{"id":"b6ca29b1-c4dd-44d0-842f-6433809c04e8","arxiv_id":"2607.25505","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"For non-axisymmetric STT/SOT MRAM devices, a 2D Fokker-Planck solver using central differencing matches 10^6-trajectory stochastic LLG write-error rates, while monotone schemes bias switching early.","lead":"This paper builds a 2D finite-volume Fokker-Planck solver on the sphere for write-error-rate prediction in STT/SOT-MRAM devices with broken symmetry, and validates it against one million stochastic LLG trajectories. It reports that central differencing matches the stochastic reference while SG/upwind schemes bias WER, recommending an adaptive hybrid scheme.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. 18 first-passage absorber is never specified and is not m_z(t)<0; until equivalence is shown, the central-vs-SG/upwind validation is comparing different observables.","rationale":"I read the paper as a numerical-methods validation: central differencing should recover the sLLGS WER. For that to be true, the FP survival probability must target exactly the sLLGS statistic. The paper never makes that identification, and the two quantities are not equal in general. The first-passage problem to an absorbing cap is a hitting-time statistic; Eq. 27 is an occupation/instantaneous statistic. The missing cap geometry makes it impossible to reproduce the calculation. This is the weakest link, more central than the repository placeholder or lack of low-WER tests, though those are real. I do not see an internal inconsistency in the FVM construction itself; conservation and adaptive blending are plausible, and the 10^6 trajectory baseline is a serious effort. If the WER-observable mismatch is repaired by specifying a cap and demonstrating insensitivity/equivalence, the paper's main numerical conclusion could stand. Thus CONDITIONAL is the right level, and my read leaves the reader's verdict unchanged.","tokens_in":16675,"tokens_out":6216,"duration_ms":70947,"concrete_test":"Using the already-generated 10^6 sLLGS trajectories, compute both (a) the occupation fraction Pr(m_z(t)<0) (Eq. 27) and (b) the first-passage fraction to each candidate cap, F_ε(t)=Pr(inf{s≤t: θ(s)≥π-ε}), for ε from π/2 down to one mesh cell (π/1024). If F_ε(t) differs from the occupation fraction by more than the Monte Carlo CI for the same ε that the FP solver would use, Eq. 18 and Eq. 27 are not equivalent and the central validation is a calibration exercise. Report the cap size that makes the best match; if it is not specified a priori, the claim of ground-truth recovery is unfalsifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Most load-bearing issue is the mismatch between the FP and sLLG WER observables. Section II-D3 defines FP WER as 1 minus cumulative probability absorbed at an unspecified 'south-pole boundary or absorbing cap near θ=π'; Section II-H defines the sLLGS WER as the fraction of 10^6 trajectories with m_z(t)<0. First-passage to a cap is not the same as occupation of the lower hemisphere at time t: a trajectory can cross m_z=0 and return, or reach the lower hemisphere without touching a small cap near the pole. The cap angular size is never given. Therefore the agreement in Figures 5-6 may be a function of the cap choice rather than of the discretization scheme. Since Scenario B and C claims are comparisons of FP WER to sLLGS WER, every central/SG/upwind conclusion inherits this unspecified calibration. This is not merely a formal quibble: the WER values shown (≈0.1-1) are in the regime where transient occupation and first-passage probabilities differ most. The paper also lists the code as github.com/IMEC/PAPERUNDERREVIEW and contains a '785.5%' typo (Sec III-E), reinforcing that the release is not yet a reproducible benchmark; but the observable mismatch is the primary obstacle.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a two-dimensional finite-volume Fokker-Planck (FP) solver on the unit sphere for write-error-rate (WER) prediction in STT and SOT MRAM devices with broken azimuthal symmetry. Four spatial discretizations are implemented: central, Scharfetter-Gummel (SG), upwind, and a Péclet-based hybrid blending. The solver is validated against 10^6-trajectory stochastic Landau-Lifshitz-Gilbert-Slonczewski (sLLGS) simulations in three scenarios: quasi-axisymmetric STT, STT with an in-plane assist field, and SOT with field-like torque. The central claim is that central differencing reproduces the sLLGS reference WER within statistical uncertainty, whereas SG and upwind introduce a systematic early-switching bias; the hybrid scheme is then recommended as a practical default.","tokens_in":17069,"tokens_out":2288,"duration_ms":27013,"significance":"If the central claim holds, the paper makes a useful contribution by demonstrating that spatial discretization is not a neutral numerical detail for 2D FP-based WER prediction in non-axisymmetric MRAM. The use of a 