{"id":"e6b62e40-7795-4327-b39f-e089841a1689","arxiv_id":"2607.21407","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A cycle-consistent U-Net surrogate for SOLPS-ITER predicts 2D edge-plasma fields in milliseconds and recovers control parameters from fields, with self-consistency checks.","lead":"The paper trains a fast neural network to mimic slow tokamak edge-plasma simulations, then adds an inverse step that recovers the five input settings from a predicted plasma field. A fusion engineer could use it for quick parameter scans, real-time control, and automated fitting—if the numbers hold on real tokamak data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that cycle-consistency regularization enables recovery of Γcore is not tested: the reported ablation optimizes exactly the metric it reports, and inverse recovery is shown only at λcyc=0.5 without a λcyc=0 control.","rationale":"The forward-model results are credible: held-out evaluation, per-field metrics, and archived code/data support the NRMSE and correlation claims. The main novel claim, however, is the causal benefit of cycle consistency for inverse recovery. The reader's weakest assumption focuses on identifiability outside the training distribution; my concern is narrower and more immediate: the paper's own ablation metric is the training objective, and the actual inverse pipeline is never run at λcyc=0. This is an evidence gap rather than an internal contradiction, and it is directly addressable with the provided code and data. If the λcyc=0 inverse control shows no degradation, the abstract overstates the role of cycle consistency; if it shows degradation, the claim is supported. The reader's rationale already mentions the missing λcyc=0 control, so there is partial agreement, but the stated weakest assumption is about well-posedness/transfer rather than this causal-attribution issue. The verdict remains CONDITIONAL pending this control and, ideally, absolute-error metrics for the inverse parameters rather than Pearson correlation alone.","tokens_in":19031,"tokens_out":6495,"duration_ms":68064,"concrete_test":"Run the full inference-time inverse pipeline of Sec. 3.2 on the same 152 held-out cases using the λcyc=0 checkpoint (forward model and Gψ trained without the cycle term), with identical warm-start, N_r=5 restarts, and 1200 Adam steps. Compare per-parameter Pearson r and, crucially, normalized RMSE/R2 and slope/intercept for all five controls against the λcyc=0.5 results. If Γcore recovery does not degrade materially, rephrase the abstract to stop attributing core-fueling recovery to cycle consistency; if it degrades, the causal claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that cycle-consistency regularization 'enables recovery of the core fueling rate'—is not supported by the experiments as reported. The ablation in Sec. 4.3 (Fig. 4, Table 3) measures 'cyclical R2' by sweeping each control parameter through Fθ and then Gψ and comparing the swept to the recovered values. This is precisely the cycle term in Eq. (7), ||c − Gψ(Fθ(c,m),m)||, that the joint training minimizes at λcyc=0.5; reporting that it rises from 0.59 to 0.99 is reporting that the training objective was optimized, not that the inference-time inverse G⋆ (Eq. 5, Sec. 3.2) recovers true parameters. The actual inverse-recovery results in Sec. 4.4 are only for the adopted λcyc=0.5 checkpoint. No λcyc=0 checkpoint is pushed through the same optimization-based inverse pipeline, so we cannot tell whether Γcore recovery is caused by the cycle term, by the warm-start pseudo-inverse Gψ, or by the forward model alone. The paper even cautions (Sec. 4.3) that improved pair self-consistency 'does not by itself imply that the forward model has become more invertible,' and Sec. 4.4 warns that good cycle reconstruction does not guarantee parameter identification. Thus the abstract's attribution to cycle consistency goes beyond the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a cycle-consistent neural surrogate for SOLPS-ITER edge-plasma simulations of a DIII-D lower-single-null configuration. A conditional U-Net Fθ maps five control parameters (Ptot, Γcore, ΓD2, D⊥, χi) plus a geometry mask to 2D fields (Te, Ti, ne, ua). A differentiable optimizer through the frozen forward model, warm-started by a learned pseudo-inverse Gψ, recovers parameters from target fields; a cycle-consistency term in joint training regularizes the forward/pseudo-inverse pair. A separate committee of MLPs predicts 1D Te/ne profiles with quality flags. On a 152-run held-out test set, the forward model reports NRMSE below 2.6% and Pearson r above 0.975 for all fields, and inverse recovery reports r ≥ 0.97 for all five parameters at the adopted cycle weight. The paper also proposes a k-d tree warm-start strategy for database generation and claims a completion rate above 95% versus roughly 30% failures in a prior cold-started ensemble.","tokens_in":19462,"tokens_out":7575,"duration_ms":80588,"significance":"If the results