{"id":"1a89766b-9e5a-41fd-aafc-39470027f6e2","arxiv_id":"2607.03772","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Level-set PINNs with interface-aware backpropagation recover sharp anomalous domains from gravity measurements in 2D and 3D.","lead":"The paper introduces level-set PINNs that recover unknown mass domains from gravity data by representing sharp interfaces as the zero level-set of a continuous neural network. This gives a practical, mesh-free solver for a well-posed subclass of the notoriously non-unique inverse gravimetry problem.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged uniqueness hypotheses.","rationale":"The reader's weakest assumption (a-priori knowledge of f and the geometric constraints of Theorem 2.1) is precisely the load-bearing condition for uniqueness; without it the inverse problem reverts to classical non-uniqueness and the PINN recovery becomes one of many possible solutions. The paper never claims otherwise. The interface-aware gradient redefinition is justified by a careful co-area / Fermi-Dirac analysis (Proposition 3.1) that shows the modified PDE loss differs from the true loss by O(τ), so the optimization target remains valid for small τ. Architecture choices are supported by Sobolev approximation theory. The only practical gaps (no public code, heuristic λ_i, noise-free data) are already noted by the reader and correctly motivate CONDITIONAL rather than unconditional ACCEPT. No stronger technical objection surfaces on a second pass.","tokens_in":44944,"tokens_out":440,"duration_ms":4278,"concrete_test":"Re-run Example 1 with additive 1 % Gaussian noise on the boundary gravity data (keeping all other hyper-parameters fixed). If the recovered zero level-set still matches the true interfaces to visual accuracy and the residual norms remain comparable to the noise-free case, the method's practical robustness is confirmed; otherwise the noise-free synthetic evidence is weaker than claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (level-set PINNs with interface-aware back-propagation recover D from surface gravity) rests on the well-posedness class of Theorem 2.1 and on the approximation |L̃_pde - L_pde| = O(τ) of Proposition 3.1. Both are internally consistent: the numerics stay inside the uniqueness regime (known non-negative f, star-shaped/convex geometries), the error analysis is rigorous under the signed-distance assumption, and the synthetic recoveries outperform the naïve PINN baselines. No hidden inconsistency or unstated assumption that would invalidate the claim under the paper's own hypotheses was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes level-set physics-informed neural networks for the domain inverse problem of gravimetry. This problem restores uniqueness (Theorem 2.1) to the severely ill-posed inverse gravimetry problem when the density contrast f is known a priori and non-negative and the anomalous domain D satisfies geometric constraints (star-shaped, convex in one direction, etc.). The unknown domain is encoded as the zero level-set of a continuous neural network ϕ_ω, yielding the density ρ_ω = f H_τ(ϕ_ω). A composite PINN loss (PDE residual, boundary gravity data, far-field condition, eikonal and length regularizers, parameter regularization) is minimized by Adam. To mitigate exploding/vanishing gradients induced by the sharp interface, an interface-aware back-propagation redefines the Heaviside derivative as δ_˜τ with ˜τ > τ while keeping a small τ in the forward pass; Lemma 3.1 and Proposition 3.1 bound the resulting loss discrepancy by O(τ). Adaptive collocation near the interface and architecture selection (ReLU for ϕ_ω, tanh for the potential U_θ) justified by W^{1,∞} approximation theory are also introduced. Synthetic 2-D and 3-D recoveries, together with standard-PINN baselines that fail, are presented as evidence of effectiveness.","tokens_in":45124,"tokens_out":1142,"duration_ms":30615,"significance":"If the claims hold, the work supplies a mesh-free neural method for a well-posed subclass of inverse gravimetry and, more broadly, for domain/interface reconstruction problems. The interface-aware gradient redefinition together with its rigorous O(τ) error analysis (Proposition 3.1) is a concrete technical contribution that can transfer to other PINN-based free-boundary or multiphase problems. Explicit approximation-capacity bounds for the chosen ReLU and tanh architectures (Corollaries 3.4–3.6) and clear ablation against naïve PINNs that ignore the level-set structure are additional strengths. The method stays strictly inside the uniqueness class of Theorem 2.1 and does not over-claim.","major_comments":[{"comment":"The loss weights λ1–λ5, interface thicknesses τ/˜τ and