acceptodds
Under review as a conference paper at ICLR 2027

Physics-supervised Convergent Born Neural Operators for Full Waveform Modeling and Inversion

20 views

Abstract

Full Waveform Inversion (FWI) is vital for subsurface resource exploration and geological hazard monitoring, but remains computationally prohibitive. Deep learning-based acceleration faces three core obstacles: scarce supervised labels, unstable physics-constrained training, and difficulty in enforcing accurate open-boundary modeling under broadband settings. We propose Physics-supervised Convergent Born Neural Operators (CBNO), a framework whose central mechanism learns wavefield mappings without paired wavefield labels by minimizing a Convergent Born Series (CBS) fixed-point discrepancy. CBS updates with Perfectly Matched Layers (PML) provide the physical training signal, avoiding the need for precomputed target wavefields. To support broadband prediction, frequency-partitioned Axial-AFNO experts combine shared low-rank projections with axis- and wavenumber-aware gates. The trained surrogate is then frozen and used in differentiable inversion, where upgoing-wave separation emphasizes weak reflection signals. Empirically, CBNO achieves highly accurate forward modeling on OpenFWI with a 10× speedup over the CBS reference solver, enables zero-shot transfer of the frozen forward surrogate to inversion on 2D Marmousi and Overthrust benchmarks, and demonstrates extensibility through separately trained 3D forward models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Full waveform inversion, Neural operators, Physics-supervised learning, Convergent Born series

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_qtq2OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The main idea that the paper introduces is a Neural Operator based on a CBS fixed-point loss which outputs the solution of a Helmholtz equation for given velocity field, source and frequency.

By doing that, no labeled data is needed for training. The loss function computes the discrepancy between the predicted wavefield and a number of CBS iterations of it.

As an application, the authors use the trained model for velocity field inversion.

Critical strengths

(1) The essencial idea of the paper is quite interesting: using CBS fixed-point updates to enable neural operator training without labeled data. This is an elegant way to incorporate physical information in the training process.

(2) The comparison with baseline methods is solid. The authors compare their method with a good set of consolidated existing architectures (PINN, PINO, among others), obtaining good results.

(3) Besides the main idea explained in the Paper Summary above, the proposed architecture has some very interesting features, such as frequency-partitioned experts (this may be particularly important for modeling wave propagation with neural networks), and a mechanism to avoid missing reflections in the inverse problem. The strategies employed by the authors have not just an emperical/experimental justification, but they specifically address some well-known difficulties in modeling wave propagation.

Critical weaknesses

None

Questions for authors

In page 4, the paragraph "Physics-Inspired Axial-AFNO backbone seems to be incomplete. Possibly, a period is missing after Fig. 3? Anyway, I think this paragraph is not very clearly written. Can you check, please?

In page 21, paragraph "Inversion Hyperparameters", you say that you used Adam with a learning rate of 10. Is this correct? It seems unconventionally high.

Do you plan to make your code available? I honestly do not think that just presenting the experimental setup is enough for reproducibility.

I miss some information regarding training time. May you add some discussion about that?

The number of CBS iterations seems to be a very important hyperparameter of your model, but it seems to be just briefly discussed in the ablation section. Maybe you could add more discussion about that? (this is just a suggestion).

3 comments
AuthorsOpenReview ↗

Response to Reviewer qtq2

Thank you very much for your positive feedback! We address your questions and summarize the corresponding revisions below.

Q1. Clarity of the Axial-AFNO paragraph.

Re: Great catch! We have merged the backbone introduction and its explanation into one continuous paragraph in Section 2.4 (p. 4), correcting the interrupted paragraph layout and punctuation.

Q2. Adam learning rate.

Re: Good question! In our FWI implementation, Adam directly optimizes velocity in m/s, so the learning rate of 10 is interpreted on this physical scale. Actual nodewise updates depend on the adaptive moment estimates and subsequent velocity clipping; this does not mean a fixed 10 m/s change at every node. Yang et al. [1] also use Adam for velocity inversion, with a learning rate of 20. We have clarified the velocity units and learning-rate interpretation in Appendix D.2 (p. 24).

Q3. Code availability.

Re: Thank you! We will publicly release our code upon acceptance to support reproduction and further research.

We have added the code-release commitment to the Reproducibility statement (p. 11).

Q4. Training cost.

Re: Great suggestion! Each of the four frequency experts receives 20,000 optimizer updates (80,000 in total). Profiling on an otherwise idle RTX 4090 gives 1.4416 s per self-supervised update at an effective batch size of 50 and CBS K=5, yielding an estimated total of 32.04 GPU-hours for network-update computation. Data preparation, transfer, validation, and checkpoint saving fall outside this timing scope. Using the 50-update inversion times of 60 s for CBNO and 720 s for CBS in Table 2, this upfront computation corresponds to roughly 175 inversions worth of runtime savings (32.04 × 3,600 / (720 − 60)). We have clarified the per-expert update budget, timing scope, and training–inversion trade-off in Section 2.7 and Appendix C.4.

Q5. Number of CBS iterations.

Re: Good question! KK controls the number of CBS corrections used to construct the training loss. We assess its effects on direct wavefield optimization and network prediction accuracy to explain our choice of K=5K=5.