hold, this is a valuable contribution to surrogate-based edge-plasma modeling: a fast, differentiable forward model, a practical inverse pipeline for parameter recovery, a self-supervised consistency check, and an uncertainty-aware 1D profile ensemble. The forward-accuracy evaluation is carefully done on a properly split held-out set, the convergence detection in Appendix B is thoughtful, and the code/data availability statements are a real strength. The authors also state important limitations (single geometry, no experimental validation, unenforced conservation laws), which is commendable. The main weakness is that the paper's central causal claim — that cycle-consistency regularization enables recovery of the core fueling rate — is not supported by the experiments as reported, because the ablation metric is essentially the training objective and no λcyc = 0 control is run through the inference-time inverse pipeline.","major_comments":[{"comment":"The average cyclical R2 reported in the ablation is computed by sweeping control parameters through Fθ and then Gψ and comparing the recovered values to the swept values. This is exactly the quantity minimized by the cycle term in Eq. (7), so increasing λcyc is expected to raise it; it is not a test of the inference-time optimizer G⋆ of Eq. (5). The abstract's claim that cycle-consistency regularization 'enables recovery of Γcore' is therefore unsupported: Sec. 4.4 reports inverse recovery only at λcyc=0.5, with no λcyc=0 checkpoint pushed through the same optimization-based inverse. Please add the missing λcyc=0 (and ideally intermediate) inverse-recovery results for all five parameters, or revise the causal claim. The manuscript's own caveats in Secs. 4.3 and 4.4 — that pair self-consistency need not imply invertibility and that good cycle reconstruction does not guarantee parameter id","section":"Sec. 4.3 / Eq. (7), Fig. 4, Table 3"},{"comment":"The paper presents Lcycle (Eq. 6) as a self-supervised quality check that 'requires no ground-truth labels at inference,' but it never validates that this metric is actually informative about parameter-recovery error. Given the explicit statement in Sec. 4.4 that good cycle reconstruction does not guarantee accurate recovery, the utility of Lcycle as a reliability diagnostic is unproven. Please provide direct evidence — e.g., a scatter or reliability diagram of Lcycle versus ∥ĉ − ctrue∥ on held-out cases — or temper the claim that the cycle loss functions as a quality check.","section":"Secs. 3.3 and 4.4"},{"comment":"The k-d tree warm-start completion-rate improvement (>95% versus ~70%) is supported only by a comparison to a different cold-started ensemble from Ref. [22], which used different runs and settings. This is not a same-design controlled comparison, so the quantitative gain attributable to the warm-start strategy is not established by the data reported here. Either run a cold-start control on the same LHS design and convergence criterion, or revise the claims to state only that the k-d tree strategy achieved >95% completion, without attributing a specific improvement relative to Ref. [22].","section":"Secs. 2 and 6"}],"minor_comments":[{"comment":"The quality score is written as 'si = k σi S', but the text and threshold 'si ≥ 1' imply si = k σi / S. Please fix the notation and define S immediately before use.","section":"Eq. (10)"},{"comment":"The ablation reports point estimates of test MSE and cyclical R2 without repeated-seed variation, even though the text notes run-to-run variation when training from scratch. Reporting mean ± std over a few seeds for each λcyc would strengthen the comparison.","section":"Table 3 / Fig. 4"},{"comment":"The window lengths Ttail and L are defined in terms of N_tail and N_win, but the numerical values of these multipliers are never stated. Please give the actual window sizes used in the convergence criterion.","section":"Appendix B"},{"comment":"Reference [8] is a disruption-prediction paper; it does not obviously support the statement about existing edge-plasma surrogates. Consider replacing it with a more direct citation to edge-plasma surrogate work.","section":"Ref. [8]"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and the forward-model contribution is solid, but the headline inverse-recovery claim needs a direct ablation at λcyc=0. If the authors can provide that control and confirm that Γcore recovery depends on the cycle term, I would be willing to support acceptance after revision. If the control shows no dependence, the abstract and conclusions should be rewritten to attribute recovery to the full forward/inverse pipeline rather than to cycle-consistency regularization. The k-d tree claim also needs a fairer baseline or more cautious wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. This is a competent, useful paper: a conditional U-Net maps five control parameters to 2D SOLPS-ITER fields, an optimization-based inverse recovers parameters through the frozen forward model, and a k-d tree warm-start makes database generation feasible. The held-out forward errors are genuine (NRMSE <2.6%, r>0.95 for all four fields), the inverse recovers all five parameters with r≥0.97 on 152 test runs, and the authors ship code and data. That's real work, properly split and carefully checked.