{"id":"ef28ca0a-d2a2-4f5b-88d2-a62a00e55f5e","arxiv_id":"1908.02494","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Collisionless fast ion losses in ITER can be estimated by a 1D advection-diffusion model using transport coefficients from a 1 ms orbit-following run, giving 100x faster parameter scans.","lead":"This paper presents computational techniques that make simulations of fast particle losses in fusion reactors faster and more reliable. It introduces loss maps to identify why particles escape and a simplified transport model that runs about one hundred times faster than full simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Advection-diffusion coefficients are extracted from a 1 ms, N=10 fit with no sensitivity study; the 16-37% benchmark errors and underprediction of high-loss scan points leave the 100x scan claim dependent on untested tuning.","rationale":"The reader's weakest assumption is the Markovian, local, one-dimensional Fokker-Planck representation of collisionless transport in rho'. My concern is the concrete manifestation of that assumption: the coefficients are estimated from a short, finite time window and the paper does not demonstrate that they are stable transport coefficients rather than finite-time fitting parameters. The absence of a sensitivity study over the explicitly admitted tuning parameters N and t is the most load-bearing gap because the headline speedup claim depends entirely on those coefficients remaining valid across the scanned phase space. The benchmark disagreements (16-37%) and the systematic underprediction of high-loss scan points are consistent with this concern, though they do not by themselves falsify the model. The paper is honest about these limitations and the loss-map methodology is independently useful, so I do not see grounds to reject the work. However, the empirical support for the advection-diffusion scan is not complete: the code and input data are only available on request, there is no independent code comparison, and the coefficient extraction is not shown to be robust. The reader's CONDITIONAL verdict therefore remains appropriate; the condition should include a sensitivity test of t and N, ideally accompanied by released artifacts for reproducibility.","tokens_in":17436,"tokens_out":7256,"duration_ms":86153,"concrete_test":"Reproduce the Section 6.2 scan for a subset of phase combinations (e.g., the two high-loss bands and two low-loss corners) using coefficient-evaluation times t = 0.5, 1, and 2 ms and averaging windows N = 5, 10, and 20, and compare the advection-diffusion loss maps against the same full slowing-down references. In addition, for the +PR case, compare the FPE-predicted loss fraction at intermediate times (5, 20, 50 ms) with direct collisionless orbit-following histories. If the inferred K,D shift by more than the quoted 16-37% error or the predicted time dependence diverges from the orbit-following history, the coefficients are fit parameters, not transport coefficients, and the scan claim needs independent validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Eq. (11) with K(rho',mu,E), D(rho',mu,E) evaluated from orbit-following can replace full slowing-down runs for ECC phase scans—requires the inferred coefficients to be genuine, time-independent Fokker-Planck coefficients. The paper's only evidence is the +PR benchmark and the 64-case scan in Section 6.2, but both use coefficients computed from a single 1 ms collisionless simulation with N=10 OMP-crossing averaging. Section 6 states that 'choosing N and t is critical for success' and that they were 'deduced with some trial and error', yet no sensitivity analysis is reported. Fig. 9 shows that after 1 ms only the first-orbit channel has fully developed; the stochastic ripple, stochastic field-line, and perturbed-banana channels contribute only through small sub-threshold displacements over tens of poloidal orbits. In that regime, ballistic or phase-correlated motion can easily produce finite-time estimates of E[Delta rho'] and Var[Delta rho'] that are not the K and D of a Markovian FPE. The benchmark itself shows 2.46 MW vs 1.79 MW (37% overestimate), and the scan systematically underestimates the high-loss cases—exactly the cases a scan is meant to find. Unless K and D are shown to be independent of the fit window (t and N) and to reproduce the loss-time history, the 100x speedup claim is a statement about a tuned fitting procedure rather than about transport.