{"id":"02b50ff9-a971-414d-bb46-bd4fa5dae95b","arxiv_id":"2507.02272","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-informed neural network solves 3-D static elastic deformation and inverts real GPS data for the 2008 Iwate-Miyagi earthquake slip, giving a pattern consistent with previous studies but a lower moment magnitude.","lead":"This paper uses physics-informed neural networks to model 3-D earthquake deformation in the Earth's crust and to estimate the fault slip behind the 2008 Iwate-Miyagi earthquake from GPS data. The method recovers a slip pattern similar to earlier models, but with a smaller moment magnitude, and it highlights how hard it is to fix rigid-body motion in forward-only simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that the method handles realistic crustal structure is not established: the strong-form loss contains no conditions at discontinuous material interfaces, and the Het-Top inversion loss is two orders of magnitude larger than the homogeneous case, indicating the physics is not being solved.","rationale":"After reading the paper, the forward validation against Okada and PyLith and the synthetic inversion tests are genuinely positive: the 'us' representation is competitive, and the synthetic inversions recover slip with small errors in dense and clean settings. The real-data inversion for the 2008 Iwate-Miyagi earthquake yields a slip pattern consistent with previous studies, which is encouraging. However, the paper's own reported loss values for the Het-Top inversion are a serious concern, and the theoretical setup reveals why. The strong-form PINN enforces the equilibrium equation and constitutive law at collocation points, but it does not include any condition at the internal boundaries where JIVSM layers meet and mu is discontinuous. In linear elasticity, the traction vector must be continuous across such interfaces; without that condition, a single smooth network cannot represent the correct piecewise-differentiable solution. The large loss in the Het-Top case is therefore not merely an optimization difficulty but a sign that the model is trying to approximate a solution that is not in the network's function space, or that the loss is dominated by unresolved interface residuals. Since the strongest claim explicitly includes 'with realistic crustal structure,' this missing interface treatment is the most load-bearing concern. The reader's weakest_assumption identified the outer-boundary and rigid-motion issue, which is relevant to the homogeneous case, but the reader's rationale also noted the large Het-Top loss. Thus I partially agree with the reader's overall conditional verdict, but the condition should be revised to require a demonstration that the method handles discontinuous material interfaces, either by adding the appropriate interface loss or by validating against a reference solution. A PyLith benchmark for the actual layered JIVSM would settle this.","tokens_in":973,"tokens_out":1378,"duration_ms":117843,"concrete_test":"Run the heterogeneous forward benchmark of Section 3.3 with the discontinuous layered JIVSM structure instead of the smooth modulus model, and compare the PINN solution against a PyLith finite-element reference. If the displacement error is concentrated near layer boundaries, this confirms that the strong-form loss is missing traction-continuity conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 incorporates JIVSM by assigning discontinuous elastic moduli (Table 6) to collocation points, with no loss term enforcing traction continuity across layer boundaries. The strong-form equilibrium equation (1) and Hooke's law (2) are only valid inside homogeneous layers; at material interfaces, traction must be continuous, but a single smooth tanh network for displacement cannot represent the discontinuous strain gradient, and no interface condition appears in the inversion loss (17). The reported loss for Het-Top is 9.73e-5, two orders of magnitude above Hom-Top (7.98e-7), and the authors attribute this to difficulty in optimization. This is not merely an optimization issue: it indicates the PDE system is not well posed as implemented for discontinuous mu. Consequently, the estimated slip for Het-Top (Figure 12c) is not a reliable solution of the stated physics, so the central claim of recovering slip with realistic crustal structure is not supported. The homogeneous forward and synthetic inversions are credible, but the heterogeneous-structure part requires explicit interface conditions or validation against a reference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a physics-informed neural network (PINN) approach for three-dimensional coseismic crustal deformation. The model uses four neural networks to represent displacement and stress in two subdomains separated by