{"id":"cdef742c-992d-4025-9c39-dd25ec980aa0","arxiv_id":"2412.00031","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TomoATT is an open-source package that implements adjoint-state traveltime tomography with Eikonal forward modeling, kernel regularization, and multi-level parallelization, validated on synthetic and real seismic datasets.","lead":"TomoATT is a new open-source software package that inverts seismic traveltimes to image Earth's interior velocity and azimuthal anisotropy. It packages Eikonal-based adjoint-state tomography with parallel computing so that regional and teleseismic inversions can run in under two hours on 64 processors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic verification is an inverse-crime test using the same solver and grid to generate and invert noise-free data; real-data benchmarks lack resolution tests, so 'effectiveness' is not yet established.","rationale":"The reader's weakest assumption concerns Eikonal-solver accuracy on coarse grids in strongly heterogeneous media and possible bias from deferred data-selection criteria. My concern overlaps but is more specific: the synthetic recovery, which the abstract explicitly cites as verification, is an inverse-crime test because the same solver and grid generate and invert the data. This makes the synthetic pillar of the effectiveness claim circular, independent of the forward-solver accuracy question. The real-data benchmarks, which are the main external evidence, are also missing resolution and uncertainty analyses, so the qualitative consistency with prior studies does not by itself establish that the package images real structures reliably. I agree with the CONDITIONAL verdict: the open-source code, forward-model checks against TauP, and consistency with previous geophysical results are real assets, but the verification claim is not yet fully supported. Adding an independent-solver synthetic test or formal resolution/uncertainty analysis would materially strengthen the paper. No reason to reject outright, and the concerns do not invalidate the software contribution; they only limit the strength of the 'effectiveness' claim as currently written.","tokens_in":18951,"tokens_out":7055,"duration_ms":69489,"concrete_test":"Repeat the Section 4.1 synthetic experiment with observational traveltimes generated by an independent forward solver, e.g., a high-order ray tracer or a wave-equation solver on a much finer grid, and with realistic picking noise (0.05-0.1 s) and irregular station/event geometry drawn from the real deployments. If the recovered velocity and anisotropy perturbations deviate from the target by more than the anomaly amplitudes, or show artifacts comparable to those in the real-data images, the 'successful recovery' is an inverse-crime artifact and the effectiveness claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central verification rests on 'the successful recovery of the synthetic model' (abstract; Section 4.1). In that experiment, 'observational traveltime data are generated using the true model and true earthquake locations' with the same Eikonal forward solver and the same discrete grid used for the inversion, and no noise is added. This is an inverse-crime setup: the inversion is asked to recover a model that lies exactly in the same parameterized solution space that generated the data, so the synthetic recovery tests internal consistency of the code but not its physical accuracy or robustness under realistic errors. The forward-model validations in Section 3.1 use a smooth linear velocity model and a 1-D AK135 model; they do not bound the error on the strongly heterogeneous 5-km/10-km inversion grids of the real cases. Consequently, the only controlled numerical evidence that TomoATT images real Earth structure is the qualitative agreement of the Parkfield and Thailand results with previous studies. Those images are presented without formal resolution or uncertainty analysis, and the data-selection criteria are deferred to Chen et al. (2023), so it is not demonstrated that the imaged anomalies are required by the data rather than imposed by regularization, kernel-density normalization (Eq. 18), or selection choices. If the synthetic recovery is circular and the real-data comparisons are not backed by resolution tests, the claim 'verifies the effectiveness of TomoATT' is stronger than the evidence supports.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents TomoATT, an open-source C++/Python package for Eikonal-equation-based adjoint-state traveltime tomography for seismic velocity and azimuthal anisotropy. It reviews the mathematical formulation (objective function with traveltime and differential arrival times, anisotropic Eikonal equation, adjoint-state kernels, multiple-grid parameterization, kernel density normalization), describes the multi-level parallelization architecture, and demonstrates the package through toy experiments, a synthetic inversion with earthquake relocation, and two real-data benchmarks (Parkfield, California; Thailand). The central claim is that the package is efficient and user-friendly, with inversions starting from three input files and completing within about two hours on 64 processors, and that synthetic recovery plus consistency with previous studies verifies the package's effectiveness.","tokens_in":19160,"tokens_out":4274,"duration_ms":36564,"significance":"If the central claims are substantiated, TomoATT would be a valuable open-source contribution: it provides a complete workflow for Eikonal-based traveltime