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces and demonstrates loss-map techniques for fast-particle loss studies in non-axisymmetric tokamaks, applied mainly to ITER. The loss map is a representation of particle birth location in (rho-prime, xi-prime) space that allows loss channels (first-orbit, ripple-induced, stochastic field-line, perturbed banana) to be identified and connected to magnetic field structure. The paper presents optimized marker initialization for improved convergence of total and peak power loads, magnetic-field-based projections for estimating losses without orbit-following, and an analysis of collisional effects. The central new claim is that collisionless fast-ion transport can be described as a one-dimensional advection-diffusion process with coefficients K(rho',mu,E) and D(rho',mu,E) evaluated from a short (1 ms) orbit-following simulation, and that the resulting model can replace full slowing-down simulations for parameter scans, specifically a 64-case ELM control coil phase scan that is claimed to be 100 times faster than full simulations while reproducing the overall loss landscape.","tokens_in":17821,"tokens_out":4610,"duration_ms":52421,"significance":"If the central claim holds, the advection-diffusion model would be a practically useful tool for fast design scans of fast-ion losses, a problem of direct relevance to ITER and future devices. The loss-map framework itself is a useful interpretive and diagnostic tool: the magnetic-field-based loss projection in Section 5.1 provides a partially independent cross-check (the +PR estimate of 1.24 MW versus 1.18 MW from orbit-following is encouraging), and the optimized marker initialization demonstrably improves peak-load convergence in Figure 7. The paper is also honest in reporting the disagreements (2.46 MW versus 1.79 MW in Section 6.1 and underestimation of high-loss scan points in Section 6.2). However, the validation of the advection-diffusion model is currently self-referential, and the extraction of the transport coefficients uses manually tuned parameters without a sensitivity study.","major_comments":[{"comment":"The transport coefficients K and D are evaluated as finite-time estimates from a single 1 ms collisionless simulation with N=10 OMP-crossing averaging, and the manuscript states that choosing N and t is critical and that they were deduced with trial and error. No sensitivity analysis with respect to t, N, or the reflecting boundary location is reported. This is load-bearing because Eqs. (11)-(14) yield genuine Fokker-Planck coefficients only if K and D are independent of the estimation window; if they are not, the advection-diffusion model is a tuned fitting procedure rather than a physics-based transport model. The point is reinforced by Figure 9, which shows that after 1 ms only the first-orbit loss channel is fully developed while the stochastic-ripple, stochastic-field-line, and perturbed-banana channels contribute only through small sub-threshold displacements. The authors should show that K and D converge as t and N are varied, and ideally that the model reproduces the time history of losses, not just the final loss map.","section":"Section 6, Eqs. (13)-(15)"},{"comment":"The benchmark accuracy is quantified only through total lost power: the advection-diffusion model gives 2.46 MW versus 1.79 MW from the collisionless orbit-following simulation (a 37% overestimate), and in the short-time comparison 1.42 MW versus 1.22 MW (a 16% overestimate). In the ECC phase scan the model systematically underestimates the losses in exactly the high-loss cases that a design scan is intended to identify. Since the stated purpose of the model is to find interesting regions in parameter space and provide rough estimates, the paper should report a quantitative accuracy metric for the scan, such as rank correlation or per-case relative error, and demonstrate that the high-loss cases are not systematically missed or suppressed. As it stands, the claim that the model is suitable for fast parameter scans is supported mainly by the qualitative visual agreement of the loss contours.","section":"Section 6.1 and Fig. 10; Section 6.2 and Fig. 11"},{"comment":"The transport coefficients are evaluated with the same orbit-following code (ASCOT5) that provides the reference losses, so the advection-diffusion benchmarks test only whether the reduced 1D model can reproduce the output of the same code; they do not test the physical fidelity of the coefficients. The genuine 1 ms-to-100 ms time extrapolation is a meaningful test of time-independence, but it remains a self-consistency check. To support the physical claim that collisionless fast-ion transport is advection-diffusive, the authors should compare the inferred coefficients with the analytic expectations for at least one channel, e.g., the stochastic-ripple diffusion coefficient of Eq. (7) or the stochastic-field-line estimate of Eq. (9), or benchmark against an independent orbit-following implementation.","section":"Section 6.1 and Section 6.2"}],"minor_comments":[{"comment":"The definition of the magnetic moment appears to be missing the factor m/2 and should be clarified; as written, mu has units of m^2/s^2 rather than the usual J/T.","section":"Eq. (2)"},{"comment":"The text says the stochastic-ripple diffusion coefficient is projected using Eq. (9), but Eq. (9) is the stochastic-field-line diffusion coefficient; the ripple diffusion coefficient is Eq. (7).","section":"Section 5.1"},{"comment":"The sentence 'comes from Ref. where this was shown' has an empty citation; the reference number is missing.","section":"Section 6, first paragraph"},{"comment":"The inverse-Gaussian first-passage-time formula is rendered ambiguously: the denominator in the exponential should be 2 c_1^2 t, and c_1, c_2 should be defined more explicitly. The current notation is hard to parse.","section":"Eq. (16)"},{"comment":"The averaging in Eq. (15) is not fully specified: it should be stated explicitly that the rho'_j are OMP-crossing averages for a single marker and that the final K and D are weighted averages over markers in each (rho',mu) bin, including how the weights are defined.","section":"Section 6, coefficient evaluation paragraph"},{"comment":"There are numerous typos and misspellings, including 'leves' for 'levels', 'quaranteed' for 'guaranteed', 'ploidal' for 'poloidal', 'extent' for 'extend', and 'absent' for 'absence'. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The advection-diffusion model is the genuinely new contribution of this paper; the loss-map methodology is largely an extension of the author's previous work. The main risk is that the central claim rests on self-consistency tests with a single code and on manually tuned coefficient-extraction parameters. I would ask for a sensitivity study of t and N, a quantitative accuracy metric for the ECC scan, and at least one independent or analytic benchmark before accepting the 100x-scan claim. The paper is otherwise well within the scope of the journal and contains useful practical techniques."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is the advection-diffusion treatment of collisionless fast-ion transport in rho' with mu and E as parameters, with K and D extracted from short orbit-following runs. The idea is attractive and the benchmarks are reported honestly, but the coefficients are generated and validated with the same code, and the averaging parameters are tuned without a sensitivity study. That leaves the 100x speedup claim less solid than the presentation suggests.\n\nWhat is actually new and good: The loss-map framework is extended in a useful way. The magnetic-field-based projections in Sec. 5.1 are clever — they connect loss regions to field structure and give a reasonable estimate of collisionless losses (1.24 MW vs 1.18 MW). The optimized marker initialization has a clear practical payoff for peak power load estimates, and the convergence curves demonstrate it. The advection-diffusion model itself is the main novelty; replacing full slowing-down runs with a 1D Fokker-Planck solve after evaluating coefficients is a worthwhile idea, and the ECC phase scan shows a plausible use case.\n\nThe soft spots are not hidden. The circularity is real: ASCOT5 supplies both the coefficients and the benchmark losses, so there is no independent physics check. The one-parameter scan uses a 1 ms, N=10 averaging window chosen by 'trial and error,' and no sensitivity analysis is reported. A 1 ms simulation only fully develops the first-orbit channel; the stochastic channels contribute through small displacements, so the finite-time coefficients could be absorbing ballistic transients rather than true Fokker-Planck transport coefficients. The scan systematically underpredicts the high-loss cases, which are the ones you most want to catch. The broken reference in Sec. 6 ('from Ref.' with no number) and the 'available on request' code/data are also weaknesses.\n\nI want to give credit where it is due. The author explicitly reports the 37% overestimate in the benchmark and the underprediction in the scan, which is honest. The time-extrapolation from 1 ms coefficients to 100 ms losses is a genuine test, and the overall shape of the loss landscape is reproduced. This is a serious methods paper, not a toy.\n\nRecommendation: send it to peer review. The advection-diffusion claim needs a referee pushing for a sensitivity study of N and t, an independent code comparison or uncertainty quantification, and the artifacts must be shipped or uploaded. If those are addressed, the method could be genuinely valuable.","headline":"Loss maps are genuinely useful and the advection-diffusion idea is promising, but the coefficients are validated with the same orbit-following code used to make them, and the missing sensitivity analysis leaves the 100x speedup claim undermoored.","tokens_in":18277,"tokens_out":3906,"would_cite":true,"duration_ms":37863,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fast-particle losses in non-axisymmetric tokamak plasmas can be reproduced by a one-dimensional advection-diffusion model