a fault surface and its extension, with loss terms for the equilibrium equations, Hooke's law, fault conditions, free-surface and contact-surface conditions, and outer-surface conditions. The authors validate forward simulations against Okada half-space solutions and PyLith finite-element solutions, examine the effect of constraining rigid motion, and then perform slip inversions using surface displacement data, first on synthetic data and then on GNSS data from the 2008 Iwate-Miyagi inland earthquake. They also incorporate realistic topography and the JIVSM layered velocity structure into the inversion. The main claims are that the method can model 3-D crustal deformation, that inversion recovers a slip distribution consistent with previous studies, and that realistic crustal structures can be incorporated into the analysis.","tokens_in":16422,"tokens_out":6492,"duration_ms":95085,"significance":"If the central claims are established, this would be a useful contribution to the growing literature on applying PINNs to geophysical inverse problems. The homogeneous forward results are quantitatively benchmarked against independent Okada and PyLith references, with relative surface-displacement errors below 5% in the unconstrained forward problem and below 1% when rigid motion is supervised. The synthetic inversions are also assessed against known slip distributions, including noisy and sparse data cases. The real-data application to the 2008 Iwate-Miyagi earthquake is a meaningful first demonstration of a PINN-based 3-D coseismic slip inversion using actual GNSS observations, and the authors are transparent about the underestimation of seismic moment and about the role of soft physics constraints. These strengths make the core of the paper worth considering. However, as detailed in the major comments, the paper's broader claim about incorporating realistic crustal structures is not supported by the heterogeneous-layer experiment, which lacks the necessary interface conditions and shows a two-order-of-magnitude loss increase.","major_comments":[{"comment":"The Het-Top inversion is not a valid solution of the stated physical problem. Equations (1)-(2) hold only inside homogeneous layers; at the JIVSM layer boundaries listed in Table 6, the correct conditions are continuity of displacement and continuity of normal traction. The inversion loss (17) contains no interface loss terms for those layer boundaries, and Lcon in Eq. (10) evaluates a single smooth tanh network's derivatives at points where layer-dependent mu and lambda are assigned. A C-infinity network cannot represent the discontinuous strain gradient required at a sharp modulus contrast, so the minimizer of (17) is not constrained by a well-posed transmission problem. The numerical signature is the reported loss: 9.73e-5 for Het-Top versus 7.98e-7 for Hom-Top. Attributing this gap to 'difficulty in optimization' is not sufficient; it is the expected consequence of missing interface conditions. Section 3.3 does not cover this case because Eq. (27) defines a smooth mu. Consequently, Figure 12c and the abstract/conclusion claim of recovering slip with realistic crustal structure are not supported. The fix is to add explicit displacement and traction-continuity losses on the fitted layer surfaces (or use a layered domain decomposition) and validate against a layered PyLith or Okada reference.","section":"Section 4.3, Eqs. (10), (17), Table 6"},{"comment":"The validation of heterogeneous structure in Section 3.3 is limited to smooth heterogeneity. The elastic modulus in Eq. (27) is a smooth logistic-plus-Gaussian function, so the PyLith comparison in Figure 8 tests only a smoothly varying medium. It does not test the discontinuous layered medium used in the Het-Top inversion of Section 4.3. The claim that the method is 'adaptable to heterogeneous structures' should be restricted to smooth heterogeneity unless the layered case is separately validated.","section":"Section 3.3"}],"minor_comments":[{"comment":"The contact-surface loss Lcs appears to contain six identical terms of the form (sigma_xx^+ - sigma_xx^-)^2. This is presumably a typesetting error: the intended expression should include the squared differences of sigma_yy, sigma_zz, sigma_xy, sigma_xz, and sigma_yz. Please correct the equation.","section":"Eq. (14)"},{"comment":"The text states that 'the loss function is modified from equation 1 as follows' before Eq. (24). It should refer to equation (8), the forward loss function, rather than equation (1).","section":"Section 3.2"},{"comment":"The inversion result for the real earthquake is reported as the average over five random seeds, and the loss value is given, but no seed-to-seed variability or uncertainty estimate for the slip distribution is reported. A brief statement of the spread across seeds would help the reader assess robustness.