acceptodds
Under review as a conference paper at ICLR 2027

Real-Data-Guided Seismic Forward Modeling in Latent Space

Abstract

Seismic forward modeling (SFM) predicts wave responses from subsurface velocity fields, but learning this mapping accurately remains challenging because seismic responses are high-dimensional and paired real-world velocity–response data are scarce. Nevertheless, we argue that real earthquake records can still provide informative waveform trends for data-driven SFM. Motivated by this insight, we apply Parallel Symbolic Enumeration (PSE) to event–station records to discover empirical relationships. After offline physics-based screening with assistance from a large language model (LLM), the retained relationships can be transformed into normalized trend templates that regularize training at no additional inference cost. Building on this real-data-derived guidance, we propose Latent Seismic Predictor (LSP), a coarse-to-refined latent framework that separates dominant response prediction from oscillatory detail recovery. LSP first maps the subsurface velocity field into a coarse response latent capturing principal arrivals, reflection geometry, and broad energy distribution, and then applies attention-guided spectral residual refinement to recover unresolved waveform components through tapered sine–cosine basis functions. Moreover, learnable tapering suppresses spurious oscillations before residual synthesis, after which the refined latent representation is decoded into the complete seismic response. Across ten OpenFWI subsets, LSP achieves the best average rank in MAE, MSE, and SSIM, reducing average MAE and MSE on the challenging Style subsets by 11.4% and 26.2%, respectively, over the strongest baseline. When transferred from Style-B to Marmousi and Overthrust without fine-tuning, it reduces average MAE by 24.5%. The leading performance on GlobalTomo also supports its extension to high-dimensional 3D acoustic wave prediction. Codes are attached.

Primary area:
learning on time series and dynamical systems
Keywords:
Seismic Forward Modeling, Scientific Machine Learning, Latent Representation Learning, Spectral Residual Refinement, Symbolic Regression

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_9ijYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LSP, a coarse-to-refined latent framework for seismic forward modeling. It uses an attention-guided sine–cosine residual module to refine predicted seismic responses and incorporates symbolic relationships extracted from real earthquake records as training-time regularization. Experiments on OpenFWI, Marmousi, Overthrust, and GlobalTomo report improvements over U-Net, FNO, and DeepONet.

Critical strengths

Interesting use of real seismic data. Extracting waveform trends from unpaired earthquake records offers a potentially useful direction for incorporating real observations into synthetic-data-trained forward models.

Critical weaknesses
  1. Insufficient methodological justification and unclear physical interpretation. The paper assigns specific physical meanings to learned representations, such as the geological latent (z_v), coarse response latent (z_c), and multi-level velocity features (f_v^{(1)},f_v^{(2)},f_v^{(3)}), without sufficiently demonstrating that these features actually capture the claimed geological structures or propagation characteristics. The paper also does not explain how the feature extraction and coarse transport modules are implemented, which raises serious concerns about the soundness of the paper. Furthermore, the motivation for applying sine–cosine functions to latent features is unclear, since these features have no explicit correspondence to physical frequencies. The claim that this refinement specifically recovers oscillatory waveform details requires stronger empirical evidence.

  2. Insufficient experimental comparisons and validation of real-data guidance. The method is compared only against standard architectures (U-Net, FNO, and DeepONet), without stronger recent or domain-specific approaches. Additionally, although removing symbolic regularization degrades performance, the paper does not establish whether the improvement comes specifically from real-data-derived relationships rather than generic waveform regularization. Comparisons against conventional physics-based constraints or synthetic-data-derived templates would be important to substantiate this contribution.

Questions for authors
  1. How are the three velocity feature representations and the coarse transport module implemented? What evidence supports their claimed physical interpretations, particularly that (z_c) represents dominant arrivals and the residual captures finer waveform details?

  2. Why are sine–cosine functions applied to learned latent features rather than physically meaningful frequency representations? Can the authors provide frequency-resolved analyses or comparisons with alternative refinement mechanisms?

  3. Can the authors compare their symbolic regularization against standard analytical physics-based constraints or waveform statistics derived directly from synthetic data to isolate the contribution of real earthquake observations?

  4. How does LSP compare against recent neural-operator or seismic forward-modeling methods beyond the three standard baselines

Reviewer_PCJAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

LSP learns to predict the full seismic response from a velocity field, in latent space. It works in two stages: a coarse stage encodes the velocity and predicts the main response structure (arrivals, reflection geometry, overall energy distribution), then a refinement stage adds back the oscillatory detail using tapered sine–cosine basis functions. Separately, it pulls trend constraints out of real earthquake records (STEAD) using symbolic regression (PSE) plus an LLM screen, and uses those as a soft training-time regularizer.

The latent coarse-to-refined idea is reasonable and the paper is well written. But the thing the paper sells hardest — the "real-data guidance" — is exactly the part that's least supported. The relations it rediscovers turn out to be standard seismological formulas, and the paper never tests whether the real-data extraction beats just writing down those formulas by hand. That's the gap that keeps this from being a strong paper.

