to EFIT.","rationale":"The reader's weakest_assumption identified the same core issue: the demonstration is in-distribution and closed-loop, with the ground truth drawn from the same database that trained the surrogate and set the priors. My stress-test concurs and sharpens the point: the single most load-bearing technical assumption is that E-Forward-NN is an accurate forward model within the intended operating regime. The paper provides no quantitative surrogate validation: no held-out error statistics, no test with magnetics generated by a different equilibrium code or with real data, and no sensitivity study showing how surrogate bias propagates to the inferred profiles and uncertainties. The MAP/MCMC agreement is genuine evidence for the inference machinery conditional on the surrogate, but it does not validate the surrogate itself. The paper's own limitations sections (IV A, VI) admit systematic errors are not modeled and the results should not be read as a quantitative ITER study. The abstract's 'statistically relevant uncertainties' is therefore stronger than what the evidence supports. A CONDITIONAL verdict is appropriate: accept the demonstrated capability, but require quantitative surrogate validation on held-out cases and, ideally, a test with data not generated by EFIT before the general ITER claim is accepted. This is a missing quantification step, not a fundamental flaw in the framework's architecture.","tokens_in":24343,"tokens_out":1919,"duration_ms":18127,"concrete_test":"Run the held-out 15% test split (or a new sample of 10,000 FUSE/EFIT equilibria not used in training) through E-Forward-NN and compare predicted magnetic signals, Chebyshev flux coefficients, axis/X-point positions, and total plasma current against EFIT. Report quantitative errors such as RMS probe-signal error relative to the Table III noise levels and maximum flux-surface displacement. If test-set errors are small compared to the assumed measurement uncertainties, the surrogate concern is largely retired; if not, the central claim and its uncertainty estimates are weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim — simultaneous kinetic profile and magnetic equilibrium reconstruction with statistically relevant uncertainties, fast enough for ITER between-discharge analysis — depends critically on the E-Forward-NN surrogate for EFIT. The entire demonstration is a closed loop: the ground truth is a FUSE/EFIT case from the same database used to train the surrogate and set the parameter transformation ranges (Section III C 1, III D 5), and the artificial magnetics data are generated by EFIT, the very code the neural network emulates. The only direct surrogate-versus-EFIT check is a single qualitative comparison in Fig. 3 for the final reconstructed case, with no quantitative error metric, no reported held-out test error, and no error bars on the flux surfaces. If the surrogate has systematic bias that a visual check can miss, then the inferred equilibrium, the flux mapping of TS/interferometry/polarimetry data, and the reported uncertainties are all jointly biased in a direction the artificial-data test cannot detect. This is not merely a lack of realism: the paper's Section IV A explicitly notes that no synthetic diagnostic models systematic errors, and Section VI cautions against interpreting the results as a quantitative study of ITER uncertainties. The claimed 'statistically relevant uncertainties' are therefore only as good as the surrogate, and that is the one unquantified link in the chain. This is the load-bearing concern because every other validated component (MAP/MCMC agreement, runtime, modularity) is conditional on the surrogate being accurate within the operating regime.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a modular Bayesian integrated data analysis (IDA) framework for tokamaks, built around IMAS data exchange, B-spline and Chebyshev parametrizations, empirical priors, and a neural-network surrogate for EFIT called E-Forward-NN. The authors demonstrate three increasingly complete workflows for an ITER-like FUSE scenario with artificial data: electron profile inference from Thomson scattering and interferometry/polarimetry; magnetic equilibrium inference from magnetics alone; and combined kinetic profile and equilibrium reconstruction. The combined MAP reconstruction, including uncertainty propagation, is reported to run in about three minutes on a single multicore node, and the MAP uncertainties are checked against MCMC. Most normalized chi-squared residuals are near unity, and the MCMC comparison shows good agreement for well-constrained quantities, with notable deviations in the magnetics-only case.","tokens_in":24638,"tokens_out":3503,"duration_ms":36513,"significance":"If validated beyond the current closed-loop demonstration, this framework addresses a