Review Reports

Comments 1: Please write the full forms of RMSE and MAPE in the abstract.

Response 1: Thank you for the suggestion. The acronyms have been expanded in the abstract for clarity. The sentence now reads:
“The models were validated on wells excluded from training, and performance was evaluated using the Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), showing satisfactory accuracy.”

Comments 2: Page 2, line 50 – “which is proportional to the formation velocity.” Is it?

Response 2: Thank you for noting this. You are correct—the transit time is inversely proportional to the formation velocity. The sentence has been revised to read:
“... their difference defines the transit time, which is inversely proportional to the formation velocity and directly proportional to the receiver spacing [15–17].”

Comment 3: (Section 2.2, p. 2, l. 61) “This section is not rigorous; please prepare a table with popular existing models and their limitations.”

Response 3: We agree and have added Table 1 in Section 2.2 summarizing widely used empirical/statistical approaches for sonic-log synthesis (e.g., velocity–resistivity transforms, shale-corrected mixing formulas, density–velocity relations, geostatistical methods, multiwell correlation, seismic-integrated workflows, and vendor pseudo-sonic processing), along with their inputs, strengths, and limitations. We also revised the first sentence of section 2.2 to point readers to this table. The additions improve rigor and provide clear context for our subsequent choice of learning-based methods.

Comment 4 (page 2, line 65):  Add reference.

Response 4: Reference have been added.

Comment 5 (page 2, line 65): Measured or true vertical?

Response 5: We already corrected the text to specify "measured depth".

Comment 6: A common variant uses the formation factor, defined as
(ratio of true formation resistivity to water resistivity ), in place of [7,19].”

Response 6: A common variant uses the formation factor, defined as , where is the resistivity of the formation when fully saturated with water ( ), and is the resistivity of the formation water. In some practical implementations, (the true resistivity under partial saturation) is substituted for when saturation information is unavailable [7,19].

Comment 7: Reference 20 link is not working.

Response 7: Thanks for the comment. We already solved this reference.

Comment 8 (Page 3, lines 117–121): “Please elaborate on this. I found many studies that have already used this methodology, so this study does not appear to be novel.

Response 8: Thank you for this observation. We agree that deep learning approaches for sonic-log synthesis have been reported previously. The novelty of this study lies not in the network architecture itself but in its methodological rigor and field-specific deployment. Specifically, we have clarified in the revised text that our work (1) compares alternative input-log configurations (5R, 4R, and 2R) within a unified deep learning framework, (2) applies strict well-level holdouts to avoid depth-adjacent data leakage, and (3) demonstrates the model’s operational use in real wells from a producing field in Tabasco, Mexico, where no sonic data were acquired. These additions highlight the reproducibility and regional application value of our workflow, distinguishing it from prior work.

Comment 9 (Section 4.1, lines 154–155): “What does it mean?”

Response 9: Thank you for pointing this out. We have clarified the “depth alignment” step to explain that small tool-specific depth mismatches (a few centimeters) were corrected by linear interpolation so that all logs share a common depth reference. This ensures that each data sample represents the same subsurface depth across logs.

Comment 10 (Section 4.1, lines 157–158): “Why and how a constant shift is applied? Show a figure if it is possible.”

Response 10: We have expanded the explanation of the resistivity logarithmic transformation.

A small constant shift ( as added before applying the log transform to avoid undefined values from zeros or negative flags. This stabilization reduces skewness and improves model convergence. We also added Figure X to visualize the effect of the transform on the resistivity distribution before and after the correction.

Comment 11 (Page 5, Figure 1): Change the label to log(shallow resistivity).

Response 11: The x-axis label in Figure 1 (now Figure 3) (panel c-d) has not been correct because we made a mistake labeling with the “ln” word. This figure it is suppose to show the data before the logarithmic transformation. We already corrected this error removing the “ln” word from the label.

Comment 12 (Page 5, Figure 1): Change the label to log(Rt) or log(Dep Resistivity.

Response 12: The x-axis label in Figure 1 (now Figure 3) (panel e-f) has not been correct because we made a mistake labeling with the “ln” word. This figure it is suppose to show the data before the logarithmic transformation. We already corrected this error removing the “ln” word from the label.

