{"id":"62e4fc8b-fc8e-49b3-9f19-5bf9b87a8234","arxiv_id":"2608.05669","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SafeDivertor maps 67 plasma signals to 116 radial heat-flux channels, beating eight time-series baselines on the new DivMPS2HF benchmark, though a large share of the gain comes from injecting the training-set mean profile into the model.","lead":"A machine-learning framework reconstructs divertor heat-flux profiles directly from fusion plasma diagnostic signals, bypassing the usual infrared-camera inversion workflow. The authors also present a 77-shot EAST dataset and report top results across five metrics, but the evaluation rests on a small test set and a prior built from training labels.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never benchmarks the trivial predictor that always outputs the training-set mean heat-flux profile (Eq. 4), so the claim that SafeDivertor's gains reflect signal-driven reconstruction is not yet established.","rationale":"I read the paper in good faith. The task is well-motivated, the dataset construction is described in detail, and the ablation study is internally consistent: each proposed component moves the reported metrics in a plausible direction. However, the strongest claim is an empirical superiority claim, and the evaluation omits the single most natural control for the main design choice. PPI injects the training-set mean of the target labels into the model, and the ablation shows that this injection alone accounts for most of the improvement over the T1 backbone. Without measuring the test performance of just outputting that same mean profile, the reader cannot distinguish between (a) a model that learns to refine a good empirical prior using plasma-state signals and (b) a model whose gain comes almost entirely from memorizing training-label statistics. The concern is not that PPI is illegitimate as a modeling choice; it is that the benchmark claim requires the prior to be representative of test shots, and the paper never tests the prior in isolation. The tiny effective test set (10 shots, 28,982 heavily overlapping windows) makes this omission more serious because it is hard to assess whether the 0.200 vs 0.211 delta is stable across shots. I do not see an internal inconsistency or a reason to reject the paper outright; the right disposition is to require the missing baseline and shot-level error bars before accepting the benchmark claim. The reader's weakest_assumption identifies the same load-bearing point, and I agree with it.","tokens_in":13894,"tokens_out":4344,"duration_ms":54795,"concrete_test":"Run the Table I evaluation with a no-learning baseline Y_hat(t) = mu (the training-set per-channel mean from Eq. 4 broadcast over all test windows), computing all five metrics. If the constant-prior MSE is below about 0.211 or within 10% of SafeDivertor's 0.200, the superiority claim is not supported; if the constant-prior MSE is substantially higher, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central benchmark claim is that SafeDivertor, mapping 67 macroscopic plasma-state channels to 116 heat-flux channels, outperforms all eight baselines on every metric. The decisive confound is PPI: the target-placeholder channels are initialized with the per-channel training-set mean heat-flux profile mu (Eq. 4-5), and Table II shows that adding PPI to the T1 backbone alone drops MSE from 0.262 to 0.211, a 19% reduction. This is exactly the statistic that a trivial constant-predictor baseline would output for every test window, yet no such baseline appears in Table I or the ablations. Because divertor heat-flux profiles are strongly dominated by a roughly fixed radial shape, the training mean may already be close to every test profile. If the constant-prior predictor achieves MSE near 0.211 or even near SafeDivertor's 0.200, then the reported 'best overall performance' mostly measures how well the training labels were memorized as a prior, not whether macroscopic signals are informative. The evaluation design compounds this: 28,982 test windows come from only 10 shots with stride-1 overlap, so the metric averages are pseudoreplicated and no shot-level confidence intervals are reported. The headline claim should be conditioned on demonstrating that a constant-prior baseline is substantially worse than SafeDivertor and that the signal channels contribute beyond the prior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DivMPS2HF, a multi-source discharge dataset for a new task: reconstructing time-resolved radial divertor heat-flux profiles (116 channels) directly from 67 macroscopic plasma-state signal channels over 0.5 s windows, without infrared thermography as input. On this dataset, the authors propose