{"id":"cebcbfca-d30f-405f-b967-637d769d9ec6","arxiv_id":"2509.06689","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A level-set joint inversion couples velocity and density through a shared interface and uses known density contrast plus a decaying weight so gravity can aid deep-structure recovery.","lead":"Seismic and gravity data are combined by forcing them to share one underground interface, tracked with a level-set function, so the gravity signal can sketch deep structures that seismic alone misses. Known density contrast and a time-varying weighting let gravity lead first and seismic refine later, giving better synthetic salt and dipping-body images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All four synthetic tests are generated inside the model class defined by Eq. (6), so the shared-interface assumption and known f/c1 priors are never stress-tested; the claim of general superiority is therefore conditional.","rationale":"The reader's weakest_assumption is the single level-set interface hypothesis in Eq. (6) and the reliance on a priori knowledge of f and c1. My stress-test concurs: this is the most load-bearing concern because it sits at the center of the proposed structural coupling. All numerical validations are generated within the model class defined by that hypothesis, with the most critical parameters supplied as ground truth. The paper never tests a model that violates the shared-interface assumption or that has an uncertain density contrast. If the method fails in such a test, the central claim that joint inversion 'combines the advantages of both FWI and gravity inversion' is only valid under a very strong, possibly unrealistic prior. The theorem in Section 2.2 is also used as motivation, but its conditions are explicitly relaxed without proof, weakening the theoretical support. The other weaknesses identified by the reader (hand-tuned hyperparameters, inverse-crime configuration) are real but secondary: they affect the strength of the empirical evidence but do not strike at the core mechanism. Since the reader already assigned CONDITIONAL and my analysis agrees with their weakest_assumption, I recommend no change to the verdict. The proposed concrete test is designed to probe whether the shared-interface assumption is actually load-bearing or simply an untested convenience.","tokens_in":21955,"tokens_out":5395,"duration_ms":62380,"concrete_test":"Run Example 3 with a modified true model: keep the velocity model as in Figure 10(a), but add a deep horizontal density-only layer (e.g., a 1 km thick band at depth 4-5 km with density contrast +0.2 g/cm3) that has no corresponding velocity anomaly, so the true model no longer satisfies Eq. (6). Keep the same acquisition geometry, f known, and the same hyperparameters as in Table 4. If the recovered velocity model shows a spurious low-velocity layer or the salt boundary is displaced, the shared-interface assumption is load-bearing; if the velocity model still matches the true salt and ignores the density-only layer, the concern is partially mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of superiority rests on the level-set parameterization (6): rho=f(r)H(phi), c=c1(r)H(phi)+c2(r)(1-H(phi)), with the same phi for both. In every example (Sections 5.2.1-5.2.4), the true models are constructed to satisfy this assumption exactly, and the density contrast f(r) and the interior velocity c1(r) are prescribed as known (e.g., c1 frozen in all examples, f=(1.8-z)*0.2 in Examples 3-4). The inversion reduces to estimating phi and c2 within the exact generative class of the synthetic data, so the demonstrations are self-fulfilling. The method is never tested on a model where velocity and density interfaces do not coincide or where f is misspecified. Theorem 2.1 is invoked as the theoretical justification for extracting unambiguous gravity information, but it requires the full modulus |nabla U| and convexity/star-shaped constraints, while the algorithm uses only g_z and explicitly relaxes these constraints (Section 2.2). Consequently, outside the tested, exactly-representable class, the gravity residual in Eq. (27) can drive the shared interface phi through the term f*delta(phi), potentially creating spurious velocity interfaces or biasing the density structure. The paper therefore does not establish that the joint inversion is robust to the structural prior itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a level-set structural method for the joint inversion of full-waveform and gravity data. The velocity and density fields are parameterized by a single level-set function and a known density-contrast function f(r), with the interior velocity c1(r) often fixed a priori. A balanced and exponentially decaying weight is introduced to let gravity dominate early iterations and full-waveform data dominate later. The method is tested on four synthetic examples, including a 2D SEG/EAGE salt model, and compared with isolated FWI and gravity inversion. The central claim is that the joint inversion integrates the strengths of both datasets and produces superior reconstructions, especially for deep or extended salt structures.","tokens_in":22310,"tokens_out":3301,"duration_ms":35696,"significance":"If the claims are robust, the paper offers a practically attractive way to impose structural coupling between velocity and density: instead of penalty terms, the shared interface is represented explicitly, and prior information such as f and c1 is