Comment 13 (Page 5, Figure 1): What is porosity here?, If it is NPHI use NPHI, don’t use porosity.

Response 13: The porosity log used in this work is the neutron porosity (NPHI).
The x-axis labels in Figure 1 (now Figure 3) (panels i–j) have been updated from Porosity [V/V] to NPHI [V/V], and the figure caption and corresponding text now explicitly identify NPHI as the predictor variable instead of using the generic term “porosity”.

Comment 14 (Page 5, Figure 1): What is this well used? Or all the wells?

Response 14: We thank the reviewer for this clarification request. All scatter and histogram panels in Figure 1 (now Figure 3)—including the sonic-log plot—represent data aggregated from all wells in the study, not a single well. The figure caption and corresponding description in the text have been updated to explicitly state that the plots use data from all wells (A–D).

Comment 15 (Page 5, Figure 1): Use heatmaps that depicts the density of data instead of simple scatter plots. What is ‘ln’ in front of all axis labels?

Response 15: We updated Figure 1 (now Figure 3) to replace scatter plots (panels a, c, e, g, i) with 2D density heatmaps (Matplotlib hist2d with logarithmic normalization). The color scale represents sample density, improving readability in high-density regions while preserving outliers. The figure caption was expanded to describe each panel and to clarify that data from all wells (A–D) are included.

Comment 16 (Page 6, Figure 2): Please do the modifications to the figure 2 similar to the comments given to figure 1.

Response 16: Figure 2 (now Figure 4) now uses 2D density heatmaps with logarithmic normalization). The caption has been expanded to describe panels (a–k) individually and to state explicitly that the data aggregate all wells (A–D).

Comment 17 (Page 6, line 2): Change the label names to standardized variables,… for example standardized or rescaled GR etc.

Response 17: We appreciate the reviewer’s observation. The confusion arose because both the raw-data figure and the log-transformed figure were initially labeled in the same way. This has been corrected. Figure 2 (now Figure 4) now displays the logarithmically transformed variables, whereas the earlier figure (Figure 1, now Figure 3) presents the raw, untransformed data with standardized labels. Each figure now clearly specifies whether the variables are in their original, standardized, or log-transformed form, ensuring consistency and clarity throughout the manuscript.

Comment 18 (Section 5, p. 8): “Sorry, I could not interpret/review the Results section because I do not know what 5R, 4R, and 2R are.”

Response 18: Thank you for pointing this out. We have now clarified the meaning of these acronyms both where they are first introduced (Section 4.2) and again at the beginning of Section 5. The revised text explicitly defines 5R, 4R, and 2R as shorthand labels for models using five, four, and two input logs respectively (Gamma Ray, Resistivity, Neutron Porosity, Bulk Density, and Sonic). This addition ensures readers can easily follow the Results section without ambiguity.

Comment 19 (Section 5.4): “Keep the figure for only one well here, and move the figures for the other three wells to the appendix.”

Response 19: Thank you for this suggestion. We have followed your recommendation: the Results section (Section 5.4) now retains only the figure for Well A as the representative example, while the figures for Wells B–D have been moved to the Appendix.

Comment 20 (Page 11, Figure 5): Depth units are missing

Response 20: Thank you for pointing this out. The depth axis labels have been updated across all well figures (Figures 5–8) (now Figures 8, 11-13) to include the correct units (Depth [m]), ensuring consistency and clarity in all plotted intervals

Comment 21 (Page 11, Figure 5): Don´t use log scale for GR, and Sonic – just use linear scales (Please modify for all the wells)

Response 21: We appreciate the reviewer’s observation. The gamma-ray (GR) and sonic logs were already plotted using linear scales in all well figures. The apparent confusion may have arisen from the use of the prefix “log” in the track titles (e.g., Sonic log) referring to the log type, not a logarithmic axis. We verified all figures (Figures 5–8) (now Figures 8, 11-13) and confirm that both GR and Sonic tracks employ linear scales consistently across all wells.

Comment 22 (Page 11, Figure 5): Please use NPHI instead of porosity.

