{"id":"09e1f49d-456b-4ea0-ad44-b3a2ac13af3f","arxiv_id":"2501.12992","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A convolutional neural network maps high-frequency elastic FWI updates to a low-wavenumber initial model, improving FWI convergence on data without low frequencies.","lead":"This paper trains a neural network to predict the smooth, low-wavenumber part of the subsurface from the rough, high-frequency updates produced by elastic full-waveform inversion when low-frequency data are absent. In tests on two benchmark models and a real marine streamer dataset, the network's prediction gives FWI a better starting model and reduces cycle-skipping artifacts compared to starting from a one-dimensional model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CNN is trained on FWI updates from an L2 misfit with 8/10 Hz stages, but applied to updates from a global-correlation misfit with 6/6/7/8/8 Hz stages; this input distribution shift is untested and load-bearing.","rationale":"The reader's verdict is CONDITIONAL, and our analysis does not change that level, but it identifies a more specific, internally verifiable vulnerability than the reader's field-transfer concern. The reader's weakest assumption focuses on external transfer (synthetic geology/acquisition covering the field case); our concern is an internal distribution shift between the FWI configuration used to generate CNN training inputs and the FWI configuration used at inference. Both are distribution-shift issues, hence 'partial' agreement. The mismatch is explicitly stated in the paper ('different from the approach for generation of CNN-input data we chose global correlation norm misfit'), making it a factual inconsistency in the workflow rather than a speculation about external conditions. The benchmark experiments do provide some evidence that the CNN still works under the shifted configuration, which is credit to the method, but they were not designed to isolate this variable. A targeted test comparing training-configuration and inference-configuration inputs on the same models would settle whether the mapping is robust or whether the improved FWI results are an artifact of the specific inference workflow. Until such a test is run, the central claim remains conditional, consistent with the reader's verdict. No new verdict level is required, so we recommend UNCHANGED, with the caveat that the paper should add this robustness check or an explicit discussion of why the shift is negligible.","tokens_in":15347,"tokens_out":5176,"duration_ms":52218,"concrete_test":"For a fixed set of validation models (e.g., Marmousi and Overthrust), generate two sets of 10 consecutive FWI model-update inputs: (A) using the training configuration from the 'Synthetic data'/'Wave propagation' sections (L2 objective, 5 iterations with <8 Hz low-pass plus 5 with <10 Hz low-pass) and (B) using the inference configuration from the 'FWI Results' section (global correlation norm, stages at 6/6/7/8/8 Hz, gradient damping 1.5/1/0.1/0.25/0.125). Run the same trained CNN on both sets, compare predicted low-wavenumber updates and final FWI R2 scores. If A and B differ materially (e.g., >0.05 in R2), the CNN is being applied outside its training distribution and the central claim is not established; if they match closely, the configuration mismatch is benign.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that a FusionNet CNN can predict low-wavenumber initial models from high-wavenumber FWI updates is only meaningful if the CNN is applied to inputs from the same distribution it was trained on. The paper's own workflow description shows this is not the case. In the 'FWI Results' section, the authors state: 'different from the approach for generation of CNN-input data we chose global correlation norm misfit as more robust, than L2 misfit,' and they run five inversion stages with low-pass cutoffs 6, 6, 7, 8, 8 Hz, with varying gradient damping. Training inputs, by contrast, were generated with the L2 objective and two stages with 8 Hz and 10 Hz low-pass filters, as described in the 'Synthetic data' and 'Wave propagation' sections. Thus, the CNN at inference receives FWI updates that differ from its training inputs in both amplitude scaling (L2 vs correlation misfit) and spectral content (two-stage 8/10 Hz vs five-stage 6/6/7/8/8 Hz). The paper provides no experiment or analysis quantifying this shift or demonstrating that the CNN is robust to it. The benchmark results (Marmousi R2 0.653 to 0.694; Overthrust R2 0.497 to 0.526) were all obtained under the shifted inference configuration, so they show the CNN can still produce useful outputs in two cases, but they do not establish that the learned mapping is the intended one under the exact conditions of use. If the CNN's predictions are sensitive to the FWI parametrization, the claimed improvement is not a reliable property of the method but an artifact of the particular inference workflow chosen. This is a controllable, internal inconsistency rather than an external transfer question, and