{"id":"3e3abe4b-0532-4153-a144-10a54bad5636","arxiv_id":"2607.04982","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A joint velocity–slope diffusion prior plus plane-wave PDE regularization reconstructs more continuous, geologically plausible velocity models from sparse wells than structural preconditioning alone.","lead":"The paper builds high-resolution subsurface velocity models from sparse well logs by combining a joint velocity–slope diffusion prior with plane-wave structural constraints. It matters because better velocity models improve seismic imaging for reservoirs, CO2 storage, and exploration under limited well control.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Volve train–test overlap plus purely visual evaluation leave the claimed improvement over structural inversion unquantified and possibly inflated by prior leakage.","rationale":"The Reader already flagged Volve leakage, purely visual evaluation, and slope-source sensitivity as the reasons for CONDITIONAL rather than ACCEPT. Those are precisely the load-bearing soft spots: without a clean quantitative comparison on an unseen model, the claim that the joint prior “improves \to geological realism” cannot be separated from prior memorization or from the benefit already provided by the PW-PDE term alone. My concrete test simply operationalizes the Reader’s weakest_assumption into a single falsifiable experiment. No stronger internal inconsistency appears in the math (the stacked system (10)–(11) and the velocity-only clean-estimate guidance (14) are standard and coherent). Therefore the verdict remains CONDITIONAL; the concern does not justify REJECT, but it does confirm that the central claim is not yet fully established.","tokens_in":14191,"tokens_out":624,"duration_ms":5849,"concrete_test":"Retrain the joint prior after completely removing every Volve-derived patch from the 5000-model corpus, then re-run the exact Volve experiment of §3.2 (same wells, same κ/λ/μ, same DDIM schedule). Report RMSE and structural-similarity of the reconstructed velocity versus the true Volve model for both the original and the decontaminated prior, and versus the pure PW-PDE structural inversion baseline. If the diffusion advantage shrinks by more than ~20 % or disappears, the central claim is inflated by train–test leakage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the joint velocity–slope diffusion prior + PW-PDE + structural preconditioning improves structural continuity and geological realism relative to conventional structurally preconditioned inversion (Abstract; §3.2–3.3). For that relative improvement to be attributable to the method rather than to memorization or visual preference, two conditions must hold: (1) the generative prior must not have already seen the target geology, and (2) the improvement must be measurable. Condition (1) is violated for the only synthetic with ground truth: §3.1 states the 5000-model training set includes Volve, and §3.2 then reconstructs “the Volve synthetic model.” No hold-out, leave-one-out, or domain-shift protocol is described. Condition (2) is unmet: all comparisons (Figs. 3, 5, 7) are visual; no RMSE/MAE/SSIM against true Volve velocity, no well-log misfit at blind Well 5 beyond overlaid profiles, and no quantitative seismic-data residual norms are reported. Consequently the strongest empirical support for the claim rests on a contaminated synthetic and qualitative field figures, so the magnitude (and even the existence) of a genuine generalization gain remains unestablished.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a unified framework for reconstructing high-resolution velocity models from sparse well logs by combining (i) plane-wave PDE regularization, (ii) structurally preconditioned Tikhonov inversion with a slope-guided spray operator S, and (iii) measurement-guided DDIM posterior sampling under a joint two-channel velocity–slope diffusion prior trained with PWD slopes. Local slopes enter both the structural operators (S and D(γ)) and the generative state; guidance is applied only to the velocity clean estimate (Eqs. 9–15), after which the reverse step continues jointly. Experiments on the Volve synthetic model and the Viking Graben field line (with Well 4 conditioning and Well 5 held out) are used to argue improved structural continuity, lateral consistency, and geological realism relative to conventional structurally preconditioned inversion, with practical inference via few-step DDIM.","tokens_in":14503,"tokens_out":1404,"duration_ms":14812,"significance":"If the claimed gains hold under proper generalization tests, the work is a useful contribution to well-constrained velocity model building: it couples a joint generative prior with explicit plane-wave structural physics rather than relying solely on RTM-derived structure (as in the authors’ prior work), and it demonstrates a field-data path with a blind well and RTM-initialized slopes. Strengths include a clear stacked least-squares formulation (Eqs. 9–11), a concrete velocity-only guidance recipe (Eqs. 13–15), and an honest Discussion of slope quality and prior representativeness. The main scientific value is methodological integration rather than a new physical principle; that value depends on showing that improvements are not artifacts of prior leakage or purely visual preference.","major_comments":[{"comment":"§3.1 states that the 5000-model training corpus includes Volve; §3.2 then reconstructs “the Volve synthetic model” as the only ground-truth synthetic showcase. No hold-out, leave-one-family-out, or domain-shift protocol is described. Because the central claim is relative improvement over structural inversion under a learned prior, this train–test overlap is load-bearing: the Volve figures cannot establish generalization of the joint prior. A non-overlapping synthetic (or explicit leave-Volve-out retraining) is needed before the synthetic results can support