{"id":"fc41a247-1e2e-486b-a650-5adff3299904","arxiv_id":"2506.10711","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PDESpectralRefiner applies blurring diffusion with a new velocity prediction formula to refine PDE rollout outputs, yielding slightly lower rollout error than the prior diffusion refiner on 2D Navier-Stokes.","lead":"A new 'spectral refiner' applies blurring diffusion with a custom velocity formula to correct neural PDE forecasts frequency by frequency, and it reports slightly lower rollout errors than the prior diffusion refiner on 2D Navier-Stokes. The result matters because long simulations of fluids typically drift over time, and high-frequency errors are the hardest part to control.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 14's derivative identity is not reproducible from the stated blurring schedule; since MSE recovery at t=1 is insensitive to Eq. 14, the reported experiments do not validate the claimed velocity.","rationale":"The reader's weakest_assumption correctly focuses on Eq. 14 and the t=1 boundary, but the t=1 boundary actually makes the first-step MSE recovery insensitive to Eq. 14, so the load-bearing part is the intermediate refinement steps. Those steps are governed entirely by Eq. 14 through Prop. 1, and by Prop. 2. The printed expression contains a π ln σ0 term and an α^2/σ^2 ratio that do not appear in a straightforward derivative of r_t^2 under either schedule mentioned in the paper. Because the appendix is absent, this cannot be checked from the manuscript. The empirical tables cannot rescue the theory: the claimed first-step reduction to MSE is automatic, and the reported NS improvements are small and sometimes within a few standard errors. The right disposition is to keep the conditional verdict, with the condition being a corrected, independently verifiable derivation and rerun experiments.","tokens_in":15022,"tokens_out":30123,"duration_ms":323541,"concrete_test":"Symbolically compute ∂r^2/∂φ from r_t^2=α_t^2 e^{-2Λτ_t}+σ_t^2 for the two schedules discussed in the paper, τ_t=(σ_B^2/2)cos^2(πt/2) and τ_t=(σ_B^2/2)sin^4(πt/2), using α_t=cos(πt/2), σ_t=sin(πt/2), and compare term-by-term with Eq. 14. If the π ln σ0 term or the α^2/σ^2 factor is absent from both, Eq. 14 is wrong as stated; then recompute the U-Net rows of Table 1 with the corrected radial term to see whether the reported PDESpectralRefiner gains persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mathematical content is Prop. 1, and its only non-trivial term is the radial component containing ∂r^2/∂φ in Eq. 14. The first-step reduction to MSE is automatic: at t=1, α_1=0 and τ_1=0 force v_1=-u_x regardless of the detailed form of Eq. 14. Thus the claimed NS improvements must come from intermediate refinement steps, which depend entirely on Eq. 14 and on Prop. 2. But Eq. 14 is given without derivation and the appendix is absent. A direct check from Eqs. 10-11 fails: for the cos^2 schedule τ_t=(σ_B^2/2)cos^2(πt/2), with φ=πt/2, α_t=cos φ, σ_t=sin φ, one has r_t^2=cos^2φ e^{-2Λτ_t}+sin^2φ, so ∂r^2/∂φ = 2 sinφ cosφ (1-e^{-2Λτ_t}) + 2Λσ_B^2 cos^3φ sinφ e^{-2Λτ_t}. This contains no π ln σ0 factor and no α^2/σ^2 cos^3 sin structure as printed in Eq. 14; using the Hoogeboom sin^4 schedule gives a similar mismatch. Therefore the velocity used for training and sampling intermediate refinement steps is built on an identity that appears incorrect as stated, and Prop. 2 inherits the error.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PDESpectralRefiner, a modification of the diffusion-based PDE refiner PDERefiner in which the forward noise process is reweighted in Fourier space through a blurring schedule, so that different frequency components are corrupted at different rates. A new \"blurring v-prediction\" is introduced (Prop. 1, Eqs. 12–14) together with a corresponding sampling formula (Prop. 2, Eq. 15). The authors claim that with a cos^2 blurring schedule the first refinement step recovers the MSE objective, and they report improved one-step and rollout MSE for 2D Navier–Stokes with U-Net and FNO backbones.","tokens_in":15402,"tokens_out":8646,"duration_ms":94347,"significance":"The problem addressed—controlling high-frequency error accumulation in learned PDE surrogates—is important, and the idea of using blurring diffusion to reweight frequencies is a natural and potentially valuable