it directly affects the validity of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deep-learning workflow for constructing low-wavenumber initial models for elastic full-waveform inversion (FWI) when low-frequency data are missing. A FusionNet-based CNN is trained on 3,904 synthetic random layered models to map ten consecutive high-wavenumber FWI model updates (plus a water taper) into a low-wavenumber model update; the target is a Gaussian-smoothed true model minus the 1D initial model. The trained network is then applied to the Marmousi II and Overthrust benchmark models and to a 2D marine streamer dataset from the Australian shelf. The authors report that FWI starting from the CNN-predicted initial models yields higher R2 scores than the 1D initial model on both benchmarks (Marmousi: 0.653 to 0.694 vs 0.569 to 0.508 for the 1D start; Overthrust: 0.497 to 0.526 vs 0.451 to 0.458), and qualitatively better matches to a single well log on field data, with fewer visible cycle-skipping artifacts.","tokens_in":15703,"tokens_out":5682,"duration_ms":57774,"significance":"Assuming the results hold, the method offers a practical, comparatively cheap way to build initial models for elastic FWI without low frequencies, because the CNN operates on FWI by-products and inference is fast once the network is trained. The paper is transparent about data generation, architecture, and training details, and it validates on two unseen benchmark models; the authors also promise to release code on GitHub. However, the significance is tempered by two weaknesses: the CNN is applied under an FWI configuration (global-correlation misfit, 6/6/7/8/8 Hz stages) that differs from the training configuration (L2 misfit, 8/10 Hz stages) without a sensitivity analysis, and the field-data demonstration relies on a single well log with no quantitative metric. These issues, rather than the core idea, are the main limitations.","major_comments":[{"comment":"The CNN is trained on input updates generated with the L2 misfit and two low-pass stages at 8 Hz and 10 Hz, as described in the 'Wave propagation' section, but at inference the paper uses a global-correlation norm misfit and five stages with low-pass cutoffs 6, 6, 7, 8, and 8 Hz, as described in the first paragraph of 'FWI Results'. The network is thus applied to input updates with different amplitude scaling and spectral content than its training distribution, and no experiment or analysis quantifies its robustness to this shift. Since all benchmark results in Figures 10 and 11 were obtained under the shifted configuration, they show empirical robustness in two cases but do not establish that the learned mapping is the intended one under the exact conditions of use. Please add a sensitivity test, for example by generating CNN inputs with the inference workflow and comparing predictions, or by retraining on the inference configuration.","section":"FWI Results / Wave propagation"},{"comment":"The field-data validation rests entirely on a single acoustic Vp well log and visual inspection; no quantitative metric (such as R2 or correlation along the well) is reported, and no error bars or training-seed variation are provided. The statement in the Conclusions that the method 'led to better FWI results' on the field dataset is therefore not quantitatively supported. Please add a quantitative well-log comparison and, ideally, a small ensemble of training runs to assess variability.","section":"Field data"},{"comment":"The R2 improvement from the CNN-predicted initial model on the Overthrust benchmark is 0.497 to 0.526, while the CNN-target initial model (the smoothed true model) actually decreases R2 from 0.678 to 0.674 after FWI. This shows that even the 'ideal' initial model does not guarantee improvement on this benchmark, so the modest gain from the CNN prediction may be within the noise of the workflow. Please discuss this result explicitly and provide additional evidence that the improvement is statistically meaningful, or temper the claim accordingly.","section":"Overthrust / Figure 11"}],"minor_comments":[{"comment":"Figure 5's caption says '12 input channels' but the text describes 11 channels (10 FWI updates plus the water taper); please correct the inconsistency.","section":"Training dataset preparation / Figure 5"},{"comment":"The outlier removal step is described only qualitatively ('we removed those models from the training dataset'); please specify the quantitative criterion used, such as the amplitude threshold on the model updates.","section":"Training dataset preparation"},{"comment":"The Gaussian smoothing kernel used to construct the CNN target is not specified; please report the kernel width or standard deviation.","section":"Synthetic data"},{"comment":"In Figures 7 and 8, the subplot labels (b) and (c) are duplicated (one for the CNN inference result and one for the target); please renumber the subplots.","section":"Testing CNN performance on Marmousi/Overthrust"},{"comment":"The conclusion that the method 'led to better FWI results' on