tomography, including azimuthal anisotropy, with validated forward modeling, multipathing-aware kernels, and multi-level parallelism. The paper's strengths include the open-source availability under GPLv3, the companion PyTomoATT module, forward-model validation against analytic solutions and TauP (Sections 3.1-3.2), and the reproduction of several independently published structural features in the real-data images. However, as argued in the major comments, the verification of the inversion's physical accuracy is not yet complete, and several reproducibility-critical parameters are not reported.","major_comments":[{"comment":"The synthetic verification is an 'inverse-crime' test. Observational data are generated by the same Eikonal solver on the same grid as used in the inversion, with no added noise. This verifies that the inversion code can invert its own forward operator, but it does not verify robustness to modeling error, data noise, or discretization mismatch. The abstract's claim that 'successful recovery of the synthetic model ... verifies the effectiveness of TomoATT' is therefore too strong. I recommend adding a synthetic test where data are generated on a different (finer) grid or with a different forward solver, or with realistic noise and outlier contamination, and showing recovery statistics.","section":"Section 4.1"},{"comment":"The real-data benchmarks lack resolution tests and uncertainty quantification. No checkerboard or spike recovery tests are shown for the Parkfield or Thailand geometries, and the images are presented without any measure of reliability such as model covariance or depth-resolution surfaces. Because the data selection criteria are deferred to Chen et al. (2023) and the regularization choices are not fully reported (see next comment), the reader cannot determine whether the imaged anomalies are required by the data. I recommend adding standard resolution tests (e.g., checkered/spike restorations with the actual station-event geometry) or at least discussing sensitivity to regularization.","section":"Sections 4.2-4.3"},{"comment":"The values of the objective-function weights alpha, beta, gamma; the kernel-density exponent zeta in Eq. (18); the multiple-grid configuration H; and the gradient-descent step size are not reported for the two real-data inversions. These parameters have a direct effect on the images (as shown for zeta in Section 3.3), and without them the claimed benchmarks cannot be reproduced, which is a central issue for an open-source package. Please provide the exact parameter files (or a table listing all parameters) for both case studies.","section":"Sections 4.2-4.3 (reproducibility)"},{"comment":"The consistency of the real-data results with 'previous studies' partly refers to the authors' own earlier papers (Chen et al. 2023a,b), since the text states that the two cases were 'detailed in our previous studies.' The comparisons to independent references (e.g., Eberhart-Phillips and Michael, 1993; Thurber et al., 2006; Yang et al., 2015) are more persuasive, but the discussion should clearly separate those independent comparisons from self-comparisons, and ideally add quantitative misfit or correlation metrics rather than qualitative agreement.","section":"Section 4"}],"minor_comments":[{"comment":"'syntehtic' should be 'synthetic'.","section":"Figure 7 caption"},{"comment":"'relative the number of processors' should be 'relative to the number of processors'.","section":"Figure 6 caption"},{"comment":"'with a standard derivation of 0.1 s' should be 'standard deviation'.","section":"Section 3.3"},{"comment":"Equation numbering is duplicated: two separate equations are labeled (4), the second being the anisotropic Eikonal equation; renumber the equations.","section":"Section 2.1"},{"comment":"'teleiseismic' is misspelled; it should be 'teleseismic'.","section":"Keywords"},{"comment":"'Figure 8. Figure 8.' is duplicated; remove the repeated text.","section":"Figure 8 caption"},{"comment":"Several equations (B5)-(B11) appear typeset with a heavy or bold font that is inconsistent with the main text; please ensure uniform notation.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The comparison to the authors' own prior results is a mild circularity but not disqualifying, since independent references are also cited. The main risk is that the inverse-crime synthetic test is presented as 'verification of effectiveness,' which could mislead readers. The paper is appropriate for a software/methods journal, but the missing parameter reporting and the lack of resolution tests are substantive and should be addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TomoATT is a real contribution, but it is a contribution of integration and packaging, not new theory. The methods (adjoint-state traveltime tomography, Eikonal solver, multiple-grid parameterization, kernel density normalization) all come from the authors' earlier papers. What is new here is a well-engineered open-source C++ package with a Python companion, multi-level MPI parallelization, and a clean three-file input interface. That is a legitimate and useful thing to publish, and the paper does it honestly: the code is GPL-licensed, the documentation is linked, and the run times on real problems are reported.\n\nThe forward modeling checks are the strongest part. Errors against an analytic linear model converge at first order, and the teleseismic differential times agree with TauP to 0.03 s standard deviation. The toy experiment in Section 3.3 includes noise and shows the multiple-grid and kernel density regularization actually do something. The parallelization benchmark is straightforward and useful for people choosing job sizes.