Physical-State-Guided Diffusion Sampling for Full-Waveform Inversion
Chen Min, Haowen Jiang, Zheng Ma, Xiongbin Yan
cs.LG
Quick verdict
PSG weds a persistent FWI velocity to a diffusion prior via a Gaussian bridge; it tops OpenFWI baselines broadly, though noise-RMSE and tuning caveats remain.
Summary
The paper proposes Physical-State-Guided Diffusion Sampling (PSG), a diffusion-based solver for full-waveform inversion (FWI). Unlike direct diffusion posterior sampling, which evaluates the wave-equation likelihood on each step's Tweedie estimate and discards it, PSG maintains a persistent physical velocity state initialized from a conventional FWI model and coupled to a pretrained diffusion prior through a Gaussian bridge. Each reverse-diffusion step refines the physical state by waveform fitting regularized toward the denoised velocity, and the refined state then guides the diffusion update, separating wave-equation adjoint gradients from denoiser Jacobians. On four OpenFWI families (100 test models each), PSG achieves the best mean MAE/RMSE/SSIM among eight methods on clean data and at every missing-trace level; under Gaussian noise it holds the highest mean SSIM in all twelve family-noise conditions and the best MAE at lower noise, though RED-DiffEq attains lower RMSE as noise grows. A 20-run ensemble study shows variability concentrated near geological interfaces and positively correlated with local error (mean Spearman 0.706). A frozen OpenFWI-trained prior transfers to Marmousi, Overthrust, and BP2004 Salt without retraining. Limitations: the coupled algorithm has no exact-posterior or convergence guarantee; the reported output switches between denoised and physical states across experiments; Salt hyperparameters were tuned on the test case itself; no code release is stated.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Coupling a persistent physical velocity to a diffusion prior through a Gaussian bridge separates wave-equation adjoint gradients from denoiser Jacobians while preserving conventional FWI initialization and accumulated optimizer history (Section 2.5).
- PSG's terminal denoised estimates achieve the best mean MAE, RMSE, and SSIM across all four OpenFWI families under clean data and at every missing-trace level (Tables 1 and 3).
- Under Gaussian noise, PSG holds the highest mean SSIM in all 12 family-noise conditions and the best MAE at lower noise, while RED-DiffEq attains lower RMSE at higher noise levels (Section 3.2.2).
- Ensemble variability across stochastic reverse trajectories concentrates near geological interfaces and is positively associated with local inversion error (mean Spearman 0.7063; Section 3.3).
- A frozen OpenFWI-trained diffusion prior supports inversion of the larger Marmousi, Overthrust, and BP2004 Salt models without retraining (Section 3.4).
Strengths
Principled decoupling of physics and prior gradients
Section 2.5 (Eqs. 37, 40-41): the wave solver runs only on the persistent physical velocity v while guidance on x_t touches only the denoiser Jacobian; unlike DPS, PSG preserves a conventional smoothed-model initialization and Adam optimizer state across the whole reverse schedule, directly addressing the three DPS failure modes enumerated in Section 2.4.
Broad, controlled benchmarking
Section 3.2, Tables 1-3: four OpenFWI families x 100 test models each, clean/Gaussian-noise (sigma = 0.1/0.3/0.5)/missing-trace (10/35/60 of 70 receivers) conditions, seven baselines sharing one EDM prior per family and the same acoustic solver; PSG is best on mean MAE/RMSE/SSIM in clean data (e.g., CF-B MAE 112.0 vs RED-DiffEq 174.1) and at every missing-trace level.
Honest negative results
Quantified ensemble/uncertainty study
Unusually thorough implementation appendix
Weaknesses
No exactness or convergence guarantee for the coupled sampler
Output selection switches across experiments
Hyperparameters tuned on the OOD test case
Thin uncertainty reporting in the main tables
Baseline fairness and budget asymmetry
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets full-waveform inversion (FWI), a PDE-constrained nonlinear inverse problem, and proposes Physical-State-Guided Diffusion Sampling (PSG): a diffusion-based solver that keeps a persistent physical velocity state coupled to a pretrained score-based prior through a Gaussian bridge (Section 2.5, Eqs. 26-41). Central claims: (i) the design separates wave-equation adjoint gradients from denoiser Jacobians and preserves conventional FWI initialization plus accumulated optimizer history, addressing three documented failure modes of direct DPS (Section 2.4: cycle-skipped early guidance, unbalanced quadratic guidance, nonphysical averaged estimates); (ii) best mean MAE/RMSE/SSIM across four OpenFWI families under clean and missing-trace acquisition, highest SSIM under noise; (iii) ensemble variability is a useful local-error indicator; (iv) a frozen OpenFWI prior transfers to Marmousi, Overthrust, and BP2004 Salt.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does PSG solve in full-waveform inversion?
It addresses the unreliability of directly coupling a diffusion model's denoiser to the nonlinear wave-equation solver. Instead of evaluating the wave likelihood on each step's blurred Tweedie estimate and discarding it, PSG keeps a persistent physical velocity (initialized like standard FWI), refines it by waveform fitting regularized toward the denoised velocity, and uses that state to guide reverse diffusion. This preserves FWI initialization and optimizer history and separates the two gradient computations.
How well does PSG perform compared to other methods?
On four OpenFWI benchmark families (100 test models each), the paper reports the best mean MAE, RMSE, and SSIM on clean data and at all missing-trace levels (10/35/60 of 70 receivers removed) against FWI, Tikhonov/TV-FWI, DPS, RED-DiffEq, DiffusionFWI, and SGDS-FWI. Under Gaussian noise it keeps the highest mean SSIM in all twelve conditions, but RED-DiffEq achieves lower RMSE at the higher noise levels, including on all four families at the strongest noise.
Does PSG need to retrain the diffusion model for new survey areas?
No. The authors show a prior trained only on OpenFWI's CurveFault-B family transfers, without retraining, to the larger Marmousi, Overthrust, and BP2004 Salt models using overlapping-patch decomposition — though Salt-specific hyperparameter adjustments were applied and tuned on that case.
Is code available for this paper?
The paper states that diffusion pretraining follows the official EDM repository and that baselines follow their official implementations, and Appendix B documents architectures, schedules, and hyperparameters in detail. However, it does not announce a released PSG codebase, so users would need to reimplement the method from the pseudocode and appendix.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Reconstruction-quality judgments rely on figure captions and in-text descriptions; the actual images (Figs. 3-8, 9-13) cannot be assessed from text extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.