genuine need: simultaneous, self-consistent kinetic profile and magnetic equilibrium reconstruction with uncertainty quantification fast enough for between-discharge analysis at ITER. The paper's strengths include a clean modular design, vectorized MAP optimization, a quantitative runtime table, and an MCMC verification step that is rarely undertaken at this scale. The main limitation is that the demonstration uses artificial data generated by the same forward-model family used for the neural-network surrogate, with the target case lying inside the training database; the only direct surrogate-versus-EFIT check is a single qualitative comparison. The significance is therefore conditional on additional out-of-sample validation of E-Forward-NN and on a clearer scoping of what the reported 'statistically relevant uncertainties' include.","major_comments":[{"comment":"The only direct comparison of E-Forward-NN with EFIT is a single qualitative flux-surface plot for the reconstructed case, with no quantitative error metric, no held-out test-set statistics, and no error bars on the flux surfaces. Since every inferred quantity and every reported uncertainty flows through this surrogate, please add quantitative test-set metrics such as root-mean-square or percentile errors of the flux matrix, synthetic magnetics, and total plasma current, and discuss the worst-case errors relative to the diagnostic noise levels.","section":"III D 5 and Fig. 3"},{"comment":"The validation is closed-loop: the target equilibrium is a FUSE case belonging to the same database used to train E-Forward-NN and to set the parameter transformation ranges covering 98% of that database, and the artificial magnetics data are produced by EFIT, the code the neural network emulates. Consequently, the test cannot detect surrogate bias or systematic deviations that would occur for real ITER conditions outside this distribution. Please add an out-of-distribution test, for example a FUSE/EFIT case excluded from the training set or a deliberately perturbed scenario, and quantify how surrogate error propagates into the inferred equilibrium and kinetic profiles.","section":"III C 1, III D 5, and IV A"},{"comment":"The authors themselves state in Section IV A that none of the synthetic diagnostics model systematic errors, and in Section VI that the results should not be interpreted as a quantitative study of expected ITER uncertainties. This directly qualifies the abstract's claim of 'statistically relevant uncertainties included.' Please reconcile the abstract and conclusions with these caveats, or add at least one systematic-error model for a representative diagnostic so that the uncertainty claim is demonstrated rather than asserted.","section":"IV A and VI"},{"comment":"The ion pressure reconstruction relies on eight artificial ion-pressure measurements because no synthetic ion diagnostic is implemented, yet the paper does not analyze the information content or sensitivity of this assumption. Since the inferred ion pressure appears as a central result in Fig. 13(h), please show how the solution and its uncertainty change when these constraints are varied in number, location, or uncertainty, or when they are removed entirely.","section":"IV D"},{"comment":"In the magnetics-only case, the MAP uncertainties for p' and f f' overestimate the MCMC uncertainties by more than 50% in some regions, and the corner plot in Fig. 17 shows clear non-Gaussian posterior structure. This means the 3-minute MAP uncertainty propagation is not universally valid. The paper should state explicitly that the fast MAP route is reliable only for the combined, well-constrained case and should give practical criteria for when MCMC verification is required.","section":"V B and Fig. 18"}],"minor_comments":[{"comment":"The figure contains garbled labels such as 'electron pro/f_iles' and 'TNe'; please regenerate the figure or fix the label rendering.","section":"Fig. 1"},{"comment":"The interferometer flat uncertainty is listed as '10 degrees' while the text discusses the vibration-compensated phase; please confirm the units and consistency with the synthetic diagnostic implementation.","section":"Table III"},{"comment":"In Eq. (2), the variables y and alpha are not fully defined in the surrounding text; please state explicitly that y is the profile value or its signed gradient and give the definition of alpha used for the ITER workflow.","section":"III G 2, Eq. (2)"},{"comment":"The sentence 'B-spline knot locations are fixed, and their locations were estimated from a training database' is redundant; please rephrase to avoid repeating 'locations'.","section":"III C"},{"comment":"In the paragraph introducing the new derived parametrizations, 'pion' should be written as 'pi' for the total ion pressure.","section":"IV D"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about