(1) Direct wavefield optimization. We analyze multi-frequency Marmousi runs, directly optimizing the real and imaginary components of UU without a neural network, while keeping velocity, sources, and frequencies fixed. We use a 256×256256\times256 physical grid with 25 m spacing, two interior sources, and frequencies sampled within 1–30 Hz at Δf=1/3\Delta f=1/3 Hz. Starting from zero, UU undergoes one million AdamW updates (learning rate 10−410^{-4}), reaching stable error plateaus. Reference fields are computed separately using CBS with a relative-update tolerance of 10−510^{-5} for evaluation. The table reports global relative L2 error over the entire batch, including PML, and the updates needed to first reach <0.1%<0.1\% error.

KKError at one million updatesUpdates to reach <0.1% error
10.5025%Not reached
20.1514%Not reached
30.0853%25,414
50.0504%16,770
100.0389%13,497
150.0374%12,132

Larger KK reduces the final error and the updates needed to reach the threshold, with diminishing gains; K=5K=5 achieves approximately 0.05% wavefield error in this direct optimization setting.

(2) Network accuracy. Table 3(a) shows that increasing KK from 1 to 5 reduces network prediction error by 29.5%, whereas increasing it from 5 to 10 yields only a further 0.52% reduction:

KKRelative error
10.2180
30.1784
50.1536
100.1528

These results support K=5K=5 as a practical default capturing most of the observed network accuracy gains. KK can be adjusted based on validation accuracy, while inference requires only one network forward pass. We have added these conditions and results in Appendix B.3 (pp. 17–19), Table 5 (p. 19), with the network-accuracy comparison in Section 3.4, Table 3(a).

References

[1] Yang, S., Alkhalifah, T., Ren, Y., Liu, B., Li, Y., & Jiang, P. (2023). Well-Log Information-Assisted High-Resolution Waveform Inversion Based on Deep Learning. IEEE Geoscience and Remote Sensing Letters, 20, 1–5. DOI: 10.1109/LGRS.2023.3234211.

Remark: Thank you again for your encouraging feedback and helpful suggestions! We hope the revisions make the presentation and experimental settings clearer. Wishing you a pleasant week!

Thank you very much for your detailed and clarifying answer!

AuthorsOpenReview ↗

Appreciation for Your Feedback

Thank you for your positive feedback and constructive suggestions! Wishing you all the best!

Reviewer_fE19OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper trains a neural surrogate for acoustic full-waveform modeling without paired wavefield targets. Instead of trying to match precomputed wavefields, it defines a physics loss from the fixed-point residual of the Convergent Born Series (CBS), and wraps this in a broadband architecture (frequency-partitioned experts over an Axial-AFNO backbone) plus a reflection-aware inversion objective. Experiments are forward modeling on OpenFWI and zero-shot transfer to Marmousi/Overthrust.

The forward results are genuinely the strongest part of the paper, and the model is more accurate than the baselines. My concern is the inversion. The paper's stated goal is to accelerate FWI, which it does — but it never tests whether its forward accuracy actually helps inversion. The inversion experiments only compare CBNO against a pure data-driven method (InversionNet/VeloGAN) and a traditional numerical solver (CBS). It does not compare against the other forward surrogates tested in Table 1 (PINO/FFNO/DPOT) under a matched inversion objective. So the reader can't tell whether the forward advantage translates to better inversion.

Critical strengths
  1. The label-free CBS supervisory signal is a sound and well-motivated idea. Proposition B.1 shows the final-iterate CBS loss has a unique zero coinciding with the physical Helmholtz solution — a genuine step beyond brittle PINN-style differential residuals for high-frequency propagation. The three motivating obstacles (label scarcity, unstable residual-based physics constraints, open-boundary modeling) are clearly mapped to the corresponding technical components.

  2. The experimental work is careful and comprehensive. The experimental setup is well designed. The paper uses a broad baseline set (CBS, PINN, PINO, FFNO, MgNO, DPOT, InversionNet, VeloGAN) with multiple metrics (RelError, RMSE, MAE, Corr, SSIM), and ablates the key components — supervision objective, backbone, frequency experts, and the inversion objective. On top of that, the forward results are strong: CBNO is the most accurate forward surrogate in Table 1, and it stays accurate across the whole 4–20 Hz band while PINN and MgNO degrade at higher frequency (Fig. 4b).

Critical weaknesses
  1. The paper never shows that its forward accuracy actually helps inversion. CBNO's forward modeling is the most accurate in Table 1, but that advantage is never carried into the inversion results. The inversion experiments in Table 2 only compare CBNO against pure data-driven methods (InversionNet, VeloGAN) and a traditional numerical solver (CBS) — not against the other forward surrogates it already benchmarks in Table 1 (PINO, FFNO, DPOT). Since swapping the forward surrogate is exactly what distinguishes these methods, a comparison under a matched inversion objective is what would show whether a better forward model leads to better inversion. Table 3(b) makes this harder to read by changing the inversion objective at the same time. The paper frames this as ranking "second only to CBS," which is true, but it also means CBNO loses to CBS. As it stands, the forward accuracy never demonstrably translates into better inversion.

  2. The reflection-aware objective is marginal, and its concept is not new. Separating up- from downgoing waves to emphasize reflections is an old idea (Claerbout, 1971), and using a frozen differentiable surrogate for gradient-based inversion is standard. The only new element is applying the separation operator to both the predicted and the observed field before a correlation loss. More importantly, Table 3(b) shows the upgoing objective gives essentially the same RMSE as the full-wavefield one, so its benefit beyond secondary metrics isn't established. If this is one of the three headline contributions, I'd expect a clearer case where it actually recovers structure the full-wavefield objective misses.

  3. The broadband parameterization reads as a combination of existing parts. Frequency-specialized networks, axial factorization, and an analytical background/residual split are all established. The one defensible novelty — cross-axis shared low-rank bases with axis/wavenumber gates instead of DPOT's grouped Fourier MLPs — is real but incremental, and mostly a parameter-efficiency change.

