acceptodds
Under review as a conference paper at ICLR 2027

Learning seismic wave propagation: from local geometry to global physics

Abstract

Seismic forward modeling is computationally expensive because a wave equation must be solved repeatedly for each source and, possibly, several velocity models. Learned surrogates could reduce this cost, but they must preserve both the local influence of heterogeneous geologic structures and the long-range coherence of propagating wavefields. We address this problem using such machine learning models as Fourier Neural Operators (FNOs), MeshGraphNets (MGNs), and Transolver. FNO provides global spectral mixing on a structured grid; MGN explicitly uses local mesh connectivity; and Transolver communicates through learned global physical states. Here, we introduce SeisGeoSolver, which composes local message passing, global physics attention, and local refinement in a shared latent state. In 2D simulations for the GeoFWI acoustic model, edge-aware models are substantially more stable in autoregressive rollout than FNO and global-only physics attention. SeisGeoSolver matches the rollout accuracy of a deeper MGN while using fewer parameters and less training time. These results confirm that explicit local geometry is essential in this setting, while global physical interaction can replace part of the local processing budget.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
wave propagation, seismic, geometric learning, gnn, transformer, physics attention, surrogate modeling, acoustic wave equation, seismology, geophysics, FWI, full wavefield inversion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QmRCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies learned surrogates for 2D acoustic seismic wave propagation (autoregressive time-stepping on the GeoFWI salt model), comparing FNO (global spectral mixing), MeshGraphNet/MGN (local message passing), and Transolver (global learned physical states), then proposes SeisGeoSolver, a hybrid that composes local MGN message passing, Transolver-style global physics attention, and a final local refinement stage on a shared latent (local → global → local). All models are trained and rolled out as autoregressive velocity-increment predictors and evaluated by space-time relative L2 rollout error, active-frame energy ratio, and wavefield/shot-gather visual fidelity, with additional validation on four unseen SEG salt patches and reporting of parameter count and training/inference cost.

The headline empirical result (Table 3) is that FNO fails catastrophically under rollout (~100% error), Transolver remains numerically stable but visually smears/misaligns reflected arrivals (43.2/33.8% error), while MGN and SeisGeoSolver both land in an ≈11-12% error band. SeisGeoSolver matches the 15-block MGN's accuracy with 5.38M vs. 9.04M parameters and less training time per epoch (Table 2, Fig. 6). The paper explicitly states in its own Discussion/Conclusion that SeisGeoSolver "has not provided yet a higher accuracy than a well-tuned MGN" — the contribution is framed as an efficiency gain at matched accuracy, not a new accuracy frontier.

Critical strengths

1. The controlled comparison across representation types (grid-spectral, local-mesh, global-state) is a clean and informative experimental design. Holding the time-stepping protocol, data, and task fixed (Eq. 2) while varying only how spatial information is exchanged (Table 1) isolates a real architectural question, and the result that explicit local connectivity is necessary for stable rollout in heterogeneous media (FNO ≈100% error, Transolver's reflected/refracted wavefronts visibly smeared in Fig. 3) is a clear, well-supported finding that is useful to the community working on seismic/PDE surrogates.

Critical weaknesses

W1. The paper's own central claim is an efficiency result, not an accuracy or capability result, and the evidence for even that efficiency claim is narrow.

Sections 8-10 are explicit that SeisGeoSolver does not exceed MGN's rollout accuracy (11.8/11.3% vs. MGN's 11.2/11.9%, Table 3) and that "these tests... do not establish that global attention improves accuracy over a fully tuned MGN" (Sec. 8). The entire positive case rests on parameter count and training cost (5.38M vs. 9.04M params, 534s vs. 759s/epoch). But the comparison is against one specific MGN depth (15 blocks) and one specific SeisGeoSolver configuration (3+8+3 blocks); there is no sweep showing where the accuracy-vs-cost Pareto frontier of MGN alone sits relative to SeisGeoSolver (e.g., does a 6- or 9-block MGN already match SeisGeoSolver's cost and accuracy?). Without that, "comparable accuracy at lower cost" is demonstrated for a single pair of configurations rather than established as a general property of the local-global-local design.

