Pith. sign in

REVIEW 6 major objections 4 minor 69 references

Geological Everything Model 3D: A Promptable Foundation Model for Unified and Zero-Shot Subsurface Understanding

T0 review · 6 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GEM proposes a single promptable generative model that replaces task-specific seismic interpretation pipelines and transfers zero-shot from synthetic seismic training to real field data and even to SHARAD radar on Mars.

desk verdict A broad and ambitious 3D promptable seismic foundation model with a genuinely surprising Mars radar transfer, but the headline zero-shot advantage over baselines is confounded by unequal prompt access at inference. read the letter →

arxiv 2507.00419 v4 pith:V5IGMSE7 submitted 2025-07-01 physics.geo-ph cs.AI

Yimin Dou , Xinming Wu , Nathan L Bangs , Harpreet Singh Sethi , Jintao Li , Hang Gao , Zhixiang Guo This is my paper
classification physics.geo-phcs.AI
keywords seismicinterpretationfoundationmodelzero-shotgeneralizationprompt-conditionedgenerationrelativegeologicaltimegeobodysegmentationSHARADradarself-supervisedpretraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces GEM, a single 3D generative model that treats every subsurface interpretation task—fault and horizon picking, full stratigraphic time, geobody segmentation, and rock-property modeling—as one operation: completing a sparse human prompt along the geological structure inferred from the image. The authors claim this prompt-conditioned completion generalizes to new surveys, new tasks, and even a different imaging modality without retraining, and they demonstrate it by interpreting reflectors in Mars radar data with a model fine-tuned only on synthetic seismic volumes. The enabling training recipe is two-stage: self-supervised masked modeling on more than 500 field seismic volumes, then adversarial fine-tuning on 1,000 synthetic volumes with mixed prompt-label pairs. If the claim holds, one expert-in-the-loop model could replace today's fragmented, survey-specific interpretation pipelines and serve equally well for energy exploration, carbon storage, hazard assessment, and planetary science.

What carries the argument

The central object is the latent structural framework—the model's internal 3D representation of faults, unconformities, marker strata, and geobodies inferred from the input subsurface volume—and the conditional generative process $\hat Y = p(Y \mid X_{\text{img}}, Y_{\text{part}})$, where $X_{\text{img}}$ is the imaging volume and $Y_{\text{part}}$ is the sparse human prompt. The carrying mechanisms are a SparK-style masked image modeling pretraining at 80% masking over 500+ field volumes; a structure-aware perceptual network trained on synthetic fault, channel, and karst labels; and adversarial fine-tuning with a relativistic GAN plus structure-aware perceptual and LPIPS losses. The architecture is a purely convolutional, HRNet-inspired 3D backbone with 0.67 billion parameters that keeps a quarter-resolution branch throughout the forward pass to preserve spatial detail.

What would settle it

Prompt GEM on a seismic volume from a structural setting absent from its synthetic training—a salt diapir province, metamorphic basement, or glacial till with no layered stratigraphy—and compare its zero-shot faults, horizons, or geobodies against independent expert picks and a task-specific supervised model trained on that survey; if GEM's outputs depart from expert interpretation by more than the supervised baseline, or if its SHARAD reflectors fail to coincide with independently published Mars stratigraphic picks, the zero-shot generalization claim is refuted.

Watch

Extended reading notes

Core claim

GEM's central claim is that structural interpretation, stratigraphic analysis, geobody segmentation, and physical property modeling are not separate problems but different instances of prompt-conditioned dimensional completion over a shared latent geological framework. Given a subsurface volume, GEM infers a structural framework of faults, unconformities, horizons, and geobodies; it then propagates whatever sparse prompt a human supplies—a few well logs, a mask, a structural sketch, or an initial relative-geological-time scale—along that framework to generate a geologically coherent 3D output. Because the same generative mechanism serves all tasks, the authors report zero-shot results across a wide span: Hikurangi subduction-zone faults reconstructed from about 2.4 prompts per surface, Martian polar stratigraphy from SHARAD radargrams, unconformities and relative geological time on the Poseidon and Delft surveys, channel and karst segmentation on Parihaka and Fort Worth data, and impedance, gamma ray, and lithology volumes from sparse wells on SEAM, Teapot Dome, F3, and Delft.

Load-bearing premise

The load-bearing premise is that the roughly 1,000 synthetic seismic volumes with labels, plus the public synthetic channel and karst data, resemble real geology closely enough that a model fine-tuned only on them can interpret unseen field seismic and Mars radar without retraining.