SafeDivertor, a T1 imputation backbone augmented by physical prior-aware initialization (PPI), input perturbation (IP), multi-scale STFT-based spectral loss (SRO), and progressive training (PT). The paper reports that SafeDivertor outperforms eight time-series baselines across all five metrics (MSE, MAE, SSIM, LSD, LSD-HF) and provides ablations, efficiency analysis, and qualitative visualizations. The code is promised on GitHub; the dataset is said to be available upon reasonable request.","tokens_in":14224,"tokens_out":4939,"duration_ms":53459,"significance":"If the claimed signal-to-flux mapping is genuinely learned, the paper makes a useful contribution: it formulates a new online-oriented reconstruction paradigm, constructs a shot-split benchmark dataset, compares against eight recent time-series models, and shows that the proposed components yield consistent quantitative and qualitative improvements over the T1 backbone. The efficiency analysis is also a strength, as SafeDivertor adds negligible overhead over T1. However, the central claim is currently conditional. The largest ablation gain comes from injecting the per-channel training-set mean of the target labels (PPI), and the paper does not benchmark the trivial predictor that always outputs that mean. Until that baseline and shot-level statistics are provided, the headline claim that the model learns a signal-driven mapping is not established.","major_comments":[{"comment":"The ablation evidence is insufficient to support the central claim that SafeDivertor learns a signal-to-flux mapping. PPI alone reduces the T1 backbone MSE from 0.262 to 0.211, a 19% drop, and PPI injects g_theta(P), where P is constructed from the per-channel training-set mean of the target heat-flux labels (Eqs. (4)–(5)). Because divertor heat-flux profiles are strongly dominated by a roughly fixed radial shape, a constant predictor that outputs the training-set mean radial profile for every test window may itself achieve an MSE close to 0.211 or even lower. Table I contains no such constant-prior baseline. Please add this trivial baseline with all five metrics, and additionally ablate the signal channels (e.g., feed noise or zeros in place of the observed X while retaining PPI) to demonstrate that the macroscopic signals contribute beyond the injected label statistic. The term 'physical prior' is also misleading: Eq. (4) is a statistical average of training labels, not a physics-based prior, and should be renamed or explicitly characterized as a label-statistic prior.","section":"§V-D, Eq. (4)–(7), Table II"},{"comment":"The evaluation is pseudoreplicated. The 28,982 test windows come from only 10 shots and are generated with a stride of one time step, so adjacent windows overlap in 499 of 500 time steps and are not independent. All metrics in Table I and Table II are pooled over these non-independent windows without shot-level aggregation or confidence intervals. With only 10 test shots, a model can obtain a spuriously good pooled score by performing well on a few shots. Please report per-shot results, shot-level means and standard errors, and state how many of the 10 test shots each improvement holds for.","section":"§V-A, §V-C"},{"comment":"No error bars or repeated-seed results are reported. Several headline differences are small in absolute terms — for example, the full model's SSIM of 0.870 equals the PPI-only ablation's SSIM of 0.870 in Table II, and the MSE difference between the full model and PPI-only is 0.200 vs. 0.211. With a single training run one cannot distinguish genuine gains from optimization noise. Please run at least three random seeds and report mean ± standard deviation for the main comparison and the ablation tables.","section":"§V-C, §V-D, Tables I–II"}],"minor_comments":[{"comment":"The multi-scale STFT notation Y_r and \\hat Y_r is not defined; please specify the STFT window sizes, hop lengths, and how the three temporal scales are constructed.","section":"Eq. (11)"},{"comment":"The paper states that the 67 input channels were 'selected from a broader pool of candidate signals' but does not report the size of the candidate pool or the selection criterion; this information is needed for reproducibility of the dataset.","section":"§IV-A"},{"comment":"The qualitative samples are described as 'randomly selected'; please state the random seed or selection procedure so that the figures can be reproduced.","section":"Figures 3 and 4"},{"comment":"The dataset is described as a benchmark but is to be made available only 'upon reasonable request', while the code is on GitHub; please clarify the data-release conditions in the final version, since a benchmark paper benefits from a public dataset repository.","section":"§V-A and §VI"},{"comment":"The latency of 11.313 ms is compared only against other models, not against the 0.5 s window duration as a real-time budget; please state the end-to-end latency (including preprocessing) and its margin relative to the online requirement.","section":"§V-G, Table V"},{"comment":"The caption introduces 'CM' and 'GP' but the table body would be clearer if the column headers repeated the labels; as printed, the reader must map the caption abbreviations to the two sub-columns under each ratio.","section":"Table IV"},{"comment":"Progressive training (PT) is never ablated in isolation; its effect is visible only by comparing rows #5 and #6, and the direction is mixed (LSD and LSD-HF worsen from 2.326/2.568 to 2.475/2.729). Please discuss whether PT is intended to trade spectral metrics for time-domain metrics.","section":"§V-D, Table II"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The core issue is an evaluation gap rather than a mathematical error. The paper is a strong candidate for publication after the authors add (i) a constant-prior baseline, (ii) shot-level statistics with confidence intervals, and (iii) repeated-seed results. I would also encourage the authors to make the dataset publicly accessible if it is positioned as a community benchmark, since 'available upon reasonable request' weakens the reproducibility claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper introduces a genuinely new task and a new dataset, and the method beats its chosen baselines. That said, the central claim is not yet proven because they never compare against the trivial predictor that always outputs the training-set mean heat-flux profile.\n\nWhat is actually new: the DivMPS2HF dataset (77 shots, 67 macroscopic signal channels to 116 radial heat-flux channels, shot-level split) and the task formulation itself. SafeDivertor is a sensible adaptation of the T1 imputation backbone with task-specific components: PPI (prior injection), input perturbation, multi-scale STFT loss, and progressive training. The paper does well by running a broad set of baselines, showing consistent ablation trends, and reporting efficiency. The SRO component clearly improves spectral metrics, as you'd expect. Code and dataset availability is also a plus.\n\nThe soft spot is real and central. PPI injects the per-channel training-set mean heat flux as a prior, and Table II shows PPI alone drops MSE from 0.262 to 0.211, a 19% reduction. Because divertor heat-flux profiles have a fairly stable radial shape, the training-set mean may already be close to every test profile. Without a constant-prior baseline in Table I, we cannot tell whether SafeDivertor's 0.200 MSE is mostly memorized label statistics or genuine signal-driven reconstruction. The evaluation design compounds this: 28,982 test windows come from only 10 shots with stride-1 overlap, so the reported averages are pseudoreplicated and no shot-level confidence intervals are given. This is not a fatal flaw, but it is a load-bearing gap in the benchmark claim.\n\nTo be fair, the method still has to map signals to flux; PPI is only an initialization, and the signal channels must carry information for the model to improve beyond the prior. A good revision could add the trivial prior baseline, report shot-level metrics, and reduce window overlap. The paper is a useful contribution that needs a more rigorous evaluation before the headline claim can be accepted.\n\nWho is this for: the ML-for-plasma and tokamak diagnostics community, especially people working on time-series reconstruction with prior injection. It deserves a serious referee; the missing baseline and evaluation details are addressable in revision. I'd send it to review.","headline":"Worth a referee: real new dataset and task, but the benchmark claim is undercut by the absence of a trivial prior-only baseline and by heavily overlapping test windows.","tokens_in":14752,"tokens_out":1845,"would_cite":false,"duration_ms":21596,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.55.Fa","52.40.Hf"],"model":"deepseek-v4-flash","headline":"Divertor heat flux can be reconstructed during a discharge from routine plasma-state signals, without infrared imaging or heat-conduction inversion; the authors' framework beats eight time-series baselines on all five metrics.","keywords":["divertor heat flux","signal-based reconstruction","multivariate time series","physical prior","short-time Fourier transform","spectral loss","tokamak plasma diagnostics","online monitoring"],"falsifier":"Run SafeDivertor with and without the physical prior on a held-out set of shots whose mean radial heat-flux profile differs clearly from the training distribution, for example a different confinement regime or a