Pith. sign in

REVIEW 3 major objections 4 minor 34 references

DiffPINN: Generative diffusion-initialized physics-informed neural networks for accelerating seismic wavefield representation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A latent diffusion model can generate PINN initialization parameters that make seismic wavefield modeling train faster and land on higher-accuracy solutions than random or meta-learned starts.

desk verdict DiffPINN is a plausible adaptation of neural network diffusion to PINNs, but the paper's compression claim is contradicted by its own numbers and the evidence is mostly qualitative. read the letter →

arxiv 2506.00471 v1 pith:ODI5IWU4 submitted 2025-05-31 physics.geo-ph cs.LGphysics.comp-ph

Shijun Cheng , Tariq Alkhalifah This is my paper
classification physics.geo-phcs.LGphysics.comp-ph
keywords physics-informedneuralnetworkslatentdiffusionmodelsparameterinitializationseismicwavefieldmodelingHelmholtzequationautoencoderout-of-distributiongeneralization
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

The paper proposes DiffPINN, a method that treats the final weights of trained physics-informed neural networks (PINNs) as data points. It compresses the weights of many PINNs, each solving the frequency-domain scattered-wave Helmholtz equation for a different velocity model, into a low-dimensional latent space with an autoencoder, then trains a diffusion model on those latent vectors conditioned on velocity models. At inference, a new velocity model is fed to the diffusion model, and the decoded parameter vector is used to initialize a fresh PINN, with a physics-loss gradient step during sampling. The authors claim this yields faster convergence and higher accuracy than random initialization and than their earlier meta-learned initialization, both for velocity models inside and outside the training distribution. The payoff, if true, is that expensive PINN retraining for each new subsurface model can be replaced by a generative warm start.

What carries the argument

The central object is the latent diffusion prior over PINN parameters. The pipeline works by flattening each converged PINN's 128,770 weights into a vector, training a 1D convolutional autoencoder to produce a 128×1590 latent code, and training a conditional 1D U-Net diffusion model—with x0 prediction and velocity/source embeddings as conditions—to generate latent codes for new velocity models. During sampling, each latent is decoded and a small gradient step on the physical loss (the Helmholtz residual) is taken with respect to the latent before the next denoising step; the final decoded vector is the PINN initialization.

What would settle it

A direct check: the autoencoder's latent tensor is 128×1590 = 203,520 numbers, larger than the 128,770-number flattened parameter vector, so the claimed compression does not reduce dimensionality as stated; one could test the manifold assumption by measuring reconstruction error on held-out converged parameter vectors and by checking whether generated weights produce a lower initial physics loss than random weights for a velocity model far outside the training distribution.

Watch

Extended reading notes

Core claim

The paper's central claim is that converged PINN parameter vectors for different velocity models can be modeled as samples from a learnable distribution, and that sampling from that distribution conditioned on a target velocity produces an initialization that is already close to a good solution. The authors demonstrate that a latent diffusion model trained on 2600 PINN parameter sets, each obtained by solving the scattered-wave Helmholtz equation for a member of the CurveVel-A class, produces weights for previously unseen velocity models—including out-of-distribution structures such as FlatFault and Marmousi—that train to higher scattered-wavefield accuracy within 500–2000 epochs than both a meta-learned initialization and a random one. They also show that injecting a physics-loss gradient into each reverse-diffusion step improves the generated weight quality, and that 10 DDIM steps strike their chosen balance between generation cost and downstream training performance.

Load-bearing premise

The method assumes that converged PINN weights for different velocity models lie on a smooth, low-dimensional manifold that an autoencoder can compress without losing the information needed to reconstruct a good wavefield solver.

Editorial extensions

If this is right

  • Warm-starting PINNs from generated weights cuts the early-training plateau that random initialization suffers, letting a 5 Hz scattered wavefield match the finite-difference reference within 2000 epochs.
  • The same pipeline works for velocity models outside the training distribution, though with a visible accuracy drop, pointing to broader training velocity diversity as the path to better generalization.
  • Physics guidance during diffusion sampling, not just the learned prior, is responsible for a meaningful share of the speedup: removing it lowers both convergence and final accuracy.
  • Ten reverse-diffusion steps give most of the benefit, while one step is markedly worse and hundreds of steps add cost for marginal gains, supporting the practical choice of a short DDIM schedule.