W2. The ablation in Table 4 does not isolate the proposed three-stage composition; it only compares against the two baselines the hybrid was built from, leaving the specific design choice (local→global→local, and 3+8+3 blocks) untested against simpler alternatives.

Table 4 shows SeisGeoSolver vs. plain Transolver (no local edges) vs. plain MGN (no global attention, deeper). This establishes that some local processing and some global processing both matter individually, which is close to given from the architectures' definitions. It does not test the paper's actual novel claim, that the specific ordering local→global→local (Eq. 4) and the specific stage split (3 local blocks before and after 8 global blocks) is better than alternatives such as global→local, a single local stage after global mixing, or other splits of the same total block budget. Since the ordering is motivated in Sec. 5 with a physical narrative ("edge features make the latent state sensitive to local contrast before global mixing... local refinement reconnects global context to the mesh"), but no ablation varies the ordering or the split while holding total compute fixed, this specific architectural claim is argued rather than tested.

Minor Comments

  • Table 3/4's rollout error for SeisGeoSolver (11.8/11.3) is very slightly worse than MGN's (11.2/11.9) on one component and better on the other; the text's framing ("matches... the accuracy band") is fair, but a significance/variance estimate (even a single repeat run) would clarify whether this is noise or a real small gap.
  • The SEG salt-patch validation ("errors increase with geologic shift," Sec. 7) is stated qualitatively in the main text with numbers deferred to Appendix D; since out-of-distribution generalization is central to the "foundation-model solver" framing in the Introduction, at least the headline SEG numbers would strengthen the main text.
  • The Limitations paragraph (Sec. 9) is appropriately candid (2D only, acoustic only, fixed 96×96 grid, single interior source, salt-specific training distribution), which is appreciated, but it also means the title's broad framing ("from local geometry to global physics") somewhat overstates the demonstrated scope.
Questions for authors
  1. MGN depth sweep. Have you evaluated MGN at parameter/cost budgets comparable to SeisGeoSolver's (e.g., a 6-9 block MGN)? This would directly test whether the hybrid's efficiency advantage holds across the cost axis or is specific to the 15-block configuration chosen as the comparison point.
  2. Ordering ablation. Is there evidence (even preliminary) that the local→global→local ordering outperforms other orderings (global→local, local→global, or a different block split) at matched total compute? This seems necessary to support Eq. (4) as a design claim rather than a post-hoc description of the configuration that happened to work.
  3. SEG generalization numbers. Could the headline SEG salt-patch rollout errors (currently only in Appendix D) be summarized in the main text, given that out-of-distribution robustness is central to the paper's motivation?
Reviewer_2QujOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SeisGeoSolver, a hybrid neural surrogate combining MeshGraphNet message passing and Transolver-style global attention for autoregressive seismic wave propagation. Experiments on 2D acoustic simulations show that it achieves accuracy comparable to a deeper MeshGraphNet while reducing parameter count and training time.

Critical strengths
  1. Potentially useful architectural direction. Combining local geometric message passing with global attention is a reasonable approach to modeling wave propagation.

  2. Promising preliminary results. The proposed model achieves accuracy comparable to a deeper MeshGraphNet with fewer parameters and lower training costs. The autoregressive evaluation also provides useful initial evidence of propagation stability.

Critical weaknesses
  1. Incomplete experimental validation and poor presentation. The work remains largely at the proof-of-concept stage. Although 100 held-out simulations are available, quantitative results are reported for only three cases, and OOD evaluation is limited to four SEG salt windows. Moreover, the baselines are limited to established architectures, including FNO, MeshGraphNet, and Transolver. The latter two also serve as component-level ablation comparisons rather than independent baselines. The presentation also requires substantial improvement. The Introduction lacks a coherent logical progression, with weak connections between sentences and paragraphs. Similar issues appear throughout the paper: for example, Section 3 introduces the acoustic pressure equation but suddenly switches to displacement-field prediction without explaining their relationship. These issues make the physical problem and methodological reasoning difficult to follow. Overall, the paper appears insufficiently developed, both experimentally and in its presentation. More comprehensive experiments are necessary to establish your hypothesis.