\n\nThe new combination — 2D cycle-consistent surrogate on an irregular mesh plus optimization-based inversion — is not in the cited literature, and the paper says its limitations plainly: single geometry, uniform transport coefficients, no experimental validation. Credit where due.\n\nThe soft spot is the central attribution. The cyclical R2 ablation in Sec. 4.3 sweeps each parameter through Fθ then Gψ and compares to the swept value. That is exactly the cycle term in Eq. (7), the objective being optimized. So '0.59 to 0.99' tells us the training objective improved, not that the forward model became more invertible or that the inference-time inverse G⋆ gets better. The paper itself says improved pair self-consistency does not imply invertibility. Yet the abstract says cycle-consistency regularization 'enables recovery of the core fueling rate.' The inverse results in Fig. 5 are only at λcyc=0.5; there is no λcyc=0 control through the same inverse pipeline. Without that, we cannot say whether Γcore recovery comes from the cycle term, the warm-started pseudo-inverse, or the forward model alone. That is a fixable but real gap.\n\nMinor issues: λcyc is selected on test MSE, which is a form of test-set tuning; the warm-start comparison is against an external prior study, not a same-setup cold-start control; λreg is not reported. None of these change the main forward results, but they should be cleaned up.\n\nWho is this for? Fusion edge-physics modelers who want fast surrogates for scans, control, or transport inference. It deserves a serious referee. I'd accept it for peer review and ask for the λcyc=0 inverse control, the λreg value, and a reworded abstract that distinguishes pair self-consistency from parameter identification. After those changes, the paper's core claims would match the evidence.","headline":"Solid forward and inverse surrogates on held-out SOLPS data, but the abstract over-attributes Γcore recovery to cycle consistency; the ablation measures the training objective, not inverse benefit.","tokens_in":19969,"tokens_out":2425,"would_cite":true,"duration_ms":24050,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.55.Fa","52.65.-y"],"model":"deepseek-v4-flash","headline":"This paper claims that a cycle-consistent surrogate—a conditional U-Net forward model paired with an optimization-based inverse—can reproduce the two-dimensional tokamak edge-plasma state from five control parameters with under 2.6 percent","keywords":["SOLPS-ITER","edge plasma","neural surrogate","cycle consistency","uncertainty quantification","inverse modeling","tokamak","U-Net"],"falsifier":"Compute the numerical Jacobian of the trained forward model with respect to the five inputs at a grid of held-out points. If the smallest singular value is near zero relative to the largest, there is a direction in parameter space that the fields do not encode; then two very different parameter vectors produce nearly identical fields, and the claimed r≥0.97 parameter recovery would fail for noisy or out-of-distribution inputs.","tokens_in":18955,"feed_emoji":"⚛️","tokens_out":9497,"duration_ms":83021,"temperature":0.7,"pith_summary":"The paper aims to make tokamak edge-plasma modeling fast enough for real-time control, parameter scans, and inverse analysis by replacing multi-hour simulations of the scrape-off layer with a neural surrogate. It claims that a conditional U-Net trained on about 760 converged runs reproduces the four key plasma fields (electron and ion temperature, density, parallel velocity) with normalized errors below 2.6 percent and Pearson correlations above 0.95. The central new idea is cycle consistency: parameters recovered from a field, when pushed back through the forward model, must reproduce that field. This raises a self-consistency score from 0.59 to 0.99 without hurting forward accuracy, and it makes the inverse recover all five control parameters (including the core fueling rate, previously the hardest) with correlation at or above 0.97. A warm-started database built with a branching k-d tree reaches a completion rate above 95 percent, and each forward evaluation takes milliseconds—five to six orders of magnitude faster than the original solver.","feed_headline":"Maps tokamak edge plasmas in milliseconds, under 2.6% error","feed_subtitle":"Also recovers all five control parameters from fields, enabling fast inverse analysis and real-time control.","key_machinery":"The load-bearing mechanism is the cycle-consistency identity linking the forward map and the inverse: a parameter vector c maps to a field Fθ(c), and the inverse G⋆(Fθ(c)) should return c, while a target field y maps through G⋆(y) to parameters whose forward pass Fθ(G⋆(y)) should