Reading between the lines

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

  • The framework is architecture-agnostic, so a similar latent diffusion prior could be built for other PINN-based PDE solvers, although the computational cost of gathering converged weights for each new problem class remains a barrier.
  • Because the conditioning vector includes source coordinates, the same generator could in principle output initializations for new source placements, not just new velocity models, something the current experiments do not isolate.
  • The finding that the autoencoder latent has more dimensions than the parameter vector suggests that the compression claim may need to be reinterpreted; the benefit could come from the diffusion model's denoising and the physics-guided sampling rather than from genuine dimensionality reduction.
  • A natural testable extension is to condition the diffusion model directly on low-frequency wavefield observations and generate initializations for full-waveform inversion, where the paper explicitly foresees an implicit-FWI use case.
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

3 major / 4 minor

Summary. The manuscript proposes DiffPINN, a method to generate initialization parameters for physics-informed neural networks (PINNs) that represent frequency-domain scattered seismic wavefields. The approach consists of four stages: (1) training many PINNs for different velocity models, (2) training a 1D convolutional autoencoder on the flattened converged parameter vectors, (3) training a conditional latent diffusion model on the autoencoder latent codes with velocity and source coordinates as conditioning, and (4) at inference, sampling a latent code for a new velocity model using a DDIM sampler with additional physics-guided gradient steps, then decoding it into a PINN initialization. The authors report experiments on 2,600 velocity models from OpenFWI, comparing DiffPINN against a meta-learned initialization (Meta-PINN) and random initialization on both in-distribution and out-of-distribution velocity models, and claim that DiffPINN accelerates training and improves accuracy.

Significance. If the claims are substantiated, DiffPINN would be a useful contribution to PINN-based seismic wavefield modeling and to the broader idea of generative initialization of neural network parameters. The study is ambitious: it involves training 2,600 PINNs, a large autoencoder, and a diffusion model, and it evaluates on external finite-difference reference solutions, which is a strength. The core concept, extending Neural Network Diffusion to physics-informed solvers with physics-guided sampling, is timely and has clear potential impact. However, the manuscript currently contains a direct internal contradiction about the claimed dimensionality reduction, and the experimental evidence is presented only as averaged curves without error bars or end-to-end timing. These issues affect the central claims of the paper.

major comments (3)
  1. [§3.5 and §3.2] Section 3.5 states that the flattened PINN parameter vector has size 1×128,770, while the autoencoder produces a latent tensor of size 128×1590, i.e., 203,520 numbers. The latent representation is therefore about 1.58 times larger than the parameter vector it is supposed to compress, contradicting the claim in Section 3.2 that the autoencoder reduces each parameter vector into a 'much lower-dimensional latent representation.' Because there is no bottleneck, the MSE loss in Eq. (10) can be minimized by a near-identity encoding, and the diffusion model is trained in a space larger than the original parameter space. This undermines the central premise that the latent space is a low-dimensional manifold on which a diffusion prior can generalize, and it makes the reported latent-map interpretations in Section 5.1 unreliable. The authors must either report the actual latent dimension after accounting for the downsampling operation, or redesign the autoencoder so that the latent representation is genuinely lower-dimensional than 128,770, and then show that this compressed representation still supports the claimed acceleration and accuracy.
  2. [§4.2 and §4.3, Figures 4, 6, and 9] The central experimental claim that DiffPINN 'significantly accelerates training and maintains high accuracy' is supported only by loss and accuracy curves averaged over five in-distribution and five out-of-distribution models, with no error bars, no per-model numerical values, and no statistical significance testing. Figure 4 shows a 'modest improvement' in PDE loss over Meta-PINN, while Figure 6 shows Meta-PINN converging faster in loss but DiffPINN being more accurate; this mixed behavior is not quantified in a way that supports the word 'significantly.' Additionally, no autoencoder reconstruction error or diffusion sampling quality metrics are reported, so the reader cannot judge whether the generated initializations are actually close to converged parameter vectors. The authors should provide per-model tables, error bars or confidence intervals, and, ideally, a statistical comparison across the ten test models.
  3. [§3.4 and §4.1] The claimed acceleration is not evaluated end-to-end. The training curves in Figures 4, 6, 9, 12, and 13 start after the initialization is already available, but the cost of generating that initialization is not included in the comparison. In particular, the physics-guided DDIM inference in Eq. (21) requires, at each of the T reverse-diffusion steps, decoding the latent through the autoencoder decoder, evaluating the physics loss Lphys on the collocation points, and backpropagating through the decoder to update the latent. With 20,000 collocation samples, this is nontrivial, and the paper does not report the wall-clock time for this generation stage or compare it against the time saved during PINN training. Without this accounting, the statement that DiffPINN 'significantly accelerates' PINN training is incomplete. The authors should report end-to-end time from a new velocity model to a converged PINN, including the diffusion sampling and physics-guidance steps, for all compared methods.
minor comments (4)
  1. [Abstract and Section 2.3] The abstract contains a duplicated word ('of of'), and the heading of Section 2.3 reads 'connection betwen' instead of 'connection between.'
  2. [Figure 3 caption and Section 3.5] The caption of Figure 3 says the autoencoder compresses the parameter vector 'into a 128 × 1590 latent vector,' which is the same number that appears in Section 3.5; the discrepancy with the claimed dimensionality reduction should be resolved and the description made consistent.
  3. [Section 3.5] The text contains formatting and grammatical errors, such as 'a last1 × 5 convolution' and 'we first applies a 1 × 5 convolution'; these should be corrected.
  4. [Section 8] The code repository is stated to be private during the review process. For reproducibility, the authors should provide an anonymized version or a detailed supplement that includes the exact network architectures, hyperparameters, and data-generation pipeline.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the diffusion-generated initialization is not a restatement of its training data, and the central evaluation is against independent finite-difference references.

