Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Physics-informed conditional diffusion model for generalizable elastic wave-mode separation

T0 review · 4 major / 6 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 The paper claims that a single physics-informed conditional diffusion model, trained once, can separate elastic wave modes as accurately as conventional numerical solvers but at much lower cost, and that this one training run generalizes…

desk verdict A plausible new diffusion-based P/S separation method whose own 18 Hz failure contradicts the central claim that it learned the frequency-independent separation operator. read the letter →

arxiv 2506.23007 v1 pith:3JWNWQGG submitted 2025-06-28 physics.geo-ph

Shijun Cheng , Xinru Mu , Tariq Alkhalifah This is my paper · ORCID
classification physics.geo-ph
keywords physics-informeddiffusionelasticwave-modeseparationconditionalmodelP-waveseismicimagingLaplacianequationsDDIMsamplinggeneralization
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

Elastic wave-mode separation is a costly step in seismic imaging: classical solvers apply the Laplacian separation operator to the wavefield, while fast neural substitutes often fail on geologies they were not trained on. This paper claims that a conditional diffusion model, trained once on simulated elastic wavefields, can directly predict the P-wave mode as accurately as numerical reference solutions but at a fraction of the cost, with no retraining when the velocity model, grid size, or snapshot time changes. The physical separation equations are embedded in two places: a training loss and a gradient-based correction during each reverse-diffusion step, and the authors attribute the model's out-of-sample success to this physics-in-the-loop design. If the claim holds, it points to a practical recipe for cheap, physically consistent wave-mode separation in large-scale seismic imaging workflows.

What carries the argument

The load-bearing identity is the Laplacian separation system of Equation 2, which expresses the Laplacian of each P-wave velocity component as mixed second derivatives of the original elastic wavefield, for example $\Delta V^p_x = \partial^2_x V_x + \partial_x \partial_z V_z$, with no explicit velocity dependence. The model is an x0-prediction conditional diffusion U-Net that estimates the clean P-wave mode directly, and the physics enters in two places: as the training loss $\mathcal{L}_{\mathrm{phys}}$ enforcing exactly those Laplacian identities, and as a per-step sampling correction $\hat{x}_{t-1} \leftarrow \hat{x}_{t-1} - \eta \nabla \mathcal{L}_{\mathrm{phys}}$ during reverse diffusion, which steers the DDIM trajectory toward separation-consistent wavefields. This combination is what lets the inference run in as few as 10 reverse steps and is credited for the generalization across unseen models.

What would settle it

Generate an elastic wavefield with a source frequency outside the training band, for example a 15 Hz Ricker wavelet on the same SEAM Arid model, run the trained PICDM, and compare the predicted P-wave mode against the numerical reference of Equation 2: the paper's claim predicts a close match, whereas significant leakage or artifacts would show the model leaned on training-distribution cues. The paper's own 18 Hz experiment already fails in this way, which bounds the claim to the training-frequency band unless multi-frequency training is added.

Watch

Extended reading notes

Core claim

The central claim is that a single physics-informed conditional diffusion model (PICDM), trained once, can replace the numerical evaluation of the separation operator in Equation 2: conditioned on the elastic wavefield $(V_x, V_z)$ and the P- and S-wave velocity models, it outputs the clean P-wave mode $(\hat{V}^p_x, \hat{V}^p_z)$, and the S-wave mode is obtained by subtraction. In tests on a homogeneous medium, SEAM Arid, Overthrust, Marmousi 2, and Otway velocity models, on a larger $300 \times 640$ grid, and on a snapshot at 2.0 s (beyond the 1.5 s training window), the predicted P-wave modes closely match the conventional numerical solutions. The authors interpret this as the network internalizing the separation operator rather than memorizing training cases, since Equation 2 contains no explicit velocity-dependent terms; their ablation shows that removing velocity conditioning keeps in-distribution accuracy but degrades out-of-distribution separation, while an untrained source frequency of 18 Hz breaks the model entirely.

Load-bearing premise

The generalization claim rests on the assumption that during training the network internalized the actual separation operator rather than patterns tied to the 12 Hz source, the 10 m grid spacing, the constant density, the 0.52 to 0.7 $v_p/v_s$ scaling, and the 1.5 s simulation window, so that it will keep separating correctly when those quantities change.

