Pith. sign in

REVIEW 3 major objections 7 minor 45 references

A physics-encoded residual network transfers across unseen seismic acquisitions and a label-free audit restores trustworthy uncertainty under real distribution shift.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 04:21 UTC pith:3QNB4T5O

load-bearing objection Solid practice-facing FWI+ML stack: geometry transfer via physics encoding is real; the label-free audit helps on long lines but is a regime-limited heuristic, not a coverage guarantee. the 3 major comments →

arxiv 2607.28535 v1 pith:3QNB4T5O submitted 2026-07-30 physics.geo-ph

Reliability-calibrated deep residual full-waveform inversion using geometry-invariant physics encoding: synthetic validation and zero-shot Marmousi-2 testing

Deepak Kumar , Jayant Nath Tripathi , Laxmidhar Behera This is my paper
classification physics.geo-ph
keywords full-waveform inversiondeep learninguncertainty quantificationconformal predictionacquisition geometryMarmousiphysics encodingresidual correction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Networks that turn raw seismic shot gathers into velocity models usually memorize the survey layout they saw in training, and their uncertainty estimates fail when the data leave that distribution. This paper argues that both failures can be fixed without replacing classical full-waveform inversion. Variable acquisitions are first mapped by classical physics operators into a fixed ten-channel model-space tensor—starting model, regularized inversion, misfit gradients, and illumination maps—so the network never sees gathers and never encodes geometry in its weights. A small heterogeneous ensemble then learns only a residual correction to that physics prior and is calibrated by physics-conditioned conformal prediction. When the target leaves the calibration distribution, a held-out-shot physics audit rescales the intervals by simulating unused shots through the predictive samples; no ground truth is required. On a thousand synthetic models the method cuts error 38 percent relative to its classical prior and shows no measurable gap on never-seen geometries; zero-shot on the full 17 km Marmousi-2 line it lowers error from 354 to 304 m/s and the audit restores near-nominal coverage across noise, wavelet error, and shot decimation while its peak height flags physics mismatch.

Core claim

A geometry-invariant physics encoding plus a residual ensemble with physics-conditioned conformal calibration transfers across unseen acquisition families without measurable degradation, and a label-free held-out-shot physics audit restores near-nominal pixel-wise coverage (about 0.89–0.91) under domain shift on full-line Marmousi-2 while the peak of data-space coverage cleanly separates physics mismatch from benign corruptions.

What carries the argument

Geometry-invariant physics encoding (GIPE): a fixed ten-channel model-space tensor built only from classical operators (starting model, ADMM-regularized FWI, two misfit gradients, six fast-marching illumination/wavenumber maps). The network is only a calibrated residual corrector on this prior; a held-out-shot physics audit then rescales interval width where simulated data-space coverage peaks, without ground truth.

Load-bearing premise

The audit works only when held-out-shot data-space coverage has a clear peak that correctly marks how much to widen the model intervals—something that requires long enough propagation paths and model error that sits above the noise floor.