Critical strengths
  1. The coarse-to-refined split is a sensible response to the real difficulty here: the input is spatial velocity, the output is a time–receiver wavefield, and they're very different objects. The ablation supports it — replacing the harmonic residual with a conv block drops SSIM from 0.9762 to 0.8957, and removing spectral refinement hurts most (0.9231).

  2. Pulling trends out of unpaired real data is a good instinct, because paired velocity–response data is exactly what you don't have. And the paper is careful to treat the discovered relations as soft constraints rather than claiming they're governing equations.

  3. The authors are honest about limits: the symbolic guidance only helps when there's a refinement mechanism to pair with it, and the 3D result is explicitly called a controlled test, not field-scale validation.

4.LSP generalizes well OOD — best MAE/MSE/SSIM on original Marmousi and Overthrust without fine-tuning. That's a real result.

Critical weaknesses
  1. The paper says the trend regularization is its headline contribution, but the ablation only checks with/without the symbolic loss. It never asks: does the trend pulled from real STEAD data beat a hand-written empirical template that encodes the same physics — standard Brune spectrum, geometric-spreading attenuation, magnitude–distance amplitude scaling? That's the whole question. If a hand-written template does just as well, then the PSE + LLM pipeline isn't buying anything over physics we already know, and the headline collapses.

  2. The three it keeps (Eq. C.4–C.6) are log10(fc)=1−M/8, γ(R)=113/512+1/(16R), log10(A)=M−log10(R)+2 — these are standard seismological formulas (Brune source spectrum, geometric spreading, magnitude–distance amplitude scaling). They're found by PSE and screened by an LLM, but the LLM only does offline plausibility checking; the final pick is decided by PSE fitting error, complexity, and physical criteria. So I don't see why you'd rediscover these well-known relations via symbolic regression and an LLM instead of just writing them down. If the contribution is really the relative-transfer mechanism and not the discovery, show that against a directly-specified relative template.

  3. On the Vel subsets FNO wins MAE/MSE (the paper says so), and LSP's advantage is on Fault/Style. OOD is even less uniform — FNO best on MarmousiSmooth, DeepONet best on SSIM. So "best average rank" is true but conceals the regime dependence. The abstract's "11.4% reduction on Style" framing ought to carry the same caveat the conclusion gives.

  4. The baseline set is thin for 2026. Just U-Net, FNO, DeepONet. The paper itself cites SVD-based operators (Koren 2026), frequency-domain scattered-wavefield (Huang & Alkhalifah 2025), and spectral refiner (Cao 2025), yet compares against none of them — odd, given the title is about latent spectral refinement. Even just an ablation against a physical Fourier-domain version of the same design would show what the latent placement buys.

Questions for authors
  1. Does real-data guidance beat writing down the physics? Add an ablation swapping the symbolic templates for hand-written ones (Brune spectrum, geometric spreading, magnitude–distance scaling), same observables and weights. This is the one experiment that decides whether PSE + LLM adds anything over known physics.

  2. What does the LLM actually change? It's only an offline plausibility check. How many candidates were kept vs. dropped, and does the result change with the LLM removed (physical criteria alone)?

  3. Is the gain really in the weak-signal regions? Add an energy- or amplitude-stratified error analysis instead of relying on SSIM and misfit pictures.

Reviewer_9NkaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a coarse2refine framework for seismic forward modelling, able to predict wave responses from subsurface velocity fields. The methodology is developed to tackle the issue of high-dimensional data and the scarsity of real-world velocity-response data. The general idea consists in a two step approach: first a empirical relationships are discovered from station records and these are used as input to an LLM able to 'discriminate' if these are physically sound; then, a latent space framework is used to predict the wave responses from the velocity fields. The authors have tested the methods against other operator learning approach on OpenFWI datasets.

Critical strengths
  • The use of real record data as input to regularize the operator learning process is an interesting idea.
  • The selected architecture appears to be solid on complex geological structures, making it more robust than available operator learning approaches.
Critical weaknesses
  • The ablation evidence is limited. Table 4 evaluates ablations only on STB, although the paper claims improvements across multiple geological regimes. I would want to know whether the contribution of real-data guidance is consistent across datasets, particularly where LSP underperforms.
  • The limitations section is weak and the conclusions are quite short. The results section is mostly describing the figures and tables, providing few interpretations and discussion of the results.
Questions for authors
  • I wonder if a combination of real-data and physics-based reduced model could be used to further improve the performance of the method. Or the connection/comparison of this approach with Data-Assimilation techniques? I would argue that this could be a more reliable baseline than purely data-driven approaches.
  • Table 1: overall your method is ranked as best compared to other operator learning approaches and it performs very well apart from FVA, FVB and FFA? Does the authors have more insights on why this is the case?
  • The ablation study is limited to STB, is there a particular reason for this choice? Have the authors considered to evaluate the performance on other datasets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.