full rationale

The paper's claimed acceleration is tested against unmodified finite-difference reference wavefields, an external benchmark, and against random and meta-learned initialization baselines. The diffusion pipeline (autoencoder encoding/decoding, conditional latent diffusion, then decoding to initial PINN weights) does not define the output as the fitted input: Eq. (16) trains the diffusion model to reconstruct latent codes from corrupted versions, and Eq. (21) applies a physics-loss gradient during sampling, but neither equation forces the final wavefield accuracy to equal the training data. The only notable self-citations (Cheng & Alkhalifah 2025b, 2025c) are used as a computational bootstrap to obtain the 2600 converged PINN parameter vectors and as suggested future extensions; they do not substitute for the experimental comparison. One non-circular internal inconsistency is present: Section 3.5 reports a latent tensor of size 128 x 1590 = 203,520 numbers while the flattened PINN vector is 1 x 128,770, so the 'much lower-dimensional latent representation' premise is numerically unsupported and the autoencoder may not actually compress. That undermines the generalization argument and leaves the acceleration claim less well supported, but it is not a circular derivation. No equation or fitted parameter is renamed as a prediction, and no load-bearing argument reduces to a self-citation chain. Overall, no circular step is exhibited.

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

The central claim rests on several modeling assumptions and hand-chosen hyperparameters. No new physical entities are introduced.

free parameters (4)
  • Latent representation dimensions (128x1590) = 128x1590 = 203,520
    Chosen by hand in Section 3.5. The paper calls it low-dimensional, but it exceeds the 128,770-dim flattened parameter vector, so this choice does not compress.
  • Reverse diffusion sampling steps (default 10) = 10
    Default in Section 4.1. Section 5.3 shows 100, 500, and 1000 steps yield better accuracy, so 10 is a cost/accuracy tradeoff.
  • Physics guidance step size gamma = not reported
    Introduced in Eq. (21), Section 3.4. No value or schedule is given, so the method cannot be reproduced.
  • Meta-learning initialization hyperparameters = inner LR 2e-3, outer LR 1.5e-3
    Inherited from the authors' prior Meta-PINN work. All 2600 training PINNs are trained from this initialization, so the parameter dataset depends on these choices.
assumptions (4)
  • domain assumption The scattered wavefield satisfies the perturbation Helmholtz equation (4) with background field (5).
    Standard in seismic scattering; the PINN loss is built on it.
  • domain assumption A meta-learned initialization (Cheng and Alkhalifah, 2025b) yields converged PINN solutions spanning the velocity model space.
    All 2600 PINNs are trained from this initialization; the diversity of the parameter dataset depends on this assumption.
  • domain assumption The autoencoder reconstructs PINN parameters with enough fidelity that decoded latent vectors remain valid initializations.
    No reconstruction error is reported. Eq. (10) is minimized but no threshold or evaluation is given.
  • ad hoc to paper A fixed-size condition vector c from sampled coordinates and velocities captures the relevant velocity model variability.
    The paper does not specify how 20,000 sampled points are aggregated into c in Eq. (15). This is a modeling assumption specific to this work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffPINN: Generative diffusion-initialized physics-informed neural networks for accelerating seismic wavefield representation." pith.science (2026). https://pith.science/paper/ODI5IWU4