whose coefficients come from short orbit-following runs, cutting scan cost a hundredfold.","keywords":["fast ions","fast particle losses","loss maps","advection-diffusion model","orbit-following simulations","ITER","magnetic field ripple","ELM control coils"],"falsifier":"Follow a collisionless orbit ensemble in a field with strong stochastic-ripple transport and test whether the radial displacement has Gaussian statistics with variance growing linearly in time and whether loss times follow the inverse-Gaussian distribution; if the inferred coefficients depend on the averaging window $N$ or on the toroidal launch phase of otherwise identical markers, the Markovian one-dimensional description is wrong. A concrete target is the paper's own 16-37\\% disagreement: a case where that gap grows with simulation time rather than staying bounded would show the model is not a reliable predictor for that regime.","tokens_in":17253,"feed_emoji":"🧲","tokens_out":13304,"duration_ms":118526,"temperature":0.7,"pith_summary":"Fast particle losses from a magnetically perturbed tokamak are usually estimated with expensive orbit-following simulations, but this paper argues that the losses can be captured with a much cheaper description. It constructs loss maps in a two-dimensional space of orbit constants — the radial coordinate $\\rho'$ and pitch $\\xi'$ at the outer mid-plane — and shows that distinct loss channels (first-orbit, ripple-induced, stochastic field-line, perturbed banana) appear as separate regions. It then treats the collisionless radial motion as a one-dimensional advection-diffusion process, evaluates the two transport coefficients with a short orbit-following run, and solves the resulting Fokker-Planck equation to predict losses on longer time scales. In ITER benchmarks the model reproduces the overall loss landscape, and a scan over ELM control coil phases runs one hundred times faster than full slowing-down simulations while preserving the locations of high-loss bands. If the approach holds, fast particle parameter scans and wall-load estimates become much cheaper and tie directly to underlying transport mechanisms.","feed_headline":"Loss maps make fast-particle loss scans 100 times faster","feed_subtitle":"Short orbit runs feed a 1D advection-diffusion model that reproduces the loss map's shape across ELM coil phases.","key_machinery":"The load-bearing object is the loss map: a histogram of lost-particle fraction in $(\\rho',\\xi')$ space, where $\\rho'$ is the normalized poloidal flux at the outer mid-plane crossing and $\\xi'$ is the pitch there. Because every orbit's topology is fixed by these two coordinates at fixed energy, loss channels appear as separated regions identifiable with known mechanisms. The second mechanism is a coefficient-estimation recipe for the advection-diffusion equation\n$$\\frac{\\partial f}{\\partial t}=-\\frac{\\partial}{\\partial\\rho'}(Kf)+\\frac{\\$partial^{2}$}{\\partial{\\rho'}^2}(Df),$$\nusing short orbit-following data: Gaussian displacement statistics for confined markers and inverse-Gaussian first-passage times for lost markers. The loss map supplies both the diagnostic that validates the model and the weighting that lets transport coefficients be evaluated only where transport actually occurs.","core_discovery":"The central claim is that collisionless fast ion transport in a three-dimensional perturbed magnetic field can be modelled to good accuracy as a one-dimensional Fokker-Planck process in the radial coordinate $\\rho'$ (flux-surface coordinate at the outer mid-plane), with magnetic moment $\\mu$ and energy $E$ treated as parameters. The advection coefficient $K(\\rho';\\mu,E)$ and diffusion coefficient $D(\\rho';\\mu,E)$ are not derived from first principles but measured from a very short orbit-following simulation, about a millisecond in these test cases, in which markers are followed for only tens of poloidal orbits. Confined markers yield $K$ and $D$ from the mean and variance of the radial displacement after averaging over ten outer mid-plane crossings; lost markers contribute through the inverse-Gaussian distribution of first-passage times. In the benchmark with the most complete magnetic field, the model predicted 2.46 MW of total lost $\\alpha$ power against 1.79 MW from the collisionless full-orbit simulation (1.42 MW versus 1.22 MW when only collisionless-time-scale losses are counted), and the $8\\times8$ scan over ELM control coil phases reproduced the two high-loss bands and their crossing point at one hundredth of the computational cost.","pith_inferences":["Beyond the paper, the same coefficient-measurement recipe could be applied to runaway electrons or other fast species; the one-dimensional Markov assumption is most plausible for strongly passing populations and would need re-testing there.","Beyond the paper, if transport coefficients