Questions for authors
  1. Can you show that your forward accuracy improves inversion? This is the key experiment I think is missing. Table 2 compares CBNO to a numerical solver (CBS) and to data-driven methods (InversionNet, VeloGAN), but not to the other forward surrogates you already evaluate in Table 1 (PINO, FFNO, DPOT). Could you run an inversion where you keep the inversion objective fixed and only swap the forward surrogate, so we can see whether your more accurate forward model actually leads to a better inversion result?

  2. What does the reflection-aware objective actually buy you? Table 3(b) shows the upgoing objective gives almost the same RMSE as the full-wavefield objective (410.74 vs. 412.76), with the gain showing up only on secondary metrics. Can you show a concrete case where the reflection-aware objective recovers structure that the full-wavefield objective misses? Without a clear example, it's hard to see why the wavefield separation is worth the added complexity.

4 comments
AuthorsOpenReview ↗

Response to Reviewer fE19

Thank you for your thoughtful review and for recognizing our forward-modeling results and CBS supervision! We address your concerns about inversion performance, the reflection-aware objective, and architectural novelty below.

W1/Q1. Impact of forward accuracy on inversion.

Re: Constructive suggestion! We compare CBNO, PINO, FFNO, and DPOT using the same inversion objective, observations, initialization, parameterization, optimizer, and update budget, with CBS as the numerical reference. Results are averaged equally over Marmousi and Overthrust.

MethodRMSE ↓MAE ↓RelError ↓Corr. ↑SSIM ↑
CBS254.80156.470.08180.96200.7370
CBNO410.74‾\underline{410.74}288.660.1293‾\underline{0.1293}0.8990‾\underline{0.8990}0.5150‾\underline{0.5150}
PINO431.47280.31‾\underline{280.31}0.13490.88640.3669
FFNO481.58338.820.14960.85520.3274
DPOT445.60285.950.14010.88510.3609

Bold and underline denote the best and second-best values, respectively; RMSE and MAE are in m/s.

CBNO performs best among the learned surrogates on four of five metrics, covering both reconstruction error and structural similarity. PINO and DPOT achieve lower MAE, while CBS leads overall. These results show that CBNO's forward-modeling advantages translate into competitive downstream inversion performance under the tested protocol. We have added this comparison alongside the Section 3.2 inversion discussion, with the full results in Appendix E.4 (pp. 28–29), Table 13 (p. 30).

W2/Q2. Novelty and structural benefits of the reflection-aware objective.

Re: Insightful question! Our contribution lies in integrating established upgoing-wave selection into the differentiable inversion objective of the frozen CBNO, with consistent processing of predicted and observed wavefields to emphasize reflections. Table 3(b) reports quantitative results for full-wavefield and upgoing objectives, averaged equally over Marmousi and Overthrust:

ObjectiveRMSE ↓MAE ↓RelError ↓Corr. ↑SSIM ↑
Full wavefield412.76396.530.14750.79570.2893
Upgoing410.74288.660.12930.89900.5150

RMSE and MAE are in m/s.

The main benefit is improved structural recovery: RMSE remains similar, while MAE decreases by 27.2%, RelError by 12.3%, and SSIM increases from 0.2893 to 0.5150. The qualitative comparison on Overthrust uses CBNO with identical initialization, observations, frequencies, optimizer, and a 70-update budget, changing only the inversion objective. Full-wavefield reconstruction captures the broad folded structure. The upgoing reconstruction exhibits more distinct layer interfaces around the upper fold crest and deeper fold trough, as shown in the two enlarged regions. Local oscillatory artifacts remain visible. We have moved this qualitative comparison to Figure 8(b) (p. 10), with the detailed settings in Appendix E.6, alongside the structural-recovery discussion in Sections 3.4–3.5 and the two-dataset quantitative results in Table 3(b).

W3. Architectural novelty and the contribution of shared low-rank bases and gates.

Re: Thank you for examining our architectural contribution! Axial-AFNO replaces DPOT's grouped Fourier MLPs with cross-axis shared low-rank bases and axis-/wavenumber-specific gates. Acoustic propagation follows common physical laws across directions, yet heterogeneous structures, sources, and boundaries produce direction-dependent wavefields. This motivates our design: shared bases reuse channel representations across axes, while the gates adapt them to directional and spatial-scale differences, providing a compact and flexible spectral representation.

To evaluate this design, Table 3(a) compares the two backbones under the same four-expert configuration and CBS supervision (K = 5), with results averaged equally over Marmousi and Overthrust:

BackboneRelError ↓Corr. ↑SSIM ↑
DPOT0.15440.97320.9259
Axial-AFNO (CBNO)0.15360.98350.9375

Axial-AFNO improves correlation and structural similarity while slightly reducing relative error. Complementing these accuracy results, Table 6 reports 12.5%–18.4% fewer parameters per model across its specified comparison configurations. Together, these comparisons support the design's ability to preserve wavefield structure with a more compact parameterization. We have clarified this rationale in Section 2.4 and expanded the supporting discussion in Section 3.4 and Appendix C.2.

Remark: Thank you again for your careful assessment! We hope the additional comparisons and structural analyses clarify our contributions and address your concerns. With our best wishes!

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer fE19,

With the discussion period coming to a close, we hope our additional experiments and analyses have helped address your concerns. The revised manuscript also highlights all corresponding changes in red.

We would be grateful to hear your thoughts before the discussion ends. Thank you again for your constructive suggestions!

Best regards,
The Authors

Reviewer_fE19OpenReview ↗

responses

Thank you for the revision. I appreciate the effort you put into addressing my comments. My main concern has been substantially alleviated.

AuthorsOpenReview ↗

Thank You for Your Follow-up

Dear Reviewer fE19,

Thank you for your follow-up! We are glad that our revisions have substantially alleviated your main concern. It would be much appreciated if you could update your score accordingly. Thank you again for your thoughtful feedback and consideration!

Best regards,
The Authors

Reviewer_g9PmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors focused on the task of learning wavefield mappings for full waveform inversion (FWI) without the need for precomputed target wavefields. They proposed a neural operator framework, named Physics-supervised Convergent Born Neural Operators (CBNO), which leverages the fixed-point consistency of the Convergent Born Series (CBS) to achieve the label-free training objective. A frequency-partitioned Axial-AFNO architecture was further introduced for broadband wavefield prediction. The trained surrogate is then frozen and incorporated into gradient-based FWI with an upgoing-wavefield-based objective. Experimental results show that their proposed method achieved competitive forward modeling accuracy and relatively faster inversion than the CBS baseline.

Critical strengths

1. Physics-supervised training strategy enables label-free training

The proposed CBS fixed-point consistency loss provides a promising alternative to the conventional supervised wavefield regression and differential PDE residual losses. The formulation is physically motivated and supported by a conditional theoretical analysis.

2. CBNO yields strong forward modeling performance

CBNO achieves consistently competitive wavefield prediction accuracy across the evaluated ID and OOD benchmarks, including Marmousi and Overthrust. The reported improvements over existing neural operators, together with the speedup relative to the CBS numerical solver, demonstrate the potential of the proposed approach as a differentiable forward surrogate.

Critical weaknesses

1. Limited inversion accuracy and insufficient evidence for practical FWI.

Although CBNO accelerates inversion compared with the CBS baseline, its inversion accuracy remains substantially lower on the OOD benchmarks. On the ID benchmarks, CBNO also underperforms the evaluated data-driven FWI methods, while requiring considerably more inference time. More importantly, the inversion experiments initialize the velocity models by smoothing the ground-truth velocity maps, which limits their ability to demonstrate robustness under realistic initialization. The paper also lacks an evaluation of the accuracy of the velocity gradients provided by the learned surrogate, which is essential for FWI.

2. Insufficient evaluation of scalability and generalization beyond the training configurations.

The experiments in the paper do not establish cross-resolution generalization or robustness to different spatial grid spacings. The 3D experiments are restricted to 1283128^3 grids, without a clear description of how the original OpenFWI 3D volumes are processed (Kimberlina 3D, 350×400×400350\times 400 \times 400). Furthermore, the frequency-specific experts are trained on 4–20 Hz, but their ability to be adpated to other frequency ranges is not demonstrated. Finally, the choice of a relatively small training set is not supported by a scaling experiment, making it difficult to assess the data efficiency and the potential benefits of additional training data.

Questions for authors

1. Inversion accuracy and robustness

Could the authors evaluate CBNO-based FWI with initial velocity models that are not derived from the ground truth? An experiment comparing the CBNO-derived velocity gradients with numerical or adjoint gradients would also be valuable.

2. Scalability and 3D dataset construction

Could the authors clarify how the 1283128^3 training volumes were obtained from OpenFWI, including the preprocessing and grid spacing? Also, I'm curious if a model trained at one resolution can be evaluated at different spatial resolutions without retraining. Results on accuracy, computational time, and GPU memory across multiple resolutions would better support the scalability claims.

3. Frequency generalization

Can the pretrained CBNO be generalized to the frequencies outside the 4–20 Hz training range? If additional frequency experts or fine-tuning are required, what is the computational cost?

4. Data efficiency and computational tradeoffs

What motivates the use of only 150 velocity samples per category? Could the authors provide a training data scaling study to demonstrate whether the performance saturates at this scale? A comparison of the total training cost, including CBS iterations, against supervised operator learning would help quantify the practical benefits of label-free training.

6 comments
AuthorsOpenReview ↗

Response to Reviewer g9Pm (Part 1/2)

Thank you for your thoughtful review and for recognizing our forward-modeling results and CBS supervision! We first address inversion accuracy and spatial adaptation, then discuss frequency adaptation and training cost in Part 2.

W1. Inversion accuracy and computational cost.

Accuracy–cost trade-off. With 50 velocity updates, CBNO takes approximately 60 s versus 720 s for CBS-based AD-FWI, giving a 12× speedup (Table 2). CBS achieves lower OOD RMSE (254.80 versus 410.74 m/s), illustrating the trade-off between runtime and reconstruction accuracy. The linear-initialization experiments below further examine this trade-off.

Acquisition flexibility. Table 2 reports OOD RMSEs of 410.74/890.17/984.97 m/s for CBNO/InversionNet/VeloGAN. Direct inversion is faster and more accurate in distribution, while CBNO performs better on the tested external benchmarks. The frozen CBNO also supports changed receiver layouts: removing approximately 30% of receivers increases RMSE by 0.82% (Appendix E.7). Geometry-specific direct inversion networks may require input adaptation or retraining.

These results are discussed in Section 3.2 (p. 8), Appendix E.5 (p. 30), and Appendix E.7 (pp. 30–31).

W1/Q1. Initialization and gradient fidelity.