field data is stronger than the evidence in the 'Field data' section supports, given the single-well, qualitative comparison; please align the wording with the evidence.","section":"Conclusions"},{"comment":"The manuscript states that open-access data to reproduce the paper 'will be available' on GitHub; please clarify the current status of the repository and provide a version or commit identifier for reproducibility.","section":"Conclusions / GitHub statement"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and presents an interesting workflow with encouraging synthetic benchmark results. The main risks are the untested train/inference distribution shift and the weak field-data validation; these are fixable with additional experiments and should be addressed before publication. The citation list is extensive and includes several works by the same group, which is unsurprising given the niche topic; I do not consider this a problem. The promise of open code is positive, but the repository should be checked for availability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper does what it says. It trains a FusionNet on synthetic elastic FWI updates from data with missing low frequencies, then shows that using the predicted low-wavenumber model as an FWI starting model beats a 1D starting model on two external benchmarks (Marmousi II, Overthrust) and gives a reasonable field-data result. The benchmarks are genuinely external—neither model is in the training set—and the R2 improvements are consistent, if modest (Overthrust goes from 0.458 to 0.526). That is the core evidence, and it holds up.\n\nThe main new thing is the extension to elastic FWI updates and to a real marine streamer dataset. The earlier acoustic work by the same group only went as far as synthetic examples. Here the synthetic training is tied to the field acquisition parameters (wavelet, offsets, water layer), which is the right way to approach transfer.\n\nThe soft spots are real but not fatal. The stress-test note is correct: the CNN is trained on L2-misfit FWI updates with 8/10 Hz two-stage low-pass filtering, but inference is run on global-correlation-misfit updates from a 6/6/7/8/8 Hz five-stage workflow. That is a distribution shift in both amplitude and spectral content, and the paper does not quantify its effect. That said, the benchmark results are all obtained under the shifted inference configuration, so the CNN demonstrably tolerates that shift in two cases. A cleaner paper would add an ablation (run the same FWI workflow for training-input generation, or run inference on L2 updates) to show the mapping is not an accident of the particular inversion setup.\n\nThe field-data section is the other weak point: one well log, qualitative comparisons, and a known velocity overshoot below 3 km. The authors are honest about this, and they also acknowledge the method does not eliminate cycle-skipping. The outlier-removal step in training-data curation is mentioned but not quantified. The promised code/data repository is not yet available, which limits reproducibility.\n\nWho should read it: anyone working on DL-based initial model building for FWI, and people thinking about train-test distribution shifts in geophysical DL. It deserves a serious referee. Send it out, but ask for (i) an experiment addressing the training/inference misfit mismatch, (ii) some quantitative field validation or at least uncertainty estimates, and (iii) the repository. If those come back, the paper goes from conditional to solid.","headline":"A solid incremental extension of the authors' earlier CNN-based low-wavenumber prediction to elastic FWI and real data; the central claim survives the benchmarks, but the training/inference mismatch and thin field validation need work.","tokens_in":16337,"tokens_out":3750,"would_cite":true,"duration_ms":36482,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Deep learning predicts low-wavenumber starting models for elastic FWI from high-frequency data alone.","keywords":["full-waveform inversion","elastic FWI","deep learning","low-wavenumber model","initial model prediction","marine streamer data","FusionNet","cycle-skipping"],"falsifier":"Take a field dataset from a different basin with unknown water-bottom geometry and salt-containing geology, generate the ten FWI updates on data above 5 Hz, apply the trained network, and compare the resulting FWI convergence and well-log match against a 1D-initial-model baseline; a failure to improve (or a larger final misfit) would falsify the claim that the CNN-predicted initial model generalizes to new marine streamer data.","tokens_in":15154,"feed_emoji":"🌊","tokens_out":10031,"duration_ms":89943,"temperature":0.7,"pith_summary":"Full-waveform inversion (FWI) needs low-frequency seismic data to recover the smooth, low-wavenumber part of the subsurface model, but such data are often missing in marine streamer surveys. This paper tries to show that a convolutional neural network can substitute for that missing information: it maps the high-wavenumber model