What would settle it

On a real marine line with well logs (for example Viking Graben), check whether the audited intervals achieve near-nominal coverage at the boreholes after zero-shot application, and whether the audit’s peak height still drops under known wavelet or elastic mismatch while staying high under pure noise or shot decimation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Acquisition geometry no longer has to be fixed or explicitly conditioned for a learned FWI corrector to transfer.
  • Uncertainty statements from synthetic conformal calibration can be transported to field data without labels by using the wave equation on held-out shots.
  • Physics mismatch (wrong wavelet, acoustic-vs-elastic) becomes detectable from the depressed ceiling of data-space coverage rather than from ground-truth error.
  • Budget-matched gather-to-model networks, even with source conditioning, are sample-inefficient relative to the physics-encoded residual route at laptop scale.
  • The same encode–correct–audit pattern is dimension-agnostic and can be tried in 3-D once classical-chain cost is managed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Any inverse problem that already owns a reliable forward operator and can reserve a few observations could adopt the same held-out physics audit instead of relying on network uncertainty alone.
  • Iterating the encode–correct cycle—feeding the corrected model back as a new prior—could lift the residual ceiling in deep, poorly illuminated zones where a single pass saturates.
  • The short-offset plateau of the audit curve supplies a practical pre-check: if paths are shorter than roughly a few kilometres at exploration frequencies, expect only conservative over-coverage rather than tight recalibration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript proposes a laptop-scale residual FWI pipeline that never feeds shot gathers to the network. Variable acquisitions are mapped into a fixed ten-channel model-space tensor (starting model, ADMM prior, two misfit gradients, six fast-marching illumination/wavenumber maps); a six-member heterogeneous ensemble then predicts a calibrated residual on that prior. Physics-conditioned (Mondrian) conformal prediction supplies finite-sample pixel-wise marginal coverage on the calibration distribution, and a label-free held-out-shot physics audit rescales interval width under domain shift by maximizing data-space coverage C(τ) of unused shots. On a 1000-model corpus spanning six acquisition families the ensemble cuts RMSE 38% relative to its classical prior and shows no geometry-specific gap on three never-trained families. Zero-shot on full-line Marmousi-2 it improves 354→304 m/s; raw coverage collapses to ~0.42 and the audit repairs it under noise, wavelet error, shot decimation, and an elastic–acoustic mismatch stress test, with peak height flagging physics mismatch. Budget-matched gather-based baselines underperform off their training geometry or everywhere.

Significance. Acquisition fragility and uncalibrated uncertainty remain the main barriers between DL-FWI demos and tools practitioners would trust. The geometry-invariant physics encoding is a clean architectural answer to the first problem; the combination of conformal calibration on synthetics with a wave-equation audit for label-free recalibration is a concrete, falsifiable answer to the second. Strengths that raise the bar for the field include: a seeded 1000-instance corpus with held-out acquisition families, budget-matched gather baselines (fixed and source-conditioned), two ablations tied to specific claims, eleven full-line corruption conditions plus elastic mismatch, explicit audit failure modes on short-offset crops, and a complete experimental matrix runnable in ~90 h on one laptop with planned Zenodo release. If the audit regime is stated carefully, this is a useful, reproducible contribution to calibrated residual FWI rather than another uncalibrated end-to-end network.