Editorial extensions

If this is right

  • A single GEM checkpoint can replace separate fault-detection, horizon-picking, geobody-segmentation, and property-inversion networks; changing survey or task requires no retraining, only new prompts.
  • With only a few interactions per target—about 2.4 prompts per fault surface and 3.3 per geobody, or 2–5 wells per property volume—an interpreter can produce complete 3D results in a human-in-the-loop loop.
  • Relative-geological-time volumes can be generated fully automatically by iteratively re-prompting at high-disagreement locations, yielding complete horizon surfaces even across faults and unconformities.
  • Sparse well logs function as prompts for generating geologically consistent impedance, gamma-ray, and lithology volumes that remain accurate at blind wells.
  • The model's zero-shot transfer to Mars SHARAD radar suggests that the same pretrained weights can be deployed on entirely new data modalities without architectural modification.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A structured distribution-shift benchmark would test the generality claim: prompt GEM on salt diapirs, metamorphic basement, or polygonal fault systems far outside its synthetic training distribution and measure whether interpretation quality degrades gracefully or collapses; the paper reports no such analysis.
  • The repeated re-prompting loop used for RGT could be generalized into an inference-time uncertainty mechanism for any task, using local disagreement between generations to decide where the next prompt should go.
  • If structure-aware latent representations are the true source of transfer, the same fine-tuning recipe might apply to non-seismic reflectivity data (ground-penetrating radar, sub-bottom profilers, borehole imaging) where labels are even scarcer than in seismic interpretation.
  • Because the quantitative field evaluation relies on roughly 20 expert-annotated slices, a multi-survey blind test with independent expert panels would separate genuine prompt-conditioned generalization from benchmark memorization on well-known public datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 4 minor

Summary. The paper introduces GEM-3D, a 0.67B-parameter convolutional generative model that reformulates structural interpretation (faults, horizons, RGT), geobody segmentation, and physical property modeling as prompt-conditioned generation over a latent geological framework. The model is pretrained with masked modeling on more than 500 field seismic volumes and then adversarially fine-tuned on 1,000 synthetic volumes plus public synthetic channel and karst datasets. The paper claims zero-shot generalization across tasks, surveys, and modalities, including transfer to Martian SHARAD radar data, and reports comparisons with UNet, DeepLabV3, HRNet, FaultNet, FaultSeg+, and SAM2 on synthetic and field cases.

Significance. If the central claims are correct, this would be a notable advance toward a general-purpose, promptable subsurface interpretation model. The manuscript has real strengths: an explicit two-stage training scheme with clearly stated losses (Eqs. 2-6), detailed architecture and inference-cost reporting (Appendix F), a large set of field demonstrations, blind-well validation in the property-modeling sections, and an interactive GUI. However, the quantitative evaluation is confounded by an unequal comparison protocol, the field sample sizes are very small, and there are internal contradictions in the reported results. The headline claim of zero-shot generalization is therefore not yet established. The contribution is potentially important, but the evaluation needs substantial revision.

major comments (6)
  1. [§2.3.1, Figures 3(a), 4(a), 5(a)] The quantitative comparison is confounded: GEM receives test-time prompts (fault traces, masks, well logs) while the baselines do not receive any prompts at test time. The statement that 'the same prompt inputs provided to GEM were incorporated as weak labels into the training sets of these baseline models' means prompts were used only as training labels, not as inference inputs. Because a prompt such as a fault trace or a well log already carries direct information about the target structure, the reported IOU and MAE advantages may reflect the information content of the conditioning rather than GEM's structural reasoning or zero-shot generalization. Please add matched evaluations: give the same prompts to the baselines at test time, and also run GEM without any prompts, reporting both settings.
  2. [§2.3.1, §2.4, Figures 3(a), 4(a)] The field quantitative evidence is very thin: the fault evaluation uses 20 expert-annotated slices and the channel evaluation uses 30 manually labeled slices, with no per-slice distributions, confidence intervals, or significance tests reported. Claims that GEM 'significantly outperforms' and 'consistently surpasses' baselines on field data are not supported at this sample size. Please report per-slice and per-survey statistics, error bars, and significance tests, or soften the claims accordingly.
  3. [§2.3.3] The Martian SHARAD result is presented as evidence of cross-modal zero-shot generalization, but there is no ground truth, no quantitative metric, and no comparison with an alternative method or independent manual interpretation. Given that this is one of the paper's most distinctive claims, the evidence should include at least a validation subset with known reflectors or an independently audited interpretation protocol; otherwise the result remains a qualitative demonstration.
  4. [Appendix D.2, Figure D5(c), Section 2.4] There is an internal contradiction: Appendix D.2 states that on karst segmentation 'traditional models perform effectively, achieving results comparable to or even surpassing those of GEM,' while Section 2.4 states that GEM 'consistently surpasses' conventional models and Figure 4(a) reports a quantitative advantage. Please reconcile these statements and clarify whether the reported karst bars include the setting in which the baselines are competitive.
  5. [Appendix C, Figure C3] The iterative RGT protocol is self-conditioning: new prompt scales are sampled from regions where GEM's own initial prediction deviates most, then re-injected as prompts. The final RGT output is therefore not a single zero-shot pass, and any systematic error in the first prediction is re-introduced as a prompt. Please report the number of iterations used for Figures 3(e,f), state whether the displayed outputs are the final iterative outputs, and describe how convergence or error propagation is controlled.
  6. [§2.3.3, §3.4, Appendix B] The model is fine-tuned only on 1,000 synthetic volumes plus public synthetic channel and karst datasets (Appendix B), yet the zero-shot claims assume transfer to field seismic and to radar. No distribution-shift analysis is provided, and Figure 3(a) itself shows only 'minor performance differences' on synthetic data. The large field-data gains therefore need additional support, such as domain-gap metrics, per-survey failure analysis, or an explicit study of how performance degrades with geological distance from the synthetic training distribution.
minor comments (4)
  1. [Abstract (opening) vs. Abstract (full text)] The two abstracts are not identical: the opening abstract says 'large-scale field seismic data' while the full-text abstract says 'over 500 field seismic volumes' and adds the phrase 'physics-informed'; please unify the wording so that the claims are consistent.
  2. [§3.4, Eq. (6)] Equation (6) does not specify the weight of the L_LPIPS term; please state the coefficient explicitly or indicate that it is 1.0.
  3. [§2.5, Figure 5(a)] For the synthetic SEAM I evaluation, the paper reports FID but does not specify how FID is computed for 3D volumes (slice-by-slice or volume-level, and at what resolution); please clarify the protocol.
  4. [References 66-67] References 66 and 67 currently point to the same URL; please verify and correct the second reference.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the central zero-shot claim rests on external benchmarks and held-out wells, with only evaluation-design caveats.

full rationale

The paper's derivation chain is a two-stage training scheme: self-supervised masked modeling on 500+ field volumes, then adversarial fine-tuning on synthetic prompt-label pairs (Section 3.2-3.4, Eqs. 3-6), followed by prompt-conditioned inference (Eq. 1). Eq. 1 is a problem formulation, not a derived prediction, so no result follows from it by construction. The evaluation does not reduce to the training inputs: the headline demonstrations are on external field surveys (Hikurangi, Baiyun, Poseidon, Delft, Parihaka, Teapot, F3) and on SHARAD radar, which are outside the synthetic fine-tuning distribution, and the property-modeling metrics are computed at blind wells explicitly excluded from both baseline training and GEM prompts (Section 2.5). Same-group citations (VFM ref. 40, FaultSeg3D refs. 14/52, CIGChannel ref. 54, KarstSeg ref. 55, Deep RGT ref. 53) supply datasets, baselines, or background, but the load-bearing generalization claims are checked on external benchmarks and are not justified solely by those citations. Two caveats are evaluation confounds rather than circular derivations: baseline models are not given prompt conditioning at inference while GEM is (Section 2.3.1), and the RGT refinement loop re-injects the model's own output as new prompts (Appendix C). These affect how much of the reported advantage is attributable to GEM's learned reasoning, but no equation or fitted parameter makes any output equivalent to its input by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

No new physical entities are postulated. The free parameters are training hyperparameters and an inference protocol. The load-bearing assumptions are domain assumptions about synthetic-to-real and cross-modal transfer; these are stated qualitatively and never stress-tested with distribution shift analyses.