updates produced by elastic FWI on band-passed data into a low-wavenumber initial model, effectively restoring the background velocity structure. The authors train a FusionNet-based CNN on 3,904 synthetic elastic models built with the field survey's acquisition parameters, then test it on the Marmousi II and Overthrust benchmarks and on a real marine streamer dataset from the North-West Australian shelf. In all three tests, elastic FWI started from the CNN-predicted initial model ends with higher $R^2$ similarity to the true model and fewer cycle-skipping artifacts than FWI started from a one-dimensional initial model. If correct, this offers a practical route to initial velocity model building that does not require low-frequency acquisition, though the paper also documents limitations such as velocity overshoot below 3 km on the field data.","feed_headline":"Deep learning builds better starting models for elastic FWI","feed_subtitle":"On synthetic benchmarks and real marine streamer data, CNN-predicted starts beat 1D models in seismic inversion.","key_machinery":"The load-bearing object is the FusionNet convolutional neural network, a fully residual image-to-image architecture with nested short skip connections inside each residual block in addition to long skip connections. The paper increases the filter count to 110, giving about 231 million trainable parameters, and inserts dropout layers after each convolutional block. The network is fed an 11-channel input: the first ten channels are the successive velocity-model updates from ten elastic FWI iterations run on data high-pass filtered above 5 Hz, and the eleventh channel is a water taper with values from 0 to 1 marking the water column. The target is the low-wavenumber model update, obtained by subtracting the initial model from a Gaussian-smoothed true model. The architecture's cumulative receptive field is what lets shallow layers see high-frequency reflectivity patterns while deeper layers integrate them into smooth background-velocity updates; the water taper supplies the known sea-floor geometry that anchors the low-wavenumber reconstruction.","core_discovery":"The central claim is that the smooth, low-wavenumber model update that would normally come from low-frequency seismic data can be learned from the high-wavenumber FWI updates obtained on data with low frequencies removed. The authors construct a supervised regression problem in which the input is a stack of the first ten elastic FWI velocity-model updates plus a water taper that encodes the known sea-floor geometry, and the target is the low-wavenumber update obtained by subtracting the initial model from a smoothed true model. A FusionNet-based fully residual convolutional network trained on 3,904 synthetic layered models, generated with the acquisition design of a real marine streamer survey, learns this mapping. On the Marmousi II and Overthrust test models, which were not in the training set, the CNN-predicted initial model improved the final FWI $R^2$ (from 0.569 to 0.694 for Marmousi and from 0.451 to 0.526 for Overthrust) relative to starting from the a priori 1D model, although it remained below the ideal case of starting from a smoothed true model. On the field data, FWI from the CNN-predicted initial model produced fewer cycle-skipping artifacts and better matched the available well log at several depths than FWI from the 1D initial model, with the paper explicitly noting velocity overshoot below 3 km.","pith_inferences":["One implication the authors leave implicit is that the same input-target construction could be carried over to other inverse problems where high-wavenumber updates are available but smooth background structure is missing, such as medical ultrasound tomography or ground-penetrating radar, as long as a training distribution can be sampled.","Because the input includes a water taper with known sea-floor geometry, the method's field transfer would be challenged by surveys with unknown or rapidly varying water-bottom topography; an extension would be to train with randomized tapers or predict the water bottom jointly.","The velocity overshoot below 3 km on the field data suggests the synthetic generator under-represents deep high-velocity geology; extending the generator with deeper high-velocity layers or salt-like bodies could reduce this bias and improve the field result.","The fact that FWI from the smoothed-true initial model still clearly outperforms FWI from the CNN-predicted model (0.870 vs 0.694 on Marmousi) indicates that the remaining bottleneck is the CNN's error relative to the smoothed true model, so further gains would come from better low-wavenumber prediction rather than from changing the FWI scheme."],"forward_implications":["FWI workflows that currently rely on low-frequency acquisition or heavy regularization can instead initialize from a CNN-predicted low-wavenumber model, reducing the need for frequencies below 5 Hz.","A network trained for one marine streamer survey design can be applied to new data from the same