inserted directly. The use of an existing open-source FWI engine (Deepwave) and the clear statement of Algorithm 1 are strengths. The main significance, however, is conditional: the four synthetic tests are generated inside the model class defined by Eq. (6), with f and c1 prescribed, so the numerical evidence does not yet establish robustness to the central structural assumption. The theoretical justification via Theorem 2.1 is also invoked in a regime the theorem does not cover.","major_comments":[{"comment":"Theorem 2.1 is the paper's principal justification for extracting 'unambiguous information' from gravity data, but the algorithm uses only the vertical component g_z (Eq. (5)) rather than the full modulus |∇U|, and it explicitly drops the geometric conditions (star-shaped, convex, or convex-in-one-direction) required by the theorem. No proof, counterexample analysis, or numerical test is given to show that the uniqueness benefit survives these relaxations. In addition, Example 3 uses f(r)=(1.8-z)*0.2, which is not constant and may violate condition (3) of Theorem 2.1 depending on the chosen direction. The theoretical claim that gravity contributes meaningful information therefore needs either an extended theorem, a precise statement of the regularity regime in which the method operates, or substantially softened wording.","section":"Section 2.2"},{"comment":"All synthetic true models satisfy the shared-interface parameterization (6) exactly, with the same f(r) and c1(r) used as known priors in the inversion. The waveform and gravity data are generated with the same numerical solvers and the same spatial mesh employed in the inversion (an inverse-crime configuration). The claim in Section 5.2.3 that the joint inversion 'produces superior solutions that closely match the true model' is therefore a demonstration within the exact generative class, not a test of the method's robustness. No experiment considers a velocity-only or density-only boundary, a misspecified f, an unknown c1, or a different forward discretization. The conclusions in Section 6 should be reframed as proof-of-concept, or additional out-of-class experiments should be added.","section":"Sections 5.2.1-5.2.4"},{"comment":"The gravity gradient term in Eq. (27) is f(r) δτ(φ) times the gravity residual. If the structural assumption fails or f is misspecified, this term still moves the shared interface φ, and through Eq. (6) it changes the velocity reconstruction via (c1-c2)δτ(φ). Thus a density anomaly with no velocity counterpart, or an incorrect density contrast, will not merely fail to help; it can actively create a spurious velocity interface. The manuscript does not analyze this mechanism or provide experiments that bound the resulting bias. Given that the paper's central novelty is the shared-interface coupling, this is a load-bearing robustness question.","section":"Equation (27)"}],"minor_comments":[{"comment":"The notation α_{v2} appears to mean α_{c2}, and λ_{c1} is absent although the optimization problem in Eq. (16) includes λ_{c1} E_{c1}. Please clarify whether c1 is always frozen and, if so, state this consistently in the text and tables.","section":"Tables 2, 3, 4, 5"},{"comment":"The terms '2-1 norm TV regularization' and '1-1 norm TV regularization' are nonstandard as written. Please define them explicitly, e.g., as the l2 norm of the gradient magnitude or the sum of absolute partial derivatives, so that the distinction is clear.","section":"Section 3.2.3"},{"comment":"There is a typo: 'repeatedlly' should be 'repeatedly'.","section":"Section 3.2.1"},{"comment":"The SSIM comparison is only reported for the velocity model; an SSIM or similar metric for the density reconstruction would help support the claim of complementary information in the noisy case, where the density result is visibly degraded (Figure 16(f)).","section":"Section 5.2.4"},{"comment":"The limitations paragraph is candid about slow convergence, hyperparameter selection, and Deepwave scalability. It would be useful to also explicitly note that the shared-interface prior and the prescription of f and c1 are untested when violated, since that is the main limitation of the numerical evidence.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: genuinely new algorithmic coupling, but the synthetic evidence is weaker than the claim because every test stays inside the model class of Eq. (6). The specific combination—one level-set function parameterizing both c and rho through a shared interface, plus a two-part weight that balances gradient scales and lets gravity lead early—is not in the prior work I know. The authors also read the gravity uniqueness theorem seriously and correctly conclude that density contrast must be known or the gravity term provides little. The algorithm is transparent: Adam updates, adjoint-state gradients via Deepwave, TV regularization on c1 and c2, reinitialization. That is solid and reproducible engineering, and the FWI-only benchmarks are fair.\n\nSoft spots, in order: (1) All four synthetic models are generated with the same shared-interface assumption and with f(r) and c1(r) prescribed. The inversion estimates phi and c2 inside the exact generative class, so the tests demonstrate consistency, not robustness to the structural prior. A wrong f or a density-only boundary could let the gravity residual move phi through f·delta(phi) and deform the velocity. (2) The paper relaxes the star-shaped/convex-in-one-direction conditions of Theorem 2.1 without proof, so the uniqueness result is more motivational than binding. (3) Data are simulated with the same solver and same mesh; that's a classic inverse-crime setup and can flatter the results. (4) Hyperparameters are hand-tuned per experiment; the authors admit this in the conclusions. No code or data are released, which makes the inverse-crime issue hard to check independently.