different strike-point configuration; the central claim would be undermined if the prior-equipped version performs worse than the no-prior baseline on that shifted set while still winning on in-distribution shots.","tokens_in":13735,"feed_emoji":"🔥","tokens_out":9134,"duration_ms":86015,"temperature":0.7,"pith_summary":"This paper claims that the time-resolved radial heat-flux profile on a tokamak divertor can be reconstructed while the discharge is running, directly from macroscopic plasma-state signals that are already recorded, with no infrared camera and no heat-conduction solver. To support the claim, the authors assemble DivMPS2HF, a dataset of 77 discharges that aligns 67 input channels with 116-channel radial heat-flux labels, and propose SafeDivertor, which fills zero-valued target placeholders with a training-set mean heat-flux prior, perturbs input channels during training, and supervises reconstruction with a multi-scale short-time Fourier transform loss. On held-out shots, SafeDivertor beats the best of eight time-series baselines on all five metrics, reducing MSE from 0.262 to 0.200 and high-frequency log-spectral distance from 3.829 to 2.729. If the claim holds, divertor heat-load monitoring and protection become an online inference problem rather than a post-shot thermal-analysis task.","feed_headline":"Heat flux reconstructed from plasma signals, no infrared needed","feed_subtitle":"SafeDivertor maps 67 tokamak signals to heat-flux profiles in 11 ms, beating eight baselines on every metric.","key_machinery":"The load-bearing mechanism is the combination of physical prior-aware initialization with spectral-aware reconstruction optimization. The prior is a per-channel mean heat-flux profile computed from the training set, whose encoded feature is added only to the target-placeholder representations, giving the model radial-distribution guidance without touching the observed channels. The spectral loss matches the log-magnitude short-time Fourier transform spectra of reconstruction and ground truth at three temporal scales, which counteracts the smoothing induced by point-wise MSE and preserves transient high-frequency dynamics. Input perturbation regularizes the model against over-reliance on any particular diagnostic channel, while progressive training introduces MSE first, perturbation second, and the spectral loss third, so that the complementary objectives do not destabilize each other.","core_discovery":"The central discovery, on the paper's own terms, is that divertor heat-flux reconstruction can be reformulated as a structured channel-level reconstruction problem: concatenate the observed plasma-state signals with zero-valued placeholders for the 116 heat-flux channels, then train a single multivariate time-series model to fill in the placeholders. The authors report that SafeDivertor does this faithfully, preserving both the overall spatiotemporal structure of the heat-flux pattern and its high-frequency transient variations, and that it establishes a new benchmark with MSE 0.200, MAE 0.291, SSIM 0.870, LSD 2.475, and LSD-HF 2.729, the best result among the nine compared models on every metric.","pith_inferences":["A consequence the paper leaves implicit is that the same signal-to-flux mapping could serve as a cross-check or fallback for infrared inversion whenever IR diagnostics are unavailable, saturated, or being serviced.","A testable extension is regime-adaptive priors: since the ablation shows the training-set mean profile drives the pointwise improvement, updating the prior per confinement regime (attached versus detached) or per discharge could be the clearest next gain in accuracy.","The structured-reconstruction scaffolding may transfer to other plasma quantities currently available only post-shot, such as radiation emissivity or particle-flux profiles, giving operators online estimates of quantities they now reconstruct offline."],"forward_implications":["Heat-flux profiles become available in a single neural forward pass at 11.313 ms per 0.5-second window, a latency compatible with window-level online monitoring during a discharge.","The multi-scale spectral loss is the main lever for transient fidelity: adding it reduces LSD from 3.521 to 2.306 and LSD-HF from 3.829 to 2.511.","The training-set mean heat-flux prior alone lowers MSE from 0.262 to 0.211 and raises SSIM from 0.848 to 0.870, showing that statistical radial guidance carries much of the pointwise gain.","Off-the-shelf forecasting and exogenous-variable models are insufficient for this task; the strongest baseline is a time-series imputation model, and SafeDivertor improves on it across all five metrics.","DivMPS2HF gives the community a fixed shot-level split and a five-metric protocol