survey without retraining, provided the acquisition parameters match.","The method complements existing initial-model building tools such as ray-based tomography and gradient filtering, since it operates on FWI by-products and can be inserted into existing workflows.","The gain is largest when the a priori 1D initial model is far from the smoothed true model; when the starting 1D model is already accurate, the CNN adds little.","Generating the training dataset is computationally expensive (about 20–40 minutes per model on 160 CPU cores), but inference on a new model is fast once the network is trained."],"supporting_citations":[{"why":"Introduced CNN-based low-wavenumber model extrapolation, the direct predecessor this work extends to elastic FWI.","marker":"Plotnitskii et al., 2019"},{"why":"Mapped high-wavenumber FWI updates to low-wavenumber ones in a synthetic acoustic example, the immediate method advanced here.","marker":"Plotnitskii et al., 2020"},{"why":"Demonstrated synthetic-to-field transfer of low-frequency extrapolation on real marine and land data, motivating the field-data test.","marker":"Ovcharenko et al., 2021b,c"},{"why":"Supplied the FusionNet architecture whose fully residual nested skip connections the paper adapts.","marker":"Quan et al., 2021"},{"why":"Provided the open-source time-domain elastic FWI code used to generate the training updates and run inversion experiments.","marker":"Köhn et al., 2012"},{"why":"Established the processing workflow for the North-West Australian marine streamer field dataset.","marker":"Kalita and Alkhalifah, 2017"},{"why":"Provided the global correlation norm misfit used in the multiscale FWI workflow.","marker":"Choi and Alkhalifah, 2012"},{"why":"Supplied the Marmousi II benchmark used to test generalization to unseen geology.","marker":"Martin et al., 2006"},{"why":"Supplied the Overthrust benchmark used to test generalization to thrusted geology.","marker":"Lecomte et al., 1994"}],"fun_headline_variants":["Deep learning supplies low-frequency missing model for elastic FWI","CNN predicts initial models that boost elastic FWI on marine data","FusionNet learns low-wavenumber updates for elastic FWI starts","DL initial models improve elastic FWI convergence without low-freq data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if the synthetic models used for training are representative enough of the real subsurface and acquisition, and if the sea-floor geometry is known; the paper's own field result showing velocity overshoot below 3 km suggests this can fail outside the training distribution.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning supplies low-frequency missing model for elastic FWI","CNN predicts initial models that boost elastic FWI on marine data","FusionNet learns low-wavenumber updates for elastic FWI starts","DL initial models improve elastic FWI convergence without low-freq data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1829,"prompt_tokens":997,"completion_tokens":832,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":759}},"tokens_in":613,"tokens_out":832,"duration_ms":8858,"temperature":1.0,"reasoning_tokens":759,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:33:06.080252+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a field dataset from a different basin with unknown water-bottom geometry and salt-containing geology, generate the ten FWI updates on data above 5 Hz, apply the trained network, and compare the resulting FWI convergence and well-log match against a 1D-initial-model baseline; a failure to improve (or a larger final misfit) would falsify the claim that the CNN-predicted initial model generalizes to new marine streamer data.","supporting_citations":[{"cited_title":"Alkhalifah, O","cited_arxiv_id":null,"evidence_quote":"Introduced CNN-based low-wavenumber model extrapolation, the direct predecessor this work extends to elastic FWI."},{"cited_title":"Kazei, O","cited_arxiv_id":null,"evidence_quote":"Mapped high-wavenumber FWI updates to low-wavenumber ones in a synthetic acoustic example, the immediate method advanced here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplied the FusionNet architecture whose fully residual nested skip connections the paper adapts."},{"cited_title":"Alkhalifah, 2017, Efficient full waveform inversion using the excitation representation of the source wavefield : Geophysical Journal International, 210 , 1581--1594","cited_arxiv_id":null,"evidence_quote":"Established the processing workflow for the North-West Australian marine streamer field dataset."},{"cited_title":"Alkhalifah, 2012, Application of multi-source waveform inversion to marine streamer data using the global correlation norm: Geophysical Prospecting, 60 , 748--758","cited_arxiv_id":null,"evidence_quote":"Provided the global correlation norm misfit used in the multiscale FWI workflow."},{"cited_title":"Campbell, and J","cited_arxiv_id":null,"evidence_quote":"Supplied the Overthrust benchmark used to test generalization to thrusted geology."}],"review_version":1}