{"id":"245432bc-f288-48e7-9942-712fc89eada8","arxiv_id":"2602.24049","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A two-parameter algorithm automatically constructs stellarator island divertors, and Bayesian optimization finds designs with peak heat flux ~3 MW/m² at far lower cost than grid scanning.","lead":"This paper presents an automated algorithm that builds stellarator island divertor shapes from just two starting points, then uses Bayesian optimization to find designs with peak heat fluxes near 3 MW/m², below the 10 MW/m² material limit. A generalist reader might care because it automates a hard plasma-exhaust engineering problem central to stellarator fusion power plants.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cost function omits vessel-wall strikes, so low qpeak may reflect power lost to the wall; 'high power fraction captured' is never quantified.","rationale":"The most load-bearing concern is not the control-surface fidelity (which the reader identified) but the absence of a power-capture term in the optimization objective. This is more immediate because it threatens the validity of the paper's headline result even under its own modeling assumptions. The paper's Sec. 3.4 shows that wall-hit fraction is a known failure diagnostic; it is simply not included in J. Without reporting the wall-hit fraction for the best design, the claim that the divertor satisfies engineering limits is incomplete. The concrete test is cheap and uses existing FLARE outputs. If the wall-hit fraction is low, the paper's central argument is substantially strengthened and the remaining issues (control surface, FLARE fidelity, exaggeration) are addressable. If it is high, the optimization has been gaming the metric by losing power. The reader's weakest_assumption focused on the control surface and the FLARE model; I partially agree with those, but the missing metric is a more fundamental internal gap. The minor 95%/92% discrepancy and the 'proven robust' overstatement are secondary. Overall, the verdict should remain CONDITIONAL: accept only if the wall-capture data are provided and, if needed, the objective is revised.","tokens_in":15025,"tokens_out":10324,"duration_ms":90415,"concrete_test":"Run FLARE's heat-load proxy for the optimized divertor (θL, θR) = (4.05, 1.21) with nominal parameters (1e6 particles, P_SOL = 8 MW, χ_i = 0.1 m²/s) and tabulate the fraction of launched particles that strike (a) the main divertor plate, (b) the end-plate, and (c) the vessel wall. If the wall-hit fraction exceeds ~5% (the threshold used in Sec. 3.4 to declare incomplete capture), then qpeak ≈ 3 MW/m² does not establish a viable divertor, and the objective in Eq. (2) must be augmented with a power-capture or wall-loss penalty before re-optimization.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim — that the algorithm finds divertors meeting the 10 MW/m² limit while maintaining high power capture — is not supported by the optimization objective. Eq. (2) defines J = W1*qpeak + W2*end-plate%, with no term for particles that miss the divertor and hit the vessel wall. The end-plate in Sec. 2.3 blocks backside strikes, but not wall losses. The authors themselves use wall-hit percentage in Sec. 3.4 as the criterion for when the divertor 'isn't capturing all the heat flux' (they say the scaling in Eq. 3 fails once wall hits exceed 5%). Yet the main results in Secs. 3.1–3.3 report only qpeak and end-plate %; the wall-hit fraction for the optimized divertor is never given. Because qpeak is derived from the density of particles actually striking the plate, a design that channels a large fraction of the launched power to the wall could show a spuriously low qpeak while also having low end-plate %, yielding a low (good) cost. The abstract's 'high power fraction captured' is therefore unquantified. The reported 95% cost reduction is also inflated (35 vs. 441 simulations is ~92%), but that is a minor numerical error compared to the missing power-capture metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an automated algorithm for constructing stellarator island divertors from two scalar inputs (θ_L, θ_R) by enforcing a maximum magnetic-field strike angle on a control surface inside the island. The resulting divertor geometries are evaluated with the FLARE field-line diffusion model, which returns q_peak and an 'end-plate hit %' metric. The authors perform a 21×21 grid scan and a Bayesian optimization run, finding essentially the same optimal divertor (q_peak ≈ 3 MW/m², end-plate ≈ 3.5%) and claim a 95% reduction in simulations. They also study robustness to the cross-field diffusion coefficient D and report a disagreement with the expected λ_q,t ∼ √D scaling at high D, attributing it to wall strikes. The abstract additionally claims that low-elongation islands are easiest to optimize, but this result does not appear in the body.","tokens_in":15344,"tokens_out":3768,"duration_ms":39292,"significance":"If the claims are fully supported, the paper would offer a genuinely low-dimensional, automated divertor-design workflow for fixed stellarator equilibria, enabling rapid exploration of island-divertor geometry. The use of FLARE for on-the-fly heat-flux evaluation and the demonstration that Bayesian optimization recovers the grid-scan optimum are valuable proof-of-principle contributions. However, the central claim that the optimized divertors 'maintain a high power fraction captured' is not currently substantiated: the cost function omits vessel-wall strikes, and the wall-hit fraction for the reported designs is never given. The abstract-body mismatch and the absence of Monte-Carlo uncertainty quantification further weaken the paper in its present form.","major_comments":[{"comment":"The cost function J = W1 q_peak + W2 end-plate% contains no term for field lines that terminate on the vessel wall. Since q_peak is computed only from particles actually hitting the divertor plate, a design that channels a large fraction of power to the wall can appear artificially good. The paper itself uses wall-hit percentage in Sec. 3.4 as a criterion for when the divertor 'isn't capturing all the heat flux' (≥5%), yet the optimized designs in Sec. 3.2 are reported only via q_peak and end-plate%. The abstract's 'high power fraction captured' is therefore unquantified. Please report the wall-hit fraction for the grid-scan and Bayesian-optimization best designs, and ideally add a wall-strike penalty to Eq. (2) or discuss why it is unnecessary for the explored parameter range.","section":"Sec. 2.3, Eq. (2); Sec. 3.2; Sec. 3.4"},{"comment":"The abstract states: 'Optimization over various islands in the equilibrium shows that low-elongation islands are the easiest to find divertors that satisfy heat flux requirements.' The full text provided contains no such optimization over multiple islands, no elongation analysis, and no comparison of different islands. This is a load-bearing discrepancy: either the result is missing from the body, or the abstract is unsupported. Please either include the multi-island study or revise the abstract to reflect the single-island results actually presented.","section":"Abstract vs. Secs. 2–4"},{"comment":"FLARE's field-line diffusion model is Monte-Carlo (1e6 particles per simulation), and q_peak is an extremal quantity. The paper provides no estimate of statistical uncertainty on q_peak or end-plate%, and the Gaussian-process optimizer is applied as if the observations were deterministic. Noise could affect both the claimed optimum and the BO-vs-grid comparison. Please add error bars (e.g., from repeated simulations or bootstrap) or at least quantify the typical run-to-run variation of the cost-function components.","section":"Secs. 3.1–3.2, Figs. 4–6"}],"minor_comments":[{"comment":"The claimed '95% reduction in computational cost' is inaccurate: 10 initial + 25 Bayesian steps = 35 simulations vs. 441 in the grid scan, which is a 92.1% reduction. Please correct the number.","section":"Abstract and Sec. 3.2"},{"comment":"The x-axis label 'field line diffusion coefficient [m]' has unclear units; the coefficient D in Eq. (3) should have units of length²/time (or the plot should specify the normalized quantity).","section":"Fig. 10"},{"comment":"There is a typo in the definition of p_L: 'p_L = (r_L0, φ, z_L0)' should be p_L = (r_L0, φ_init, z_L0), and the subsequent 'p_L = (r_R0, ...)' should be p_R. Please proofread the coordinate definitions.","section":"Sec. 2.1"},{"comment":"The statement that the optimized divertor is 'proven to be robust' is stronger than what the single-D-scan supports; the authors themselves note the scaling breaks down at high D. Suggest softening to 'shown to be robust over the tested D range'.","section":"Sec. 3.4"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the missing wall-strike metric in the cost function and the lack of wall-hit percentages for the optimized designs; this directly bears on the abstract's 'high power fraction captured' claim. The abstract-body mismatch regarding multi-island optimization is also a red flag and must be resolved. The proposed algorithm and BO framework are promising, but the manuscript needs these load-bearing fixes before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's new: the algorithm itself. Two user-picked starting angles on a control surface generate a full V-shaped divertor with a prescribed max strike angle. That's a genuinely lower-dimensional parameterization than Davies et al.'s semi-automated tilting, and coupling it to Bayesian optimization is a sensible way to search it. The grid scan and the BO run land in the same valley, and the Pareto analysis is straightforward. The λq,t ~ √D check is also a good-faith robustness test, and it honestly shows the scaling breaks at high D when particles start hitting the wall.