\n\nThe soft spots are real but not disqualifying. The synthetic case study in Section 4.1 generates noise-free data with the same Eikonal solver and the same grid used in the inversion, so it is an inverse-crime test. It verifies that the code is internally consistent and that the relocation loop works, but it does not by itself demonstrate robustness to real data errors. The real-data images for Parkfield and Thailand reproduce structures seen in independent studies, which is meaningful, but neither case includes formal resolution tests, checkerboard recovery on the actual acquisition geometry, or uncertainty estimates. The 'strict data selection criteria' are deferred to prior papers, so a reader cannot fully audit the input data. And the kernel density normalization exponent zeta is a free parameter with a known noise-amplification trade-off, which the paper acknowledges but does not quantify for the real cases.\n\nNone of this undermines the package's usefulness. The paper is exactly what a software paper should be: it tells the community what the code does, shows the components work, and points to the code. The verification could be stronger, but for a package whose methods have already been peer-reviewed, this is enough to justify publication. I would send it to review with a request for a resolution/uncertainty discussion in the real-data sections and a note in Section 4.1 clarifying that the synthetic test is a code-consistency check, not a physical validation.\n\nWorth citing if you work in traveltime tomography, and worth bringing to a reading group that cares about reproducible seismic imaging.","headline":"A well-engineered open-source tomography package; the synthetic validation is internally consistent but thin, and the real-data sections need resolution tests before 'effectiveness' is fully earned.","tokens_in":19740,"tokens_out":2495,"would_cite":true,"duration_ms":23839,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TomoATT packages Eikonal-based adjoint-state tomography for velocity and azimuthal anisotropy into an open-source, three-file workflow that runs regional and teleseismic inversions in under two hours on 64 processors.","keywords":["TomoATT","adjoint-state traveltime tomography","Eikonal equation","azimuthal anisotropy","seismic tomography","fast sweeping method","teleseismic tomography","regional tomography"],"falsifier":"A controlled synthetic inversion on the same grid spacings as the real-data cases (for example, $5$ km $\\times$ $5$ km $\\times$ $1$ km) using a strongly heterogeneous model with sharp velocity contrasts and realistic noise levels would settle the accuracy question: if reconstructed anomalies are smeared or displaced beyond the grid scale, or if traveltime residuals exceed the forward-model errors reported in Section 3.1, the claimed effectiveness for real Earth would be undercut.","tokens_in":18690,"feed_emoji":"🌍","tokens_out":6218,"duration_ms":54487,"temperature":0.7,"pith_summary":"The paper presents TomoATT, an open-source package that performs regional and teleseismic traveltime tomography by solving the anisotropic Eikonal equation for synthetic traveltimes and using the adjoint-state method to compute volumetric sensitivity kernels. The central claim is that this implementation makes adjoint-state tomography practical: an inversion starts from three input files, runs to completion in under two hours on 64 processors, and recovers both isotropic velocity structure and azimuthal anisotropy. The claim is supported by toy experiments on forward modeling, multipathing, regularization, and parallelization, plus a synthetic inversion and two real-data benchmarks in central California near Parkfield and in Thailand. The paper argues that agreement of the real-data images with regional tectonics and earlier studies verifies the package's effectiveness.","feed_headline":"Open-source package images crust and anisotropy in under 2 hours","feed_subtitle":"Eikonal-based adjoint-state solver tests on Parkfield and Thailand, mapping velocity and azimuthal anisotropy from traveltimes.","key_machinery":"The central object is the adjoint-state traveltime tomography workflow, built on the anisotropic Eikonal equation in spherical coordinates, $\\nabla T^t M(\\mathbf{x};\\xi,\\eta)\\nabla T = s^2(\\mathbf{x})$, solved with the fast sweeping method and multiplicative factorization. The adjoint equation $\\nabla\\cdot(P(\\mathbf{x}) M(-\\nabla T)) = \\sum_m R_m \\delta(\\mathbf{x}-\\mathbf{x}_m)$ turns receiver misfits into volumetric sensitivity kernels $K_s, K_\\xi, K_\\eta$, which are then projected onto coarse multiple-grid basis functions and normalized by kernel density. This combination is what lets the package handle multipathing, avoid ray-path undersampling, and regularize uneven data coverage.","core_discovery":"TomoATT's core claim is that Eikonal equation-based adjoint-state traveltime tomography can be packaged as an efficient, user-friendly tool for routine seismic imaging. It solves for slowness $s(\\mathbf{x})$ and azimuthal anisotropy parameters $\\xi(\\mathbf{x})$, $\\eta(\\mathbf{x})$ by minimizing a weighted sum of traveltime, common-source differential, and common-receiver differential misfits. Each iteration solves the anisotropic Eikonal equation for every source (or, via reciprocity, every receiver), solves the adjoint equation to obtain