closed-loop validation lands: the central claim depends on E-Forward-NN accuracy, and the manuscript provides only a single qualitative check. This is not a circularity in the derivation but a validation-design limitation that can be addressed by adding held-out test metrics and an out-of-distribution case, and by scoping the uncertainty claims. The paper is otherwise within the journal's scope and contains a substantial technical contribution. I recommend major revision rather than rejection because the load-bearing issue is fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a credible, well-scoped proof-of-principle for a fast, modular simultaneous kinetic-profile and magnetic-equilibrium IDA framework, but the validation is in-distribution and the neural-network surrogate's bias is unquantified. Treat the ITER-readiness claims as provisional.\n\nWhat's genuinely new: the paper assembles known pieces—B-spline profile parametrization, a neural-network Grad–Shafranov surrogate, coil-current inference, MAP with Hessian-based uncertainties, and MCMC verification—into one configurable workflow that speaks IMAS. That integration for ITER geometry, with the full diagnostic set (magnetics, TS, interferometry, polarimetry) solved jointly in about three minutes on a multicore CPU, is a real step forward relative to waterfall approaches like CAKE. The MCMC comparison is the strongest part: for well-constrained quantities, MAP uncertainties match MCMC well, and the paper is honest about where they don't (notably ff'). The authors also explicitly state that their synthetic diagnostics do not model systematic errors and that the results should not be read as a quantitative ITER uncertainty study. That candor earns credit.\n\nThe soft spots are proportionate to a proof-of-principle, but they are real. The entire chain depends on E-Forward-NN, and the only direct surrogate-vs-EFIT check is one qualitative flux-surface comparison (Fig. 3). There is no held-out error metric, no out-of-distribution test, and the ground truth comes from the same FUSE/EFIT database used to set the parameter transformation ranges and priors. The artificial magnetics are generated by EFIT itself, the code the network emulates. So the demonstration is in-distribution by construction; surrogate bias, if any, would propagate silently into the equilibrium, flux mapping, and the reported uncertainties. The abstract's phrase 'statistically relevant uncertainties' is stronger than the evidence, given the MAP/MCMC disagreement for poorly constrained profiles. Reproducibility is also limited: no code or trained models are released.\n\nWho should read it: anyone working on Bayesian IDA, equilibrium reconstruction, or NN surrogates for plasma state inference. It's a useful existence proof of the modular design and of how to run a MAP/MCMC consistency check.\n\nRecommendation: yes, send it to serious peer review. A good referee should push for quantitative surrogate validation (held-out and out-of-distribution error), code/model release, and a more measured abstract. The framework is sound on its own terms; the limitations are about validation depth, not a load-bearing flaw.","headline":"A solid, honest proof-of-principle for fast simultaneous profile+equilibrium IDA, with a load-bearing but unquantified NN surrogate and closed-loop validation.","tokens_in":25215,"tokens_out":3072,"would_cite":true,"duration_ms":26689,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.55.Fa","52.70.Kz","02.50.Tt"],"model":"deepseek-v4-flash","headline":"Neural-network Bayesian analysis joins kinetic profiles and magnetic equilibrium in one fast inference.","keywords":["Bayesian inference","integrated data analysis","tokamak equilibrium reconstruction","kinetic profile reconstruction","neural network surrogate","Grad-Shafranov equation","uncertainty quantification","ITER"],"falsifier":"Take an equilibrium deliberately outside the training ranges, or use measured data from a current tokamak with an independent high-resolution equilibrium constraint, and compare E-Forward-NN outputs to a full Grad–Shafranov solve and to independent diagnostics; if flux surfaces or synthetic magnetics deviate by more than the reported uncertainty bands, the surrogate is biased and the central claim weakens.","tokens_in":24106,"feed_emoji":"⚛️","tokens_out":9240,"duration_ms":78105,"temperature":0.7,"pith_summary":"Fusion reactors will have fewer, noisier diagnostics than today's machines, so plasma-state reconstruction must extract maximum information from limited data and report honest uncertainties. This paper presents a Bayesian integrated data analysis framework that reconstructs the magnetic equilibrium and the kinetic profiles (electron density, temperatures, pressures) simultaneously in a single inference step. In a test on a reactor-scale