  2. Limited novelty and insufficient methodological justification. The proposed method primarily combines existing MeshGraphNet and Transolver architectures without introducing a substantially new learning mechanism. More importantly, the experiments do not convincingly demonstrate the necessity or advantages of global attention, as the proposed model only achieves comparable accuracy to MeshGraphNet. Stronger parameter-matched ablations are needed to isolate the contributions of individual components. Additionally, comparisons are limited to basic architectures, with MeshGraphNet and Transolver serving more as component-level ablations than independent state-of-the-art baselines.

Questions for authors
  1. Can the authors provide controlled, parameter-matched comparisons to establish the necessity of global attention over simpler or shallower MeshGraphNet architectures?

  2. Why are quantitative results reported for only three of the 100 held-out simulations? Can the authors provide aggregate results and statistical variability across the complete test set?

  3. How does SeisGeoSolver compare with more recent hybrid graph-transformer or multiscale neural-operator approaches?

Reviewer_QAJWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SeisGeoSolver, a hybrid neural surrogate for seismic wave propagation that combines MeshGraphNet (MGN) message passing with Transolver-style physics attention in a local–global–local architecture. The model incorporates mesh connectivity, material contrasts, and global feature interactions to predict wavefield evolution autoregressively. Experiments on 2D acoustic GeoFWI salt models compare SeisGeoSolver against FNO, Transolver, and MGN. The proposed method achieves comparable rollout accuracy to a deeper MGN while reducing model parameters and training cost. Additional evaluations examine long-term wavefield fidelity and zero-shot generalization to SEG salt models.

Critical strengths
  1. The combination of local graph message passing and global physics attention is reasonable for heterogeneous seismic wave propagation. Incorporating local geometric and material information helps preserve wavefront structures that are difficult for global-only architectures to capture.

  2. SeisGeoSolver achieves accuracy comparable to a 15-block MGN while using approximately 40% fewer parameters and 30% less training time per epoch. The evaluation also considers autoregressive stability, wavefield snapshots, seismic gathers, and computational efficiency.

Critical weaknesses
  1. The proposed architecture mainly combines existing MGN message-passing blocks and Transolver physics-attention modules. The paper does not introduce a fundamentally new message-passing mechanism, attention formulation, or physics-informed learning principle.

  2. The ablation studies are insufficient to isolate the contribution of global attention. In particular, comparisons against a parameter-matched MGN, a shallower MGN, and alternative local–global configurations are missing.

  3. The experiments are restricted to fixed-resolution 96×96 grids, 2D acoustic propagation, and a single interior source. Generalization across spatial resolutions, source configurations, wavelet frequencies, and more complex geological structures remains unverified.

  4. Despite the motivation from full-waveform inversion, the paper does not evaluate derivatives with respect to velocity, inversion accuracy, or the impact of surrogate modeling errors on FWI. Since wavefield errors are concentrated near strong interfaces, their influence on inversion gradients could be significant.

Questions for authors
  1. Can the authors compare SeisGeoSolver with a parameter-matched MGN and shallower MGN variants? Would the proposed method still offer a significant advantage under comparable computational budgets?

  2. What happens when the global attention stage, initial local processing stage, or final local refinement stage is removed independently? How sensitive are the results to the number and ordering of these blocks?

  3. Why are the main quantitative results reported for only three simulations despite having 100 held-out test samples? Can the authors provide full-test-set statistics, standard deviations, and results across multiple random seeds?

  4. How does the model perform under different spatial resolutions, source locations, wavelet frequencies, or substantially different velocity distributions? What explains the considerable performance degradation on the SEG salt models?

  5. Have the authors examined the accuracy of wavefield derivatives with respect to velocity? Can the learned solver produce sufficiently accurate gradients for FWI, particularly near high-contrast salt boundaries?

  6. Are the reported SPECFEM2D and neural surrogate runtimes measured under comparable hardware and computational settings? How do preprocessing, graph construction, and data-transfer costs affect the actual end-to-end speedup?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.