@misc{pith2026250600471,
  author       = {Pith},
  title        = {Pith review of: DiffPINN: Generative diffusion-initialized physics-informed neural networks for accelerating seismic wavefield representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODI5IWU4}},
  note         = {Machine review of arXiv:2506.00471}
}
read the original abstract

Physics-informed neural networks (PINNs) offer a powerful framework for seismic wavefield modeling, yet they typically require time-consuming retraining when applied to different velocity models. Moreover, their training can suffer from slow convergence due to the complexity of of the wavefield solution. To address these challenges, we introduce a latent diffusion-based strategy for rapid and effective PINN initialization. First, we train multiple PINNs to represent frequency-domain scattered wavefields for various velocity models, then flatten each trained network's parameters into a one-dimensional vector, creating a comprehensive parameter dataset. Next, we employ an autoencoder to learn latent representations of these parameter vectors, capturing essential patterns across diverse PINN's parameters. We then train a conditional diffusion model to store the distribution of these latent vectors, with the corresponding velocity models serving as conditions. Once trained, this diffusion model can generate latent vectors corresponding to new velocity models, which are subsequently decoded by the autoencoder into complete PINN parameters. Experimental results indicate that our method significantly accelerates training and maintains high accuracy across in-distribution and out-of-distribution velocity scenarios.

Figures

Figures reproduced from arXiv: 2506.00471 by the authors.

Figure 1
Figure 1. (a) Illustration of the PINN optimization process, starting from a random initialization [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An overview of our latent diffusion approach for generating PINN initialization parameters. (a) PINN Training: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. An illustration of network architectures. (a) Autoencoder compresses the flatted PINN parameter vector [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Performance comparison among DiffPINN (blue), Meta-PINN (orange), and vanilla PINN (yellow), averaged [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Comparison of the real part of the scattered wavefield solutions at 5 Hz for an in-distribution test velocity [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison among DiffPINN (blue), Meta-PINN (orange), and vanilla PINN (yellow), averaged [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Similar with Figure [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Comparison of DiffPINN performance on in-distribution (blue) versus out-of-distribution (orange) velocity [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Initial real part of the scattered wavefield predicted directly from the different weights (epoch 0): (a) [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Latent representations z0 of the weights for each initialization method. (a), (b), and (c) correspond to DiffPINN, Meta-PINN, and vanilla PINN, respectively. Each latent map is of size 128 × 1590, with the vertical axis indexing latent dimensions and the horizontal ax…
Figure 12
Figure 12. Figure 12: Performance comparison between physics-guided and unguided generated initializations, averaged over five [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Impact of reverse-diffusion sampling step size on PINN performance, averaged over five in-distribution and [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Seismic modeling

    Jose M Carcione, G \'e rard C Herman, and APE Ten Kroode. Seismic modeling. Geophysics, 67 0 (4): 0 1304--1325, 2002

    work page 2002

  2. [2]

    Full seismic waveform modelling and inversion

    Andreas Fichtner. Full seismic waveform modelling and inversion. Springer Science & Business Media, 2010

    work page 2010

  3. [3]

    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

  4. [4]

    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

  5. [5]

    3d heterogeneous staggered-grid finite-difference modeling of seismic motion with volume harmonic and arithmetic averaging of elastic moduli and densities

    Peter Moczo, Jozef Kristek, V \'a clav Vavrycuk, Ralph J Archuleta, and Ladislav Halada. 3d heterogeneous staggered-grid finite-difference modeling of seismic motion with volume harmonic and arithmetic averaging of elastic moduli and densities. Bulletin of the Seismological Society of America, 92 0 (8): 0 3042--3066, 2002

    work page 2002

  6. [6]