major comments (3)
  1. [Abstract; Table 2; §5.5; §7] Abstract and §7 claim the audit “restores coverage to 0.89–0.91 across eleven corruption conditions.” Table 2 contradicts that band: SNR 2 yields Cov.aud = 0.971 (τ_audit = 8.0 vs oracle 4.24); SNR 32 yields 0.859; the 32-shot row has no audit by construction; elastic mismatch (§5.4) repairs only to 0.83. The body (§5.5–5.6) already documents these cases and the short-offset plateau failure. The abstract/conclusion range should be revised to match Table 2 (e.g., “typically 0.86–0.91, with documented overshoot when the noise floor dominates and under-repair under physics mismatch”) so the central deployment claim is not overstated.
  2. [§3.4 Eqs. (12)–(14); §5.6; Eq. (15); Table S2] The deployment half of the strongest claim rests on Eqs. (12)–(14): maximizing held-out data-space coverage C(τ) under the assumed forward operator is taken as a proxy for the model-space inflation that restores marginal coverage. §5.6 and Eq. (15) correctly show that C(τ) is identifiable only when paths are long enough and model error dominates the predictive band; on 4 km crops the curve plateaus and the plateau-right rule runs to the grid edge (Table S2, Fig. 12), producing only conservative over-coverage. This is an empirically useful recalibrator inside a documented regime, not a general ground-truth-free coverage guarantee. The abstract and conclusions should state the aperture/noise preconditions with the same clarity as the Discussion, and preferably give a checkable pre-deployment diagnostic (e.g., pre-arrival noise vs. predicted data residual, or a minimum path-length criterion
  3. [§5.3; Figs. 9–10; Table S5] §5.3 reports that both budget-matched gather baselines (InversionNet-style fixed geometry and a 35M source-conditioned FiLM/Fourier-DeepONet-style net) land at or worse than the classical prior. The paper appropriately limits the claim to “at the matched 600-instance budget.” To keep the comparison load-bearing rather than a straw man, the main text should state the training protocol more explicitly (same augmentation, output grid, early-stopping, and that neither baseline received the GIPE channels or the ADMM prior as input) and note whether a residual-on-prior or multi-scale gather baseline was considered. Without that, readers may discount the geometry-invariance claim as an artifact of weak baselines rather than of representation.
minor comments (7)
  1. [§5.1; Fig. 4] Fig. 4 and §5.1 report Spearman 0.76 and AUSE 0.118 for σ ranking error; add the corresponding curves or a one-sentence definition of the sparsification protocol so the number is interpretable without the supplement.
  2. [§3.3 Eq. (8); §6] Eq. (8) sets λ = 0.1 in normalized units with no sensitivity. A one-row ablation or brief note on whether λ was tuned on validation would help, given the Discussion’s own remark that the residual scale caps deep-section gains.
  3. [§3.3] Mondrian strata are “illumination quartiles and depth halves” (§3.3). State how many calibration pixels fall in the worst stratum and whether empty/near-empty strata are pooled; finite-sample quantile validity is sensitive to stratum size.
  4. [§1; References] Self-citation to Kumar & Tripathi (2026) is listed as “Under review.” Clarify what is inherited (ADMM prior, reweighted-ℓ1) versus new (GIPE, ensemble, conformal+audit, pre-stack velocity) so novelty is unambiguous.
  5. [§4.3; Table 1] Table 1 and §4.3: “PyTorch 2.13” is likely a typo (current public releases are 2.x with different minor numbering); correct for reproducibility.
  6. [Fig. 3; Eq. (6)] Fig. 3 caption and Eq. (6): channel order and units (normalized vs. physical) are not fully specified; a short table or colorbar units would help re-implementation.
  7. [Title block; References] Minor prose: “Simut˙ e” encoding in the Fichtner references; “Deepak Kumara” vs. “Kumar” in the author line; ensure consistent author spelling before production.

Circularity Check

0 steps flagged

No significant circularity: empirical pipeline validated on held-out synthetics and public Marmousi-2; self-citation is an upstream extension, not a load-bearing premise.

full rationale

This is a methods-and-validation paper, not a first-principles derivation. The geometry-invariant encoding is an explicit architectural choice (ten model-space channels from classical operators), the residual ensemble is trained on a seeded train split and evaluated on held-out test families and never-seen Marmousi-2, and conformal quantiles are computed on a separate calibration split under the usual exchangeability claim. The held-out-shot audit rescales intervals from data-space coverage of unused shots simulated under the assumed forward operator; it does not fit τ to ground-truth velocity and then relabel that fit as a prediction. The only self-citation (Kumar & Tripathi 2026) states that the work extends a prior impedance pipeline; none of the Marmousi RMSE, coverage, or geometry-transfer numbers are forced by that citation, uniqueness theorems, or fitted inputs renamed as results. Weaknesses of the audit (plateau on short apertures, τ vs oracle mismatch at extreme SNR) are empirical regime limits, not circular reductions. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard acoustic wave physics, standard conformal exchangeability on the calibration split, and several modeling/hyperparameter choices (residual scale, audit perturbation fields, classical ADMM prior quality). No new physical entities are postulated. Free parameters are optimization and audit knobs that affect interval width and residual magnitude but are partially stress-tested in sensitivity tables.