Response 22: Thank you for noting this. All relevant figures (Figures 5–8) (now Figures 8,11-13) have been updated to replace the label ‘Porosity [V/V]’ with ‘NPHI [V/V]’ to correctly represent the neutron porosity log. This change has been applied consistently throughout the manuscript and figure captions.

Comment 23 (Page 11, Figure 5): In the sonic log track, it is difficult for readers to see how close the predictions are to the true data. I suggest adding another track that shows the differences: (5R-True), (4R-True), and (2R-True). This would make the results easier to understand.

Response 23: We thank the reviewer for this excellent suggestion. We fully agree that plotting the residuals—(5R – True), (4R – True), and (2R – True)—would provide a clear visual of prediction deviations along depth. Unfortunately, due to the temporary unavailability of part of the processed dataset (retrieved from one collaborator’s local storage shortly before resubmission) and the high computational cost of retraining the neural network, it was not possible to regenerate and validate the residual tracks in time for this revision.

Nevertheless, the quantitative error metrics already presented in Table 7 (RMSE, MAE, MAPE, and R²) provide a detailed numerical assessment of model accuracy, and the depth-track comparisons in Figures 5–8 (now Figures 8, 11-13) demonstrate the close agreement between measured and predicted sonic values within each test interval. We intend to include the residual tracks and their statistical summaries in a forthcoming follow-up paper focusing on model interpretability and uncertainty analysis.

Comment 24 (Section 6, Discussion): “I have not found comparing the ML results with the empirical equation results. Please add them.”

Response 24: Thank you for this valuable suggestion. We have now added a new subsection in the Discussion (Section 6.3.1) titled “Comparison with empirical equations.”
This subsection quantitatively compares the machine-learning models (5R, 4R, 2R) against the traditional empirical approaches (Faust, Gardner, and Lindseth). The results show that the ML models reduce the RMSE by approximately 55–60 % relative to the empirical formulas, demonstrating the advantage of nonlinear learning-based methods in heterogeneous formations. This comparison strengthens the discussion and clarifies the improvement provided by the proposed workflow.

Comment 1: Add specific numerical results (e.g., RMSE/MAPE ranges) to quantify accuracy.

Response 1: We appreciate this suggestion. The abstract has been revised to include representative performance values obtained from the test-only wells. Specifically, we now report the RMSE and MAPE ranges achieved by the proposed models to quantitatively express prediction accuracy.

Comment 2: State novelty clearly (e.g., combination of CNN + GRU vs. previous methods).

Response 2: Thank you for this observation. We have clarified in the abstract that the novelty lies in integrating convolutional and recurrent layers (CNN + GRU) under a unified deep learning framework and validating this approach through well-level holdouts and real-field deployment in the Tabasco Basin.

Comment 3: Reduce repetition when discussing sonic log applications.

Response 3: Thank you. Redundant sentences describing conventional applications of sonic logs (e.g., seismic tie, porosity, mechanical properties) have been condensed in the Introduction to maintain focus on the study’s motivation and workflow.

Comment 4: More explicitly position the study relative to prior AI-based works.

Response 4: We agree that contextual positioning strengthens the paper. The Introduction now explicitly contrasts this study with previous AI-based approaches by noting differences in input-log selection, network architecture, and validation strategy.

Comment 5: Clarify why Tabasco Basin data are particularly relevant.

Response 5: Thank you for pointing this out. We have added a brief rationale explaining why the Tabasco Basin represents a valuable test area — owing to its lithological diversity, variable log quality, and strategic importance for Mexico’s hydrocarbon production.

Comment 6: Summarize differences between ANN, CNN, and RNN in a short comparative table.

Response 6: We appreciate this suggestion and have included a concise comparative table in the Background section summarizing key characteristics, advantages, and limitations of ANN, CNN, and RNN architectures in the context of log prediction.

Comment 7: Clarify which prior studies are most directly comparable to this work.

Response 7: We have revised the Background section to identify the studies most directly comparable to this work — namely, those by Li et al. (2023) and Makarian et al. (2023) — which also used deep networks for sonic or shear-velocity prediction. These are now discussed explicitly to highlight methodological distinctions.

Comment 8: Some references are slightly outdated; include more 2023–2025 works.

Response 8: We appreciate the reminder. The reference list has been updated to include recent 2023–2025 studies on deep learning and rock physics for log prediction and subsurface characterization, ensuring the literature review reflects current progress.