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The homogeneous and smooth-heterogeneous forward results and the synthetic inversions are credible and represent the main strength of the paper. The real-data application is useful and mostly well executed. The key problem is that the Het-Top experiment is presented as demonstrating realistic crustal-structure modeling, but the missing interface conditions make the result unreliable as a solution of the stated physics. This is fixable within the scope of the paper by adding layer-interface losses and validating against a layered reference solution. I would support publication after that revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is the first 3-D PINN formulation for coseismic deformation, with a four-network displacement–stress representation on two subdomains and the first application to real GNSS data (2008 Iwate–Miyagi). The forward validation is genuinely solid: against Okada solutions and PyLith, surface displacement errors drop below 1% once rigid motion is supervised. Synthetic slip inversions are also credible, with errors mostly under 6% and stable seismic moment even for sparse/noisy data. The real-data slip distribution is consistent with earlier studies. That part deserves a careful read.\n\nThe soft spots are real but concentrated. The biggest one is the Het–Top result in Section 4.3. The layered JIVSM structure has discontinuous elastic moduli, but the strong-form loss contains no interface condition enforcing traction continuity across layer boundaries. A continuous tanh network cannot represent the required stress jumps, and the loss for Het–Top is two orders of magnitude above Hom–Top. The authors call this an optimization difficulty; it is actually a missing physical constraint. So the central claim that the method handles realistic crustal structure is not established. The forward heterogeneous test (Section 3.3) uses smooth moduli, so it does not validate the layered case. This needs either explicit interface terms or validation against a FEM reference for the same JIVSM layers.\n\nOther issues are minor by comparison. The real-data inversion has no same-data baseline against Okada or FEM, so readers cannot judge whether the PINN adds value for source estimation. There is no uncertainty quantification on the inferred slip. Equations 9 and 14 contain typos that obscure the loss definition—eq. 9 has sigma_xz,x twice in the last term, and eq. 14 repeats sigma_xx three times. The artificial outer boundary at 100x fault depth is well analyzed in the paper; the rigid-motion discussion is honest and the inversion loss sensibly drops the OS term. Self-citation is heavy but the prior 2-D papers are the direct predecessors, so that is appropriate.\n\nWho is this for? People working on PINNs for geophysical inverse problems, and seismologists who want a mesh-free alternative to FEM-based inversion. The homogeneous and smooth-heterogeneous results are worth taking seriously; the layered-structure claim needs more work before it is used.\n\nRecommendation: send to peer review. The paper is inventive, mostly careful, and the forward machinery is reproducible. A serious referee should focus on the interface condition and the lack of a baseline in the real-data inversion. With revisions, this could be a useful reference.","headline":"First credible 3-D PINN treatment of coseismic deformation with real GNSS data, but the heterogeneous-structure claim is not yet supported because the loss omits interface conditions.","tokens_in":17030,"tokens_out":1558,"would_cite":true,"duration_ms":22111,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["86A17","86A22","68T07","74B05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Physics-informed neural networks with four subdomain networks solve and invert 3-D crustal deformation, recovering the 2008 Iwate–Miyagi fault slip from surface GNSS data.","keywords":["physics-informed neural networks","crustal deformation","earthquake fault slip inversion","dislocation model","semi-infinite domain","2008 Iwate-Miyagi earthquake","GNSS geodesy","elastic equilibrium equations"],"falsifier":"Run the synthetic inversion with the same 51-station geometry and noise level as the real GNSS data: if the PINN systematically returns a moment magnitude more than 0.1 below the truth while an independent conventional slip inversion on identical synthetic data returns the truth, the underestimation is a property of the PINN method; if both are low, the discrepancy is a data-resolution effect.","tokens_in":16013,"feed_emoji":"🌏","tokens_out":10561,"duration_ms":109012,"temperature":0.7,"pith_summary":"This paper tries to establish that physics-informed neural networks (PINNs)—neural networks trained to satisfy