\n\nWhere it's soft: the cost function in Eq. (2) has no term for particles that miss the divertor and hit the vessel wall. The paper itself says in Sec. 3.4 that once wall hits exceed 5%, the divertor \"isn't capturing all the heat flux.\" But the main results never report the wall-hit fraction for the optimized divertor. Without that number, you can't tell whether qpeak ≈ 3 MW/m² is a genuine engineering solution or partly an artifact of dumping power on the wall. This is the main issue, and it's fixable: add a wall-loss term to J, or at least report wall-hit fractions for every design on the Pareto frontier.\n\nMinor points: the \"95% reduction\" claim is 35 vs 441 simulations, which is ~92%, not 95. The Monte Carlo qpeak comes from 1e6 particles with no error bars, so it's unclear how much of the fine structure in the Pareto frontier is noise. There's also only one equilibrium, one P_SOL, and one assumed χ⊥; the abstract says \"proven robust,\" which overstates a single-parameter scan. The control-surface assumption is acknowledged and reasonable for a first pass, but it isn't tested against a stochastic separatrix.\n\nNo hidden circularity here: qpeak is evaluated by FLARE, not fitted, and the scaling test is independent. Self-citations point to the relevant FLARE and design papers, which is appropriate.\n\nBottom line: this is a solid proof-of-principle for a useful design tool. It deserves a serious referee, and the main revision should be about accounting for and reporting power to the wall. I'd cite it if I worked on island divertors.","headline":"Automated two-parameter island divertor design is a genuine step forward, but the optimization omits wall-strike losses so the headline heat-flux numbers may be flattered.","tokens_in":15888,"tokens_out":2131,"would_cite":true,"duration_ms":20977,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.55.Hc","52.55.Rk"],"model":"deepseek-v4-flash","headline":"An automated two-parameter algorithm designs stellarator island divertors whose peak heat flux lands near 3 MW/m², under the 10 MW/m² material limit.","keywords":["stellarator","island divertor","divertor shape optimization","heat flux","field line diffusion","Bayesian optimization","strike angle","power exhaust"],"falsifier":"A direct test would be to run a higher-fidelity plasma and neutral transport simulation on the best optimized divertor in the same equilibrium and check whether the predicted peak heat flux on the actual divertor mesh remains below 10 MW/m² under the same scrape-off-layer power. If the reduced model's roughly 3 MW/m² result rises above the engineering limit by a factor of about three, the central claim that this design satisfies material heat-load limits is falsified.","tokens_in":14868,"feed_emoji":"⚛️","tokens_out":4591,"duration_ms":44472,"temperature":0.7,"pith_summary":"The paper presents a fully automated way to construct an island divertor for a stellarator from just two starting points on a magnetic island's separatrix. By stepping toroidally and enforcing a maximum magnetic-field strike angle on the plate, the algorithm generates V-shaped curved divertor plates that spread heat flux. Using a field-line diffusion model to estimate heat loads, the authors scan all two-input configurations and find divertors whose peak heat flux is about 3 MW/m², below the 10 MW/m² engineering limit for tungsten, with only about 3.5% of field lines hitting undesirable end-plates. They then show Bayesian optimization finds essentially the same best design after roughly 35 simulations instead of 441, a 95% reduction in cost. A sympathetic reader would care because this gives stellarator power-plant designers a cheap, low-dimensional search tool for divertor geometry in a fixed magnetic equilibrium.","feed_headline":"Two-angle recipe designs divertors at 3 MW/m²","feed_subtitle":"Bayesian search matches a 441-run scan with ~35 simulations, making island-divertor design cheap enough for stellarator optimization loops.","key_machinery":"The key mechanism is the strike-angle constraint step: at each toroidal step, the algorithm searches on a smooth control surface inside the island for a point where the unit magnetic-field vector at the midpoint satisfies b̂·d̂ = cos(α), so the plate is locally oriented to keep the field incidence at or below a user-set maximum angle. Combined with a spline representation of the island surface and a V-shaped geometry with a rounded junction, this converts divertor design into two scalar inputs and creates curved plates that spread heat flux. The paper evaluates designs with a source-free field-line