the Abstract claim.","section":null},{"comment":"§3.2–3.3 and Figs. 3–8 support “improved structural continuity, lateral consistency, and geological realism” almost entirely by visual comparison and overlaid well profiles. No RMSE/MAE/SSIM (or equivalent) versus true Volve velocity, no quantitative blind-well misfit at Well 5, and no seismic residual norms for Figs. 4, 6, 8 are reported. Without such metrics—and preferably an ablation isolating PW-PDE, structural preconditioning alone, and the joint slope channel—the magnitude and even existence of a genuine gain over conventional preconditioned inversion remain unestablished.","section":null},{"comment":"The joint velocity–slope prior is a stated contribution (Abstract; §2.1; contribution bullets), yet guidance never corrects the slope channel (Eq. 14: γ̃0 = γ̂0), and Discussion §4 notes that generated slopes often stay close to the initial structural estimate. The manuscript does not show that joint training improves reconstructions relative to a velocity-only diffusion prior with the same S/D(γ) operators. A controlled comparison (joint vs velocity-only prior; fixed vs recomputed slopes) is needed to justify the two-channel design as load-bearing rather than incidental.","section":null},{"comment":"Free parameters κ, λ, μ, η, warm-start t/T, and LSQR iteration count are fixed by statement (§3.2: κ=10−6, λ=0.01, μ=0.6, η=0.3, T=20) without sensitivity or stability analysis. Given that Discussion §4 already stresses strong dependence on slope quality (initial-velocity PWD vs RTM), the relative ranking of methods in Figs. 5 and 7 could shift under modest retuning. At least a limited sensitivity study on κ and μ (and slope source) should accompany the field claims.","section":null}],"minor_comments":[{"comment":"Throughout: “V olve” appears with a spurious space (Abstract, §1, §3.2, figure captions); fix consistently to “Volve”.","section":null},{"comment":"§3.3: “trevltime tomography” → “traveltime tomography”; “full-wavefor inversion” → “full-waveform inversion”.","section":null},{"comment":"§6 Acknowledgment: “DeepWave sponsors fort their support” → “for their support”.","section":null},{"comment":"Eq. (1)–(2) and surrounding text: γγγ notation is heavy; a single bold γ would improve readability without loss of meaning.","section":null},{"comment":"Fig. 3i / 5g–h / 7g–h: well profiles would be clearer with a residual panel or tabulated misfit; currently the eye must judge “higher resolution agreement.”","section":null},{"comment":"§2.3 Eq. (8): the plane-wave PDE is written as an approximate equality; state the discrete residual norm used in D(γ) more explicitly (finite-difference stencil, boundary treatment).","section":null},{"comment":"References [43] is cited as arXiv:2603.01231 (future-dated relative to this manuscript’s stamp); ensure citation metadata and priority relative to the present work are accurate.","section":null},{"comment":"§3.1: “500 training epochs” in Discussion vs “50 epochs” in the training paragraph—reconcile the training budget statement.","section":null}],"recommendation":"major_revision","confidential_remarks":"The methodological integration is publishable after revision, but the Volve-in-training issue and purely qualitative evaluation are the two points that currently prevent a fair assessment of the Abstract claim. I would not reject on novelty grounds: the joint slope channel plus PW-PDE guidance is a reasonable incremental step beyond the authors’ prior guided-DDIM VMB work. Fit for a geophysics methods journal is good if quantitative hold-out results are added. No integrity concerns beyond the undisclosed train–test overlap, which should be fixed transparently rather than treated as misconduct."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a clean methods paper that puts a joint two-channel velocity–slope DDPM inside measurement-guided DDIM, with velocity-only Gauss–Newton updates and plane-wave PDE regularization in one loop. That combination is what is actually new. Structural spray, PWD slopes, and DPS-style guidance already exist in the literature they cite, including their own prior work.\n\nWhat they do well. The linear algebra (Eqs. 9–15) is standard and consistently applied. The reverse process is described carefully enough to reimplement. The Discussion is unusually honest about slope quality and prior sensitivity. Viking with a held-out well, plus the RTM-versus-initial-slope comparison, is useful field evidence. Figures look better than pure structural inversion, and they are open that RTM slopes help more than slopes from a smooth background. Inference cost after training is practical via DDIM.\n\nSoft spots, in proportion. Volve sits in the 5000-model training corpus and is then the only synthetic with ground truth. That is a real leakage risk for the showcase, not a footnote. Evaluation is almost entirely visual—no RMSE/SSIM against true Volve, no residual norms, no uncertainty. Free parameters (κ, λ, μ, η) are hand-set. No code or weights. Those gaps do not break the method; they leave the size of the claimed improvement unmeasured. The stress-test on train–test overlap is fair for the synthetic; the field blind well still carries weight.