the governing elastic equations and boundary conditions—can model three-dimensional crustal deformation from earthquake faults in a semi-infinite elastic Earth, both as forward solvers and as fault-slip inversions. It argues that the right architecture is four networks that represent displacement and stress on the two sides of the fault surface, because this displacement–stress form is more accurate and efficient than a displacement-only network. The payoff is flexibility: faults of arbitrary geometry, heterogeneous structure, and real satellite-positioning observations can be handled by one optimizer without building a mesh. The headline demonstration is a slip inversion of the 2008 Iwate–Miyagi inland earthquake whose peak slip agrees with earlier GNSS and strong-motion studies, although its moment magnitude comes out lower.","feed_headline":"AI recovers 3-D fault slip from quake surface data","feed_subtitle":"Physics-informed neural networks match prior slip for the 2008 Iwate-Miyagi quake, though they put its magnitude slightly low.","key_machinery":"The load-bearing object is a displacement–stress PINN built from four fully connected networks: one for displacement and one for stress in each of two subdomains divided by the fault plane and its extension. Treating displacement and stress as independent outputs keeps the governing system first-order and makes the boundary conditions algebraic, at the cost of satisfying Hooke's law only approximately through a dedicated loss term. The loss function is a sum of residuals for equilibrium, the constitutive law, the displacement discontinuity and traction continuity on the fault, the free-surface condition, continuity across the artificial contact surface, and, in forward runs, zero displacement and stress on the outer surface; in inversion mode the prescribed-slip term is replaced by a shear-only normal-continuity constraint plus a data-misfit term at observation sites.","core_discovery":"Working with the linear-elastic equilibrium equations and representing the fault as a displacement discontinuity, the paper shows that four fully connected networks (displacement and stress on each side of a fault-bounded split of the domain) can solve the 3-D forward problem in a homogeneous half-space, in a heterogeneous structure with topography, and—after removing the prescribed-slip loss and adding a surface data-misfit term—can invert surface displacements for slip. In the homogeneous half-space, forward surface-displacement errors are about 2–3% near the fault, and once rigid-body motion is anchored by a few known displacement values the errors fall below 1%. The paper itself documents the central caveat: the elastic equations determine internal deformation but not rigid translation or rotation, so forward solutions inherit offsets from the finite outer boundary, whereas in inversion the observed displacements resolve the ambiguity. Applied to satellite-positioning (GNSS) data from the 2008 Iwate–Miyagi inland earthquake, the inversion gives a reverse-dominated slip with a peak of 4.50 m, consistent with the 3.5–6.2 m range from previous studies, but a moment magnitude of about 6.73, below the previously reported 6.9. The paper concludes that PINNs are capable of 3-D crustal deformation analysis and can incorporate realistic layered velocity structure, making large-scale modeling with real observations feasible.","pith_inferences":["If rigid motion is fixed only by observed surface displacements, any reference-frame error or common-mode bias shared by all stations would be absorbed into the slip estimate; this follows from the paper's rigid-motion finding but is not quantified there.","The same four-network construction should extend to multi-segment or nonplanar faults and to time-dependent deformation, because the domain split follows the fault geometry and the loss is boundary-driven; that extension is not tested in the paper.","The stability of total moment alongside unstable peak slip suggests aggregate quantities are the more trustworthy output of PINN slip inversions; a checkerboard-resolution study on the same station geometry could make this explicit.","The larger loss value for the layered heterogeneous model hints that sharp material boundaries are hard for smooth activation functions, pointing toward piecewise parameterizations or explicit interface constraints as a testable improvement."],"forward_implications":["The displacement–stress PINN solves 3-D static elastic problems with finite faults in a semi-infinite domain without meshing, with accuracy in the homogeneous half-space adequate for modeling internal deformation.","Static equilibrium equations alone leave rigid translation and rotation unconstrained, so any forward model needs anchors—either observed displacements or an exact far-field