    Viscoelastic finite-difference modeling

    Johan OA Robertsson, Joakim O Blanch, and William W Symes. Viscoelastic finite-difference modeling. Geophysics, 59 0 (9): 0 1444--1456, 1994

    1994

  7. [7]

    Low and high order finite element method: experience in seismic modeling

    Enrico Padovani, E Priolo, and G Seriani. Low and high order finite element method: experience in seismic modeling. Journal of Computational Acoustics, 2 0 (04): 0 371--422, 1994

    work page 1994

  8. [8]

    Finite-element simulation of seismic ground motion with a voxel mesh

    Kazuki Koketsu, Hiroyuki Fujiwara, and Yasushi Ikegami. Finite-element simulation of seismic ground motion with a voxel mesh. Pure and Applied Geophysics, 161 0 (11-12): 0 2183--2198, 2004

    work page 2004

Show all 34 references
  1. [9]

    Modeling acoustic wave propagation in heterogeneous attenuating media using decoupled fractional laplacians

    Tieyuan Zhu and Jerry M Harris. Modeling acoustic wave propagation in heterogeneous attenuating media using decoupled fractional laplacians. Geophysics, 79 0 (3): 0 T105--T116, 2014

    2014

  2. [10]

    Propagating seismic waves in vti attenuating media using fractional viscoelastic wave equation

    Ning Wang, Guangchi Xing, Tieyuan Zhu, Hui Zhou, and Ying Shi. Propagating seismic waves in vti attenuating media using fractional viscoelastic wave equation. Journal of Geophysical Research: Solid Earth, 127 0 (4): 0 e2021JB023280, 2022

    2022

  3. [11]

    A graphics processing unit implementation of time-domain full-waveform inversion

    Pengliang Yang, Jinghuai Gao, and Baoli Wang. A graphics processing unit implementation of time-domain full-waveform inversion. Geophysics, 80 0 (3): 0 F31--F39, 2015

    2015

  4. [12]

    Cu q-rtm: A cuda-based code package for stable and efficient q-compensated reverse time migration

    Yufeng Wang, Hui Zhou, Xuebin Zhao, Qingchen Zhang, Poru Zhao, Xiance Yu, and Yangkang Chen. Cu q-rtm: A cuda-based code package for stable and efficient q-compensated reverse time migration. Geophysics, 84 0 (1): 0 F1--F15, 2019

    2019

  5. [13]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378: 0 686--707, 2019

    2019

  6. [14]

    Wavefield solutions from machine learned functions constrained by the helmholtz equation

    Tariq Alkhalifah, Chao Song, Umair bin Waheed, and Qi Hao. Wavefield solutions from machine learned functions constrained by the helmholtz equation. Artificial Intelligence in Geosciences, 2: 0 11--19, 2021

    2021

  7. [15]

    Solving the frequency-domain acoustic vti wave equation using physics-informed neural networks

    Chao Song, Tariq Alkhalifah, and Umair Bin Waheed. Solving the frequency-domain acoustic vti wave equation using physics-informed neural networks. Geophysical Journal International, 225 0 (2): 0 846--859, 2021

    2021

  8. [16]

    Pinneik: Eikonal solution using physics-informed neural networks

    Umair Bin Waheed, Ehsan Haghighat, Tariq Alkhalifah, Chao Song, and Qi Hao. Pinneik: Eikonal solution using physics-informed neural networks. Computers & Geosciences, 155: 0 104833, 2021

    2021

  9. [17]

    Wavefield reconstruction inversion via physics-informed neural networks

    Chao Song and Tariq A Alkhalifah. Wavefield reconstruction inversion via physics-informed neural networks. IEEE Transactions on Geoscience and Remote Sensing, 60: 0 1--12, 2021

    2021

  10. [18]

    Physics-informed neural networks (pinns) for wave propagation and full waveform inversions

    Majid Rasht-Behesht, Christian Huber, Khemraj Shukla, and George Em Karniadakis. Physics-informed neural networks (pinns) for wave propagation and full waveform inversions. Journal of Geophysical Research: Solid Earth, 127 0 (5): 0 e2021JB023120, 2022

    2022

  11. [19]

    Pinnup: Robust neural network wavefield solutions using frequency upscaling and neuron splitting