\n\nWho it is for: people already doing structure-oriented VMB or diffusion-for-geophysics. A serious editor should send it to referees. I would engage if I were working on well-guided velocity building; I would not treat the Volve panels as decisive until they re-run with held-out geology and report numbers. Peer review, with quantitative metrics and a cleaner train–test split.","headline":"Solid methods stitch of known pieces into a joint velocity–slope diffusion prior; Volve train–test overlap and purely visual metrics leave the claimed gain unquantified.","tokens_in":15132,"tokens_out":494,"would_cite":true,"duration_ms":12059,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A joint velocity–slope diffusion prior, guided by sparse wells and plane-wave structure, reconstructs more continuous high-resolution subsurface velocity models than classical structural preconditioning alone.","keywords":["velocity model building","diffusion models","plane-wave destruction","structural preconditioning","well-log interpolation","DDIM","joint velocity-slope prior","geophysical inversion"],"falsifier":"On a hold-out field line with blind wells, if diffusion-guided models with plane-wave PDE fail to improve blind-log match and reflection continuity over classical structural preconditioning—especially when training geology mismatches the target—the claimed gains in continuity and realism would be falsified.","tokens_in":15026,"feed_emoji":"🗺️","tokens_out":852,"duration_ms":18307,"temperature":0.7,"pith_summary":"High-resolution subsurface velocity models are needed for reservoir work and monitoring, yet surface seismic data are band-limited and wells are sparse. This paper presents a unified reconstruction that paints well values along local geological dips, enforces a plane-wave PDE so velocity stays constant along those dips, and steers a diffusion sampler with a generative prior trained on paired velocity and slope fields. Guidance from the well-fitting inverse problem corrects only the velocity channel during DDIM sampling; the joint prior keeps slope and structure consistent. On a Volve synthetic model and Viking Graben field data, the combination improves lateral continuity, structural realism, and blind-well agreement relative to conventional structurally preconditioned inversion, while inference stays practical once the prior is trained.","feed_headline":"Diffusion fills sparse wells with continuous velocity structure","feed_subtitle":"Plane-wave constraints and a learned slope prior beat classical structural smoothing on field data","key_machinery":"Joint velocity–slope diffusion prior with velocity-only measurement guidance: the network learns (velocity, plane-wave-destruction slope) pairs; each DDIM reverse step forms a clean estimate, solves a structurally preconditioned Tikhonov system (spray operator S plus plane-wave PDE operator D(γ)) to correct only the velocity channel, then recomposes the next noisy state so structure remains coupled.","core_discovery":"Coupling plane-wave PDE regularization and structurally preconditioned least-squares well fitting with measurement-guided diffusion posterior sampling under a joint velocity–slope generative prior yields velocity models with better structural continuity, lateral consistency, and geological realism from sparse well logs than conventional structural preconditioning, at practical DDIM inference cost.","pith_inferences":["The same joint prior could condition multiparameter elastic or time-lapse velocity updates if training pairs include those attributes.","Steep dips and faulted zones are natural stress tests of whether the PDE and spray operators still dominate the generative prior.","Training-set diversity is the practical bottleneck; domains far from the training mix may need adaptation rather than pure transfer.","Explicitly updating slopes from the evolving velocity during reverse sampling, as the discussion suggests, is a direct next experiment for fuller structural coupling."],"forward_implications":["Sparse well logs can be propagated farther along geological structure without isotropic oversmoothing.","Slopes from a migrated image supply stronger structural guidance than slopes from a smooth interpolated starting model when the background lacks dip detail.","After one-time prior training, DDIM inference cost becomes comparable to classical preconditioned inversion.","Explicit PDE structural constraints plus a learned prior can reduce early reliance on migration images that themselves depend on an accurate background velocity.","Joint generative modeling of velocity and dip offers a template for other ill-posed seismic inversions that need structural consistency."],"fun_headline_variants":["Joint velocity-slope diffusion builds continuous models from sparse wells","Plane-wave PDE and diffusion prior reconstruct velocity with dip continuity","Diffusion sampling guided by slopes fills sparse wells geologically","Structurally preconditioned diffusion yields consistent velocity models","Measurement-guided joint prior improves lateral velocity realism from logs"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes the trained velocity–slope prior and the slopes used to build the structural operators correctly represent the target geology, so well information is painted along the right dips rather than the wrong ones.","fun_headline_variants_meta":{"raw":{"variants":["Joint velocity-slope diffusion builds continuous models from sparse wells","Plane-wave PDE and diffusion prior reconstruct velocity with dip continuity","Diffusion sampling guided by slopes fills sparse wells geologically","Structurally preconditioned diffusion yields consistent velocity models","Measurement-guided joint prior improves lateral velocity realism from logs"]},"model":"grok-4.5","effort":"low","cost_usd":0.003626,"raw_usage":{"total_tokens":1093,"prompt_tokens":685,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":36260000,"prompt_tokens_details":{"text_tokens":685,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":345,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":685,"tokens_out":63,"duration_ms":2941,"temperature":1.0,"reasoning_tokens":345,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T10:33:35.294556+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a hold-out field line with blind wells, if diffusion-guided models with plane-wave PDE fail to improve blind-log match and reflection continuity over classical structural preconditioning—especially when training geology mismatches the target—the claimed gains in continuity and realism would be falsified.","supporting_citations":[],"review_version":1}