refinement width C_refine are chosen purely heuristically (Table 1) and held fixed for every example. No sensitivity study or adaptive-weighting experiment is supplied. Because λ1 = λ2 = 20 dominate the data terms, the reported geometric recoveries could be fragile; this is load-bearing for the claim that the method is effective.","section":"Section 3.2 and Table 1"},{"comment":"The O(τ(c−1))+O(τ³) bound that justifies the interface-aware strategy assumes that ϕ_ω is a signed-distance function near ∂D and that collocation points are uniformly distributed. In the actual algorithm the eikonal penalty is only soft (λ3 = 1) and adaptive refinement concentrates points, so the hypothesis of Proposition 3.1 is not guaranteed along the optimization trajectory. A numerical verification of |˜L_pde − L_pde| would close the gap.","section":"Proposition 3.1"},{"comment":"With the arbitrary constant c0 = 1 the guaranteed approximation error for the tanh network U_θ is ε ≈ 0.27 (Section 3.5.3). While the authors remark that practice outperforms the bound, an error of this magnitude relative to typical potential values leaves the theoretical capacity argument for the Laplacian-sensitive potential incomplete.","section":"Section 3.5.3"}],"minor_comments":[{"comment":"The surrogate derivative is written H′_τ(s) := δ_˜τ(s) (Eq. 3.18), overwriting the true derivative. A distinct symbol (e.g., δ^surr) would avoid notational confusion.","section":"Equation (3.18)"},{"comment":"All examples employ noise-free synthetic data generated from the exact integral equation with perfectly known f. A brief noise-robustness or f-perturbation test would better support the practical claim, even though the theoretical setting of Theorem 2.1 is respected.","section":"Section 4"},{"comment":"Pre-training of both networks (Section 4.1) is presented as essential because of initialization sensitivity, yet no ablation without pre-training is shown.","section":"Section 4.1"},{"comment":"The discussion of related PINN-interface papers (Section 1) correctly notes that they treat forward problems; a one-sentence explicit contrast of the inverse setting would sharpen the novelty claim.","section":"Section 1"}],"recommendation":"minor_revision","confidential_remarks":"Solid contribution that genuinely extends the authors’ earlier classical level-set work on the same inverse problem into the PINN regime. No citation or novelty-disclosure issues. Suitable for a math.NA / scientific-ML journal once the hyper-parameter robustness and the gap between Proposition 3.1 and the actual training dynamics are addressed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful piece here is the interface-aware back-propagation: they keep a small τ for a sharp Heaviside on the forward pass and replace the derivative with a wider δ_˜τ (˜τ > τ) only on the backward pass, then prove that the modified PDE loss stays within O(τ) of the true one under a signed-distance assumption (Lemma 3.1 + Prop. 3.1). That, plus the explicit ReLU/tanh architecture recommendations grounded in W^{1,∞} approximation theory, is what is actually new. Level-set inverse gravimetry and interface PINNs already exist; the combination with the error analysis and the architecture study is not in the cited literature.\n\nThey do the rest carefully. The loss is the standard PINN residual plus eikonal and length regularizers; the adaptive collocation near the zero level-set is simple and sensible; the 2-D/3-D synthetics recover the geometries (circle+parallelogram, four spheres, mixed shapes, depth-dependent f) and match the gravity data, while the naïve PINN baselines (plain ReLU density or sigmoid-constrained density) produce essentially useless reconstructions. Pre-training of both networks is described and helps. Citations to Isakov, the classical level-set papers, and the recent PINN interface literature look complete and fair.\n\nSoft spots are real but secondary. All λ_i, τ, ˜τ, C_refine and the network size (L=6, width 300) are chosen by hand; there is no public code; the experiments stay inside the uniqueness class of Theorem 2.1 (known non-negative f, star-shaped or convex domains) and use noise-free data. Those are the usual limitations of a first methods paper, not hidden contradictions. The central claim holds under the paper’s own hypotheses.\n\nThis is for people who already work on free-boundary PINNs or geophysical domain inversion and want a mesh-free solver that actually recovers sharp interfaces. It deserves a serious referee; I would send it out. I would cite the interface-aware gradient trick and the architecture discussion if I were writing a related PINN paper in the next year.","headline":"Solid engineering of level-set PINNs for a well-posed subclass of inverse gravimetry; the interface-aware backprop and its O(τ) analysis are the real additions.","tokens_in":45713,"tokens_out":553,"would_cite":true,"duration_ms":6117,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N21","49Q10","68T07","86A22"],"pacs":[],"model":"grok-4.5","headline":"Level-set PINNs recover unknown mass domains from surface gravity by treating the interface as a continuous network whose zero level set is the boundary.","keywords":["level-set method","physics-informed neural networks","domain inverse problem","inverse gravimetry","interface-aware back-propagation","signed-distance function"],"falsifier":"A synthetic experiment in which the true density contrast is deliberately misspecified, or the true domain violates the star-shaped/convexity hypotheses, and the recovered zero level-set fails to match the true boundary while the data residual remains small.","tokens_in":45842,"feed_emoji":"🌐","tokens_out":600,"duration_ms":5141,"temperature":0.7,"pith_summary":"Gravity inversion is classically ill-posed: many different underground mass distributions can produce the same surface measurements. Restricting the unknown to a domain D of known density contrast restores uniqueness under mild geometric conditions. The paper shows that this domain inverse problem can be solved by physics-informed neural networks once the discontinuous characteristic function of D is replaced by a continuous level-set function whose zero contour is the interface. The key technical device is an interface-aware back-propagation rule that widens the numerical support of the interface derivative during training, preventing vanishing or exploding gradients while still keeping the recovered density sharp. Adaptive collocation near the interface and carefully chosen network architectures (ReLU for the level-set, tanh for the potential) complete the scheme. Synthetic two- and three-dimensional tests recover circles, parallelograms, spheres and mixed bodies whose predicted gravity matches the data, demonstrating that the method works in practice.","feed_headline":"Level-set PINNs recover mass domains from surface gravity","feed_subtitle":"A continuous network tracks the interface; wider back-prop gradients keep training stable.","key_machinery":"Interface-aware back-propagation: the numerical Heaviside is kept thin (parameter τ) in the forward pass, but its derivative is replaced by a thicker kernel (˜τ > τ) in the backward pass, so that a broader set of collocation points can drive interface motion without destroying sharpness.","core_discovery":"A continuous neural-network level-set function, trained by a physics-informed loss whose interface derivative is deliberately widened during back-propagation, recovers the unknown domain of a volume-mass anomaly from surface gravity measurements while preserving a sharp interface.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Level-set PINNs recover mass domains via widened interface gradients","Continuous level-set nets solve domain inverse gravimetry stably","Interface-aware backprop lets PINNs track sharp gravity mass boundaries","Widened derivatives enable level-set PINNs to recover volume-mass domains","Adaptive collocation PINNs reconstruct mass anomalies from surface gravity"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The density contrast is known in advance and non-negative, and the unknown domain satisfies one of the geometric constraints that make the domain inverse problem unique.","fun_headline_variants_meta":{"raw":{"variants":["Level-set PINNs recover mass domains via widened interface gradients","Continuous level-set nets solve domain inverse gravimetry stably","Interface-aware backprop lets PINNs track sharp gravity mass boundaries","Widened derivatives enable level-set PINNs to recover volume-mass domains","Adaptive collocation PINNs reconstruct mass anomalies from surface gravity"]},"model":"grok-4.5","effort":"low","cost_usd":0.004758,"raw_usage":{"total_tokens":1327,"prompt_tokens":705,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":47580000,"prompt_tokens_details":{"text_tokens":705,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":551,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":705,"tokens_out":71,"duration_ms":4050,"temperature":1.0,"reasoning_tokens":551,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:03:11.053524+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A synthetic experiment in which the true density contrast is deliberately misspecified, or the true domain violates the star-shaped/convexity hypotheses, and the recovered zero level-set fails to match the true boundary while the data residual remains small.","supporting_citations":[],"review_version":1}