free parameters (5)
  • residual scale λ = 0.1
    Multiplies network residual in normalized units before adding to c_ADMM (Eq. 8); chosen as 0.1, caps how far the ensemble can move from the prior.
  • audit correlation length of ξ_i = 60 m (default)
    Spatial correlation of perturbation fields in Eq. 12; sensitivity table shows it is the most sensitive audit knob and moves τ_audit and repaired coverage.
  • audit sample count M and plateau tolerance δ = M=12, δ=0.005
    M=12 predictive simulations and δ=0.005 in Eqs. 12–14 control empirical percentiles and τ selection; varied in Table S6.
  • ADMM/FWI regularization and step schedule = e.g. ρ=0.02, μ=0.2; Marmousi bands 3/5/8/12 Hz
    μ, ρ, reweighting, frequency bands, and step lengths define c_ADMM prior quality that the residual learner inherits; set by recipe not learned end-to-end.
  • conformal level and Mondrian strata = primarily 90% intervals; 8 strata
    Nominal 1−α and illumination-quartile × depth-half strata define which quantile is applied; physics-conditioned choice is a design decision.
axioms (4)
  • domain assumption 2D constant-density acoustic wave equation adequately generates and inverts the synthetic and Marmousi experiments (elastic used only as mismatch stress test).
    Forward model Eq. 1 and entire classical chain; field elastic/3D effects are acknowledged as future work in §6.
  • standard math Finite-sample marginal conformal coverage holds under exchangeability of calibration and test pixels/instances on the synthetic calibration distribution.
    Invoked in §3.3 with Vovk/Angelopoulos; paper correctly does not claim simultaneous image-level coverage.
  • ad hoc to paper Held-out shots unused by inversion are exchangeable enough with the inverted acquisition that maximizing data-space coverage C(τ) yields a useful model-space width rescale without labels.
    Core of the physics audit §3.4; supported empirically on full-line conditions but fails identifiability on short crops (§5.6).
  • domain assumption Classical multiscale FWI + reweighted-TV ADMM prior is a sufficient base so that a small residual network can improve error without replacing inversion.
    Framing in Introduction and residual form Eq. 8; gains saturate where prior is uninformative (deep section, dead prior under wavelet error).
invented entities (2)
  • Geometry-invariant physics encoding (GIPE) ten-channel tensor independent evidence
    purpose: Map variable acquisitions into fixed model-grid inputs so the network never sees gathers or geometry structure.
    Defined in Eq. 6–7 as composition of existing operators (models, gradients, fast-marching coverage); engineering construct rather than new physics.
  • Held-out-shot physics audit (τ_audit via C(τ)) independent evidence
    purpose: Label-free recalibration and physics-mismatch diagnostic under deployment shift.
    Introduced in §3.4 as physics analogue of conformal risk control; falsifiable via coverage vs oracle and peak-height ranking on controlled corruptions.

pith-pipeline@v1.2.0-daily-grok45 · 22298 in / 3880 out tokens · 76736 ms · 2026-07-31T04:21:07.151513+00:00 · methodology

0 comments
read the original abstract

Neural networks that map seismic data directly to velocity models tend to memorize the acquisition geometry they were trained on, and their uncertainty estimates are rarely trustworthy once the data drift away from the training distribution. We describe a laptop-scale pipeline that addresses both problems. The network never sees shot gathers. Instead, variable acquisition geometries are mapped into a fixed model-space representation computed by classical physics operators: the starting model, a regularized classical inversion, two misfit-gradient images, and six illumination and wavenumber-coverage maps from fast-marching traveltimes. The learned component is not a replacement for FWI; it is a calibrated residual corrector applied to this physics-derived prior. A six-member heterogeneous ensemble with per-pixel variance heads is calibrated by physics-conditioned conformal prediction, giving finite-sample pixel-wise marginal coverage on the calibration distribution. Beyond that distribution, a held-out-shot physics audit simulates shots the inversion never used through samples of the predictive distribution and rescales interval width where their data-space coverage peaks; no ground truth is involved. On a corpus of 1000 synthetic models spanning six acquisition families, the ensemble reduces error by 38% relative to its classical prior and transfers to never-seen geometries without measurable degradation. Applied zero-shot to the full 17 km Marmousi-2 line, it lowers the error from 354 to 304 m/s while raw coverage collapses to 0.42; the audit restores coverage to 0.89-0.91 across eleven corruption conditions covering noise, wavelet error, and shot decimation, and its peak height cleanly separates physics mismatch from benign corruptions. Budget-matched gather-based baselines underperform substantially off their training acquisition. Code and checkpoints will be archived on Zenodo.