free parameters (5)
  • L1 loss weight alpha = cosine annealed from 30.0 to 2.0
    Hand-scheduled balance between voxel-level and adversarial/perceptual losses (Eq. 6); chosen without derivation and central to training behavior.
  • LPIPS loss weight = 0.01
    Fixed constant in the final loss (Eq. 6), set by hand.
  • Masking ratio = 80%
    Chosen because 3D seismic reconstruction is easier than 2D; affects the quality of pretrained representations (Section 3.2).
  • Fine-tuning schedule = 300k iterations at 320x400x128, then 200k at 512x512x128
    Hand-set training schedule that determines model convergence (Section 3.4).
  • RGT iterative prompts = initial uniform scale plus 4 new scales per refinement round
    Hand-designed inference protocol for RGT estimation (Appendix C); no principled stopping criterion.
assumptions (4)
  • domain assumption Seismic and radar reflectivity share sufficient structural commonality, such as interface continuity and layered geometry, that synthetic seismic training transfers to SHARAD radar without radar training.
    Invoked in Section 2.3.3 to explain the Martian results.
  • domain assumption The 1,000 synthetic volumes and public synthetic channel and karst datasets represent the structural diversity of real subsurface geology.
    Appendix B; this underlies all zero-shot transfer claims.
  • domain assumption Sparse prompts such as well logs, masks, and sketches uniquely specify the task and enough boundary conditions for geologically consistent generation.
    Used in Equation 1 and Section 3.1 to justify the unified formulation.
  • ad hoc to paper A perceptual network trained only on faults, channels, and karst provides a universal structural prior for all dense outputs, including RGT and property volumes.
    Equation 5 applies this network to tasks whose structures it was never trained to segment.
invented entities (1)
  • latent geological framework
    purpose: Claimed internal representation along which prompts are propagated to unify all subsurface tasks
    No physical entity is proposed, but this construct is central to the explanation of transfer. Its existence is inferred from PCA visualizations (Figure 2a), which provide no interventional or falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geological Everything Model 3D: A Promptable Foundation Model for Unified and Zero-Shot Subsurface Understanding." pith.science (2026). https://pith.science/paper/V5IGMSE7

@misc{pith2026250700419,
  author       = {Pith},
  title        = {Pith review of: Geological Everything Model 3D: A Promptable Foundation Model for Unified and Zero-Shot Subsurface Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5IGMSE7}},
  note         = {Machine review of arXiv:2507.00419}
}
read the original abstract

Understanding Earth's subsurface is critical for energy transition, natural hazard mitigation, and planetary science. Yet subsurface analysis remains fragmented, with separate models required for structural interpretation, stratigraphic analysis, geobody segmentation, and property modeling-each tightly coupled to specific data distributions and task formulations. We introduce the Geological Everything Model 3D (GEM), a unified generative architecture that reformulates all these tasks as prompt-conditioned inference along latent structural frameworks derived from subsurface imaging. This formulation moves beyond task-specific models by enabling a shared inference mechanism, where GEM propagates human-provided prompts-such as well logs, masks, or structural sketches-along inferred structural frameworks to produce geologically coherent outputs. Through this mechanism, GEM achieves zero-shot generalization across tasks with heterogeneous prompt types, without retraining for new tasks or data sources. This capability emerges from a two-stage training process that combines self-supervised representation learning on large-scale field seismic data with adversarial fine-tuning using mixed prompts and labels across diverse subsurface tasks. GEM demonstrates broad applicability across surveys and tasks, including Martian radar stratigraphy analysis, structural interpretation in subduction zones, full seismic stratigraphic interpretation, geobody segmentation, and property modeling. By bridging expert knowledge with generative reasoning in a structurally aware manner, GEM lays the foundation for scalable, human-in-the-loop geophysical AI-transitioning from fragmented pipelines to a vertically integrated, promptable reasoning system. Project page: https://douyimin.github.io/GEM

Figures

Figures reproduced from arXiv: 2507.00419 by the authors.

Figure 1
Figure 1. Overview of GEM-3D, a promptable foundation model for unified and zero-shot subsurface modeling. (a) GEM supports a wide range of subsurface imaging tasks, including structural and stratigraphic interpretation (e.g., faults, horizons, unconformities, relative geological time), geobody segmentation (e.g., channels, salt bodies), and physical property modeling (e.g., impedance, gamma ray, lithology). By conditioning o… view at source ↗
Figure 2
Figure 2. Internal structure-aware reasoning mechanism of GEM. (a) Comparison between expert interpretations and GEM’s internal representations of seismic volumes. Dashed lines highlight how geologists identify key geological structures—such as faults, unconformities, and marker strata. Without using these annotations as inputs, GEM’s latent features capture similar patterns, revealing structure-aware reasoning consistent wit… view at source ↗
Figure 3
Figure 3. GEM demonstrates strong zero-shot performance across diverse subsurface interpretation tasks. (a) Quantitative evaluation of fault interpretation using intersection-over-union (IOU) on both synthetic and field datasets. GEM consistently outperforms baseline models across multiple fault types, without task-specific retraining. (b) Schematic illustrating how expert prompts (e.g., BSR and fault sketches) are combined w… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: GEM demonstrates generalizable geobody segmentation across diverse geological settings. (a) Quantitative comparison on synthetic channel and karst datasets. GEM significantly outperforms conventional models on real seismic data, demonstrating strong generalization with…
Figure 5
Figure 5. Figure 5: GEM enables prompt-guided, multi-physical-property modeling with minimal well control. (a) Quantitative comparison of impedance and gamma ray modeling on synthetic (SEAM I) and field datasets (F3, Teapot). In zero-shot settings using only sparse well-log prompts, GEM o…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 59 canonical work pages

  1. [1]

    communications15, 2057 (2024)

    Yu, P.et al.Crustal permeability generated through microearthquakes is constrained by seismic moment.Nat. communications15, 2057 (2024)

    work page 2024

  2. [2]

    Adv.11, eadr6771 (2025)

    Liu, Q.et al.Natural hydrogen in the volcanic-bearing sedimentary basin: Origin, conversion, and production rates.Sci. Adv.11, eadr6771 (2025)

    work page 2025

  3. [3]

    & Krevor, S

    Zhang, Y., Jackson, C. & Krevor, S. The feasibility of reaching gigatonne scale co2 storage by mid-century.Nat. Commun.15, 6913 (2024)

    work page 2024

  4. [4]

    Earth & Environ.5, 333 (2024)

    Creasy, N.et al.Co2 rock physics modeling for reliable monitoring of geologic carbon storage.Commun. Earth & Environ.5, 333 (2024). 5.Wang, T.et al.Earthquake forecasting from paleoseismic records.Nat. communications15, 1944 (2024)

    work page 2024

  5. [6]

    & Zhan, Z

    Li, J., Zhu, W., Biondi, E. & Zhan, Z. Earthquake focal mechanisms with distributed acoustic sensing.Nat. Commun.14, 4181 (2023)

    work page 2023

  6. [7]

    Li, C.et al.Layered subsurface in utopia basin of mars revealed by zhurong rover radar.Nature610, 308–312 (2022)

    work page 2022

  7. [8]

    Astron.8, 69–76 (2024)

    Zhang, L.et al.Buried palaeo-polygonal terrain detected underneath utopia planitia on mars by the zhurong radar.Nat. Astron.8, 69–76 (2024)

    work page 2024

  8. [9]

    Cui, X., Li, Z. & Hu, Y. Similar seismic moment release process for shallow and deep earthquakes.Nat. Geosci. 16, 454–460 (2023). 10.Bergen, K. J., Johnson, P. A., de Hoop, M. V. & Beroza, G. C. Machine learning for data-driven discovery in solid earth geoscience.Science363, eaau0323 (2019)

    work page 2023

Show all 69 references
  1. [11]

    J.et al.Himalayan valley-floor widths controlled by tectonically driven exhumation.Nat

    Clubb, F. J.et al.Himalayan valley-floor widths controlled by tectonically driven exhumation.Nat. Geosci.16, 739–746 (2023). 12.Mousavi, S. M. & Beroza, G. C. Deep-learning seismology.Science377, eabm4470 (2022)

    2023

  2. [13]

    communications15, 10025 (2024)

    Laurenti, L.et al.Probing the evolution of fault properties during the seismic cycle with deep learning.Nat. communications15, 10025 (2024)

    2024

  3. [14]

    & Fomel, S

    Wu, X., Liang, L., Shi, Y. & Fomel, S. Faultseg3d: Using synthetic data sets to train an end-to-end convolutional neural network for 3d seismic fault segmentation.Geophysics84, IM35–IM45 (2019). 17/28

    2019

  4. [15]

    IEEE Transactions on Geosci

    Wu, W.et al.Mtl-faultnet: Seismic data reconstruction assisted multitask deep learning 3-d fault interpretation. IEEE Transactions on Geosci. Remote. Sens.61, 1–15 (2023)

    2023

  5. [16]

    & Zheng, Y

    Gao, K., Huang, L. & Zheng, Y. Fault detection on seismic structural images using a nested residual u-net. IEEE Transactions on Geosci. Remote. Sens.60, 1–15 (2021)