10^6-trajectory sLLGS reference and the conservative finite-volume formulation are strengths, as is the systematic Péclet-number analysis across three physically distinct scenarios. However, the current manuscript does not yet establish the equivalence between the FP observable and the sLLGS observable used for validation, and the validation is restricted to relatively high WER values. These gaps must be resolved before the central claim can be accepted.","major_comments":[{"comment":"The FP WER is defined as first-passage probability to an absorbing 'south-pole boundary or absorbing cap near θ=π', while the sLLGS WER is the fraction of trajectories with m_z(t)<0. These are different observables: a trajectory can cross m_z=0 and return, or reach the lower hemisphere without hitting a small polar cap. The cap size is never specified. Since every reported agreement between FP and sLLGS depends on equating these two quantities, the central validation is not well-posed as written. Please either specify the cap and justify equivalence, or change the FP WER definition to P(m_z(t)<0) computed from the full density, and re-run the comparisons.","section":"§II-D3 vs §II-H, Eq. (18) vs Eq. (27)"},{"comment":"The abstract and introduction motivate WER prediction 'especially below 10^-6', but the validation results in Figures 5 and 6 cover only WER values in the approximate range 0.1–1. In this transient regime, first-passage and occupation probabilities differ most, and the regime where WER is a rare-event probability is not tested. The paper should either extend validation to lower WER (e.g., lower current or longer time) or explicitly limit the claims to the demonstrated range.","section":"Abstract, §I, §III-D (Figs. 5–6)"},{"comment":"The hybrid adaptive blending uses hand-picked thresholds Pe<1 and Pe>2 at which central differencing is selected exclusively, so the hybrid is constructed to prefer central. It is then used as evidence that 'central differencing is geometrically optimal'. This is circular as a validation of central differencing, and the reported diagnostic contains an impossible value: 'Scenario 3: ∼785.5% central differencing and ∼14.5% blended region' does not sum to 100%. Please provide a sensitivity analysis for the thresholds and correct the reporting.","section":"§III-E, Eq. (24)"},{"comment":"The paper claims central differencing remains stable and accurate for Pe≫1 because the transport is 'rotational' on the sphere, but no formal argument or mesh-convergence evidence is provided. Central differencing is non-monotone and can produce spurious oscillations at high Péclet number; the claim that flux-balance preservation outweighs monotonicity needs quantitative support, such as error norms, density positivity checks, or a resolved-mesh convergence study. Without this, the central-vs-SG/upwind conclusion is not fully established.","section":"§III-D2, §III-F3"}],"minor_comments":[{"comment":"The code repository is given as 'github.com/IMEC/PAPERUNDERREVIEW', which is a placeholder and not a reproducible benchmark. Please provide a permanent DOI or repository URL before final submission.","section":"§II-H, §IV"},{"comment":"The phrase '∼785.5% central differencing' is a clear typo; presumably 78.5% was intended. Please correct and re-verify all percentages in the hybrid diagnostic.","section":"§III-E"},{"comment":"Figure captions refer to 'no numerical diffusion' for central and hybrid schemes; this is misleading because central differencing can introduce phase errors even if it is less dissipative. Consider rephrasing to 'minimal numerical diffusion'.","section":"Figure captions, Figs. 4–7"},{"comment":"The blending function is described as Hermite interpolation, but the formula as written (3x^2-2x^3) is a smoothstep, not a Hermite interpolation between two functions. Please clarify the relationship to the cited hybrid scheme.","section":"§II-E4, Eq. (24)"},{"comment":"Minor typographical issues include 'araising' in §III-B and inconsistent use of 'P ´eclet' spacing. These should be corrected.","section":"§III-B, Fig. 3"},{"comment":"The empirical WER is defined as a fraction of trajectories with m_z<0, but the paper also compares 'switching-time distributions'. Please define the switching time for a trajectory, since m_z<0 alone does not specify a unique switching time in the presence of thermal fluctuations.","section":"§III-D1, Eq. (27)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the computational effort behind the sLLGS baseline is substantial. However, the mismatch between the FP first-passage observable and the sLLGS m_z<0 observable is load-bearing for the validation, and the current manuscript does not resolve it. The high-WER-only validation and the placeholder code URL also weaken reproducibility. These issues are fixable, so I recommend major revision rather than rejection. I would also ask the editor to ensure the hybrid-blending section is rewritten