Figures

Figures reproduced from arXiv: 2607.28535 by Deepak Kumar, Jayant Nath Tripathi, Laxmidhar Behera.

Figure 1
Figure 1. Figure 1: Training and evaluation corpus: one instance per acquisition family, showing the [PITH_FULL_IMAGE:figures/full_fig_p022_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Marmousi-2 benchmark data: velocity model with the 32-shot marine acquisition [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The ten-channel geometry-invariant physics encoding of Eq. [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Synthetic-corpus evaluation of the calibrated ensemble: reliability across nominal [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Zero-shot application to the full 17 km Marmousi-2 line: true model, classical [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Physics-mismatch experiment: elastic Marmousi-2 data inverted with acoustic [PITH_FULL_IMAGE:figures/full_fig_p025_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Held-out-shot physics audit on zero-shot Marmousi-2: data-space coverage [PITH_FULL_IMAGE:figures/full_fig_p026_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full-line robustness spine: accuracy of FWI, the ADMM prior, and the zero [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Fixed-geometry gather-to-model baseline: the blue bar is the network evaluated [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Source-conditioned gather network at matched budget: worse than the classical [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Crop-scale sweeps on two 4×2 km Marmousi crops: RMSE and coverage across the noise, wavelet-frequency, wavelet-phase, and shot-count axes. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Crop-scale audit: C(τ ) curves whose broad plateau reflects the short-offset identifiability limit, the peak-height severity indicator, and the conservative coverage repair. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 13 canonical work pages

  1. [11]

    Geophysics 83, R63–R75

    Assessing uncertainties in velocity models and images with a fast nonlinear uncertainty quantification method. Geophysics 83, R63–R75. doi:10.1190/geo2017-0321.1. Esser, E., Guasch, L., van Leeuwen, T., Aravkin, A.Y., Herrmann, F.J.,

    work page doi:10.1190/geo2017-0321.1

  2. [12]

    SIAM Journal on Imaging Sciences 11, 376–406

    Total variation regularization strategies in full-waveform inversion. SIAM Journal on Imaging Sciences 11, 376–406. doi:10.1137/17M111328X. Fichtner, A., Simut˙ e, S.,

    work page doi:10.1137/17m111328x

  3. [13]

    Journal of Geophysical Research: Solid Earth 123, 2984–2999

    Hamiltonian Monte Carlo inversion of seismic sources in complex media. Journal of Geophysical Research: Solid Earth 123, 2984–2999. doi:10.1002/2017JB015249. Gal, Y., Ghahramani, Z.,

    work page doi:10.1002/2017jb015249

  4. [17]

    Geophysics 86, R711–R721

    Mapping full seismic waveforms to vertical velocity profiles by deep learning. Geophysics 86, R711–R721. doi:10.1190/geo2019-0473.1. Kendall, A., Gal, Y.,

    work page doi:10.1190/geo2019-0473.1

  5. [18]

    ArXiv:1703.04977

    What uncertainties do we need in Bayesian deep learning for computer vision?, in: Advances in Neural Information Processing Systems. ArXiv:1703.04977. Kumar, D., Tripathi, J.N.,

    Pith/arXiv arXiv

  6. [20]

    ArXiv:1612.01474

    Simple and scalable predictive uncertainty estimation using deep ensembles, in: Advances in Neural Information Processing Systems. ArXiv:1612.01474. van Leeuwen, T., Herrmann, F.J.,

    Pith/arXiv arXiv

  7. [23]

    Math- ematical Geosciences 52, 53–79

    Stochastic seismic waveform inversion using generative adversarial networks as a geological prior. Math- ematical Geosciences 52, 53–79. doi:10.1007/s11004-019-09832-6. Osypov, K., Yang, Y., Fournier, A., Ivanova, N., Bachrach, R., Yarman, C.E., You, Y., Nichols, D., Woodward, M.,

    doi:10.1007/s11004-019-09832-6

  8. [24]