could be estimated from magnetic-field-based projections rather than short orbit runs, the approach might eventually skip orbit-following entirely, although the field-based loss-map projections shown here are close but not yet accurate enough for that.","Beyond the paper, the systematic overestimate of lost power in the benchmark suggests the model may be collapsing several mechanisms into a single effective diffusion coefficient; checking whether the inferred coefficients are independent of the averaging window and simulation time would test whether the Markov description is genuinely valid."],"forward_implications":["ELM control coil phase scans that would normally require millions of markers per configuration can be reduced to short coefficient runs plus a cheap one-dimensional solve, making it practical to map the full parameter space of coil phases and currents.","Because loss maps can be projected from magnetic field structure alone, one can cross-check an orbit-following result or estimate losses without running a dedicated simulation.","Marker initialization can be concentrated on the loss channels identified in the map, so peak power load estimates converge with one or two orders of magnitude fewer markers.","Orbit-averaged transport codes can incorporate three-dimensional-field fast-ion transport by adding the measured advection and diffusion coefficients rather than following orbits in full geometry.","The collisionless approximation captures most alpha losses at birth energy, so collisions matter mainly near thermal energies where neoclassical transport takes over."],"supporting_citations":[{"why":"It introduced the loss-map analysis and identified perturbed banana transport, and the whole paper builds on that construction.","marker":"[7]"},{"why":"It supplies the method for evaluating advection-diffusion coefficients from the mean and variance of marker displacement.","marker":"[24]"},{"why":"It established that loss channels can be identified in constants-of-motion space, motivating the radial-coordinate and pitch parameterization.","marker":"[8, 9]"},{"why":"It provides the stochastic-ripple threshold and diffusion coefficient used to project ripple loss regions from the magnetic field.","marker":"[10]"},{"why":"It defines the ripple-well parameter used to identify ripple-trapping regions in field-based loss maps.","marker":"[11]"},{"why":"It gives the stochastic field-line diffusion coefficient used to project stochastic field-line loss channels.","marker":"[12]"}],"fun_headline_variants":["Loss maps make fast-ion loss scans 100x faster","Advection-diffusion model reproduces fast-ion losses at 1/100 cost","Fast particle loss scans cut to 1/100 cost with loss maps","Loss maps tie perturbed fields to fast particle losses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that collisionless radial transport is Markovian and local, so that a particle's random walk in $\\rho'$ is fully described by one advection and one diffusion coefficient at each point; if the motion has memory, depends on toroidal or poloidal phase, or needs more than two coefficients, the model fails.","fun_headline_variants_meta":{"raw":{"variants":["Loss maps make fast-ion loss scans 100x faster","Advection-diffusion model reproduces fast-ion losses at 1/100 cost","Fast particle loss scans cut to 1/100 cost with loss maps","Loss maps tie perturbed fields to fast particle losses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1765,"prompt_tokens":1006,"completion_tokens":759,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":684}},"tokens_in":622,"tokens_out":759,"duration_ms":8484,"temperature":1.0,"reasoning_tokens":684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:41:42.756594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Follow a collisionless orbit ensemble in a field with strong stochastic-ripple transport and test whether the radial displacement has Gaussian statistics with variance growing linearly in time and whether loss times follow the inverse-Gaussian distribution; if the inferred coefficients depend on the averaging window $N$ or on the toroidal launch phase of otherwise identical markers, the Markovian one-dimensional description is wrong. A concrete target is the paper's own 16-37\\% disagreement: a case where that gap grows with simulation time rather than staying bounded would show the model is not a reliable predictor for that regime.","supporting_citations":[{"cited_title":"Illustration of the toroidal magnetic field ripple in reduced field scenarios","cited_arxiv_id":null,"evidence_quote":"It supplies the method for evaluating advection-diffusion coefficients from the mean and variance of marker displacement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the ripple-well parameter used to identify ripple-trapping regions in field-based loss maps."}],"review_version":1}