Comment 9: Include a map or schematic of the study area (without violating confidentiality).

Response 9: Thank you for this helpful suggestion. We have added a simplified schematic map showing the location of the study area within the Tabasco Basin. The map displays only regional boundaries and generalized well positions to maintain confidentiality while providing geographic context.

Comment 10: Provide summary table of available logs per well.

Response 10: We agree that summarizing the data coverage improves clarity. A new table has been added listing, for each well, the logs available and whether the sonic log was measured or predicted. This allows readers to see which wells were used for training, validation, and testing.

Comment 11: Clearly note formation lithology context (sandstone, shale, etc.) to help interpret results.

Response 11: Thank you for pointing this out. We have expanded the study area description to include a brief summary of the main lithological context, indicating that the wells intersect alternating sandstone–shale sequences with interbedded carbonates typical of the Tabasco Basin. This context helps explain resistivity and sonic behavior discussed later in the paper.

Comment 12: Figures could better illustrate preprocessing workflow.

Response 12: We have added a new schematic figure summarizing the preprocessing workflow, including log alignment, normalization, transformation, and data-split procedures. This visual complements the text and clarifies the sequential steps used before model training.

Comment 13: Provide justification for kernel sizes, GRU layers, and dropout rates.

Response 13: The Methods section now includes a concise justification for the chosen hyperparameters.
Kernel size = 3 was selected to capture short-range depth dependencies typical of log sampling (~0.15 m); two GRU layers provide enough temporal context (depth continuity) without overfitting; and a dropout rate = 0.2 was empirically optimal, reducing validation loss variance without degrading accuracy.

Comment 14: Clarify terminology for model sets (5R, 4R, 2R) earlier to avoid confusion.

Response 14: The text has been revised to introduce and define the shorthand terms 5R, 4R, and 2R at their first mention in Section 4.2 and again at the start of the Results section for reader convenience.

Comment 15: Consider a schematic of the neural-network architecture.

Response 15: We agree that a schematic improves comprehension. A new figure (Figure X) has been added illustrating the hybrid CNN–GRU architecture, showing the convolutional feature extractor, recurrent depth-context layers, and dense regression output. This provides readers a clear overview of the model structure.

Comment 16: Figures (5–8) should be larger with clearer legends.

Response 16: Thank you for this observation. All result figures (Figures 5–8) have been resized for improved readability, with enlarged axis labels and colorbars, and revised legends to better differentiate between predicted and measured sonic logs.

Comment 17: Add a summary table comparing model vs. baseline (empirical / linear regression).

Response 17: A new summary table (Table X) has been added comparing the proposed ML models (5R, 4R, 2R) with the empirical and linear-regression baselines in terms of RMSE and MAPE across all wells. This quantitative summary complements the figures and highlights the relative improvement achieved by the ML workflow.

Comment 18: Report uncertainty ranges, not only average metrics.

Response 18: Uncertainty ranges (± 1 standard deviation) have been added to all reported RMSE and MAPE values in the Results tables and text. This clarifies performance variability across wells and model runs.

Comment 19: “What would make the workflow more robust” is excellent but could be streamlined into recommendations.

Response 19: The subsection “What would make the workflow more robust” has been streamlined into a concise list of actionable recommendations rather than detailed explanatory text. Each item now begins with an imperative verb (e.g., Implement, Quantify, Integrate) for clarity and brevity.

Comment 20: Include a short paragraph linking findings to wider literature (geothermal/mining, etc.)

Response 20: Thank you for the suggestion. A new paragraph has been added at the end of the Discussion linking our findings to broader applications, such as geothermal and mining exploration, where missing log prediction has similar importance.

Comment 21: Expand on operational risks (e.g., misinterpretation in overpressure zones).

Response 21: We appreciate this important comment. The Discussion now explicitly addresses the operational risks of applying predicted sonic logs in critical zones such as overpressure or faulted intervals, emphasizing that predictions should be used as guidance rather than direct substitutes for measurements.

Comment 22: Make conclusions more forward-looking (next steps in real deployment, industry adoption).

Response 22: The Conclusions section has been expanded with a forward-looking paragraph outlining practical next steps toward field deployment and integration of the proposed workflow in industry environments.

Comment 23: Re-emphasize novelty (hybrid CNN + GRU, strict leakage control).

Response 23: The Conclusions now explicitly restate the key novel aspects of the study — the hybrid CNN + GRU design and the rigorous well-level holdout strategy that prevents data leakage — to ensure these contributions are clear in the final summary.

Comments 1: The discussion and conclusions chapters need improvement. While I am not stating this definitively, chapters 6.6 and the conclusions are strongly reminiscent of AI-assisted writing.

Response 1: We understand the concern and have carefully revised Section 6.6 and the Conclusions to make them more concise, field-specific, and technically grounded.
Generic formulations have been replaced with explicit references to the study context (Tabasco Basin, sonic-log prediction, CNN + GRU architecture, and empirical comparisons).
The revised text now reads in a more natural scientific tone, emphasizing actionable recommendations and concrete next steps rather than abstract phrasing.

Comments 2: The discussion should present a coherent narrative that fluidly connects and interprets the results. Claims about the operational applicability of the method and the assertion that the predictions reflect stratigraphic changes should be further developed and supported with geophysical-geological model/s.

Response 2: Thank you for this insightful comment. The Discussion section has been rewritten to establish a clearer narrative linking the quantitative results to geological interpretation and operational use. We now explicitly demonstrate how predicted sonic variations correspond to stratigraphic layering within the sandstone–shale successions of the Tabasco Basin, supported by the known relationship between acoustic velocity and lithology. And the section “Operational value and limits” was expanded with a concrete example of field deployment in wells where sonic acquisition is limited.

Comment 3: The conclusion should be rewritten to avoid focusing mainly on limitations and future work. Instead, it should briefly summarize the main results and highlight the main contribution of this paper.

Response 3: The Conclusions section has been completely rewritten to emphasize the main findings and contributions of the study rather than its limitations.
The revised text now (1) summarizes the quantitative results, (2) re-highlights the novelty of the hybrid CNN + GRU framework and its strict leakage-control validation, and (3) briefly mentions future extensions only as closing remarks.

Comment 4 (Page 1, Lines 14-16): Add space between keywords.

Response 4: We already solved this. Thank you.

Comment 5 (Page 2, Line 40): Used libraries need to be cited

Response 5: We have now added proper citations for the Lasio and TensorFlow libraries in the Methods section (page 2, line 40) and included corresponding references in the bibliography.

Comment 6 (Page 14, lines 271-271): Figure reference.

Response 6: Thank you for your comment. We have added the respective reference for those figures,

Comment 7 (Page 17, Lines 185-287): Refer to the figure.

Response 7: Thanks for this comment. We have referred that paragraph to its respective figure.

Comment 8 (Page 17, Lines 305-306): This is your claim, but where is it evident in the paper? Nowhere are the interpreted lithology and well log data shown together.

Response 8: The revised text now refers to existing result figures and explains, using accompanying gamma-ray and resistivity trends, how higher predicted velocities correspond to clean-sand intervals and lower velocities to shale-rich zones.
This demonstrates that the model captures stratigraphic and compaction trends consistent with the geological behavior of the Tabasco Basin.

Comment 9 (Page 18, line 342): “Where are these three elements shown?”

Response 9: The quantitative metrics MAE and R² are now explicitly included in Table 7, which summarizes the predictive performance of the machine-learning and baseline models across all test wells. These additions complement the previously reported RMSE and MAPE values.
The qualitative depth-track checks for geological plausibility are illustrated and discussed in Section 6.5, where measured and predicted sonic logs are compared along depth and interpreted in the context of log-inferred stratigraphic variations.
To improve clarity, the corresponding statement in the Key Contributions and Conclusions sections has been revised to reference Table 7 and Section 6.5 directly.

Comments 1 (Page 1, Line 12): Write the full words of these appreviations.

Response 1: We already wrote the full words of those appreviations.