for comparing future signal-based divertor heat-flux reconstruction methods."],"supporting_citations":[{"why":"Supplies the T1 backbone, the multivariate time-series imputation architecture that SafeDivertor extends with its task-driven components.","marker":"[30]"},{"why":"Represents the conventional thermographic reconstruction pipeline whose post-shot heat-flux outputs serve as the supervised labels.","marker":"[5]"},{"why":"Defines the divertor heat-flux analysis problem and the offline finite-element solver that the proposed paradigm aims to bypass.","marker":"[1]"},{"why":"Motivates online-oriented heat-flux estimation by bringing physics-informed neural networks to real-time heat-flux calculation.","marker":"[3]"},{"why":"Establishes that macroscopic plasma-state signals carry predictive information about plasma behavior, supporting the choice of input channels.","marker":"[9]"},{"why":"Defines the structural similarity metric used to evaluate spatiotemporal reconstruction quality.","marker":"[33]"},{"why":"Defines log-spectral distance, the metric used to quantify spectral fidelity in the benchmark.","marker":"[34]"}],"fun_headline_variants":["SafeDivertor reconstructs divertor heat flux from plasma signals without infrared","AI model maps 67 tokamak signals to heat flux profiles in 11 milliseconds","Plasma signals alone can reconstruct divertor heat flux in 11 milliseconds","No infrared needed: SafeDivertor maps plasma signals to heat flux in 11 ms","SafeDivertor sets a benchmark for signal-based divertor heat flux reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the training-set mean radial heat-flux profile, used as the physical prior, is representative of the test shots; if test conditions shift the mean heat-load shape, the prior will bias the reconstruction instead of guiding it.","fun_headline_variants_meta":{"raw":{"variants":["SafeDivertor reconstructs divertor heat flux from plasma signals without infrared","AI model maps 67 tokamak signals to heat flux profiles in 11 milliseconds","Plasma signals alone can reconstruct divertor heat flux in 11 milliseconds","No infrared needed: SafeDivertor maps plasma signals to heat flux in 11 ms","SafeDivertor sets a benchmark for signal-based divertor heat flux reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001099,"raw_usage":{"total_tokens":4600,"prompt_tokens":973,"completion_tokens":3627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":3522}},"tokens_in":589,"tokens_out":3627,"duration_ms":27598,"temperature":1.0,"reasoning_tokens":3522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:25:39.854904+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SafeDivertor with and without the physical prior on a held-out set of shots whose mean radial heat-flux profile differs clearly from the training distribution, for example a different confinement regime or a different strike-point configuration; the central claim would be undermined if the prior-equipped version performs worse than the no-prior baseline on that shifted set while still winning on in-distribution shots.","supporting_citations":[{"cited_title":"T1: One-to-one channel-head binding for multivariate time-series imputation,","cited_arxiv_id":null,"evidence_quote":"Supplies the T1 backbone, the multivariate time-series imputation architecture that SafeDivertor extends with its task-driven components."},{"cited_title":"Thermographic studies of outer target heat fluxes on kstar,","cited_arxiv_id":null,"evidence_quote":"Represents the conventional thermographic reconstruction pipeline whose post-shot heat-flux outputs serve as the supervised labels."},{"cited_title":"The development of a three-dimensional finite element method code for the heat flux analysis of tungsten monoblock divertor on east,","cited_arxiv_id":null,"evidence_quote":"Defines the divertor heat-flux analysis problem and the offline finite-element solver that the proposed paradigm aims to bypass."},{"cited_title":"Physics informed neural networks towards the real-time calculation of heat fluxes at w7-x,","cited_arxiv_id":null,"evidence_quote":"Motivates online-oriented heat-flux estimation by bringing physics-informed neural networks to real-time heat-flux calculation."},{"cited_title":"Predicting disruptive instabilities in controlled fusion plasmas through deep learning,","cited_arxiv_id":null,"evidence_quote":"Establishes that macroscopic plasma-state signals carry predictive information about plasma behavior, supporting the choice of input channels."},{"cited_title":"Audio super-resolution with latent bridge models,","cited_arxiv_id":null,"evidence_quote":"Defines log-spectral distance, the metric used to quantify spectral fidelity in the benchmark."}],"review_version":1}