    Geophysical Prospecting 61, 1114–1134

    Model-uncertainty quantification in seismic tomography: method and applications. Geophysical Prospecting 61, 1114–1134. doi:10.1111/1365-2478.12058. Ovcharenko, O., Kazei, V., Kalita, M., Peter, D., Alkhalifah, T.,

    arXiv

  9. [25]

    Geophysics 84, R989–R1001

    Deep learning for low-frequency extrapolation from multioffset seismic data. Geophysics 84, R989–R1001. doi:10.1190/geo2018-0884.1. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.,

    work page doi:10.1190/geo2018-0884.1

  10. [26]

    ArXiv:1912.01703

    PyTorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems. ArXiv:1912.01703. Peters, B., Herrmann, F.J.,

    Pith/arXiv arXiv 1912

  11. [27]

    Geophysical Journal International 167, 495–503

    A review of the adjoint-state method for computing the gradient of a functional with geophysical applications. Geophysical Journal International 167, 495–503. doi:10.1111/j.1365-246X.2006.02978.x. Pratt, R.G.,

    Pith/arXiv arXiv 2006

  12. [29]

    Journal of Computational Physics 378, 686–707

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686–707. doi:10.1016/j.jcp.2018.10.045. Richardson, A.,

    doi:10.1016/j.jcp.2018.10.045 2018

  13. [30]

    arXiv preprint arXiv:1801.07232

    Seismic full-waveform inversion using deep learning tools and techniques. arXiv preprint arXiv:1801.07232 . Romano, Y., Patterson, E., Candès, E.J.,

    Pith/arXiv arXiv

  14. [31]

    ArXiv:1905.03222

    Conformalized quan- tile regression, in: Advances in Neural Information Processing Systems. ArXiv:1905.03222. Ronneberger, O., Fischer, P., Brox, T.,

    Pith/arXiv arXiv 1905

  15. [34]

    Journal of Machine Learning Research 9, 371–421

    A tutorial on conformal prediction. Journal of Machine Learning Research 9, 371–421. ArXiv:0706.3188. Sirgue, L., Pratt, R.G.,

    Pith/arXiv arXiv

  16. [36]

    Geophysics 85, R87–R99

    A theory-guided deep-learning formulation and optimization of seismic waveform inversion. Geophysics 85, R87–R99. doi:10.1190/geo2019-0138.1. Tarantola, A.,

    work page doi:10.1190/geo2019-0138.1

  17. [38]

    Geophysics 74, WCC1–WCC26

    An overview of full-waveform inversion in exploration geophysics. Geophysics 74, WCC1–WCC26. doi:10.1190/1. 3238367. 20 Vovk, V.,

    doi:10.1190/1

  18. [39]

    Ma- chine Learning 92, 349–376

    Conditional validity of inductive conformal predictors. Ma- chine Learning 92, 349–376. doi:10.1007/s10994-013-5355-6. Vovk, V., Gammerman, A., Shafer, G.,

    doi:10.1007/s10994-013-5355-6

  19. [41]

    Computers & Geosciences 155, 104833

    PIN- Neik: Eikonal solution using physics-informed neural networks. Computers & Geosciences 155, 104833. doi:10.1016/j.cageo.2021.104833. Warner, M., Guasch, L.,

    arXiv 2021

  20. [42]

    Geo- physics 81, R429–R445

    Adaptive waveform inversion: Theory. Geo- physics 81, R429–R445. doi:10.1190/geo2015-0387.1. Wu, Y., Lin, Y.,

    work page doi:10.1190/geo2015-0387.1

  21. [43]

    IEEE Transactions on Computational Imaging 6, 419–433

    InversionNet: An efficient and accurate data-driven full waveform inversion. IEEE Transactions on Computational Imaging 6, 419–433. doi:10.1109/TCI.2019.2956866. Yang, F., Ma, J.,

    arXiv 2019

  22. [44]