extension of PDERefiner. The paper also makes a falsifiable methodological claim: the new v-prediction is the correct velocity for the spectrally reweighted process. I credit the authors for explicitly acknowledging limitations, including the poor FNO cooperation and the need for non-FFT projectors on non-uniform grids. However, the central derivation is not available in the submitted manuscript, the empirical gains are small and partly negative, and the experimental design has selection issues. If the derivation is corrected and the experiments are validated, the approach could be a useful tool, but in its current form the central claims are not verifiable.","major_comments":[{"comment":"The load-bearing identity for ∂r_t^2/∂φ_t is stated without proof, and the promised derivation is deferred to an appendix that is not present in the submitted manuscript. Prop. 1 (Eqs. 12–13) and Prop. 2 (Eq. 15) both depend on this identity, so no intermediate refinement step of the proposed method can be reproduced from the main text. Moreover, a direct calculation from Eqs. (10)–(11) with the standard blurring-diffusion parameterization α_t=cos φ, σ_t=sin φ, φ=πt/2 and the paper's proposed τ_t=(σ_B^2/2)cos^2 φ gives ∂r^2/∂φ = 2 sinφ cosφ [1 − exp(−σ_B^2 Λ cos^2 φ) + σ_B^2 Λ cos^2 φ exp(−σ_B^2 Λ cos^2 φ)], which does not match the structure of Eq. (14): the π/ln σ0 factor and the α^2/σ^2 · cos^3 sin prefactor are absent. The same mismatch occurs for the Hoogeboom sin^4 schedule. The authors need to supply the full derivation and state the exact α_t, σ_t, τ_t parameterization; otherwise the velocity used for training and sampling is not defined.","section":"§4.1, Eq. (14)"},{"comment":"The t=1 boundary argument does not validate Eq. (14). At t=1, α_1=0 and σ_1=1 force the first term in Eq. (12) to −u_x for any schedule with τ_1=0, while the radial term in Eq. (13) vanishes only if the unverified condition ∂r_1^2/∂φ_1=0 holds. Thus the reported one-step MSE recovery is the standard DDPM v-prediction property and provides no evidence for the correctness of the new radial term at intermediate times. A numerical check of ∂r^2/∂φ against automatic differentiation, or an ablation that removes the radial term, is needed before the method can be considered validated.","section":"§4.1, \"Resuming MSE training objective\""},{"comment":"The empirical support is weak and partially negative. On U-Net, the best relative gains over PDE-Refiner are about 4–5% (one-step MSE 1.86e-3 vs 1.95e-3; unrolled loss 2.34e-1 vs 2.45e-1 for step 3), with no significance testing and no ablation that isolates the effect of the blurring schedule from other changes. For FNO, both refinement approaches degrade performance relative to plain MSE training (3.88e-1 vs 3.11e-1), and the paper itself concedes poor FNO cooperation; the abstract's claim that PDESpectralRefiner is \"more accurate for both one-step MSE loss and rollout loss\" is therefore not supported in the FNO case. In addition, the blurring direction and maximum blur σ_B appear to be selected after inspecting the DDPM refiner's spectral errors; a validation-based selection protocol is needed to avoid overfitting the test set.","section":"§5.1, Table 1"}],"minor_comments":[{"comment":"The notation α_t is heavily overloaded: Eq. (10)–(11) write α_t = α_t · d_t using the same symbol for the scaled and unscaled coefficients; introduce a distinct symbol such as \\bar α_t for the scaled coefficient throughout.","section":"§3.3–§4"},{"comment":"The d_min term in the definition d_t = (1−d_min)exp(−Λτ_t)+d_min is silently dropped in §4's d_t = exp(−Λτ_t); the relationship between these definitions and the role of d_min should be stated explicitly.","section":"§3.3–§4"},{"comment":"The frequency-spectrum plots are essential to the motivation and analysis, but in the submitted text they appear as unreadable glyphs rather than legible figures; they should be regenerated and embedded correctly.","section":"Figures 2 and 3"},{"comment":"There are numerous typos and terminological inconsistencies, including \"Navior Stokes\" for \"Navier–Stokes\", \"stabling rollout\" for \"stable rollout\", and \"resume MSE training objective\" where \"recover\" seems intended; the manuscript needs careful proofreading.","section":"Throughout"},{"comment":"The sampling formula in Eq. (15) should clarify whether u_x denotes the