diffusion model, tracing many particles from the last closed flux surface onto the plate mesh","core_discovery":"The central claim is that the entire useful shape space of an island divertor in a fixed stellarator equilibrium can be parameterized by two angles on one toroidal plane, and that this low-dimensional map lets optimization find divertors that satisfy engineering heat-load limits. The authors' algorithm builds each plate by following field lines from the two starting points and, at each toroidal step, locating the point on the island surface where the field direction meets a prescribed maximum strike angle of 3°. The best grid-scan divertor has a peak heat flux of 2.915 MW/m² and a 3.487% end-plate hit fraction; Bayesian optimization recovers a nearly identical design (3 MW/m², 3.53%) with ab","pith_inferences":["Editorial inference: Because the algorithm needs only a control surface and a magnetic-field solve, the same two-input construction likely applies to other resonant islands and to multiple divertor modules per field period, possibly enabling whole-device exhaust optimization rather than single-island designs.","Editorial inference: The 95% savings suggest a testable extension: running the same Bayesian optimizer with strike angle, toroidal position, and toroidal extent as free parameters, then checking whether the predicted cost savings persist in higher dimensions.","Editorial inference: The robustness scan could be extended to detached-plasma conditions by replacing the assumed 8 MW scrape-off-layer power with a detached profile; a design that keeps q_peak below the limit under those profiles would be considerably stronger evidence for the engineering claim.","Editorial inference: Since the strike-angle constraint is enforced only at discrete toroidal steps, a natural convergence test is to repeat the construction with finer toroidal resolution and verify that no local strike-angle violations appear between steps."],"forward_implications":["If the claim holds, island-divertor design in a fixed stellarator equilibrium reduces to a two-parameter search, making divertor optimization feasible alongside equilibrium optimization.","The modeled peak heat flux of about 3 MW/m² sits below the 10 MW/m² tungsten limit with a factor-of-three margin, leaving room for additional physics not modeled here, such as neutrals and impurities.","The 95% reduction in simulations from Bayesian optimization implies that adding more design parameters, such as strike angle, toroidal location, and divertor extent, remains computationally practical.","The Pareto-frontier analysis shows that reducing end-plate hits below about 5% costs at least q_peak ≈ 3 MW/m², quantifying the unavoidable trade-off between divertor viability and peak load.","The robustness of the optimized design across cross-field diffusivities supports using a single plasma parameter set for initial divertor screening."],"fun_headline_variants":["Two-angle recipe hits 3 MW/m² divertor design","Bayesian cuts divertor design cost by 95%","Two angles, one plane: cheap island divertor design","Island divertors optimized in just two parameters","Fast divertor design: two angles, 95% fewer runs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The construction assumes that a smooth control surface just inside the island can stand in for the real, stochastic island separatrix, and that the quoted heat fluxes come from a source-free field-line diffusion model with assumed scrape-off-layer power and cross-field diffusivity; if either fails, plates shaped by the algorithm may not intercept the heat flux as modeled.","fun_headline_variants_meta":{"raw":{"variants":["Two-angle recipe hits 3 MW/m² divertor design","Bayesian cuts divertor design cost by 95%","Two angles, one plane: cheap island divertor design","Island divertors optimized in just two parameters","Fast divertor design: two angles, 95% fewer runs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000684,"raw_usage":{"total_tokens":2930,"prompt_tokens":722,"completion_tokens":2208,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":2126}},"tokens_in":466,"tokens_out":2208,"duration_ms":13823,"temperature":1.0,"reasoning_tokens":2126,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:02:49.310879+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to run a higher-fidelity plasma and neutral transport simulation on the best optimized divertor in the same equilibrium and check whether the predicted peak heat flux on the actual divertor mesh remains below 10 MW/m² under the same scrape-off-layer power. If the reduced model's roughly 3 MW/m² result rises above the engineering limit by a factor of about three, the central claim that this design satisfies material heat-load limits is falsified.","supporting_citations":[],"review_version":1}