gradient kernels, and updates model parameters by step-size-controlled gradient descent with multiple-grid parameterization and kernel density normalization. The synthetic experiment recovers staggered velocity and anisotropy perturbations, while the two real-data inversions reproduce features such as the velocity contrast across the San Andreas Fault and the high-velocity Khorat Plateau lithosphere. From these results the paper concludes that TomoATT is a verified, efficient tool for both regional and teleseismic tomography.","pith_inferences":["If the claimed two-hour turnaround holds on typical HPC nodes, adjoint-state Eikonal tomography could become a routine replacement for ray-based tomography in regional networks, where kernel accuracy and multipathing have been the main barriers.","The kernel density normalization trades noise amplification against anomaly recovery; the paper shows that high $\\zeta$ values can create artifacts in poorly constrained regions, so an adaptive or data-driven choice of $\\zeta$ would be a natural extension.","The solver validation suggests that users on coarser grids or in very strong-contrast media should first run the package's traveltime-error tests on their own models before trusting recovered anomalies.","Because the package is open source and supports YAML, HDF5, and TEXT formats, it could serve as a building block for joint inversions or full-waveform approaches that need a fast, trustworthy traveltime prior."],"forward_implications":["Regional and teleseismic inversions can be set up from three files, so the package removes much of the scripting burden in traveltime tomography.","The source-receiver reciprocity trick reduces the Parkfield inversion to 607 Eikonal solves per iteration instead of 32,721, making 80 iterations feasible in 57 minutes.","Multiple-grid parameterization and kernel density normalization let the inversion recover anomalies in sparsely sampled regions while suppressing parameter-count artifacts.","The same machinery relocates hypocenters, so simultaneous inversion can avoid the artifacts shown when source locations are inaccurate.","Users can image azimuthal anisotropy as well as isotropic velocity from first P arrivals, enabling stress and fabric interpretation near faults."],"supporting_citations":[{"why":"Defines the adjoint-state traveltime tomography method for azimuthally anisotropic media in spherical coordinates that TomoATT implements.","marker":"(J. Chen, G. Chen, et al., 2023)"},{"why":"Presents the teleseismic adjoint-state tomography method and the Thailand application that TomoATT benchmarks.","marker":"(J. Chen, S. Wu, et al., 2023)"},{"why":"Establishes the adjoint-state traveltime tomography formulation for central California near Parkfield used as a regional benchmark.","marker":"(Tong, 2021a)"},{"why":"Provides the fast sweeping method, the Eikonal solver whose first-order accuracy and O(N) complexity the package relies on.","marker":"(Zhao, 2005)"},{"why":"Supplies the multiplicative factorization technique that removes source singularity in the Eikonal solver.","marker":"(Luo & Qian, 2012)"},{"why":"Defines the AK135 reference model used to validate teleseismic differential traveltime accuracy.","marker":"(Kennett et al., 1995)"},{"why":"Provides the TauP software against which the Eikonal differential traveltimes are compared.","marker":"(Crotwell et al., 1999)"},{"why":"Introduces multiple-grid parameterization, the regularization that reduces artifacts from excessive inversion variables.","marker":"(Tong et al., 2019)"},{"why":"Derives the adjoint-state differential arrival time tomography and adjoint source forms used in the objective function.","marker":"(Tong et al., 2023)"}],"fun_headline_variants":["TomoATT: open-source Eikonal tomography for velocity and anisotropy","Efficient adjoint-state traveltime tomography, now open-source","Under 2 hours: open-source tomography with anisotropy","Eikonal adjoint tomography for crust and anisotropy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fast sweeping Eikonal solver, validated only on simple synthetic models and the AK135 teleseismic test, remains accurate enough on the moderately coarse inversion grids used for the real-data benchmarks to produce trustworthy sensitivity kernels in strongly heterogeneous crust and mantle.","fun_headline_variants_meta":{"raw":{"variants":["TomoATT: open-source Eikonal tomography for velocity and anisotropy","Efficient adjoint-state traveltime tomography, now open-source","Under 2 hours: open-source tomography with anisotropy","Eikonal adjoint tomography for crust and anisotropy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000707,"raw_usage":{"total_tokens":3188,"prompt_tokens":949,"completion_tokens":2239,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2170}},"tokens_in":565,"tokens_out":2239,"duration_ms":15638,"temperature":1.0,"reasoning_tokens":2170,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:51:01.555786+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled synthetic inversion on the same grid spacings as the real-data cases (for example, $5$ km $\\times$ $5$ km $\\times$ $1$ km) using a strongly heterogeneous model with sharp velocity contrasts and realistic noise levels would settle the accuracy question: if reconstructed anomalies are smeared or displaced beyond the grid scale, or if traveltime residuals exceed the forward-model errors reported in Section 3.1, the claimed effectiveness for real Earth would be undercut.","supporting_citations":[],"review_version":1}