Fourier-transformed field and whether all operations are elementwise in spectral space; the text alternates between u_s and u_x for the same quantity, which is confusing.","section":"§4, Prop. 2 / Eq. (15)"},{"comment":"The terms \"Step 1\" and \"Step 3\" and the relation of the first refinement step to the diffusion timestep t=1 should be defined precisely in the main text, since the interpretation of the reported losses depends on this convention.","section":"§5.1, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not in a publishable state in its current form: the central derivation is missing, parts of the main text are unreadable, and the empirical support is marginal. However, the underlying idea is tractable and the experimental setup is appropriate, so the paper could become a useful contribution after a major revision that supplies a correct derivation of Eq. (14), includes the missing appendix, and adds a validation-based schedule-selection protocol. I therefore recommend major revision rather than rejection, but I would not be able to endorse acceptance without seeing the corrected derivation and re-validated experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me cut to the chase. The idea—apply blurring diffusion to PDERefiner and design a v-prediction for it—is a reasonable, even attractive, next step. The paper does a few things well: it connects the blurring schedule to the observed over-degradation of high frequencies in Navier-Stokes, it reports honest negative results for FNOs, and it lays out the spectral intuition clearly. That's not nothing.\n\nBut the load-bearing math in Prop. 1 does not check out as printed. I re-derived ∂r²/∂φ from the paper's own definitions (α_t = cos φ, σ_t = sin φ, τ_t = (σ_B²/2)cos²φ). You get 2 sinφ cosφ (1 − e^{−2Λτ}) + 2Λ σ_B² cos³φ sinφ e^{−2Λτ}. Eq. 14, however, has a second term with cos⁶ and a factor π/ln σ0 that does not appear in this derivative. The mismatch is substantial. It's possible the appendix (which the arXiv version does not contain) defines something differently, but with the main text as given, the central formula appears wrong.\n\nWorse, the paper's headline claim—recovering MSE on the first refinement step—doesn't actually test Eq. 14. At t=1, α_1=0 and τ_1=0, so v_1 = −u_x regardless of the derivative term. The experiments compare one-step and three-step refinement, so any real validation of the new velocity has to come from the intermediate steps, and those depend entirely on the suspect Eq. 14 and Prop. 2. The reported gains are small: U-Net unrolled loss 2.45e−1 to 2.34e−1, FNO 3.93e−1 to 3.88e−1, and the FNO refiner still underperforms plain MSE training. Also, the blurring direction and maximum blur are selected after seeing the failure mode, which is fine for a design study but weakens the claim that the method 'matches the spectrum' in a predictive sense.\n\nI want to be fair: the paper is readable, the limitations are acknowledged, and the underlying hypothesis—that controllable spectral reweighting can help diffusion-based PDE refinement—is worth exploring. But a referee cannot verify the central identity because the derivation is in a missing appendix, and my own direct check says it's wrong as stated. That needs to be fixed before I'd trust the empirical claims.\n\nVerdict: yes, I would send this to peer review, because the idea merits referee time and the issue might be a typo or a missing definition. But it needs major revision, and the authors must show the derivation or correct the formula. I would not cite it in its current form.","headline":"A plausible spectral-refinement idea undermined by an unverified, likely incorrect v-prediction formula and thin empirical gains.","tokens_in":15876,"tokens_out":7720,"would_cite":false,"duration_ms":77184,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that reweighting frequency components in a blurring-diffusion refiner reduces one-step and rollout error for PDE surrogates compared with the DDPM-based refiner, and that a new blurring v-prediction formula makes the…","keywords":["PDE surrogate models","long rollouts","diffusion refinement","blurring diffusion","v-prediction","spectral reweighting","Navier-Stokes equations","frequency spectrum"],"falsifier":"Compute $r_t^2 = \\alpha_t^2 \\exp(-2\\Lambda\\tau_t)+\\sigma_t^2$ directly with $\\tau_t = \\sigma_B^2/2 \\cos^2(\\pi t/2)$ and compare $\\partial r^2/\\partial\\phi$ with Eq. 14, evaluating at $t=1$; if $r_1 \\neq 1$ or if the radial derivative fails to vanish there, then $v_1 = -u_x$ does not hold and the first step is not the MSE objective, undoing the paper's central reduction.","tokens_in":14795,"feed_emoji":"🌊","tokens_out":6088,"duration_ms":61596,"temperature":0.7,"pith_summary":"This paper tries to establish that the failure mode of diffusion-based PDE refiners on harder problems is a spectral one, and that it can be fixed by weighting frequencies explicitly. Standard DDPM refinement over-degrades high-frequency content for a problem like 2D Navier-Stokes, where high frequencies carry more energy than in 1D Kuramoto-Sivashinsky. The paper adopts blurring diffusion, which operates in Fourier space and multiplies each frequency by a per-frequency coefficient, and contributes a blurring v-prediction that reduces to the usual MSE objective at the first refinement step. It reports lower one-step and rollout error than DDPM-based PDERefiner on 2D Navier-Stokes for U-Net and neural-operator backbones. If right, the result turns the blurring schedule into a practical spectral dial for PDE rollout stability.","feed_headline":"Spectral reweighting sharpens diffusion refiners for PDE rollouts","feed_subtitle":"A blurring v-prediction makes the first refiner step match MSE training and beats DDPM on 2D Navier-Stokes.","key_machinery":"The central object is the blurring v-prediction formula, a velocity for a diffusion trajectory whose radius $r_t = \\sqrt{\\alpha_t^2 \\exp(-2\\Lambda\\tau_t)+\\sigma_t^2}$ varies in time, so the trajectory is an ellipse rather than a circle. The formula splits the velocity into a perpendicular component $\\alpha_t u_\\epsilon - \\sigma_t \\exp(-\\Lambda\\tau_t) u_x$ and a radial component proportional to $\\partial r^2/\\partial\\phi$, where $\\phi$ parameterizes the data-to-noise rotation. The radial term is what makes sampling and training consistent under a non-constant radius, and the boundary behavior of the chosen $\\cos^2$ blurring schedule is what collapses the first step to the MSE prediction $v_1 = -u_x$.","core_discovery":"In the paper's own terms, PDESpectralRefiner extends the PDERefiner idea by replacing the isotropic DDPM noise schedule with a frequency-dependent blurring schedule: in Fourier space the signal coefficient becomes $\\alpha_t \\cdot \\exp(\\pm\\Lambda\\tau_t)$ while the noise coefficient stays $\\sigma_t$. Because this breaks the variance-preserving circle $\\alpha_t^2+\\sigma_t^2=1$, the paper derives a new velocity formula (Eqs. 12-14) whose first term is the usual perpendicular velocity and whose second term is a radial correction proportional to $\\partial r^2/\\partial\\phi$. With the schedule $\\tau_t = \\sigma_B^2/2 \\cos^2(\\pi t/2)$, the radius satisfies $r_1=1$ and the radial derivative vanishes at $t=1$, so the first predicted velocity is $v_1=-u_x$: the first refinement step is exactly MSE training. A companion sampling formula (Prop. 2) recovers $u_x$ from the predicted velocity, and the paper reports that on 2D Navier-Stokes this spectral reweighting improves both one-step MSE and unrolled rollout loss relative to the DDPM-based refiner for U-Net, and improves the degraded FNO refinement results as well.","pith_inferences":["If the derivative identity in Eq. 14 holds generally, the blurring v-prediction should extend to other frequency schedules beyond the $\\cos^2$ and $\\sin^2$ cases the paper discusses, giving a family of refiners indexed by a spectral mask.","The paper's own FNO results suggest a limit: refining already hurts Fourier operators with mode cutoff because noise breaks inter-frequency interactions; a natural testable extension is to mask noise only on high-frequency modes that the operator keeps, rather than the full spectrum.","Because the claimed MSE recovery depends on a boundary calculation deferred to an appendix, a reader can settle the paper's core mechanism by differentiating $r_t^2$ directly before committing to the method.","The anisotropic spectra of $v_x$ and $v_y$ in the Navier-Stokes data point toward direction-dependent blurring schedules as an obvious next test."],"forward_implications":["With the down-scaling schedule $d_t=\\exp(-\\Lambda\\tau_t)$, the model matches the high-frequency spectrum of 2D Navier-Stokes better than DDPM