    2021

  6. [17]

    & AlRegib, G

    Alaudah, Y., Michałowicz, P., Alfarraj, M. & AlRegib, G. A machine-learning benchmark for facies classification. Interpretation7, SE175–SE187 (2019)

    2019

  7. [18]

    & Gao, J

    Gao, Z., Wang, K., Wang, Z. & Gao, J. Optimizing seismic facies classification through differentiable network architecture search.IEEE Transactions on Geosci. Remote. Sens.62, 1–12 (2024)

    2024

  8. [19]

    & Geng, Z

    Yang, J., Wu, X., Bi, Z. & Geng, Z. A multi-task learning method for relative geologic , horizons, and faults with prior information and transformer.IEEE Transactions on Geosci. Remote. Sens.61, 1–20 (2023)

    2023

  9. [20]

    & Fan, Y

    Xu, Z., Li, K., Huang, Z., Yin, R. & Fan, Y. 3d salt body segmentation method based on multi-view co-regularization.IEEE Transactions on Geosci. Remote. Sens.(2024)

    2024

  10. [21]

    P.et al.Deep-salt: Complete three-dimensional salt segmentation from inaccurate migrated subsurface offset gathers using deep learning.Geophys

    Muller, A. P.et al.Deep-salt: Complete three-dimensional salt segmentation from inaccurate migrated subsurface offset gathers using deep learning.Geophys. Prospect.72, 2186–2199 (2024)

    2024

  11. [22]

    Yang, L.et al.Salt3dnet: A self-supervised learning framework for 3d salt segmentation.IEEE Transactions on Geosci. Remote. Sens.(2024)

    2024

  12. [23]

    & Liu, G

    Gao, H., Wu, X. & Liu, G. Channelseg3d: Channel simulation and deep learning for channel interpretation in 3d seismic images.Geophysics86, IM73–IM83 (2021)

    2021

  13. [24]

    Yu, J. & Wu, B. Attention and hybrid loss guided deep learning for consecutively missing seismic data reconstruction.IEEE Transactions on Geosci. Remote. Sens.60, 1–8 (2021)

    2021

  14. [25]

    Saad, O. M. & Alkhalifah, T. F-siamesefwi: A novel deep-learning framework for multisource full-waveform inversion.Geophysics90, R221–R230 (2025)

    2025

  15. [26]

    Wu, Y. & Ma, J. How does neural network reparametrization improve geophysical inversion?J. Geophys. Res. Mach. Learn. Comput.2, e2025JH000621 (2025)

    2025

  16. [27]

    Li, Y., Alkhalifah, T., Huang, J. & Li, Z. Self-supervised pretraining vision transformer with masked autoencoders for building subsurface model.IEEE Transactions on Geosci. Remote. Sens.61, 1–10 (2023)

    2023

  17. [28]

    neural information processing systems33, 1877–1901 (2020)

    Brown, T.et al.Language models are few-shot learners.Adv. neural information processing systems33, 1877–1901 (2020)

    2020

  18. [29]

    Touvron, H.et al.Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  19. [30]

    Bi, X.et al.Deepseek llm: Scaling open-source language models with longtermism.arXiv preprint arXiv:2401.02954(2024)

    2024 arXiv

  20. [31]

    InProceedings of the IEEE/CVF International Conference on Computer Vision, 4015–4026 (2023)

    Kirillov, A.et al.Segment anything. InProceedings of the IEEE/CVF International Conference on Computer Vision, 4015–4026 (2023)

    2023

  21. [32]

    neural information processing systems35, 35946–35958 (2022)

    Feichtenhofer, C., Li, Y., He, K.et al.Masked autoencoders as spatiotemporal learners.Adv. neural information processing systems35, 35946–35958 (2022)

    2022

  22. [33]

    & Abubakar, A

    Di, H., Truelove, L., Li, C. & Abubakar, A. Accelerating seismic fault and stratigraphy interpretation with deep cnns: A case study of the taranaki basin, new zealand.The Lead. Edge39, 727–733 (2020)

    2020

  23. [34]

    T., Chen, Y

    Schuster, G. T., Chen, Y. & Feng, S. Review of physics-informed machine-learning inversion of geophysical data.Geophysics89, T337–T356 (2024)

    2024

  24. [35]

    & Saygin, E

    Chen, Y. & Saygin, E. Seismic inversion by hybrid machine learning.J. Geophys. Res. Solid Earth126, e2020JB021589 (2021)