    Geophysics 89, A23–A28

    WISE: full- waveform variational inference via subsurface extensions. Geophysics 89, A23–A28. doi:10.1190/geo2023-0744.1. Zhu, M., Feng, S., Lin, Y., Lu, L.,

    doi:10.1190/geo2023-0744.1

  23. [45]

    Computer Methods in Applied Mechanics and Engineering 416, 116300

    Fourier-DeepONet: Fourier-enhanced deep operator networks for full waveform inversion with improved accuracy, generalizability, and robustness. Computer Methods in Applied Mechanics and Engineering 416, 116300. doi:10.1016/j.cma.2023.116300. 21 Figure 1: Training and evaluation corpus: one instance per acquisition family, showing the velocity model with i...

    arXiv 2023

  24. [1984]

    Geophysics 49, 1259–1266

    Inversion of seismic reflection data in the acoustic approximation. Geophysics 49, 1259–1266. doi:10.1190/1.1441754. Virieux, J., Operto, S.,

    doi:10.1190/1.1441754

  25. [1992]

    Physica D: Nonlinear Phenomena 60, 259–268

    Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena 60, 259–268. doi:10.1016/0167-2789(92)90242-F. Sethian, J.A.,

    doi:10.1016/0167-2789(92)90242-f

  26. [1995]

    Geophysics 60, 1457–1473

    Multiscale seismic waveform inversion. Geophysics 60, 1457–1473. doi:10.1190/1.1443880. Candès, E.J., Wakin, M.B., Boyd, S.P.,

    doi:10.1190/1.1443880

  27. [1996]

    Proceedings of the National Academy of Sciences 93, 1591–1595

    A fast marching level set method for monotonically advancing fronts. Proceedings of the National Academy of Sciences 93, 1591–1595. doi:10.1073/pnas.93.4.1591. Shafer, G., Vovk, V.,

    doi:10.1073/pnas.93.4.1591

  28. [1999]

    Geophysics 64, 888–901

    Seismic waveform inversion in the frequency domain, part 1: Theory and verification in a physical scale model. Geophysics 64, 888–901. doi:10.1190/1.1444597. 19 Raissi, M., Perdikaris, P., Karniadakis, G.E.,

    doi:10.1190/1.1444597

  29. [2004]

    Geophysics 69, 231–248

    Efficient waveform inversion and imaging: A strategy for selecting temporal frequencies. Geophysics 69, 231–248. doi:10.1190/1.1649391. Sun, J., Niu, Z., Innanen, K.A., Li, J., Trad, D.O.,

    work page doi:10.1190/1.1649391

  30. [2005]

    Springer, New York

    Algorithmic Learning in a Random World. Springer, New York. doi:10.1007/b106715. bin Waheed, U., Haghighat, E., Alkhalifah, T., Song, C., Hao, Q.,

    doi:10.1007/b106715

  31. [2006]

    The Leading Edge 25, 156–166

    Marmousi2: An elastic upgrade for Marmousi. The Leading Edge 25, 156–166. doi:10.1190/1.2172306. Mosser, L., Dubrule, O., Blunt, M.J.,

    doi:10.1190/1.2172306

  32. [2008]

    Journal of Fourier Analysis and Applications 14, 877–905

    Enhancing sparsity by reweighted ℓ1 minimization. Journal of Fourier Analysis and Applications 14, 877–905. doi:10.1007/s00041-008-9045-x. Deng, C., Feng, S., Wang, H., Zhang, X., Jin, P., Feng, Y., Zeng, Q., Chen, Y., Lin, Y.,

    doi:10.1007/s00041-008-9045-x

  33. [2009]

    Geophysics 74, WCC105–WCC118

    Seismic imaging of complex onshore structures by 2D elastic frequency-domain full-waveform inversion. Geophysics 74, WCC105–WCC118. doi:10.1190/1.3215771. Bunks, C., Saleck, F.M., Zaleski, S., Chavent, G.,

    work page doi:10.1190/1.3215771

  34. [2011]