\n\nCredit where it's earned: the weighting strategy is simple and plausible, the level-set coupling is a neat way to enforce structural similarity directly rather than through a penalty, the noise tests in Example 2 are a good start, and the SSIM numbers in Example 4 give a quantitative edge over FWI. The paper is not overclaiming to the extreme; the conclusion is careful about limitations.\n\nThis deserves a serious referee. In review, I'd ask for one experiment where the interface is not shared, or where f is wrong, to see whether the joint inversion still helps or starts to hurt. If the authors can add that, the paper would be solid. As it stands, it's a conditional acceptance with a clear path.\n\nFor readers: joint-inversion people in exploration geophysics and anyone using level sets in inverse problems. It won't change how FWI is done tomorrow, but it's a useful piece of the toolkit.","headline":"A genuinely new algorithmic coupling of level-set FWI and gravity with a sensible weighting strategy, but the synthetic evidence is weaker than the prose suggests because the tests never leave the generative class of the model.","tokens_in":22793,"tokens_out":2575,"would_cite":true,"duration_ms":26717,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["86A22","35R30","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Gravity data can be made to rescue the deep structures that full-waveform inversion misses, provided the density contrast of the target body is known and velocity and density share one level-set interface.","keywords":["level-set method","joint inversion","full-waveform inversion","gravity inversion","salt structure imaging","multi-physics inversion","inverse gravimetry","balanced-decaying weight"],"falsifier":"Build a synthetic model with two separate anomalous bodies: one with a velocity anomaly and zero density contrast, the other with a density contrast and zero velocity anomaly. Run the proposed joint inversion with the density-contrast value of the second body imposed. If the shared level-set coupling is correct, the velocity reconstruction should ignore the density-only body; if the velocity model is pulled toward the density-only interface, the central claim fails. A second decisive experiment is to image a U-shaped or disconnected anomaly, where the uniqueness theorem's geometric conditions","tokens_in":21853,"feed_emoji":"🧂","tokens_out":8749,"duration_ms":98403,"temperature":0.7,"pith_summary":"This paper argues that gravity data can do more than ride on the coattails of full-waveform inversion: if the density contrast of the target body is known in advance, gravity measurements can actively rescue the deep or extended structures that seismic waves fail to illuminate. The mechanism is a level-set function whose zero set is the shared boundary of both the wave-velocity model and the anomalous-density model, so a single moving interface is recovered from both datasets at once. To make the two datasets cooperate, the authors schedule a balanced, decaying weight that lets gravity dominate the early iterations, when it quickly sketches the overall body, and lets full-waveform inversion dominate later, when its high resolution sharpens the interface. In synthetic tests including salt models, the joint inversion recovers deep salt bodies and noisy data more accurately than either method alone.","feed_headline":"Joint inversion recovers deep salt bodies FWI alone misses","feed_subtitle":"Gravity sketches the deep body first; seismic data then refine it, outperforming either method alone.","key_machinery":"The central object is the shared level-set interface: one scalar field φ whose zero level set is the boundary of the anomalous body, appearing in both the velocity parameterization c(r)=c1(r)H(φ(r))+c2(r)(1−H(φ(r))) and the density parameterization ρ(r)=f(r)H(φ(r)), with H a smoothed Heaviside and δτ its derivative. It enforces structural similarity directly rather than through gradient penalties. The second mechanism is the balanced-decaying weight ω(n)=ω1(n)·ω2(n), set from the ratio of waveform and gravity gradient magnitudes in φ so that the two data misfits act comparably on the interface, and decaying exponentially so gravity dominates the early iterations and full-waveform data domina","core_discovery":"The paper claims that in the joint inversion of seismic and gravity data, the density function should be written as ρ(r)=f(r)H(φ(r)) — a known density-contrast value f times the indicator of a region D — and the velocity as c(r)=c1(r)H(φ(r))+c2(r)(1−H(φ(r))). Because the same level-set function φ carries both parameterizations, the boundary of the salt body or other anomaly is the single object being inverted. The authors further claim that with the weight ω(n)=ω1(n)ω2(n), where ω1 scales the gravity gradient to the waveform gradient and ω2 decays exponentially, the gravity data first establish the large-scale deep structure and the waveform data then refine it. This reverses the usual one-w","pith_inferences":["Inference: The real test of the shared-interface assumption is geology where