Editorial extensions

If this is right

  • Wave-mode separation for large or 3D surveys could run at the cost of a few DDIM steps instead of pseudo-spectral operator evaluations, making high-resolution imaging pipelines substantially cheaper.
  • The model can be applied to geological models it has never seen, including faulted structures and realistic heterogeneous velocity distributions, without retraining or fine-tuning.
  • Since the S-wave mode is computed by subtracting the predicted P-wave mode from the elastic wavefield, a single pass yields both modes in a mutually consistent way.
  • Physics-guided sampling keeps separation quality stable with as few as 10 reverse steps, reducing inference time enough for interactive or high-throughput workflows.
  • The demonstrated temporal extrapolation to a 2.0 s snapshot suggests the trained operator generalizes beyond the training window, though the 4.0 s failure bounds that reach.

Reading between the lines

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

  • The 18 Hz failure is a built-in test of the learned-operator story, because Equation 2 contains no frequency terms; training on multi-peak-frequency data would both test and likely extend the model's envelope.
  • The same recipe, conditioning a diffusion model on a known constraint and correcting samples by gradient backpropagation of that constraint, could be applied to other expensive wavefield operators such as up/down-going decomposition or attenuation compensation.
  • Because the velocity-free ablation degrades only out of distribution, the velocity inputs may function as a contextual key that aligns wavefields and velocity structures with the training manifold; testing on $v_p/v_s$ ratios outside the 0.52 to 0.7 range would clarify whether the network truly learned an operator or a conditional lookup.
  • The bounded temporal extrapolation (success at 2.0 s, artifacts at 4.0 s) suggests the generalization envelope is set by the variety of wavefront complexity in the training window, so longer simulations or more snapshots per model are a concrete path to widening it.
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

4 major / 6 minor

Summary. The paper proposes a physics-informed conditional diffusion model (PICDM) for separating P-wave modes from elastic wavefields in 2D isotropic media. The method trains an x0-prediction diffusion model conditioned on the full elastic wavefield components and P/S velocity models, with a training loss that combines an ℓ2 data term and a physics-informed term based on the Laplacian separation equations of Zhu (2017). At inference, the same physics loss is used as a gradient-based correction during DDIM sampling. The authors evaluate the model on homogeneous media, SEAM Arid, Overthrust, Marmousi 2, Otway, a larger 300×640 model, and temporal extrapolation, and they include ablations without velocity conditioning and with an 18 Hz source. The paper claims that a single trained model achieves near-perfect agreement with numerical references across all these scenarios and that the model has effectively learned the frequency-independent separation operator of Eq. (2).

Significance. If the central claims hold, the contribution is potentially useful: a single diffusion model that replaces costly pseudo-spectral separation and generalizes across velocity models, model sizes, and time windows would be a practical advance, and the idea of embedding physics both in training and sampling is appealing. The paper also deserves credit for using a standard diffusion formulation, building on an open-source architecture, providing an ablation of velocity conditioning, and honestly disclosing the 18 Hz failure. However, the current evidence does not support the strongest claims: there are no quantitative error metrics or baseline comparisons, and the 18 Hz result directly contradicts the proposed operator-learning mechanism. The work is promising but needs substantial revision before the claims can be accepted.

major comments (4)
  1. [Section 4, Figure 16] The claim that the model 'has effectively learned the underlying separation operator described in Equation 2' is directly contradicted by the 18 Hz experiment. Equation 2 is a purely spatial differential relation with no dependence on source frequency, so a model that had internalized this operator should separate an 18 Hz wavefield as accurately as a 12 Hz one. The complete failure shown in Figure 16 indicates instead that the learned mapping is tied to the 12 Hz training distribution. This undermines the mechanism proposed in support of the 'learn once, apply everywhere' paradigm. Please either revise the mechanism claim or train with multi-frequency data and demonstrate that the operator-level generalization actually holds.
  2. [Sections 3.2–3.7 and Section 5] The central quantitative claims of 'almost perfectly match', 'near-perfect agreement', and 'minimal residual discrepancies' are supported only by qualitative residual images. No numerical error metrics (e.g., MSE, SNR, or normalized error) are reported for any test case, no error bars or repeated-run statistics are given, and no comparison is made against the existing learned baselines cited in the introduction (e.g., Wang and Ma 2019; Kaur et al. 2021; Meng et al. 2025). Please report per-experiment accuracy metrics, state the color scale values used in the residual plots, and benchmark against at least one prior learned separation method.
  3. [Sections 2.3–2.4, Eqs. (8)–(11)] The physics loss is not an independent check of physical consistency. The training labels are generated by the same Equation 2 that defines Lphys, so Ldata and Lphys are correlated by construction. More importantly, Equation 9 constrains only the Laplacian of the predicted P-wave field, and many different fields share the same Laplacian; the inference correction in Equation 11 minimizes the same residual and therefore cannot by itself select the correct P-wave field. The actual selection is performed by the data-driven prior learned from the training distribution. Please analyze this non-uniqueness explicitly and clarify what the physics term adds beyond a regularizer.
  4. [Section 3.7, Figures 12–13] The temporal extrapolation claim should be qualified quantitatively. The paper states that the method 'continues to accurately recover' the P-wave mode at 2.0 s but shows visible artifacts at 4.0 s; without error metrics it is unclear where the reliable extrapolation limit lies and how this relates to the 1.5 s training window. Please provide a time-sweep error curve or otherwise specify the maximum reliable extrapolation time supported by the data.
minor comments (6)
  1. [Section 3.1] The framework is referred to as 'PIGDM' in the opening of Section 3.1; this appears to be a typo for 'PICDM' and should be corrected.
  2. [Section 3.3] In the description of Figure 6, the text says 'following the same panel layout as in Figure 6' when it should refer to Figure 5; the panel layouts of the two figures differ because Figure 6 omits the velocity-model panels.
  3. [Section 3.2 and Figure 4] The accuracy metric in Figure 4 is described as MSE, but the y-axis labels and units are not provided, and no numerical values are quoted in the text; please add axis labels and state the numerical ranges.
  4. [Sections 2.4 and 3.1] The hyperparameters λ in Eq. (10) and η in Eq. (11) are never specified, and no sensitivity analysis is reported; please provide their values and, if possible, a brief sensitivity study.
  5. [Abstract and Section 3.3] The claim of 'reduced cost' is not supported by any wall-clock comparisons; please provide inference-time measurements against the numerical reference solver for at least the larger test cases.
  6. [Section 7] The code repository is currently private; to facilitate review, please provide an anonymized copy or specify a concrete release date.

Circularity Check

1 steps flagged · score 2.0 of 10

Central derivation is self-contained; only the physics-loss demonstration is tautological because Eq. 11 directly minimizes Eq. 9, while the 18 Hz failure is a generalization limitation, not circularity.

  1. self definitional [Section 2.4, Eq. (11); Figure 3; Section 3.2]
    "The intermediate wavefield may still violate Equation (2). Therefore, we further apply a physics-guided correction to the intermediate prediction ˆxt−1 = (Vˆ p x , Vˆ p z ) by evaluating the physics-informed loss Lphys (e.g., Equation 9) and backpropagating its gradient: (Vˆ p x , Vˆ p z ) ← (Vˆ p x , Vˆ p z ) − η ∇Lphys. ... We can find that the diffusion model with physics guidance converges to a very low physical residual in far fewer steps."

    Equation 11 is a gradient-descent step on Lphys, so the reduction in Lphys during physics-guided sampling (Figure 3) is true by construction; it shows the optimizer minimizing its own objective, not an independent check that the model has learned physics. Moreover, Lphys (Eq. 9) is the residual form of the same Zhu 2017 Laplacian operator (Eq. 2) used to generate the training labels, so the physics constraint and the reference solution are not fully independent. The main accuracy claim, however, is measured as MSE against the numerical reference on held-out wavefields, which is an external benchmark, so the central generalization result does not reduce to this loss.

full rationale

The paper's central claim—a single trained PICDM separates P-wave modes on in-distribution, out-of-distribution, larger, and temporally extrapolated models without retraining—is supported by held-out comparisons against numerical reference solutions obtained with Zhu's method. Those comparisons are legitimate supervised benchmarks, and no parameter is fitted to the test cases. The physics-informed loss (Eq. 9) and inference correction (Eq. 11) restate the same Laplacian separation operator (Eq. 2) that generates the labels; this is a standard physics-informed setup, but it means the physical-loss plots are a definitional consequence of the algorithm rather than independent evidence. That is a minor circularity in a supporting metric, not in the derivation itself. The paper's admitted 18 Hz failure (Figure 16) and the velocity-free ablation (Figures 14-15) undercut the speculation that the model 'has effectively learned the underlying separation operator' in a frequency- and velocity-independent sense, but these are generalization limitations honestly disclosed by the authors, not circular reasoning. The self-citations (SeparationPINN, Cheng et al. 2025) are motivational or analogical and do not carry the argument.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on standard diffusion-model mathematics and on Zhu's Laplacian separation operator as a definition of the P-wave target. No new physical entities are introduced. The main uncharged assumptions are that the numerical labels are correct ground truth and that the 150k-example training distribution supports the claimed generalization.

free parameters (4)
  • lambda (physics loss weight, Eq. 10) = not reported
    Hand-chosen hyperparameter balancing Ldata and Lphys; central to the 'physics-informed' training claim.
  • eta (physics correction step size, Eq. 11) = not reported
    Controls gradient correction strength during inference; impacts accuracy and sampling cost.
  • DDIM sampling steps = 10
    Chosen after the analysis in Figure 4 to balance accuracy and speed; all main results use 10 steps.
  • S-wave velocity scaling range = 0.52 to 0.7 of vp
    Used to generate training S-velocity models; Otway provides one independent test, but the training distribution is narrow.
assumptions (5)
  • domain assumption Zhu 2017 Laplacian equations (Eq. 2) define the correct P-wave mode and provide accurate training labels.
    Invoked in Sections 2.1 and 3.1; labels for training and the physics loss both come from this operator.
  • domain assumption The first-order velocity-stress equations (Eq. 1) and the 10th-order finite-difference scheme produce physically accurate training wavefields.
    Invoked in Sections 2.1 and 3.1; all wavefields are synthetic.
  • standard math The x0-prediction diffusion model with DDIM sampling approximates samples from the target conditional distribution.
    Standard results from Ho et al. 2020 and Song et al. 2020, used in Sections 2.2 to 2.4.
  • domain assumption A network trained on 150k examples from 6000 velocity models generalizes to unseen velocity models, larger grids, and later times.
    This is the paper's central generalization claim; supported only by a handful of qualitative tests and challenged by the 18 Hz failure.
  • domain assumption Constant density, fixed 12 Hz source, and vp/vs derived by scalar multiplication are representative of the target application.
    Training set uses these; Otway is the only independent vp/vs test, and no density variation is tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-informed conditional diffusion model for generalizable elastic wave-mode separation." pith.science (2026). https://pith.science/paper/3JWNWQGG

@misc{pith2026250623007,
  author       = {Pith},
  title        = {Pith review of: Physics-informed conditional diffusion model for generalizable elastic wave-mode separation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3JWNWQGG}},
  note         = {Machine review of arXiv:2506.23007}
}
read the original abstract

Traditional elastic wavefield separation methods, while accurate, often demand substantial computational resources, especially for large geological models or 3D scenarios. Purely data-driven neural network approaches can be more efficient, but may fail to generalize and maintain physical consistency due to the absence of explicit physical constraints. Here, we propose a physics-informed conditional diffusion model for elastic wavefield separation that seamlessly integrates domain-specific physics equations into both the training and inference stages of the reverse diffusion process. Conditioned on full elastic wavefields and subsurface P- and S-wave velocity profiles, our method directly predicts clean P-wave modes while enforcing Laplacian separation constraints through physics-guided loss and sampling corrections. Numerical experiments on diverse scenarios yield the separation results that closely match conventional numerical solutions but at a reduced cost, confirming the effectiveness and generalizability of our approach.

Figures

Figures reproduced from arXiv: 2506.23007 by the authors.

Figure 1
Figure 1. Schematic of our conditional diffusion model for P-wave mode separation. Forward diffusion (solid arrow) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of snapshots at t = 0.4 s of the P-wave mode separation results for a source at the center of the homogeneous model. The original (a) horizontal and (e) vertical elastic wavefields. The separated P-wave mode in the horizontal (b) and vertical (f) directions obtained using the traditional numerical method. The separated P-wave mode in the horizontal (c) and vertical (g) directions obtained using the our me… view at source ↗
Figure 3
Figure 3. Comparison of physics loss Lphys during the DDIM sampling process for cases without (orange) and with (blue) physics guidance. 3.3 SEAM Arid model Next, we test our PICDM method on a more complex in-distribution scenario using the SEAM Arid velocity model [Oristaglio, 2015]. We select a layered region from the SEAM Arid model that is excluded from the training dataset, obtaining a distinct P-wave velocity profile. T… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: The impact of DDIM sampling step count (1, 5, 10, 25, 50, and 100 steps) on wavefield separation quality. (a) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison of snapshots at t = 1 s of the P-wave separation results for a source in the middle of the top surface in the SEAM Arid model. (a) P- and (f) S-wave velocities of the layered profiles extracted from SEAM Arid model. The original (b) horizontal and (g) vertic…
Figure 6
Figure 6. Figure 6: Separation performance for SEAM Arid model with the source placed at the center of the model at [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Except for velocity models, the arrangement of other panels is the same as in Figure [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Separation performance for a larger and more realistic Marmousi model (size [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Separation performance for the modified Marmousi 2 model (see Figures [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

  1. [1]

    2d and 3d elastic wavefield vector decomposition in the wavenumber domain for vti media

    Qunshan Zhang and George A McMechan. 2d and 3d elastic wavefield vector decomposition in the wavenumber domain for vti media. Geophysics, 75 0 (3): 0 D13--D26, 2010

    work page 2010

  2. [2]

    Fast algorithms for elastic-wave-mode separation and vector decomposition using low-rank approximation for anisotropic media

    Jiubing Cheng and Sergey Fomel. Fast algorithms for elastic-wave-mode separation and vector decomposition using low-rank approximation for anisotropic media. Geophysics, 79 0 (4): 0 C97--C110, 2014

    work page 2014

  3. [3]

    Quantitative seismology

    Keiiti Aki and Paul G Richards. Quantitative seismology. 2002

    work page 2002

  4. [4]

    Elastic wave-mode separation for vti media

    Jia Yan and Paul Sava. Elastic wave-mode separation for vti media. Geophysics, 74 0 (5): 0 WB19--WB32, 2009

    work page 2009

  5. [5]

    Wave-field separation in two-dimensional anisotropic media

    J Dellinger and J Etgen. Wave-field separation in two-dimensional anisotropic media. Geophysics, 55 0 (7): 0 914--919, 1990

    work page 1990

  6. [6]

    Improving the efficiency of elastic wave-mode separation for heterogeneous tilted transverse isotropic media

    Jia Yan and Paul Sava. Improving the efficiency of elastic wave-mode separation for heterogeneous tilted transverse isotropic media. Geophysics, 76 0 (4): 0 T65--T78, 2011

    work page 2011

  7. [7]

    Simulating propagation of decoupled elastic waves using low-rank approximate mixed-domain integral operators for anisotropic media

    Jiubing Cheng, Tariq Alkhalifah, Zedong Wu, Peng Zou, and Chenlong Wang. Simulating propagation of decoupled elastic waves using low-rank approximate mixed-domain integral operators for anisotropic media. Geophysics, 81 0 (2): 0 T63--T77, 2016

    work page 2016

  8. [8]

    Elastic wavefield separation based on the helmholtz decomposition

    Hejun Zhu. Elastic wavefield separation based on the helmholtz decomposition. Geophysics, 82 0 (2): 0 S173--S183, 2017

    work page 2017

Show all 29 references
  1. [9]

    Elastic wavefield separation in anisotropic media based on eigenform analysis and its application in reverse-time migration

    Jidong Yang, Houzhu Zhang, Yang Zhao, and Hejun Zhu. Elastic wavefield separation in anisotropic media based on eigenform analysis and its application in reverse-time migration. Geophysical Journal International, 217 0 (2): 0 1290--1313, 2019

    2019

  2. [10]

    Attenuation-compensated viscoelastic reverse time migration with finite difference operators

    Xinru Mu, Tariq Alkhalifah, and Jianping Huang. Attenuation-compensated viscoelastic reverse time migration with finite difference operators. Geophysics, 90 0 (3): 0 1--53, 2024

    2024

  3. [11]

    Ps decomposition of isotropic elastic wavefields using cnn-learned filters

    W Wang and J Ma. Ps decomposition of isotropic elastic wavefields using cnn-learned filters. In 81st EAGE Conference and Exhibition 2019, volume 2019, pages 1--5. European Association of Geoscientists & Engineers, 2019

    2019

  4. [12]

    A fast algorithm for elastic wave-mode separation using deep learning with generative adversarial networks (gans)

    Harpreet Kaur, Sergey Fomel, and Nam Pham. A fast algorithm for elastic wave-mode separation using deep learning with generative adversarial networks (gans). Journal of Geophysical Research: Solid Earth, 126 0 (9): 0 e2020JB021123, 2021