    Foundations and Trends in Machine Learning 3, 1–122

    Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning 3, 1–122. doi:10.1561/2200000016. Brossier, R., Operto, S., Virieux, J.,

    doi:10.1561/2200000016

  35. [2013]

    Geophysical Journal International 195, 661–667

    Mitigating local minima in full- waveform inversion by expanding the search space. Geophysical Journal International 195, 661–667. doi:10.1093/gji/ggt258. Martin, G.S., Wiley, R., Marfurt, K.J.,

    work page doi:10.1093/gji/ggt258

  36. [2016]

    1050–1059

    Dropout as a Bayesian approximation: Rep- resenting model uncertainty in deep learning, in: International Conference on Machine Learning, pp. 1050–1059. ArXiv:1506.02142. Gebraad, L., Boehm, C., Fichtner, A.,

    Pith/arXiv arXiv

  37. [2017]

    1321–1330

    On calibration of modern neural networks, in: International Conference on Machine Learning, pp. 1321–1330. ArXiv:1706.04599. Kazei, V., Ovcharenko, O., Plotnitskii, P., Peter, D., Zhang, X., Alkhalifah, T.,

    Pith/arXiv arXiv

  38. [2018]

    The Leading Edge 37, 58–66

    Deep-learning tomography. The Leading Edge 37, 58–66. doi:10.1190/tle37010058.1. Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J.,

    doi:10.1190/tle37010058.1

  39. [2019]

    Geophysical Journal International 218, 855–872

    Implementing bound con- straints and total-variation regularization in extended full-waveform in- version with the alternating direction method of multiplier: application to large contrast media. Geophysical Journal International 218, 855–872. doi:10.1093/gji/ggz189. Angelopoulos, A.N., Bates, S.,

    work page doi:10.1093/gji/ggz189

  40. [2020]

    Journal of Geophysical Research: Solid Earth 125, e2019JB018428

    Bayesian elastic full-waveform inversion using Hamiltonian Monte Carlo. Journal of Geophysical Research: Solid Earth 125, e2019JB018428. doi:10.1029/2019JB018428. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.,

    work page doi:10.1029/2019jb018428

  41. [2021]

    IEEE Signal Processing Magazine 38, 89–119

    Deep learning for seismic inverse problems: Toward the acceleration of geophysical analysis workflows. IEEE Signal Processing Magazine 38, 89–119. doi:10.1109/MSP.2020.3037429. Aghamiry, H.S., Gholami, A., Operto, S.,

    arXiv 2020

  42. [2022]

    6007–6020

    OpenFWI: Large-scale multi-structural benchmark datasets 17 for full waveform inversion, in: Advances in Neural Information Processing Systems, pp. 6007–6020. ArXiv:2111.02926. Ely, G., Malcolm, A., Poliannikov, O.V.,

    Pith/arXiv arXiv

  43. [2023]

    Foundations and Trends in Machine Learning 16, 494–591

    Conformal prediction: A gentle in- troduction. Foundations and Trends in Machine Learning 16, 494–591. doi:10.1561/2200000101. Angelopoulos, A.N., Bates, S., Fisch, A., Lei, L., Schuster, T.,

    doi:10.1561/2200000101

  44. [2024]

    ArXiv:2208.02814

    Conformal risk control, in: International Conference on Learning Representations. ArXiv:2208.02814. Araya-Polo, M., Jennings, J., Adler, A., Dahlke, T.,

    Pith/arXiv arXiv

  45. [2026]

    IEEE Transactions on Geoscience and Remote Sensing Under review; EarthArXiv preprint, doi:10.31223/X56J49

    ADMM-guided physics-informed deep learning for two-dimensional acoustic impedance inversion with reweighted ℓ1 sparse regularization. IEEE Transactions on Geoscience and Remote Sensing Under review; EarthArXiv preprint, doi:10.31223/X56J49. 18 Lakshminarayanan, B., Pritzel, A., Blundell, C.,

    work page doi:10.31223/x56j49