Linear-background initialization. Our original experiments use GT-smoothed initial models, following prior neural-operator FWI work [1]. To examine more challenging initialization, both cases use the same laterally constant profile, increasing linearly from 1500 to 4500 m/s over 0–690 m depth, independently of the target model. All velocity nodes are updated, including the upper layers, with no fixed water layer or target-derived mask. Low-to-high frequency continuation followed by refinement at 20 Hz yields:

CaseInitial RMSE ↓CBS RMSE ↓CBNO RMSE ↓
Marmousi582.30267.09419.65
Overthrust801.98282.82306.89

RMSE is in m/s. Both methods recover the main velocity structures from linear backgrounds. CBS resolves layer boundaries more clearly, while CBNO captures the main velocity variations with localized artifacts, particularly in shallow Overthrust.

Gradient fidelity. To investigate the remaining accuracy gap, we compare CBNO and CBS gradients at identical velocity states along CBS inversion trajectories, using the same observations and upgoing correlation objective:

CaseStep 0Step 10Step 50
Marmousi0.8566460.6713810.450404
Overthrust0.9075330.7843690.745800

Gradient alignment weakens along the trajectories, especially for Marmousi, suggesting that derivative errors contribute to the inversion gap. Motivated by this finding, we have implemented joint physics-based self-supervision for forward predictions and gradients, requiring neither wavefield nor gradient labels, and are validating this extension for release in the near term.

Settings and results appear in Sections 3.2 and 3.5 and Appendix E.5 (Tables 14–15 and Fig. 8(a)).

W2/Q2. Spatial resolution, grid spacing, and 3D data.

Spatial adaptation. Starting from a model trained on a 70² grid with 10 m spacing, we rediscretize the same 690 m domain at different spacings. Network weights remain frozen, while input-size checks, PML padding, and output resizing are adjusted, with positional-embedding interpolation accommodating each grid.

GridSpacingID RelL2 ↓OOD RelL2 ↓
47²15 m68.66%71.25%
70²10 m14.56%15.36%
139²5 m104.25%101.08%

Changing grid spacing substantially degrades accuracy, showing that runtime size adaptation alone is insufficient for reliable cross-resolution generalization. Appendix E.9 (pp. 31–32) reports these results alongside runtime and memory scaling. They motivate grid-independent encoding/decoding and multiresolution CBS supervision.

3D preprocessing. We randomly select 50 OpenFWI3D volumes for training. Both Overthrust and OpenFWI3D volumes are reordered to (z, y, x) and trilinearly resampled onto a uniform 128³ physical grid with isotropic 25 m spacing. Each axis spans 3,175 m between the first and last nodes; coordinates are in meters. We then sample frequencies and source positions to construct c/(ωh) and the source mask. The 3D model is trained separately through CBS supervision, without precomputed wavefield labels. Appendix D.2 (p. 25) documents this pipeline.

References

[1] Huang, X., Wang, F., & Alkhalifah, T. (2025). Physics-Informed Waveform Inversion Using Pretrained Wavefield Neural Operators. IEEE Transactions on Geoscience and Remote Sensing, 63, 1–14.

AuthorsOpenReview ↗

Response to Reviewer g9Pm (Part 2/2):

We now turn to frequency adaptation and the data and computational requirements of training.

W2/Q3. Frequency range and adaptation cost.

The main broadband evaluation covers 4–20 Hz. To extend beyond this range, we evaluate the existing 16–20 Hz expert at 20, 22, and 24 Hz before and after 2,000 label-free fine-tuning updates. Fine-tuning updates this expert alone, preserving the expert count and routing within 4–20 Hz. Three-seed mean relative L2 errors are:

Evaluation frequenciesBeforeAfter
20, 22, 24 Hz20.18%15.58%

The mean error decreases by 22.8%, demonstrating label-free adaptation beyond the expert's original training band. At the measured update time of 1.4416 s on an RTX 4090 (Appendix C.4), 2,000 updates correspond to approximately 48.05 minutes of network-update computation, excluding data preparation, validation, and checkpoint saving. We have added these details in Appendix E.9 (pp. 31–32), Table 20 (p. 32).

W2/Q4. Dataset size and training cost.

1. Dataset size. Our training set contains 150 base velocity models per subset across six subsets (900 in total). Physics-based self-supervision enables online velocity augmentation through swap, flip, and blur, together with continuous frequency–source sampling, without generating additional wavefield labels. This expands the training conditions available from a relatively small collection of base models. To assess the effect of base-model count, we compare 50, 150, and 200 models per subset using the same 4–8 Hz expert, CBS K=5, and 30,000-update budget across three seeds:

Base models per subsetID validation RelL2 ↓
5017.35% ± 0.74 pp
15014.51% ± 1.74 pp
20014.49% ± 1.63 pp

Values are means ± sample standard deviations; pp denotes percentage points. Increasing the count from 50 to 150 reduces mean error by 2.84 percentage points, while increasing it to 200 yields a further 0.02-point reduction, much smaller than the across-seed standard deviations. These diminishing gains support 150 models per subset as a practical balance between data requirements and endpoint accuracy under this budget. Evaluation uses 24 fixed ID validation conditions, separately from the broadband OOD benchmarks. We report this study in Appendix E.10 (p. 32), Table 21 (p. 32).

2. Training cost. For the main broadband models, each of the four frequency experts receives 20,000 optimizer updates (80,000 in total). Profiling on an otherwise idle RTX 4090 gives 1.4416 s per self-supervised update, yielding an estimated 32.04 GPU-hours of network-update computation. Using the 50-update inversion times of 60 s for CBNO and 720 s for CBS in Table 2, this upfront computation corresponds to roughly 175 inversions worth of runtime savings (32.04 × 3,600 / (720 − 60)).