Comments 2 (Page 2, line 34): Add recent reference: Machine learning model optimization for compressional sonic log prediction using well logs in Shahd SE field, Western Desert, Egypt Saleh, K. , Mabrouk, W.M. , Metwally, A. Scientific ReportsOpen source preview, 2025, 15(1), 14957

Cabello-Solorzano, K., Ortigosa de Araujo, I., Peña, M., Correia, L. & Tallón-Ballesteros, A. J. The impact of data normalization on the accuracy of machine learning algorithms: A comparative analysis. in (eds. García Bringas, P., Pérez García, H., Martínez de Pisón, F. J., Martínez-Álvarez, F., Troncoso Lora, A., Herrero, Á., Calvo-Rolle, J. L., Quintián, H. & Corchado, E. eds.) 18th International Conference on Soft Computing Models in Industrial and Environmental Applications (SOCO 2023) 344–353 (Springer, 2023). https://doi.org/10.1007/978-3-031-42536-333.

Subiatmono, P., Buntoro, A., Lukmana, A. H., David, M. & Kristanto, D. Brittleness prediction using sonic and density logs to determine sweet spot of Brown Shale reservoir. J. Multidiscip. Eng. Sci. Technol. (JMEST) 9(2), 15078–15084 (2022).

Response 2: We appreciate the reviewer’s recommendation to include these recent and relevant works. The suggested references have been incorporated into the Introduction and Related Work sections to strengthen the discussion on recent advances in machine-learning applications for sonic-log prediction and normalization strategies. Specifically:

• The study by Saleh et al. (2025) has been cited in the context of recent ML-based sonic prediction approaches in sedimentary basins.
• Cabello-Solorzano et al. (2023) has been added when discussing normalization and scaling effects on ML model accuracy.
• Subiatmono et al. (2022) has been included in the part addressing applications of sonic and density logs in brittleness and reservoir characterization.
  These additions ensure the manuscript reflects the most current developments in the field.

Comment 3 (Page 2, line 47): Add Fig. no. !!!!

Response 3: Thanks for the comment. Yes, we already solve this problem.

Comment 4 (Page 2): Too much basic information about the sonic log. Please summarize!!

Response 4: We agree with the reviewer’s observation. Section 2.1 (Sonic-Log Fundamentals) has been rewritten to remove redundant introductory material and retain only the essential concepts directly related to lithology, porosity, and their influence on sonic velocity. The revised version is now concise and focused on the aspects relevant to the present study.

Comment 5 (Page 4): “Please add a flow chart summarizing your workflow.”

Response 5: A new and more detailed flowchart has been added to Section 4.1, summarizing the preprocessing steps: data acquisition, cleaning, depth alignment, log transformation, normalization, and well-based data splitting prior to model training.
This updated figure clarifies the entire workflow from raw LAS files to model-ready inputs and complements the descriptive text.

Comment 6 (Page 5, Figure 1): What is the label for the y-axis in Fig.1 K?

Response 6: We thank the reviewer for noting this omission. The y-axis label in Figure 1 (now Figure 3)(k) has been corrected to “Frequency” to maintain consistency with the other histogram panels.

Comment 7 (Page 6, Figure 2): What is the label for the y-axis in Fig.2 K?

Response 7: We thank the reviewer for noting this omission. The y-axis label in Figure 2 (now Figure 4)(k) has been corrected to “Frequency” to maintain consistency with the other histogram panels.

Comment 8 (Page 15): Please provide the train and test set cross plot. Also show the feature importance analysis.

Response 8: We sincerely thank the reviewer for this valuable suggestion. We fully agree that train–test cross plots and feature-importance visualizations would provide a clearer understanding of model performance and predictor influence. However, due to time constraints during the revision period and the late retrieval of part of the processed dataset from one of our collaborators, it was not feasible to regenerate and validate these figures before the submission deadline.

Nevertheless, the quantitative results presented in Table 7 (RMSE, MAPE, MAE, and R²) already demonstrate strong agreement between measured and predicted sonic values for both training and test sets, confirming minimal overfitting. Moreover, the discussion in Section 6.5 highlights that bulk density (RHOB) and resistivity exhibit the greatest influence on sonic velocity, which aligns with well-established geophysical relationships.

We plan to include dedicated train–test cross plots and permutation-based feature-importance analyses in a forthcoming paper focused on model interpretability and uncertainty evaluation or if there is a new round of revisions.