    2021

  25. [36]

    Wu, X.et al.Sensing prior constraints in deep neural networks for solving exploration geophysical problems. Proc. Natl. Acad. Sci.120, e2219573120 (2023)

    2023

  26. [37]

    & Molder, A

    Heir, A., Aghayev, S., Tran, C. & Molder, A. Inversion with stratigraphy-guided deep learning.Geophysics89, R377–R386 (2024)

    2024

  27. [38]

    In Second EAGE Digitalization Conference and Exhibition, vol

    Tilke, P.et al.Stratigraphic forward modeler for artificial intelligence and machine learning workflows. In Second EAGE Digitalization Conference and Exhibition, vol. 2022, 1–5 (European Association of Geoscientists & Engineers, 2022). 18/28

    2022

  28. [39]

    & Ovcharenko, O

    Kanfar, R., Alali, A., Tonellot, T.-L., Salim, H. & Ovcharenko, O. Intelligent seismic workflows: The power of generative ai and language models.The Lead. Edge44, 142–151 (2025)

    2025

  29. [40]

    Gao, H.et al.A foundation model enpowered by a multi-modal prompt engine for universal seismic geobody interpretation across surveys.arXiv preprint arXiv:2409.04962(2024)

    2024 arXiv

  30. [41]

    Liu, Q. & Ma, J. Foundation models for geophysics: Review and perspective.arXiv preprint arXiv:2406.03163 (2024)

    2024 arXiv

  31. [42]

    arXiv preprint arXiv:2504.17384(2025)

    Sheng, H.et al.On the workflow, opportunities and challenges of developing foundation model in geophysics. arXiv preprint arXiv:2504.17384(2025)

    2025 arXiv

  32. [43]

    Sheng, H.et al.Seismic foundation model (sfm): a next generation deep learning model in geophysics.Geophysics 90, 1–64 (2024)

    2024

  33. [44]

    Dou, Y., Li, K., Zhu, J., Li, X. & Xi, Y. Attention-based 3-d seismic fault segmentation training by a few 2-d slice labels.IEEE Transactions on Geosci. Remote. Sens.60, 1–15 (2021)

    2021

  34. [45]

    L.et al.Slow slip along the hikurangi margin linked to fluid-rich sediments trailing subducting seamounts.Nat

    Bangs, N. L.et al.Slow slip along the hikurangi margin linked to fluid-rich sediments trailing subducting seamounts.Nat. Geosci.16, 505–512 (2023)

    2023

  35. [46]

    J., Putzig, N

    Foss, F. J., Putzig, N. E., Campbell, B. A. & Phillips, R. J. 3d imaging of mars’ polar ice caps using orbital radar data.The Lead. Edge36, 43–57 (2017). 47.dGB. Opendtect projects. https://terranubis.com/datalist/free

    2017

  36. [48]

    & Brox, T

    Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, 234–...

    2015

  37. [49]

    & Adam, H

    Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. InProceedings of the European conference on computer vision (ECCV), 801–818 (2018)

    2018

  38. [50]

    Intell.43, 3349–3364, DOI: 10.1109/TPAMI.2020.2983686 (2021)

    Wang, J.et al.Deep high-resolution representation learning for visual recognition.IEEE Transactions on Pattern Analysis Mach. Intell.43, 3349–3364, DOI: 10.1109/TPAMI.2020.2983686 (2021)

    2021

  39. [51]

    Dou, Y.et al.Md loss: Efficient training of 3-d seismic fault segmentation network under sparse labels by weakening anomaly annotation.IEEE Transactions on Geosci. Remote. Sens.60, 1–14 (2022)

    2022

  40. [52]

    & Wang, Q

    Li, Y., Wu, X., Zhu, Z., Ding, J. & Wang, Q. Faultseg3d plus: A comprehensive study on evaluating and improving cnn-based seismic fault segmentation.Geophysics89, N77–N91 (2024)

    2024

  41. [53]

    Bi, Z., Wu, X., Geng, Z. & Li, H. Deep relative geologic time: a deep learning method for simultaneously interpreting 3-d seismic horizons and faults.J. Geophys. Res. Solid Earth126, e2021JB021882 (2021)

    2021

  42. [54]

    & Zhang, W

    Wang, G., Wu, X. & Zhang, W. cigchannel: A massive-scale 3d seismic dataset with labeled paleochannels for advancing deep learning in seismic interpretation.Earth Syst. Sci. Data Discuss.2024, 1–27 (2024)