so that it does not appear to validate central differencing by construction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the thing to know: this is a genuinely useful numerical study, but its headline claim is currently held up by an observable mismatch. The FP WER in Eq. 18 is survival probability against an absorbing south-pole cap; the sLLGS WER in Eq. 27 is the fraction of trajectories with m_z(t)<0. Those are different objects. A trajectory can dip below m_z=0 and return, or enter the lower hemisphere without hitting a small cap near θ=π. Since the cap size is never given, every agreement in Figs. 5–6 is formally a function of an unspecified free parameter. The stress-test note is right, and the reader's weakest assumption is the correct one.\n\nWhat is actually new and good: the systematic comparison of central, Scharfetter–Gummel, upwind, and hybrid adaptive fluxes for a 2D finite-volume FP solver on the sphere is new in this MRAM context. The Péclet analysis is sensible, and the choice of scenarios—quasi-axisymmetric STT, in-plane-field STT, and SOT with field-like torque—covers the relevant symmetry-breaking mechanisms. The 10^6-trajectory sLLGS baseline is appropriate, and the conservative finite-volume formulation looks sound. The authors also deserve credit for showing that the 1D reduction fails in these scenarios.\n\nSoft spots: the observable mismatch is not minor; it is load-bearing. Related, validation is only in the WER 0.1–1 range, far from the <1e-6 regime the abstract motivates, and first-passage versus occupation differences are largest exactly there. The hybrid 'self-validation' is circular by construction—the blending thresholds are hand-picked, and using the hybrid's regime percentages to conclude central is optimal is not independent evidence. There is also a typo (785.5% central in Scenario 3) and the code repository is a placeholder, github.com/IMEC/PAPERUNDERREVIEW. All fixable in revision.\n\nMy take: the scheme ranking may survive a proper boundary definition, but right now the central-versus-SG/upwind conclusion is conditional. This does not look like a desk reject; it is worth a serious referee. I would not cite it until the boundary is specified and low-WER validation appears.\n\nRecommendation: send to peer review, ask the authors to define the absorber (or use the same observable for both solvers), add quantitative error metrics, show at least one case below 1e-4 WER, and release the actual code.","headline":"Useful comparison of FP discretization schemes for non-axisymmetric MRAM, but the central validation compares a first-passage absorber to an m_z<0 occupation metric without establishing equivalence.","tokens_in":17548,"tokens_out":3065,"would_cite":false,"duration_ms":33600,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M08","82C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that a 2D finite-volume Fokker-Planck solver with central differencing reproduces stochastic LLG write-error rates for non-axisymmetric MRAM devices, while Scharfetter-Gummel and upwind schemes systematically predict earli","keywords":["write error rate","Fokker-Planck","stochastic Landau-Lifshitz-Gilbert","finite volume method","MRAM","spin-transfer torque","spin-orbit torque","unit sphere"],"falsifier":"Run the same STT scenario with the FP solver using central differencing while varying the absorbing-cap angular radius from a small cap to a full hemisphere, and compare each resulting WER curve to the sLLG m_z<0 reference; if no cap size reproduces the reference within the 95% confidence interval, the validation result is an artifact of the absorbing-boundary definition rather than a genuine property of the flux scheme.","tokens_in":16553,"feed_emoji":"🧲","tokens_out":3837,"duration_ms":39265,"temperature":0.7,"pith_summary":"The paper tackles the problem of predicting write error rates (WER) in STT and SOT MRAM devices when azimuthal symmetry is broken by in-plane fields or field-like torques. It develops a 2D finite-volume Fokker-Planck solver on the unit sphere and validates it against one million stochastic LLG trajectories. The central claim is that the choice of spatial flux discretization controls WER accuracy: central differencing matches the stochastic reference within statistical uncertainty, whereas Scharfetter-Gummel and upwind schemes introduce an early-switching bias. The paper argues that preserving flux balance matters more than monotonicity for coupled drift-diffusion on the sphere. If correct, this makes reliable WER prediction possible for next-generation non-axisymmetric devices without expensive trajectory simulations.","feed_headline":"Central differencing recovers true MRAM write-error rates","feed_subtitle":"A 2D finite-volume solver beats 1D projections, while upwind and SG schemes bias switching times.","key_machinery":"The central object is the finite-volume discretization of the Fokker-Planck equation in conservative flux form on a structured (phi, theta) mesh on the unit sphere. The four face-flux schemes -- central difference, Scharfetter-Gummel, upwind, and a Peclet-based hybrid adaptive blend -- differ only in how the drift-diffusion flux across each control-volume face is approximated. The hybrid blend uses Hermite interpolation to weight central versus SG fluxes based on the local Peclet number, and in the tested scenarios it automatically selects mostly central differencing. The spherical metric is handled exactly through geometric face lengths and cell areas, avoiding pole singularities and preser","core_discovery":"The paper shows that for a spin-transfer-torque device with an in-plane assist field and for a spin-orbit-torque device with intrinsic field-like torque asymmetry, the 2D Fokker-Planck solution using central differencing overlaps the 10^6-trajectory sLLG WER reference within Monte Carlo statistical uncertainty. In contrast, the Scharfetter-Gummel and upwind schemes shift WER curves to earlier switching times, with the bias growing in the SOT case where azimuthal circulation is stronger. The authors conclude that in these 2D non-axisymmetric regimes, preserving the coupled drift-diffusion flux balance is more important than the monotonicity that SG and upwind were designed to enforce, and tha","pith_inferences":["If the first-passage absorbing-cap WER is truly equivalent to the m_z<0 switching convention, the same solver architecture could be extended to non-macrospin models, though those would require a higher-dimensional Fokker-Planck state space.","The observed bias of monotone schemes on curved manifolds suggests that classical upwinding may systematically distort first-passage statistics in other surface-drift-diffusion problems, not just magnetism.","A testable extension is to sweep the absorbing-cap angular radius near the south pole and compare against the m_z<0 threshold; this would pin down the boundary detail that the current validation leaves unspecified.","The hybrid blending recommendation is based on three test geometries; a broader suite of anisotropy landscapes and temperature ranges would show how far the 'central-differencing-first' heuristic generalizes."],"forward_implications":["1D Fokker-Planck reductions are unreliable for any MRAM geometry with broken azimuthal symmetry; a full 2D treatment on the sphere is required for accurate WER.","Central-differencing WER matches 10^6-trajectory sLLG within statistical uncertainty across the measured range, offering a trajectory-free route to write-error prediction.","Scharfetter-Gummel and upwind schemes systematically lower WER by predicting earlier switching, so discretization choice should be reported as a model parameter in MRAM reliability studies.","The hybrid adaptive blending scheme recovers central results while diagnosing local Peclet regimes, making it a practical default for production use.","The exact spherical metric and conservative finite-volume formulation preserve probability to machine precision, enabling robust rare-event WER calculations."],"fun_headline_variants":["2D solver: central differencing nails MRAM write-error rates","Central vs upwind: 2D FP solver exposes WER bias","For MRAM, central differencing matches sLLG","Scharfetter-Gummel skews MRAM WER; central doesn't","2D Fokker-Planck: choose central for true MRAM WER"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper equates the write-error rate to first-passage probability through an absorbing boundary near the south pole (Eq. 18) while the stochastic reference counts trajectories with m_z below zero (Eq. 27), without proving their equivalence or specifying the cap's angular size, so every reported agreement rests on this unexamined boundary identification.","fun_headline_variants_meta":{"raw":{"variants":["2D solver: central differencing nails MRAM write-error rates","Central vs upwind: 2D FP solver exposes WER bias","For MRAM, central differencing matches sLLG","Scharfetter-Gummel skews MRAM WER; central doesn't","2D Fokker-Planck: choose central for true MRAM WER"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001182,"raw_usage":{"total_tokens":4727,"prompt_tokens":759,"completion_tokens":3968,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":3870}},"tokens_in":503,"tokens_out":3968,"duration_ms":28559,"temperature":1.0,"reasoning_tokens":3870,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:14:20.242789+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same STT scenario with the FP solver using central differencing while varying the absorbing-cap angular radius from a small cap to a full hemisphere, and compare each resulting WER curve to the sLLG m_z<0 reference; if no cap size reproduces the reference within the 95% confidence interval, the validation result is an artifact of the absorbing-boundary definition rather than a genuine property of the flux scheme.","supporting_citations":[],"review_version":1}