To examine the cost of the supervision itself, we also profile label-supervised and CBS-supervised updates with effective batch size 50 and CBS K=5:

QuantityLabel supervisionCBS supervision
Label generation per wavefield, measured0.508 sNot required
Time per update, measured1.430 s1.442 s
Training time for 10,000 updates, projected3.97 h4.00 h

CBS supervision and its wavefield derivative account for only 0.88% of the profiled self-supervised update time, while eliminating wavefield-label generation. For scale, 150 base models × 97 frequencies × 5 sources yield 72,750 labels per subset, or 436,500 across six subsets. Extrapolating from 128 timed labels gives approximately 10.26 h of CBS computation per subset (61.6 h in total), excluding data preparation, transfer, and storage. These label counts cover only the base models and sampled frequency–source combinations. Including additional samples from online augmentation would further increase label-generation time and storage requirements, whereas CBS supervision accommodates these samples directly without precomputed wavefield labels. The 10,000-update projection provides a common comparison basis; doubling it gives approximately 8.01 GPU-hours per expert, consistent with the 32.04 GPU-hour total above. We have added the cost breakdown and timing conditions in Appendix C.4 (p. 22), Table 9 (p. 22).

Remark: Thank you again for your constructive suggestions! We hope these additional results clarify the method's capabilities, current limitations, and training requirements. Best wishes!

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer g9Pm,

As the rebuttal deadline approaches, we would like to kindly follow up on our responses. We have also incorporated the corresponding revisions into the updated manuscript, with changes highlighted in red.

We would appreciate any further feedback or questions. Thank you for your time and consideration!

Best regards,
The Authors

Reviewer_M3qZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Physics-supervised Convergent Born Neural Operators (CBNO) for full-waveform modeling and inversion. The method introduces CBS fixed-point consistency for training neural operators without paired wavefield labels. It further combines frequency-partitioned Axial-AFNO experts with reflection-aware inversion. Experiments demonstrate promising forward modeling accuracy, computational efficiency, and generalization on several 2D benchmarks, along with preliminary 3D results.

Critical strengths

1.The CBS-based fixed-point loss provides a physically motivated alternative to conventional supervised neural operators, eliminating the need for precomputed wavefield labels.

  1. The frequency-partitioned Axial-AFNO combines low-rank spectral representations and frequency-specific experts, improving wavefield prediction across different frequencies.

  2. CBNO achieves strong forward modeling accuracy on unseen Marmousi and Overthrust models, demonstrating encouraging out-of-distribution generalization.

Critical weaknesses
  1. Despite its strong forward modeling performance, CBNO still exhibits a noticeable inversion accuracy gap compared with conventional CBS-based FWI, particularly on OOD benchmarks.

  2. Although CBNO is faster than the CBS solver, its inference latency exceeds that of several existing neural operators (Appendix C.3). The overall efficiency advantage, including offline training costs, remains unclear.

  3. The proposed upgoing-wave separation is effective in the evaluated setting but may be difficult to apply directly to field observations, where complete wavefield information is generally unavailable.

  4. The fixed-point consistency analysis establishes uniqueness under specific discrete-operator assumptions, but does not directly guarantee neural-network training convergence or quantify the relationship between small physics loss and wavefield prediction error.

Questions for authors
  1. What are the primary reasons for the remaining accuracy gap between CBNO-based inversion and conventional CBS-based FWI? Could this gap be reduced through additional physical constraints or surrogate refinement?

  2. Could the authors provide additional experiments on independent 3D geological models or more complex velocity distributions to further demonstrate generalization?

  3. What is the total offline training cost of CBNO, and how many forward simulations or inversion runs are required to amortize this cost relative to conventional numerical solvers?

  4. How sensitive is its performance to complex acquisition geometries and incomplete wavefield information?

This paper presents an interesting and physically motivated neural-operator framework for accelerating waveform modeling and inversion. The CBS-based supervision and broadband neural architecture are meaningful contributions, supported by promising experimental results. Although the inversion accuracy, computational trade-offs, and practical generalization warrant further investigation, I consider the overall contribution sufficient to support a weak acceptance.

4 comments
AuthorsOpenReview ↗

Response to Reviewer M3qZ (Part 1/2)

Thank you for your thoughtful review and constructive suggestions! We first address gradient accuracy, computational efficiency, and receiver-layout flexibility, then discuss the remaining questions in Part 2.

W1/Q1. Accuracy and gradients.

Insightful question! Inversion requires accurate velocity derivatives as well as accurate wavefield predictions [1,2]. To examine this relationship, we compared CBNO–CBS gradient cosine similarities at identical velocity states along CBS inversion trajectories, using the same observations and upgoing correlation objective:

CaseStep 0Step 10Step 50
Marmousi0.8566460.6713810.450404
Overthrust0.9075330.7843690.745800

Gradient alignment weakens along the tested trajectories, particularly for Marmousi, suggesting that gradient errors may contribute to the inversion accuracy gap. Motivated by this finding, we have implemented joint physics-based self-supervision for forward predictions and gradients, requiring neither paired wavefield nor gradient labels. Further validation is underway ahead of a planned release in the near term.

We have added the gradient analysis in Section 3.2 (p. 8), Appendix E.5 (p. 30), Table 15 (p. 30).

W2/Q3. Latency and training cost.