    2024

  43. [55]

    & Zeng, H

    Wu, X., Yan, S., Qi, J. & Zeng, H. Deep learning for characterizing paleokarst collapse features in 3-d seismic images.J. Geophys. Res. Solid Earth125, e2020JB019685 (2020)

    2020

  44. [56]

    & Xiao, Y

    Dou, Y., Li, K., Lv, W., Li, T. & Xiao, Y. Contrasinver: Ultra-sparse label semi-supervised regression for multi-dimensional seismic inversion.IEEE Transactions on Geosci. Remote. Sens.(2024)

    2024

  45. [57]

    Wu, X., Yan, S., Bi, Z., Zhang, S. & Si, H. Deep learning for multidimensional seismic impedance inversion. Geophysics86, R735–R745 (2021)

    2021

  46. [58]

    Wu, B., Xie, Q. & Wu, B. Seismic impedance inversion based on residual attention network.IEEE Transactions on Geosci. Remote. Sens.60, 1–17 (2022)

    2022

  47. [59]

    pytorch-fid: FID Score for PyTorch

    Seitzer, M. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid (2020). Version 0.3.0

    2020

  48. [60]

    C., Sheikh, H

    Wang, Z., Bovik, A. C., Sheikh, H. R. & Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 600–612 (2004). 61.Goodfellow, I.et al.Generative adversarial networks.Commun. ACM63, 139–144 (2020)

    2004

  49. [62]

    InThe Eleventh International Conference on Learning Representations

    Tian, K.et al.Designing bert for convolutional networks: Sparse and hierarchical masked modeling. InThe Eleventh International Conference on Learning Representations. 19/28

  50. [63]

    A., Shechtman, E

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E. & Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, 586–595 (2018)

    2018

  51. [64]

    The relativistic discriminator: a key element missing from standard gan.arXiv preprint arXiv:1807.00734(2018)

    Jolicoeur-Martineau, A. The relativistic discriminator: a key element missing from standard gan.arXiv preprint arXiv:1807.00734(2018)

    2018 arXiv

  52. [65]

    & Hochreiter, S

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B. & Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Adv. neural information processing systems30(2017)

    2017

  53. [66]

    The national archive of marine seismic surveys, u.s

    USGS. The national archive of marine seismic surveys, u.s. geological survey. https://walrus.wr.usgs.gov/namss/ search/

  54. [67]

    Resource and energy georeference databases, south australian resources information gateway

    SARIG. Resource and energy georeference databases, south australian resources information gateway. https: //walrus.wr.usgs.gov/namss/search/

  55. [68]

    Dutch oil and gas portal, netherlands oil and gas exploration and production information

    NLOG. Dutch oil and gas portal, netherlands oil and gas exploration and production information. https: //www.nlog.nl/en. 69.SEG. Open data on the seg wiki, society of exploration geophysicists. https://wiki.seg.org/wiki/Open_data/. 70.Ravi, N.et al.Sam 2: Segment anything in i...

    2024 arXiv

  56. [69]

    For downstream tasks, the first 300,000 steps used an input size of320× 400 × 128with a batch size of 8, and each step 27/28 took approximately 1.7 seconds

    The model was trained for 500,000 steps, each taking approximately 1.8 seconds, totaling around 10 days. For downstream tasks, the first 300,000 steps used an input size of320× 400 × 128with a batch size of 8, and each step 27/28 took approximately 1.7 seconds. The remaining 2...

  57. [72]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976–11986 (2022)

    Liu, Z.et al.A convnet for the 2020s. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976–11986 (2022)

    2022

  58. [73]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16133–16142 (2023)

    Woo, S.et al.Convnext v2: Co-designing and scaling convnets with masked autoencoders. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16133–16142 (2023)

    2023

  59. [74]

    Tan, M. & Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning, 6105–6114 (PMLR, 2019)

    2019

  60. [75]

    Tan, M. & Le, Q. Efficientnetv2: Smaller models and faster training. InInternational conference on machine learning, 10096–10106 (PMLR, 2021)

    2021

  61. [76]

    & Kweon, I

    Woo, S., Park, J., Lee, J.-Y. & Kweon, I. S. Cbam: Convolutional block attention module. InProceedings of the European conference on computer vision (ECCV), 3–19 (2018). 20/28 Appendices A Challenges in Subsurface Imaging Tasks Figure A1.Challenges and strategies for generaliz...

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.