Advanced Elastic and Reservoir Properties Prediction through Generative Adversarial Network

Round 1

Reviewer 1 Report

This manuscript describes a study focused on predicting subsurface properties in petroleum geosciences. The authors propose a new method using a deep generative adversarial network to predict these properties in a completely data-driven manner. Overall, interesting work on using deep learning methods in geoscience. Considering the first example is based on a synthetic data, it is strongly suggested that the authors share a demo that describing the details (input data, code, output, etc.) so that readers can follow step by step.

 

Figures:

-       Quality is really bad for some figures. Please consider improving resolution.

-       Please explain each figure in detail in captions. Readers may only glance through the figures and captions without reading the context.

-       Figure 6 caption: “decreasing” is not spelled correctly.

-       Figure 6 & 7: why is the accuracy changing zigzagged with increasing epoch? Please explain.

-       Figure 23: even with the tolerance of 15% error, almost half of the prediction is “wrong”. If this is the case, how can authors trust the prediction of this model in field dataset?

-       Figure 24 & Table 3: “Prediction accuracy” seems to be not a good metric to evaluation the prediction performance.

o   Table 3 shows the prediction accuracy for density is 99.6 and 99.9. However, the predicted density plots in Figure 24 look very different than the ground truth density. How can the authors demonstrate 99.6% or 99.9% represent good and realiable prediction? 

o   The porosity prediction accuracy increased from 48.0% to 74.7%. However, in the figure, the MS pix2pix still fails to capture the reservoir rock (>20%). If this the case, how can the authors use the results for further exploration?

-       There is no big difference between original Pix2Pix (top row) and MS-Pix2Pix.

 

Experiments:

-       Line 186: How are the 1000 inlines selected? Are they selected randomly?

-       Line 189: Why do you use 90% training and 10% validation? Will this lead to overfitting problem? Have the authors try less training dataset such as 70% or 80%? Will the results be different?

-       Line 217: Is “qc” quality check? Please spell it out.

-       Line 332: How to solve the “over prediction” issue?

 

Conclusions:

-       Conclusions are too weak. “although the results seem good, they can still be improved further”. How can they improved? Which aspects should be focused in the next steps?

 

Comments for author File: Comments.pdf

Author Response

Dear Reviewer,

We would to thank for the constructive comments and feedback. Please see the attachment for the detail answers.

Thank you.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript "Advanced Elastic and Reservoir Properties Prediction through Generative Adversarial Network" presents a deep learning approach to estimates quantitative model parameters of the subsurface from the observed seismic data.  In particular, the authors consider a pix2pix model as base for introduced a modified version. The subject addressed is of interest in several areas, especially in exploration geophysics. Although the manuscript addresses an important issue, the authors must present the work accurately. The authors superficially explain various vital points. Several contents must be appropriately discussed, with the due background. Also, the presentation is sometime confusing, with missing or poorly motivated definitions and a (unnecessarily) repeated figure. The findings need to be better presented. Please see some more comments below.

 

(1) In lines 35-36, the authors state that: "Another widely used approach is Full Waveform Inversion (FWI) [2] which calculates the misfit between the observed and calculated data." This statement is partially true, as several objective functions associated with the FWI can be calculated differently. For example, in reference [https://0-doi-org.brum.beds.ac.uk/10.1093/gji/ggw485] the comparison between data is computed through the correlation between observed and calculated data. A second example is given in reference [https://0-doi-org.brum.beds.ac.uk/10.1103/PhysRevE.106.034113] where the graphs of the observed data are compared with the graphs of the calculated data than the data itself. I suggest the authors replace the phrase "which calculates the misfit between the observed and calculated data" with something similar to "which matches calculated data with observed data by considering amplitude and travel-times information." in line 36 and also mention these two references (or others that the author finds attractive).

 

(2) The beginning of section 3. Pix2Pix and Multiscale PatchGAN of Pix2Pix (lines 115-127) needs further development. The presentation of this one is confusing, with missing or poorly motivated definitions. Besides, the dimensions of variables and operators must be described and equations numbered. Also a more detailed and accurate explanation of these equations must be included. The authors must describe mathematical quantities, such as E and lcGAN.  

 

(3) The artifacts next to the salt top in the synthetic numerical experiments are worrying. Although the authors suggested a possible cause for the problem, the authors should perform a new synthetic experiment using a simpler model. After all, synthetic numerical tests often represent best-case or, in some cases, ideal circumstances. In this regard, excellent results are expected. 

 

(4) Figures 1-5 should be described in more detail and connected with the variables described in the first two paragraphs of section 3. For example, Fig. 4 is not even referenced in the manuscript body. Figures should help the reader understand the methodology. If it is not essential, the authors must remove it from the manuscript.

 

(5) The first paragraph of Section 4 (lines 179-180) must be clarified. Could the authors briefly describe the contents of the SEG Seam salt dataset database and attached models?

 

(6) In Fig. 05, what does the WS mean? IL must mean inline? Please describe in detail in the manuscript the diagram depicted in Fig. 5. Moreover, it is worth noting that the abbreviations must be inserted in parentheses right after the written-out means when defined for the first time. This one should be made at its first appearance in the main text. I suggest that the authors review the abbreviations entered in the manuscript. For example, please replace "inline" with "inline (IC)". Please double-check the abbreviations.

 

(7) In the caption of Fig. 5, what does "Normalization from 0 to 1" mean? It is understandable that when a quantity is normalized, the values of these quantities are between 0 and 1, but not necessarily in the operator used to perform the normalization. The authors should also clarify how the normalization process was performed. Is it normalizing the maximum value of each quantity? By the sum of all discrete parts?

 

(8) The upper part (outside the rectangle with dashed lines) of Fig. 9 is the same as Fig. 5. Please consider condensing both figures or representing the top part of Fig. 9 in a single block.

 

(9) The body text of the manuscript between lines 282-334 are completely disconnected. By the way, should they be connected? Or are the contents of lines 326-334 part of table 2?

 

(10) Discussions on the results of the present manuscript should begin on lines 282-286. However, the authors do not discuss and mention the figures without any motivation. Accuracy metrics should also be explained. For example, what is the best result when the SSIM is employed as an accuracy metric? Many readers will know that the SSIM ideal value is 1. However, many other readers may not remember. Authors must mathematically define cross-correlation and SSIM index. In the latter's case, I suggest the authors mention the following work [https://0-doi-org.brum.beds.ac.uk/10.1109/TIP.2003.819861]. 

 

Overall, the authors do not present the results accurately. Figures should be better explained and connected to the text. The present manuscript should address the above points for the work to be interesting for several readers, in addition to a more detailed and accurate explanation of the methodology, especially around equations and figures.

Author Response

Dear Reviewer,

We would like to thank for the constructive comments and feedback. Please see the attachment for the detail answers.

Thank you.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors addressed all questions raised in the first round of review.