Good question! The comparison with learned surrogates reveals an accuracy–latency trade-off. Table 8 reports the following inference times and Marmousi errors:

MethodLatency (ms/sample)Marmousi RelError ↓
CBNO29.1830.1464
FFNO17.4260.6035
PINO11.4390.2707
DPOT25.0040.6597

Measurements use an RTX 4090, batch size 1, and FP32. CBNO achieves the lowest error among these learned surrogates, with a higher inference latency. Against the numerical solver, the timings in Appendix C.3 are 300.980 ms per sample for CBS and 29.183 ms for CBNO, giving a 10.31× forward-modeling speedup.

Repeated model use can amortize the upfront training computation. Each of the four frequency experts receives 20,000 optimizer updates (80,000 in total). Profiling on an otherwise idle RTX 4090 measures 1.4416 s per self-supervised update at effective batch size 50 and CBS K=5, yielding an estimated 32.04 GPU-hours of network-update computation, excluding data preparation, transfer, validation, and saving. Using the 50-update CBNO/CBS times of 60/720 s in Table 2, this computation corresponds to roughly 175 inversions worth of runtime savings (32.04 × 3,600 / (720 − 60)). We have clarified the per-expert training budget and computational cost estimates in Section 2.7 and Appendix C.4.

Q4. Receiver layouts.

Valuable question! The frozen CBNO accommodates the tested receiver layouts by adjusting observation sampling, without retraining. We evaluate this flexibility on Overthrust, using the same initial model, sources, frequencies, optimizer, and 70-update budget for CBS and CBNO. The layouts include interior receivers; irregular layouts perturb regular receiver positions by up to 10 m along each axis. Entries below are CBS/CBNO, with perturbed-layout results averaged over three random seeds.

LayoutRMSE ↓ (m/s)SSIM ↑Corr. ↑
Shared initial model419.640.30450.8628
Regular288.87/293.870.5675/0.56410.9367/0.9328
Irregular289.44/295.850.5669/0.56130.9363/0.9321
10% missing289.13/295.110.5668/0.56450.9364/0.9326
20% missing291.74/297.180.5613/0.56210.9353/0.9315
30% missing291.34/296.260.5575/0.56370.9352/0.9321

Both methods remain stable across the tested perturbations. Removing approximately 30% of receivers increases RMSE by only 0.82% for CBNO and 0.86% for CBS relative to their regular-layout baselines. This supports reuse of the frozen forward model across the evaluated layouts. For data-supervised direct inversion methods such as InversionNet, geometry changes may require input adaptation or retraining. We have added the acquisition settings and results in Appendix E.7 (pp. 30–31), Table 16 (p. 31), with the main discussion in Section 3.2.

References

[1] Czarnecki, W. M., Osindero, S., Jaderberg, M., Swirszcz, G., & Pascanu, R. (2017). Sobolev Training for Neural Networks. In Advances in Neural Information Processing Systems, 30, 4278–4287.

[2] O'Leary-Roseberry et al. Derivative-Informed Neural Operator: An Efficient Framework for High-Dimensional Parametric Derivative Learning. J. Comput. Phys. 496:112555, 2024.

AuthorsOpenReview ↗

Response to Reviewer M3qZ (Part 2/2)

We now turn to incomplete observations, the fixed-point analysis, and 3D generalization.

W3. Incomplete observations.

Valuable point! Frequency-domain upgoing-wave decomposition provides a convenient way to emphasize reflection energy in CBNO inversion. Table 3(b) shows that the full-wavefield objective also recovers velocity structure, with an RMSE of 412.76 m/s, close to 410.74 m/s with the upgoing objective. The main benefit of decomposition is improved structural recovery, with SSIM increasing from 0.2893 to 0.5150.

For practical time-domain receiver data, direct-arrival muting can also be considered to suppress strong direct arrivals and emphasize reflections. This provides a potential approach to reflection-enhancing preprocessing for incomplete observations. We have clarified this practical option in Section 2.6 (pp. 5–6), with the structural benefits of the upgoing objective documented in Appendix E.6 (p. 30).

W4. Fixed-point analysis.

Helpful suggestion! Appendix B.3 establishes the physical solution as the unique zero-loss wavefield under the stated discrete-operator assumptions. This characterizes the loss at the wavefield level; attaining a sufficiently low loss during network training additionally depends on model capacity, initialization, and optimization.

To examine how loss reduction relates to wavefield accuracy, we fixed the Marmousi velocity model and used two sources and 87 frequencies, giving 174 source–frequency pairs. Each complex wavefield was initialized to zero, with its grid-point real and imaginary values serving directly as optimization variables. We minimized the discrepancy between the current wavefield and the result of applying five CBS updates to it. The converged CBS solution served exclusively as the evaluation reference.

UpdatesPhysics lossRelative L2 error ↓
08.6204e-1100%
1,0002.2803e-168.4244%
5,0008.5916e-417.8089%
10,0003.6992e-55.5987%
20,0002.3894e-60.05170%
100,0002.3659e-60.05043%
500,0002.3668e-60.05043%

Loss and error are evaluated at the same iterates. Relative error covers the full domain, including PML, against a CBS reference computed with a stopping tolerance of 1e-5. As the physics loss decreases, wavefield error falls substantially and stabilizes near 0.0504%, providing numerical evidence that loss reduction accompanies improved wavefield accuracy in this experiment.

We have consolidated these experiments in Appendix B.3, presenting the shared setup, loss–error trajectories, and dependence on CBS depth K together. See Section 2.5 (pp. 4–5), Appendix B.3 (pp. 17–19), Tables 4–5 (pp. 18–19), and Figures 10–11 (pp. 18–19).