tokamak scenario with artificial data from magnetics, Thomson scattering, interferometry, and polarimetry, the maximum-a-posteriori (MAP) solution, including uncertainty propagation, took about three minutes on a multicore server and agreed with the known ground truth within reasonable tolerance. The framework replaces the expensive equilibrium solver with a neural-network surrogate, making gradient-based optimization fast enough for routine use between discharges.","feed_headline":"Joint Bayesian analysis reconstructs tokamak plasma in minutes","feed_subtitle":"Simultaneous kinetic profile and magnetic equilibrium inference with uncertainties, verified by MCMC, in minutes.","key_machinery":"The load-bearing component is E-Forward-NN, a set of neural networks trained on roughly 376,000 equilibria that maps low-dimensional B-spline coefficients of $p'$ and $f f'$, plus poloidal coil currents, to Chebyshev coefficients of the plasma flux, magnetic axis and x-point positions, synthetic magnetic measurements, and total plasma current. Around it, the framework combines profile parametrizations (B-splines on normalized poloidal flux, exponentials for positive quantities), a Chebyshev parametrization of the flux matrix, analytic derivatives, vectorized and parallelized posterior evaluation, and a gradient-based optimizer (BFGS) for MAP, with the inverse Hessian used to propagate uncertainties. This machinery turns a high-dimensional non-linear inverse problem into one solvable in minutes and verifiable by MCMC.","core_discovery":"The central claim is that the coupled inverse problem of equilibrium and kinetic profile reconstruction can be solved jointly, rather than in the usual sequential waterfall of separate fits, at a speed that makes integrated data analysis practical for reactor operation. The authors demonstrate this by inferring B-spline coefficients for the pressure gradient $p'$ and the function $f f'$, the poloidal field coil currents, the electron density and temperature profiles, and the derived ion and total pressures from a single posterior. A machine-learned surrogate for the Grad–Shafranov equilibrium solver, called E-Forward-NN, predicts the plasma flux response, separatrix positions, magnetic sensor readings, and total plasma current from the profile coefficients and coil currents, which makes the forward model differentiable and fast. Compared with magnetics-only reconstruction, adding the kinetic diagnostics sharply reduces the uncertainty of $p'$, $f f'$, and the flux surfaces. The MAP uncertainties were verified against Markov chain Monte Carlo (MCMC) sampling: agreement is good for well-constrained quantities, while in the magnetics-only case the posterior is markedly non-Gaussian and MAP overestimates uncertainties by more than 50% in some regions.","pith_inferences":["If the surrogate's training coverage is broadened and its bias characterized on out-of-sample equilibria, the same architecture could move from between-discharge analysis toward reactor control.","The sharp uncertainty reduction from combining kinetic and magnetic data suggests that spending on a few high-quality profile diagnostics may be as valuable for equilibrium knowledge as adding magnetics.","A testable next step is comparing inferences against independent equilibrium constraints on existing tokamak discharges, which would expose surrogate bias that a closed-loop artificial-data test cannot.","Adding Motional Stark Effect or poloidal polarimetry should make the $f f'$ posterior closer to Gaussian, improving MAP uncertainty accuracy, as the authors note."],"forward_implications":["Routine between-discharge reconstruction on a reactor can include all major diagnostics in one self-consistent analysis, because the combined MAP inference runs in about three minutes with uncertainty propagation included.","Kinetic measurements tighten the equilibrium: adding Thomson scattering, interferometry, and polarimetry shrinks the uncertainty of $p'$, $f f'$, and the flux surfaces compared to magnetics-only reconstruction.","Uncertainties in flux-surface mapping flow into profile uncertainties, closing a gap in conventional waterfall analyses where the equilibrium is treated as fixed.","Quantitative uncertainty statements for a burning-plasma device will require high-fidelity synthetic diagnostics that include systematic errors; the framework is modular so new forward models can be added.","MCMC remains necessary to audit MAP in data-poor regimes, because the magnetics-only posterior is non-Gaussian and MAP can misstate error bars."],"supporting_citations":[{"why":"Supplies the Grad–Shafranov equilibrium solver whose outputs are emulated by E-Forward-NN and used to generate artificial magnetics data for the validation scenario.","marker":"[9,10]"},{"why":"Integrated design tool that generated the transport