    Xinquan Huang and Tariq Alkhalifah. Pinnup: Robust neural network wavefield solutions using frequency upscaling and neuron splitting. Journal of Geophysical Research: Solid Earth, 127 0 (6): 0 e2021JB023703, 2022

    2022

  12. [20]

    Helmholtz-equation solution in nonsmooth media by a physics-informed neural network incorporating quadratic terms and a perfectly matching layer condition

    Yanqi Wu, Hossein S Aghamiry, Stephane Operto, and Jianwei Ma. Helmholtz-equation solution in nonsmooth media by a physics-informed neural network incorporating quadratic terms and a perfectly matching layer condition. Geophysics, 88 0 (4): 0 T185--T202, 2023

    2023

  13. [21]

    Physics-informed neural wavefields with gabor basis functions

    Tariq Alkhalifah and Xinquan Huang. Physics-informed neural wavefields with gabor basis functions. Neural Networks, 175: 0 106286, 2024

    2024

  14. [22]

    Modeling multisource multifrequency acoustic wavefields by a multiscale fourier feature physics-informed neural network with adaptive activation functions

    Xintao Chai, Zhiyuan Gu, Hang Long, Shaoyong Liu, Taihui Yang, Lei Wang, Fenglin Zhan, Xiaodong Sun, and Wenjun Cao. Modeling multisource multifrequency acoustic wavefields by a multiscale fourier feature physics-informed neural network with adaptive activation functions. Geop...

    2024

  15. [23]

    Robust data driven discovery of a seismic wave equation

    Shijun Cheng and Tariq Alkhalifah. Robust data driven discovery of a seismic wave equation. Geophysical Journal International, 236 0 (1): 0 537--546, 2024

    2024

  16. [24]

    Discovery of physically interpretable wave equations

    Shijun Cheng and Tariq Alkhalifah. Discovery of physically interpretable wave equations. Surveys in Geophysics, 46: 0 119--144, 2025 a

    2025

  17. [25]

    Meta learning for improved neural network wavefield solutions

    Shijun Cheng and Tariq Alkhalifah. Meta learning for improved neural network wavefield solutions. Surveys in Geophysics, 46: 0 145--167, 2025 b

    2025

  18. [26]

    Multiple wavefield solutions in physics-informed neural networks using latent representation

    Mohammad H Taufik, Xinquan Huang, and Tariq Alkhalifah. Multiple wavefield solutions in physics-informed neural networks using latent representation. IEEE Geoscience and Remote Sensing Letters, 2024

    2024

  19. [27]

    Multi-frequency wavefield solutions for variable velocity models using meta-learning enhanced low-rank physics-informed neural network

    Shijun Cheng and Tariq Alkhalifah. Multi-frequency wavefield solutions for variable velocity models using meta-learning enhanced low-rank physics-informed neural network. arXiv preprint arXiv:2502.00897, 2025 c

    2025 arXiv

  20. [28]

    Neural network diffusion

    Kai Wang, Dongwen Tang, Boya Zeng, Yida Yin, Zhaopan Xu, Yukun Zhou, Zelin Zang, Trevor Darrell, Zhuang Liu, and Yang You. Neural network diffusion. arXiv preprint arXiv:2402.13144, 2024

    2024 arXiv

  21. [29]

    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

  22. [30]

    Richards

    Keiiti Aki and Paul G. Richards. Quantitative seismology: Theory and methods. 1980. URL https://api.semanticscholar.org/CorpusID:58794764

    1980

  23. [31]

    Cold diffusion: Inverting arbitrary image transforms without noise

    Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. Advances in Neural Information Processing Systems, 36, 2024

    2024

  24. [32]

    Openfwi: Large-scale multi-structural benchmark datasets for full waveform inversion

    Chengyuan Deng, Shihang Feng, Hanchen Wang, Xitong Zhang, Peng Jin, Yinan Feng, Qili Zeng, Yinpeng Chen, and Youzuo Lin. Openfwi: Large-scale multi-structural benchmark datasets for full waveform inversion. Advances in Neural Information Processing Systems, 35: 0 6007--6020, 2022

    2022

  25. [33]

    Decoupled weight decay regularization

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

    2017 arXiv

  26. [34]

    Denoising diffusion implicit models

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

    2010 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.