approximate y. The forward model is a U-Net (a convolutional network that preserves fine spatial detail via skip connections) conditioned on the five scalars through FiLM layers and restricted to the plasma domain by a binary mask; the optimization-based inverse G⋆ uses Adam through the frozen network, warm-started by a learned convolutional pseudo-inverse Gψ and refined with random restarts. The cycle constraint re","core_discovery":"The paper introduces a three-stage surrogate for edge plasmas: a conditional U-Net forward model maps five scalar control parameters (input power, core fueling, gas puff, and two cross-field transport coefficients) to the two-dimensional fields Te, Ti, ne, and parallel velocity; an optimization-based inverse recovers the parameters from a target field by differentiating through the frozen forward model; and a cycle-consistency stage passes the recovered parameters back through the forward model, giving a label-free quality check. On a held-out test set, all four fields achieve Pearson correlation above 0.95 and NRMSE below 2.6 percent, with parallel velocity the hardest (r=0.975). Cycle-cons","pith_inferences":["Beyond the paper: the same cycle-consistency round trip could serve as an out-of-distribution detector for any expensive surrogate—a large cycle error on a new query would signal regime shift even when no ground truth is available.","A testable extension is to train a second pseudo-inverse directly on the cycle-error gradient, potentially cutting the 1200-step Adam descent and making real-time inversion practical.","Because the inner divertor's near-strike-point electron temperature is bimodal across the database, active-learning acquisition focused there might resolve the observed R2≈0.89 and simultaneously improve inverse identifiability of parameters that set detachment onset.","The k-d tree warm-start is effectively numerical continuation in five dimensions; replacing the Euclidean distance with a physics-informed metric (e.g., weighting parameters by local sensitivity in the forward model) could further reduce the residual failure rate."],"forward_implications":["Parameter scans and sensitivity analysis across the five-dimensional operating space become feasible in milliseconds per evaluation, enabling broad exploration of detachment regimes.","The differentiable inverse offers an automated route to infer cross-field transport coefficients and fueling rates from measured or simulated fields, replacing hand-tuned iterative fitting.","Real-time plasma control loops (1–10 ms) could embed edge-plasma state estimation, since a forward pass takes about 16 ms on a single CPU core and sub-millisecond on a GPU.","The cycle-consistency round-trip error acts as a self-supervised reliability flag: at inference, predictions whose reconstructed fields deviate from the target can be marked untrustworthy without ground-truth parameters.","The committee ensemble flags the inner divertor near the strike point as the least certain regime (about 13 percent of points flagged), directing new expensive simulations to where they are most needed."],"fun_headline_variants":["Cycle-consistent AI maps tokamak edges in milliseconds","Tokamak plasma surrogate: fast, accurate, reversible","ML surrogate for tokamak edges: 2.6% error, ms speed","Uncertainty-aware AI for tokamak edge plasma control","Surrogate recovers tokamak inputs from fields, ms-fast"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The five control parameters are assumed to determine the two-dimensional steady-state fields uniquely enough that optimizing through the frozen forward model recovers the true parameters; the paper itself notes that the forward mapping is locally insensitive to some parameter combinations and that a clean cycle reconstruction does not by itself prove the parameters are identified.","fun_headline_variants_meta":{"raw":{"variants":["Cycle-consistent AI maps tokamak edges in milliseconds","Tokamak plasma surrogate: fast, accurate, reversible","ML surrogate for tokamak edges: 2.6% error, ms speed","Uncertainty-aware AI for tokamak edge plasma control","Surrogate recovers tokamak inputs from fields, ms-fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3342,"prompt_tokens":882,"completion_tokens":2460,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":2384}},"tokens_in":626,"tokens_out":2460,"duration_ms":17774,"temperature":1.0,"reasoning_tokens":2384,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:29:43.026943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the numerical Jacobian of the trained forward model with respect to the five inputs at a grid of held-out points. If the smallest singular value is near zero relative to the largest, there is a direction in parameter space that the fields do not encode; then two very different parameter vectors produce nearly identical fields, and the claimed r≥0.97 parameter recovery would fail for noisy or out-of-distribution inputs.","supporting_citations":[],"review_version":1}