REVIEW 3 major objections 5 minor 2 references
Deep Recurrent Architectures for Seismic Tomography
T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read GRU and LSTM cells sharpen salt-body prediction in seismic tomography
desk verdict First recurrent-network tomography comparison, but the architecture changes are confounded and the gains are small; worth a close look, not a strong claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (3)
- [Deep Recurrent Neural Network Architectures for Seismic Tomography, Table 1] 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.
- [Performance Evaluation, Table 2 and Figure 4] 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.
- [Performance Evaluation, Table 2] 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.
minor comments (5)
- [Introduction] 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'.
- [Performance Evaluation] 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.
- [References] 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.
- [Figure 4] 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.
- [Table 1] 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.
Circularity Check
No significant circularity: empirical comparison with external metrics; baseline self-citation is not load-bearing.
full rationale
The paper is an empirical benchmark, not a derivation. The claimed advantage of GRU/LSTM over CNN is supported by external quality metrics (SSIM and MSE) computed on a held-out test set of 2,400 synthetic velocity models, with network weights learned by gradient descent from synthetic gathers. No parameter is fitted to the test data and then renamed a prediction; no uniqueness theorem or prior result is invoked to force the architecture choice; and the recurrent cells are standard published formulations (LSTM and GRU) whose equations are stated in the paper. The only self-citation is the use of the authors' earlier CNN (Araya-Polo et al., 2018) as the non-recurrent baseline, which is a natural experimental control rather than a load-bearing derivation step. Concerns that the ablation changes depth and pooling alongside recurrence, or that the salt-body claim lacks a salt-region metric, are methodological and correctness issues, not circularity. Accordingly, no step in the paper's reasoning reduces to its own input.
Assumptions & free parameters
free parameters (3)
- Recurrent hidden state count =
512
- Number of training epochs =
50
- Mini-batch size =
100
assumptions (3)
- domain assumption The synthetic velocity models and generated seismic gathers are representative of real subsurface data for VMB evaluation
- domain assumption Feeding seismic traces sequentially into recurrent cells is a suitable inductive bias for velocity model building
- standard math Standard deep learning training (RMSprop, MSE loss) converges to solutions that support the comparison
Cite this review
Pith. "Pith review of Deep Recurrent Architectures for Seismic Tomography." pith.science (2026). https://pith.science/paper/TEKQBQXS
@misc{pith2026190807824,
author = {Pith},
title = {Pith review of: Deep Recurrent Architectures for Seismic Tomography},
year = {2026},
howpublished = {\url{https://pith.science/paper/TEKQBQXS}},
note = {Machine review of arXiv:1908.07824}
}
read the original abstract
This paper introduces novel deep recurrent neural network architectures for Velocity Model Building (VMB), which is beyond what Araya-Polo et al 2018 pioneered with the Machine Learning-based seismic tomography built with convolutional non-recurrent neural network. Our investigation includes the utilization of basic recurrent neural network (RNN) cells, as well as Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) cells. Performance evaluation reveals that salt bodies are consistently predicted more accurately by GRU and LSTM-based architectures, as compared to non-recurrent architectures. The results take us a step closer to the final goal of a reliable fully Machine Learning-based tomography from pre-stack data, which when achieved will reduce the VMB turnaround from weeks to days.
Figures
Reference graph
Works this paper leans on
-
[1]
M. Araya-Polo, J. Jennings, A. Adler and T. Dahlke [2018]. Deep Learning Tomography, The Leading Edge, pp. 58-66, 37(1). Presented at the 81st EAGE Conference & Exhibition, 3-6 June 2019, London, UK K. Greff, R. K. Srivastava, J. Koutník, B. R. Steunebrink and J. Schmidhuber [2017]. LSTM: A Search Space Odyssey, in IEEE Trans. on Neural Networks and Learn...
work page 2018
-
[10]
Figure 4 VMB results of 4 models with salt bodies: 1st row ground truth (GT); 2nd row CNN (non- recurrent); 3rd row RNN; 4th row LSTM; and 5th row GRU results. The dimensions of each image are 70 x 90 pixels, representing depth (vertical axis) and lateral offset (horizontal axis). C. Zhang, C. Frogner, M. Araya-Polo and D. Hohl [2014]. Machine-learning Ba...
arXiv 2014
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.