equilibria forming the surrogate training database and the ground-truth scenario for the test.","marker":"[35]"},{"why":"Source of the neural-network training techniques for equilibrium surrogates, adapted here to predict Chebyshev flux coefficients.","marker":"[34]"},{"why":"Earlier neural-network accelerated Bayesian inference of profiles with self-consistent MHD equilibria; this work extends that idea to tokamak current and ion profiles.","marker":"[22]"},{"why":"Prior Bayesian inference of axisymmetric equilibrium with Gaussian processes on a tokamak; showed feasibility but impractical cost, motivating this faster surrogate approach.","marker":"[21]"},{"why":"Provides the interferometer and polarimeter formulas used as forward models for the TIP and DIP synthetic diagnostics.","marker":"[36]"},{"why":"Grounds the Bayesian likelihood and error-distribution treatment for fusion diagnostics, including the inadequacy of simple normal assumptions.","marker":"[11]"},{"why":"EMCEE ensemble sampler used for the MCMC verification runs that audit the MAP uncertainty estimates.","marker":"[26]"},{"why":"Established integrated data analysis for profile diagnostics on a tokamak; the framework generalizes this to joint equilibrium and profile reconstruction.","marker":"[5]"}],"fun_headline_variants":["Bayesian joint fit of profiles and equilibrium in minutes","One Bayesian pass: plasma state with uncertainties in minutes","Simultaneous kinetic and equilibrium inference for ITER","Fast Bayesian IDA: kinetic profiles plus equilibrium","Bayesian reconstruction for ITER: minutes, not hours"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that the plasma encountered in reality is similar to the database of roughly 376,000 computed equilibria used to train the neural-network stand-in for the equilibrium solver, and that the artificial test data behave like real measurements; if a real scenario falls outside that training distribution, or the stand-in's bias is larger than the single comparison shown, the reconstructed equilibrium and its error bars will be off.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian joint fit of profiles and equilibrium in minutes","One Bayesian pass: plasma state with uncertainties in minutes","Simultaneous kinetic and equilibrium inference for ITER","Fast Bayesian IDA: kinetic profiles plus equilibrium","Bayesian reconstruction for ITER: minutes, not hours"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000545,"raw_usage":{"total_tokens":2645,"prompt_tokens":1024,"completion_tokens":1621,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":1545}},"tokens_in":640,"tokens_out":1621,"duration_ms":10157,"temperature":1.0,"reasoning_tokens":1545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:05:27.110478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an equilibrium deliberately outside the training ranges, or use measured data from a current tokamak with an independent high-resolution equilibrium constraint, and compare E-Forward-NN outputs to a full Grad–Shafranov solve and to independent diagnostics; if flux surfaces or synthetic magnetics deviate by more than the reported uncertainty bands, the surrogate is biased and the central claim weakens.","supporting_citations":[{"cited_title":"Arbon , author J","cited_arxiv_id":null,"evidence_quote":"Integrated design tool that generated the transport equilibria forming the surrogate training database and the ground-truth scenario for the test."},{"cited_title":"Jardin ,\\ 10.1201/EBK1439810958 title Computational Methods in Plasma Physics ,\\ Vol.\\ volume 1st ed","cited_arxiv_id":null,"evidence_quote":"Source of the neural-network training techniques for equilibrium surrogates, adapted here to predict Chebyshev flux coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier neural-network accelerated Bayesian inference of profiles with self-consistent MHD equilibria; this work extends that idea to tokamak current and ion profiles."},{"cited_title":"Smith , author Z","cited_arxiv_id":null,"evidence_quote":"Prior Bayesian inference of axisymmetric equilibrium with Gaussian processes on a tokamak; showed feasibility but impractical cost, motivating this faster surrogate approach."},{"cited_title":"McClenaghan , author C","cited_arxiv_id":null,"evidence_quote":"Provides the interferometer and polarimeter formulas used as forward models for the TIP and DIP synthetic diagnostics."},{"cited_title":"Nishizawa , author R","cited_arxiv_id":null,"evidence_quote":"EMCEE ensemble sampler used for the MCMC verification runs that audit the MAP uncertainty estimates."},{"cited_title":"Walsh , author P","cited_arxiv_id":null,"evidence_quote":"Established integrated data analysis for profile diagnostics on a tokamak; the framework generalizes this to joint equilibrium and profile reconstruction."}],"review_version":1}