refinement and gives the best three-step refinement results on the U-Net backbone.","With the up-scaling schedule $d_t=\\exp(\\Lambda\\tau_t)$, the model concentrates on high frequencies immediately, which is the better one-step choice when only a single refinement step is available.","For Fourier neural operators, where refinement with DDPM hurts accuracy, spectral adjustment partially recovers the loss; the paper attributes this to preserving signal-to-noise ratio and frequency-mode interactions under mode cutoff.","The blurring schedule becomes a per-problem design parameter: the same framework can use different $\\Lambda$ for different physical quantities whose spectra differ.","Spectral reweighting changes only the amplitudes of the complex Fourier modes, not their phases, so it emphasizes frequencies without distorting their alignment."],"supporting_citations":[{"why":"Provides the PDERefiner baseline whose v-prediction first step recovers MSE training; the new method must beat it.","marker":"Lippe et al. 2024"},{"why":"Supplies the blurring diffusion framework and the blurring schedule that the paper modifies with its own scaling choices.","marker":"Hoogeboom and Salimans 2023"},{"why":"Introduces v-prediction and the circular-motion intuition that the paper rederives for elliptic trajectories.","marker":"Salimans and Ho 2022"},{"why":"Defines the DDPM forward and reverse processes and the time indexing used throughout the refiner formulation.","marker":"Ho, Jain, and Abbeel 2020"},{"why":"Connects diffusion-style generation to heat dissipation and high-frequency amplitude reduction, which motivates the Fourier analysis.","marker":"Rissanen, Heinonen, and Solin 2023"},{"why":"Provides the 2D Navier-Stokes dataset through PDEArena and the modern U-Net backbone used in the experiments.","marker":"Gupta and Brandstetter 2022"},{"why":"Defines the Fourier neural operator backbone used as a second architecture for testing the method.","marker":"Li et al. 2021"}],"fun_headline_variants":["Spectral adjustment boosts diffusion refiner for PDE rollouts","Frequency-aware refiner makes PDE rollouts more accurate","Blurring diffusion v-prediction sharpens PDE rollout accuracy","Spectral reweighting beats standard diffusion refiners on 2D Navier-Stokes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands on the derivative identity in Eq. 14 being correct and on the chosen $\\cos^2$ blurring schedule having $r_1=1$ with zero radial derivative at $t=1$; the paper defers that calculation to an appendix absent from the main text, and both the first-step MSE recovery and the sampling formula depend on it.","fun_headline_variants_meta":{"raw":{"variants":["Spectral adjustment boosts diffusion refiner for PDE rollouts","Frequency-aware refiner makes PDE rollouts more accurate","Blurring diffusion v-prediction sharpens PDE rollout accuracy","Spectral reweighting beats standard diffusion refiners on 2D Navier-Stokes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1845,"prompt_tokens":1038,"completion_tokens":807,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":732}},"tokens_in":654,"tokens_out":807,"duration_ms":9373,"temperature":1.0,"reasoning_tokens":732,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:19:51.330068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $r_t^2 = \\alpha_t^2 \\exp(-2\\Lambda\\tau_t)+\\sigma_t^2$ directly with $\\tau_t = \\sigma_B^2/2 \\cos^2(\\pi t/2)$ and compare $\\partial r^2/\\partial\\phi$ with Eq. 14, evaluating at $t=1$; if $r_1 \\neq 1$ or if the radial derivative fails to vanish there, then $v_1 = -u_x$ does not hold and the first step is not the MSE objective, undoing the paper's central reduction.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PDERefiner baseline whose v-prediction first step recovers MSE training; the new method must beat it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the blurring diffusion framework and the blurring schedule that the paper modifies with its own scaling choices."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces v-prediction and the circular-motion intuition that the paper rederives for elliptic trajectories."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Connects diffusion-style generation to heat dissipation and high-frequency amplitude reduction, which motivates the Fourier analysis."}],"review_version":1}