Time-Lapse Seismic Inversion Using a 3D Neural Operator

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

As I understand it, this work deals not with the full formulation of full-wavefield inversion, but rather with seismic monitoring. However, since the variability of the medium's parameters is quite low, the required computational resources can be significantly reduced by limiting the approach to a linearized formulation—also known as the Born approximation. This allows for a substantial reduction in computational resource requirements.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript entitled “Time-Lapse Seismic Inversion Using a 3D Neural Operator” is within the scope of this journal. This manuscript presents a 3D neural-operator-based framework for time-lapse post-stack seismic inversion. The method is demonstrated on both the Kimberlina synthetic CO₂ monitoring dataset and the Sleipner field dataset, with additional comparisons against conventional Tikhonov and double-difference inversion. Overall, the manuscript is clearly motivated, technically sound, and relevant to time-lapse seismic monitoring and CO₂ storage applications. I believe the manuscript is suitable for publication after minor revision. Some sections could benefit from more explanation or clarification.

1. The forward mapping implemented in this study is based on post-stack reflectivity and wavelet convolution. It would be helpful to state the neural operators more explicitly in the Introduction and Discussion so that readers clearly understand that the current neural operator approximates a post-stack convolutional forward operator rather than a full wave-equation propagator.
2. Please provide more details on the two-channel input strategy. During inversion the first channel is iteratively updated and the second channel remains the initial model? What’s the useful for the unchanged second channel?
3. The construction of the initial impedance models should be described more explicitly. The training-data generation procedure should provide the ranges of impedance, layer thickness, fault displacement, fold geometry, and noise levels.
4. The ground truth is available for the Kimberlina synthetic experiment, the authors may consider reporting one or two quantitative metrics, such as relative error, RMSE, or correlation coefficient, for the recovered impedance or time-lapse impedance differences?
5. It would be useful to show inversion-loss curves for representative Kimberlina and Sleipner examples.
6. Some figure-label descriptions should be made consistent throughout the manuscript.
7. Page 7, Equations (8)–(10): There seems to be a notational inconsistency between the equations and the actual network input.
8. Page 14, Figure 9 caption: “inlince” is a typographical error?
9. Page 16, Lines 367–374: The expressions “bypasses this trade-off entirely” and “parameter-free behavior” are somewhat too strong.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

•  This is extended version of https://www.earthdoc.org/content/papers/10.3997/2214-4609.202610122.  What is the added value of this document wrt the published one?
• I'm missing accuracy metrics, all data is either synth or it has vintages that can be used as ground-truth. Further, post-stack gathers is a common practice for QA, and it is missing here.
• Despite computational costs are the main motivator, I don't see a section with that analysis, just one mention about 15 minutes elapsed time for the proposed method, but does compare with a efficient competitive FWI? remember to add training into the comparison, and because this method will not generalize across fields, that part will be needed to be repeated for every new case.
• Figure 14 is hard to read, can you zoom into the interesting sections?
• The following reference is relevant: https://arxiv.org/abs/2310.04430

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf