{"id":"4fe085c3-58ad-4f2f-b4d0-7f52ef5435e0","arxiv_id":"1908.07824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Recurrent network cells, particularly GRU and LSTM, produce more accurate velocity models from synthetic seismic gathers than the prior CNN baseline, with the RNN using up to 78% fewer parameters.","lead":"This paper applies recurrent neural networks (RNNs, LSTMs, GRUs) to velocity model building from seismic data, replacing part of a convolutional network. The recurrent models, especially GRU and LSTM, achieve slightly better image quality on synthetic tests while using many fewer parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported GRU/LSTM advantage is confounded: Table 1 changes layer count, pooling, and parameter count alongside recurrence, so the result does not isolate recurrence; the salt-body claim lacks a salt-region metric.","rationale":"The reader's conditional verdict is appropriate: the paper reports a plausible synthetic evaluation but does not yet establish that recurrence, rather than other architectural differences, causes the improvement. The weakest assumption identified by the reader already flags the non-ablation nature of replacing layers 3 and 4 with a recurrent layer and the different parameter counts. I agree with that reading but sharpen it: the removal of maxpool is a separate structural change that is explicitly mentioned in the architecture table, and the salt-body claim is separately under-quantified. These concerns do not refute the empirical finding that the specific GRU/LSTM architectures perform better on this synthetic test set; they only show that the causal attribution to recurrence and the salt-body-specific claim are not yet supported. Therefore the verdict should remain conditional: the paper should be accepted only if the authors provide a matched non-recurrent baseline and a quantitative salt-region metric. A rejection would be too harsh for an extended abstract whose primary contribution is architectural exploration, and an unconditional acceptance would overstate the evidence. No formal verification, code release, or independent replication is present, so the burden of additional controls lies with the authors. The proposed concrete test directly addresses the main confound and can be run with modest additional compute on the same dataset.","tokens_in":3230,"tokens_out":5002,"duration_ms":56570,"concrete_test":"Train a capacity-matched non-recurrent baseline with the same layer structure as the GRU/LSTM networks, i.e., Conv, a 512-unit non-recurrent fully connected layer, FC, Upscale, and the final Conv stack, with maxpool removed and total parameters matched to the GRU architecture. Use the same 9,600/2,400 split, 50 epochs, mini-batch size 100, and RMSprop MSE training. If this matched CNN matches or exceeds the GRU/LSTM SSIM and MSE, the reported advantage is due to capacity or architecture, not recurrence; if GRU/LSTM still win, the confound is at least partly resolved. In the same evaluation, additionally compute SSIM and MSE restricted to the salt-body pixels of the test models to test the salt-body-specific claim directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that recurrent cells improve VMB accuracy, not just parameter efficiency, is not identifiable from the reported comparison. In Table 1, moving from the CNN baseline to each recurrent network changes three variables at once: layers 3 and 4 are removed, a recurrent cell is inserted, and the maxpool is dropped (with the comment that it is 'not required'). Total coefficients also change substantially, from 7,182,728 for the CNN to 1,557,896–3,525,512 for the recurrent networks. Thus the lower MSE and higher SSIM of GRU and LSTM, relative to the CNN, could be due to reduced capacity, different depth, or the removed pooling operation, rather than to recurrence itself. The paper does not provide a matched non-recurrent baseline, so the effect of recurrence is not isolated. Additionally, the abstract's more specific assertion that salt bodies are 'consistently predicted more accurately' by GRU and LSTM is not directly supported by the reported numbers: Table 2 gives only global SSIM and MSE averaged over 2,400 models, and Figure 4 is illustrative. There is no quantitative salt-region metric, such as intersection-over-union or SSIM restricted to salt pixels, so the salt-body claim remains qualitative. The architecture comparison and the salt-body claim are the two load-bearing pillars of the paper; neither is adequately controlled or measured.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This short EAGE conference paper proposes recurrent neural network (RNN, LSTM, GRU) architectures for Velocity Model Building (VMB) from synthetic seismic gathers, building on an earlier CNN baseline of Araya-Polo et al. (2018). The authors replace two fully-connected layers of the CNN with a recurrent layer of 512 hidden units, then compare reconstruction quality on 2,400 synthetic test velocity models using SSIM and MSE, also reporting total network coefficients. They report that the GRU and LSTM variants achieve lower MSE and higher SSIM than the CNN, that salt bodies are 'consistently predicted more accurately' by these variants, and that the RNN saves about 78% of coefficients. The paper is a brief conference-style report with one architecture table, one results table, one illustrative figure, and four references to related work, but no code, no statistical significance testing, and no per-region quantitative evaluation.","tokens_in":3544,"tokens_out":2750,"duration_ms":29760,"significance":"If the central claim—that recurrent cells improve velocity-model accuracy beyond a parameter-efficiency benefit—were established, this would be a useful contribution to the emerging machine-learning tomography literature and a natural extension of the authors' prior CNN work. The paper has several strengths: the evaluation is performed on a reasonably large synthetic dataset (12,000 models, 2,400 test examples), the metrics (SSIM, MSE) are standard external image-quality measures, and the comparison against the authors' own earlier CNN is an appropriate baseline rather than a strawman. The coefficient-count reduction (21.68%–49.08% of the CNN coefficients) is directly documented. However, the significance as presented is limited because the experimental design does not isolate recurrence from other architecture changes, the salt-body claim is not supported by any quantitative region-specific metric, and the reported accuracy differences lack error bars or significance tests. The potential practical impact (reducing VMB turnaround from weeks to days) is stated but not quantified.","major_comments":[{"comment":"The comparison does not isolate the effect of recurrence. Moving from the CNN baseline to each recurrent network removes layers 3 and 4, inserts a recurrent layer with 512 hidden states, and removes the Maxpool layer all at the same time, while the total coefficient count changes from 7,182,728 to 1,557,896–3,525,512. The lower MSE and higher SSIM of the GRU and LSTM variants could therefore be attributable to reduced capacity, different depth, or the removed pooling operation rather than to the recurrent cell itself. A matched non-recurrent baseline (e.g., a CNN with the same layer count, pooling configuration, and coefficient budget) is needed to make the accuracy claim identifiable.","section":"Deep Recurrent Neural Network Architectures for Seismic Tomography, Table 1"},{"comment":"The abstract and conclusions assert that salt bodies are 'consistently predicted more accurately' by GRU and LSTM architectures, but Table 2 reports only global SSIM and MSE averaged over 2,400 test models, and Figure 4 is a set of illustrative examples. No quantitative metric restricted to salt regions (e.g., intersection-over-union, boundary error, or SSIM computed only over salt pixels) is provided. As written, the salt-body claim is qualitative and is not directly supported by the reported numbers.","section":"Performance Evaluation, Table 2 and Figure 4"},{"comment":"The results are reported without any measure of variability. Each network appears to be trained once over 50 epochs, yet the reported differences are small in absolute terms (SSIM 0.8414 vs. 0.8199; MSE 0.0013 vs. 0.0018). Without multiple training runs, confidence intervals, or a paired significance test over the 2,400 test models, the differences could be within run-to-run training variability. The word 'consistently' in the abstract is stronger than the evidence supports.","section":"Performance Evaluation, Table 2"}],"minor_comments":[{"comment":"There are several typographical and spacing errors: 'comput ing', 'suit able', and 'MSE a loss function' should read 'computing', 'suitable', and 'MSE as the loss function'.","section":"Introduction"},{"comment":"The dataset description is under-specified: the paper mentions 12,000 velocity models but then gives a split of 9,600 training pairs and 2,400 test pairs; it should clarify whether each pair corresponds to one model and how many gathers are generated per model. Acquisition geometry and forward modeling parameters are not stated, which limits reproducibility.","section":"Performance Evaluation"},{"comment":"The Chung et al. reference lists the authors as 'J. Chung and Ç. Glçehre and K. Cho and Y. Bengio'; the second author name appears garbled and should be corrected.","section":"References"},{"comment":"Figure 4 would benefit from clearly labeled axes, color bars with velocity units, and a statement of whether the displayed examples were selected randomly or chosen to illustrate the claimed improvement, since it is otherwise anecdotal evidence.","section":"Figure 4"},{"comment":"The table gives layer types but not layer dimensions, kernel sizes, strides, or pooling parameters; readers are forced to consult Araya-Polo et al. (2018) and the details of the recurrent-layer input shaping are not given, making the architecture comparison hard to reproduce.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a short EAGE paper that takes the CNN tomography architecture from Araya-Polo et al. 2018 and replaces the final layers with RNN/LSTM/GRU cells. That is genuinely new for velocity-model building, and the direct comparison of the three cell types on 12,000 synthetic models is useful. The problem is that the comparison does not isolate recurrence. As the stress-test note says, Table 1 changes three variables at once: layers 3 and 4 are removed, a recurrent cell is inserted, and the maxpool is dropped. The recurrent networks also have far fewer coefficients. So the lower MSE and higher SSIM for GRU and LSTM could come from reduced capacity, different effective depth, or the missing pooling rather than from recurrence per se. There is no matched non-recurrent baseline with the same layer count, pooling, and parameter budget. That is the main soft spot.\n\nSecond, the numbers are close: SSIM 0.8414 vs 0.8199, MSE 0.0013 vs 0.0018. There are no error bars, no repeated runs, no significance test, and no code or data release. On a test set of 2,400 models the differences might be stable, but we cannot tell from the paper. The salt-body claim in the abstract is illustrative — Figure 4 shows four examples, and Table 2 gives only global metrics. No salt-region IoU or SSIM is reported, so that specific claim is not supported quantitatively.\n\nWhat the paper does well: it frames seismic traces as sequential data, which is reasonable, and it reports parameter reduction (up to 78% fewer coefficients with the RNN) as a separate benefit. The dataset generation is described well enough to replicate, and the baseline is the authors' own earlier CNN, which is appropriate. The writing is clear for a three-page conference paper.\n\nIs the central claim refuted? No. Recurrence might genuinely help, and the authors are not hiding the architecture changes. But the evidence as reported does not establish that recurrence is the cause. A serious referee would ask for matched baselines, variance estimates, and a salt-specific metric. As an EAGE abstract this is acceptable; as a journal paper it would need that extra work.\n\nI would send it to peer review — the topic is timely and the comparison, once cleaned up, is worth publishing. But I would not cite the current version as evidence that recurrent cells improve tomography.\n\nBest","headline":"First recurrent-network tomography comparison, but the architecture changes are confounded and the gains are small; worth a close look, not a strong claim.","tokens_in":4021,"tokens_out":2490,"would_cite":false,"duration_ms":25616,"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":"GRU and LSTM cells sharpen salt-body prediction in seismic tomography","keywords":["velocity model building","seismic tomography","recurrent neural networks","LSTM","GRU","salt body prediction","deep learning","seismic gathers"],"falsifier":"Train recurrent and non-recurrent architectures with matched total parameter counts on the same synthetic gathers, then test them on real or more realistic field-style gathers with noise and irregular geometry; if GRU and LSTM no longer beat the CNN on SSIM or salt-body accuracy, the paper's causal claim is falsified.","tokens_in":3063,"feed_emoji":"🌍","tokens_out":7684,"duration_ms":73641,"temperature":0.7,"pith_summary":"This paper sets out to show that recurrent neural networks—simple RNN, LSTM, and GRU cells—can build subsurface velocity models from seismic gathers more accurately than the convolutional network used as a baseline, with the largest gains on salt bodies, the high-velocity geological formations that are hardest to image. The authors replace two fully connected reconstruction layers of a CNN with a single recurrent layer of 512 hidden states and train on 9,600 synthetic model-gather pairs. On 2,400 test models, the GRU variant gives the best quality (structural similarity 0.8414, mean squared error 0.0013) followed closely by LSTM, while recurrent networks cut total coefficients to between 21.68% and 49.08% of the CNN count. If the result holds, machine-learning tomography could produce velocity models much faster than wave-equation inversion, shrinking turnaround from weeks to days.","feed_headline":"GRU and LSTM sharpen salt-body prediction in seismic tomography","feed_subtitle":"Recurrent networks cut coefficients by up to 78% while improving velocity-model accuracy on synthetic tests.","key_machinery":"The recurrent cell is the load-bearing component: the RNN, LSTM, and GRU cells process seismic gathers sequentially, maintaining a hidden state that carries information across time steps. GRU uses update and reset gates, LSTM adds a cell state with input, forget, and output gates, and the basic RNN keeps only a single hidden state. The argument rests on seismic traces being time series, so recurrence is a natural inductive bias for this inverse problem. In the evaluated architectures, the recurrent layer replaces the CNN's two fully connected reconstruction layers, and the paper connects RNNs to residual networks, which train well on image-like tasks.","core_discovery":"The paper's central claim is that recurrence itself, rather than simply a different network size, improves velocity-model reconstruction from seismic data. Using the same convolutional feature extractor and super-resolution tail as the baseline CNN, the authors insert a recurrent cell with 512 hidden states in place of layers 3 and 4 and report test-set means: the GRU variant scores SSIM 0.8414 and MSE 0.0013, the LSTM variant scores SSIM 0.8378 and MSE 0.0014, the RNN variant scores SSIM 0.8210 and MSE 0.0019, while the non-recurrent CNN scores SSIM 0.8199 and MSE 0.0018. They further report that salt bodies, whose sharp high-velocity boundaries are the hardest part of velocity model building, are consistently resolved better by the GRU and LSTM variants, and that recurrent architectures reduce total coefficients to between 21.68% and 49.08% of the CNN count. The paper states this as a step toward fully machine-learning-based tomography from pre-stack data.","pith_inferences":["The architecture swap changes both recurrence and parameter count, so a matched-parameter ablation would test whether the GRU and LSTM gain comes from gating memory or simply from a leaner model; the paper does not report such a control.","Because the evaluation is entirely synthetic with 4 to 8 layers and fixed velocity ranges, transfer to field data is the unstated risk; fine-tuning on a small set of real gathers is a natural way to see whether the recurrent advantage survives.","The cited equivalence between RNNs and residual networks suggests the recurrent layer may be acting partly as a deep feature extractor; if so, a well-tuned residual CNN with comparable coefficients might reproduce part of the gain, which would change the interpretation of the result.","The authors' own conclusion says the reason for the higher LSTM and GRU quality is not yet fully understood, and they list multiple recurrent layers and analysis of long-term dependencies as future work, leaving the mechanism open."],"forward_implications":["On the 2,400-model test set, GRU and LSTM achieve lower MSE and higher SSIM than the non-recurrent CNN, making them the best of the tested architectures for the salt-heavy synthetic regime.","Because the RNN variant uses 21.68% of the CNN's coefficients at similar SSIM, recurrent architectures can shrink the memory footprint of machine-learning tomography without sacrificing overall quality.","Salt-body boundaries, the most difficult features, are the place where LSTM and GRU gains are clearest, pointing to recurrence as a way to improve geological edge reconstruction.","The reduction in parameters while preserving quality makes deployment on limited hardware more plausible, which matters if machine-learning tomography moves closer to field-side processing."],"supporting_citations":[{"why":"Supplies the baseline non-recurrent CNN tomography architecture and workflow that the recurrent networks modify and compare against.","marker":"Araya-Polo et al, 2018"},{"why":"Provides the LSTM cell design and sequence-modelling analysis that motivates using LSTM cells for seismic traces.","marker":"Greff et al, 2017"},{"why":"Supplies the empirical result that GRU performs comparably to LSTM, justifying the GRU variant in the comparison.","marker":"Chung et al, 2014"},{"why":"Establishes the equivalence between RNN cells and residual networks, the theoretical motivation for expecting recurrent layers to train well.","marker":"Liao and Poggio, 2016"},{"why":"Defines SSIM, the structural similarity metric used to rank reconstructed velocity-model quality.","marker":"Wang et al, 2004"},{"why":"Documents RNN time unfolding and the vanishing and exploding gradient problems that motivate LSTM and GRU gates.","marker":"Goodfellow et al, 2016"}],"fun_headline_variants":["Recurrent nets sharpen salt-body imaging in seismic tomography","GRU and LSTM beat CNN for salt-body velocity models","Recurrent tomography cuts parameters while boosting salt-body accuracy","GRU/LSTM slash parameters, sharpen salt bodies","Recurrence wins: better salt mapping with fewer network parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on replacing two CNN layers with one recurrent layer being a fair test of recurrence, and on the synthetic 12,000-model dataset standing in for real subsurface conditions; if the gains come from the changed number of parameters or from the synthetic data itself rather than from recurrence, the central claim falls.","fun_headline_variants_meta":{"raw":{"variants":["Recurrent nets sharpen salt-body imaging in seismic tomography","GRU and LSTM beat CNN for salt-body velocity models","Recurrent tomography cuts parameters while boosting salt-body accuracy","GRU/LSTM slash parameters, sharpen salt bodies","Recurrence wins: better salt mapping with fewer network parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3764,"prompt_tokens":889,"completion_tokens":2875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":2795}},"tokens_in":505,"tokens_out":2875,"duration_ms":21523,"temperature":1.0,"reasoning_tokens":2795,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:53:21.105830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train recurrent and non-recurrent architectures with matched total parameter counts on the same synthetic gathers, then test them on real or more realistic field-style gathers with noise and irregular geometry; if GRU and LSTM no longer beat the CNN on SSIM or salt-body accuracy, the paper's causal claim is falsified.","supporting_citations":[],"review_version":1}