condition—to remove the resulting offsets.","In slip-inversion mode, sparse and noisy surface data can halve the inferred peak slip, but the total seismic moment is still recovered within about 8%.","For the 2008 Iwate–Miyagi earthquake, the PINN inversion yields a reverse-slip pattern with a peak of 4.50 m in a half-space model, matching the 3.5–6.2 m range reported earlier; the estimated moment magnitude of 6.73 is lower than earlier values around 6.9.","Adding realistic topography and a layered velocity structure changes the inferred slip: topography narrows and sharpens the peak, while the heterogeneous model lowers both peak slip and moment magnitude."],"supporting_citations":[{"why":"Supplies the two-subdomain, fault-split PINN formulation and inversion loss design that this paper extends from 2-D inplane to 3-D problems.","marker":"[26]"},{"why":"Earlier PINN crustal-deformation modeling in an antiplane setting that this work generalizes to 3-D.","marker":"[24]"},{"why":"Semi-infinite-domain elastic-wave PINN that motivates modeling displacement and stress with separate networks.","marker":"[18]"},{"why":"Provides the semi-analytical half-space dislocation solutions used as ground truth in forward tests and as the data generator for synthetic inversions.","marker":"[31]"},{"why":"Provides the finite-element reference solution used to validate the heterogeneous-structure forward test.","marker":"[32]"},{"why":"Supplies the initial fault-plane geometry and a previous slip model of the 2008 earthquake used for comparison.","marker":"[34]"},{"why":"A dense-GPS coseismic fault model of the same earthquake used as a comparison for slip and moment.","marker":"[35]"},{"why":"A strong-motion rupture model of the same earthquake used as a comparison for slip and moment and noted for its less smooth slip pattern.","marker":"[36]"},{"why":"Supplies the layered velocity structure and surface topography incorporated into the realistic inversion.","marker":"[37]"}],"fun_headline_variants":["PINN solves 3D quake deformation, inverts for slip","Neural networks invert quake data for 3D fault slip","AI pinpoints 3D quake slip from surface data, magnitude low","Physics-informed nets solve 3D crustal deformation","Fault slip from 2008 quake recovered by AI, but magnitude low"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an artificial outer boundary at 100 times the fault depth, with zero displacement and stress, adequately stands in for the semi-infinite Earth; the paper's own forward tests show rigid-body offsets there, and in the real-data inversion that term is removed, so the observed surface data alone must fix the rigid motion.","fun_headline_variants_meta":{"raw":{"variants":["PINN solves 3D quake deformation, inverts for slip","Neural networks invert quake data for 3D fault slip","AI pinpoints 3D quake slip from surface data, magnitude low","Physics-informed nets solve 3D crustal deformation","Fault slip from 2008 quake recovered by AI, but magnitude low"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001349,"raw_usage":{"total_tokens":5522,"prompt_tokens":1031,"completion_tokens":4491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":4397}},"tokens_in":647,"tokens_out":4491,"duration_ms":34348,"temperature":1.0,"reasoning_tokens":4397,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:33:36.623562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the synthetic inversion with the same 51-station geometry and noise level as the real GNSS data: if the PINN systematically returns a moment magnitude more than 0.1 below the truth while an independent conventional slip inversion on identical synthetic data returns the truth, the underestimation is a property of the PINN method; if both are low, the discrepancy is a data-resolution effect.","supporting_citations":[{"cited_title":"X., Sun, H., & Liu, Y","cited_arxiv_id":null,"evidence_quote":"Semi-infinite-domain elastic-wave PINN that motivates modeling displacement and stress with separate networks."},{"cited_title":"(2008), Crustal deformation and seismic fault model of the Iwate -Miyagi nairiku earthquake in 2008, GSI Journal, 117, 79 –80 (in Japanese)","cited_arxiv_id":null,"evidence_quote":"Supplies the initial fault-plane geometry and a previous slip model of the 2008 earthquake used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A dense-GPS coseismic fault model of the same earthquake used as a comparison for slip and moment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A strong-motion rupture model of the same earthquake used as a comparison for slip and moment and noted for its less smooth slip pattern."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the layered velocity structure and surface topography incorporated into the realistic inversion."}],"review_version":1}