velocity and density boundaries do not coincide; in such settings a single level-set function would either flatten the velocity image or create a spurious interface, and a multi-level-set extension would be needed.","Inference: Because the density contrast f is imposed a priori, the method does not estimate the magnitude of the density anomaly; treating f as unknown with its own regularization would make the method applicable where salt or basement densities are not well constrained.","Inference: The balanced-decaying weight is essentially a continuation schedule on the data terms; the same schedule could be ported to other multi-physics inversions, with the early dataset chosen as the one providing the fastest large-scale constraint.","Inference: Since the uniqueness theorem's geometric conditions are relaxed in the algorithm, a U-shaped or disconnected anomaly is the natural stress test for whether the gravity information remains reliable outside the proven regime."],"forward_implications":["Deep and laterally extended structures that full-waveform inversion alone cannot resolve become recoverable from surface data, because gravity with a known density contrast supplies large-scale information through the shared interface.","The recovered velocity model remains nearly correct even when gravity noise degrades the density reconstruction, so the velocity and density outputs complement each other.","The method requires no explicit velocity-density formula; structural coincidence replaces petrophysical composition, which is useful where rock-physics relations are unknown.","The interface-penalty and reinitialization steps keep the level-set evolution stable, so the joint inversion can be run from a very general initial guess."],"supporting_citations":[{"why":"Supplies the inverse-gravimetry uniqueness theorem that justifies treating gravity as informative when the density contrast f is known a priori.","marker":"[16, 26, 5]"},{"why":"Introduces the level-set representation of moving interfaces that the shared-boundary parameterization is built on.","marker":"[33]"},{"why":"Provides the level-set gravity imaging formulation and the depth-dependent salt density-contrast formula used in the salt examples.","marker":"[25]"},{"why":"Provides the numerical wave-equation forward modeling and gradient backpropagation used to evaluate the full-waveform misfit gradient.","marker":"[36]"},{"why":"Gives the adjoint-state method used to compute the waveform misfit derivative under the wave-equation constraint.","marker":"[35]"},{"why":"Provides the Adam update rule that the optimization algorithm modifies with per-parameter adjustment coefficients.","marker":"[21]"},{"why":"Defines the standard salt benchmark model used in the final numerical test case.","marker":"[3]"}],"fun_headline_variants":["Gravity first, seismic second: joint inversion nails deep salt","Level-set joint inversion beats FWI alone on deep salt","Joint gravity-seismic inversion finds deep salt that FWI misses","One level set, two data types: deep salt imaged","Gravity cues seismic in level-set joint inversion for salt"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole scheme hinges on two priors: the density contrast f(r) is known ahead of time, and the wave-velocity anomaly and the density anomaly are exactly the same body with one common boundary; if either fails in the field—say a velocity-only sediment boundary or an unknown salt density—the gravity data can pull the reconstructed interface to a false location.","fun_headline_variants_meta":{"raw":{"variants":["Gravity first, seismic second: joint inversion nails deep salt","Level-set joint inversion beats FWI alone on deep salt","Joint gravity-seismic inversion finds deep salt that FWI misses","One level set, two data types: deep salt imaged","Gravity cues seismic in level-set joint inversion for salt"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00055,"raw_usage":{"total_tokens":2461,"prompt_tokens":741,"completion_tokens":1720,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1650}},"tokens_in":485,"tokens_out":1720,"duration_ms":13710,"temperature":1.0,"reasoning_tokens":1650,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:14:35.105734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a synthetic model with two separate anomalous bodies: one with a velocity anomaly and zero density contrast, the other with a density contrast and zero velocity anomaly. Run the proposed joint inversion with the density-contrast value of the second body imposed. If the shared level-set coupling is correct, the velocity reconstruction should ignore the density-only body; if the velocity model is pulled toward the density-only interface, the central claim fails. A second decisive experiment is to image a U-shaped or disconnected anomaly, where the uniqueness theorem's geometric conditions","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the level-set representation of moving interfaces that the shared-boundary parameterization is built on."},{"cited_title":"A level set method for imaging salt structures using gravity data.Geophysics, 81(2):G35–G51, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the level-set gravity imaging formulation and the depth-dependent salt density-contrast formula used in the salt examples."},{"cited_title":"Aminzadeh, J","cited_arxiv_id":null,"evidence_quote":"Defines the standard salt benchmark model used in the final numerical test case."}],"review_version":1}