    2021

  5. [13]

    Deep learning-based p-and s-wave separation for multicomponent vertical seismic profiling

    Yanwen Wei, Yunyue Elita Li, Jingjing Zong, Jizhong Yang, Haohuan Fu, and Mengyao Sun. Deep learning-based p-and s-wave separation for multicomponent vertical seismic profiling. IEEE Transactions on Geoscience and Remote Sensing, 60: 0 1--16, 2021

    2021

  6. [14]

    P/s-wave separation of multicomponent seismic data at the land surface based on deep learning

    He Huang, Tengfei Wang, and Jiubing Cheng. P/s-wave separation of multicomponent seismic data at the land surface based on deep learning. Geophysics, 88 0 (3): 0 V233--V247, 2023

    2023

  7. [15]

    Up/down and p/s decomposition of distributed acoustic sensing-vertical seismic profiling data using multitask deep-learning method

    Tao Meng, Tengfei Wang, Jiubing Cheng, Pengfei Duan, and Zhonglin Cao. Up/down and p/s decomposition of distributed acoustic sensing-vertical seismic profiling data using multitask deep-learning method. Geophysics, 90 0 (4): 0 V283--V296, 2025

    2025

  8. [16]

    Separationpinn: Physics-informed neural networks for seismic p- and s-wave mode separation

    Xinru Mu, Shijun Cheng, and Tariq Alkhalifah. Separationpinn: Physics-informed neural networks for seismic p- and s-wave mode separation. IEEE Transactions on Geoscience and Remote Sensing, pages 1--1, 2025

    2025

  9. [17]

    Sh-wave propagation in heterogeneous media: Velocity-stress finite-difference method

    Jean Virieux. Sh-wave propagation in heterogeneous media: Velocity-stress finite-difference method. Geophysics, 49 0 (11): 0 1933--1942, 1984

    1933

  10. [18]

    P-sv wave propagation in heterogeneous media: Velocity-stress finite-difference method

    Jean Virieux. P-sv wave propagation in heterogeneous media: Velocity-stress finite-difference method. Geophysics, 51 0 (4): 0 889--901, 1986

    1986

  11. [19]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  12. [20]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162--8171. PMLR, 2021

    2021

  13. [21]

    The 2004 bp velocity benchmark

    FJ Billette and Sverre Brandsberg-Dahl. The 2004 bp velocity benchmark. In 67th EAGE Conference & Exhibition, pages cp--1. European Association of Geoscientists & Engineers, 2005

    2004

  14. [22]

    3-d salt and overthrust seismic models

    Fred Aminzadeh. 3-d salt and overthrust seismic models. 1996

    1996

  15. [23]

    Survey design and implementation of a multiple wide-azimuth towed streamer seismic acquisition strategy at the tiber discovery; deep water gulf of mexico, usa

    John C Naranjo, Eric J Ekstrand, John Etgen, Kenneth L Hargrove, James G Brewton, Oswaldo Garcia, Georgiy Astvatsaturov, Damian Hite, Bill Howieson, Bing Bai, et al. Survey design and implementation of a multiple wide-azimuth towed streamer seismic acquisition strategy at the ...

    2011

  16. [24]

    Application of the perfectly matched absorbing layer model to the linear elastodynamic problem in anisotropic heterogeneous media

    Francis Collino and Chrysoula Tsogka. Application of the perfectly matched absorbing layer model to the linear elastodynamic problem in anisotropic heterogeneous media. Geophysics, 66 0 (1): 0 294--307, 2001

    2001

  17. [25]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  18. [26]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  19. [27]

    Seam update: The arid model—seismic exploration in desert terrains

    Michael Oristaglio. Seam update: The arid model—seismic exploration in desert terrains. The Leading Edge, 34 0 (4): 0 466--468, 2015

    2015

  20. [28]

    Marmousi2: An elastic upgrade for marmousi

    Gary S Martin, Robert Wiley, and Kurt J Marfurt. Marmousi2: An elastic upgrade for marmousi. The leading edge, 25 0 (2): 0 156--166, 2006

    2006

  21. [29]

    Seismic wavefield solutions via physics-guided generative neural operator

    Shijun Cheng, Mohammad H Taufik, and Tariq Alkhalifah. Seismic wavefield solutions via physics-guided generative neural operator. arXiv preprint arXiv:2503.06488, 2025

    2025 arXiv

Pith tools

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