Q2. 3D generalization.

Valuable suggestion! To examine performance across geological structures, we evaluated three Kimberlina subvolumes from OpenFWI3D and three volumes generated using StructuralGeo [3]. The StructuralGeo volumes use different random seeds and span complex geological structures, including dipping layers, folds, and faults. All supplementary evaluation grids use 25 m spacing in each direction, with spatial coordinates in meters. The following results characterize performance on these specified volumes, measured by relative L2 wavefield error against CBS:

Kimberlina4 Hz5 Hz6 Hz
Subvolume 112.63%12.98%14.21%
Subvolume 214.91%16.74%19.19%
Subvolume 316.99%20.32%20.64%
StructuralGeo4 Hz7 Hz10 Hz
Model 196.72%94.81%102.09%
Model 292.67%88.94%100.95%
Model 396.13%93.54%101.86%

CBNO achieves substantially lower errors on the evaluated OpenFWI3D Kimberlina subvolumes, whose velocity structures are relatively simple and similar, consisting mainly of flat or gently dipping layers. Performance degrades on the more structurally diverse and complex StructuralGeo volumes. This contrast motivates more diverse physics-supervised 3D training to broaden geological coverage. We have added these evaluations and the discussion in Section 3.3 and Appendix E.8 (p. 31).

References

[3] Ghyselincks, S., Okhmak, V., Zampini, S., Turkiyyah, G., Keyes, D., and Haber, E. Synthetic Geology: Structural Geology Meets Deep Learning. Journal of Geophysical Research: Machine Learning and Computation, 2026.

Remark: Thank you again for your thoughtful review! We hope these additional analyses clarify the accuracy, efficiency, and scope of our method. Wishing you all the best!

Thanks for your rebuttal. I will keep the score as is.

AuthorsOpenReview ↗

Thank You for Your Review

Thank you for your time and valuable feedback, which has helped improve our manuscript. Wishing you a pleasant day!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses

Dear Reviewers, Area Chairs, and Program Chairs,

Thank you for your time and constructive feedback! Our central contribution is learning wavefield operators through CBS fixed-point consistency without paired wavefield labels. This provides a flexible way to incorporate physics into training and supports online sampling of media, frequencies, and sources. All four reviewers positively recognized this core idea. Guided by their suggestions, we have added experiments and analyses that strengthen the following aspects of the manuscript.

1. Stronger evidence for downstream inversion.

Under matched inversion settings, CBNO leads the compared learned surrogates on four of five metrics (Reviewer fE19: W1/Q1). Additional experiments recover the main velocity structures from target-independent linear backgrounds (Reviewer g9Pm: W1/Q1), while the reflection-aware objective improves structural similarity and produces clearer layer interfaces (Reviewer fE19: W2/Q2). These results strengthen the connection between our forward-modeling advantages and downstream inversion performance.

2. Clearer support for physical supervision and model design.

Direct wavefield optimization shows that decreasing the physics loss accompanies improved wavefield accuracy, complementing the conditional fixed-point analysis (Reviewer M3qZ: W4). Further experiments support K=5 as a practical training choice, while inference requires only one network forward pass (Reviewer qtq2: Q5). Backbone comparisons also support Axial-AFNO's ability to preserve wavefield structure with a compact parameterization (Reviewer fE19: W3).

3. Quantified computational benefits.

Our timing results show 10.31× faster forward modeling and 12× faster inversion at a fixed 50-update budget, relative to CBS (Reviewer M3qZ: W2/Q3; Reviewer g9Pm: W1). The four-expert training budget corresponds to an estimated 32.04 GPU-hours of network-update computation (Reviewer qtq2: Q4), with CBS supervision and its wavefield derivative accounting for only 0.88% of the profiled update time. Generating the illustrative fixed label set would require approximately 61.6 hours of CBS computation, before accounting for additional augmented samples. Physics supervision directly accommodates these samples and continuously sampled frequency–source combinations without generating or storing paired wavefield labels (Reviewer g9Pm: W2/Q4).

4. Broader evidence for data efficiency and adaptation.

The training-data study supports the chosen base-model count under the evaluated budget, with online augmentation enriching the available training conditions (Reviewer g9Pm: W2/Q4). Label-free fine-tuning reduces error beyond the original frequency band by 22.8% (Reviewer g9Pm: W2/Q3), and the frozen model accommodates the tested receiver-layout changes without retraining (Reviewer M3qZ: Q4). These experiments demonstrate complementary forms of flexibility in training and deployment.

5. Better understanding of the remaining challenges.

Gradient comparisons clarify a potential contributor to the remaining inversion accuracy gap (Reviewer M3qZ: W1/Q1; Reviewer g9Pm: W1/Q1). CBS retains higher final inversion accuracy, while CBNO provides substantial runtime savings. Joint forward-and-gradient physics supervision has been implemented and is undergoing validation. Evaluations under changed grid spacing (Reviewer g9Pm: W2/Q2) and complex 3D geology (Reviewer M3qZ: Q2) further motivate multiresolution supervision and more diverse 3D training.

We have incorporated these experiments and clarifications into the revised manuscript, with all revisions highlighted in red. We have also detailed the 3D preprocessing and grid-scaling results (Reviewer g9Pm: W2/Q2). Our code will be publicly released upon acceptance (Reviewer qtq2: Q3).

Remark: We sincerely thank the reviewers, Area Chairs, and Program Chairs for your time and consideration. Your feedback has helped us strengthen the manuscript. We welcome further discussion and wish you all the best!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.