Optimization of Abnormal Hydraulic Fracturing Conditions of Unconventional Natural Gas Reservoirs Based on a Surrogate Model

Round 1

Reviewer 1 Report

The review of the manuscript entitled “Optimization of Abnormal Hydraulic Fracturing Conditions of Unconventional Natural Gas Reservoirs Based on Surrogate Model”. The work is very interesting and in the scope of the journal. The work is easy to follow and well-written. The work is novel and has the useful information. The work can be published after responding to the following comments:

1.      Please add the controlling factors of abnormal conditions in the abstract.

2.      A brief detail of presented machine learning-based method should be provided in the abstract.

3.      Also it is recommended to add some applications of the findings in the abstract.

4.      It is recommended to mention the other methods of optimization such as RSM (response surface method). To do this, the next work can be used in the revision stage:
https://doi.org/10.1007/s13202-023-01679-2
https://doi.org/10.1016/j.egyr.2021.05.012

5.      Please support the last paragraph of page 3 (The K-Nearest Neighbor (KNN) classification algorithm is a relatively mature classification method in theory, and it is also a machine learning algorithm ......)

6.      Table 6, on what basis was the algorithm chosen (Table 6)?

7.      Page 11, accuracy of 77% for casing deformation is not high.

8.      It is recommended to present the optimization results with a figure, as the title of manuscript has “optimization”, thus this section should be improved.

Author Response

We are grateful for your valuable comments and suggestions. The manuscript has been revised carefully, and the revised parts were highlighted in the revised version. The detailed responses are listed below.

1. Please add the controlling factors of abnormal conditions in the abstract.

A: Thanks for your suggestion. We have added the controlling factors of abnormal conditions in the Abstract.

2. A brief detail of presented machine learning-based method should be provided in the abstract.

A: Thanks for your advice. A brief detail of machine learning-based method is added in the Abstract.

3. Also it is recommended to add some applications of the findings in the abstract.

A: Thanks for your suggestion. We have added some applications of the findings in the Abstract.

4. It is recommended to mention the other methods of optimization such as RSM (response surface method). To do this, the next work can be used in the revision stage.

A: Thanks for your comments. We study the other methods of optimization, including RSM, PSO and DE. Then to conform to the style of the research paper, we have added some relevant literature on optimization algorithms such as RSM [34], GA [35], PSO [36] and DE [37] into the Section 1 Introduction.

[34] Veza I, Spraggon M, Fattah I M R, et al. Response surface methodology (RSM) for optimizing engine performance and emissions fueled with biofuel: Review of RSM for sustainability energy transition[J]. Results in Engineering, 2023: 101213.

[35] Khatri K C A, Shah K B, Logeshwaran J, et al. Genetic algorithm based techno-economic optimization of an isolated hybrid energy system[J]. CRF, 2023, 8: 1447-1450.

[36] Du W, Ma J, Yin W. Orderly charging strategy of electric vehicle based on improved PSO algorithm[J]. Energy, 2023, 271: 127088.

[37] Song Y, Zhao G, Zhang B, et al. An enhanced distributed differential evolution algorithm for portfolio optimization problems[J]. Engineering Applications of Artificial Intelligence, 2023, 121: 106004.

5. Please support the last paragraph of page 3 (The K-Nearest Neighbor (KNN) classification algorithm is a relatively mature classification method in theory, and it is also a machine learning algorithm .....)

A: Thanks for your comments. To support the KNN classification algorithm, we have added the partial details of KNN by studying the related literatures [20-21], shown in Page 3.

[21] Huang A, Xu R, Chen Y, et al. Research on multi-label user classification of social media based on ML-KNN algorithm[J]. Technological Forecasting and Social Change, 2023, 188: 122271.

[22] Huang L, Song T, Jiang T. Linear regression combined KNN algorithm to identify latent defects for imbalance data of ICs[J]. Microelectronics Journal, 2023, 131: 105641.

6. Table 6, on what basis was the algorithm chosen (Table 6)?

A: Thanks for your comments. For the algorithms selected in Table 6, we mainly select them on the basis of algorithm accuracy. According to the prediction accuracy of different algorithms for abnormal working conditions, the algorithm is screened in the experimental results.

7. Page 11, accuracy of 77% for casing deformation is not high.

A: We would be really grateful for your comments. We have reexamined the Section 2, especially Subsection 2.3 of the original manuscript. We re-performed the prediction experiment of complex conditions, filled the sample set, and processed the data set. It can be found that the accuracy of predicting casing deformation has increased by 6% than before. We know that it is still not good enough, but as the complexity of underground conditions, it is really difficult to obtain accurate enough results. In the future, we will employ more information, such as the mechanism or microsesmic monitoring, to increase the accuracy.

8. It is recommended to present the optimization results with a figure, as the title of manuscript has optimization" thus this section should be improved.

A: Thanks for your suggestion. We have added a figure, i.e., Figure 8. Comparison of the probability before and after optimization, to present the optimization results, which illustrate the occurring probability of abnormal condition before and after optimization.

Reviewer 2 Report

The manuscript demonstrates excellent writing and is suitable for publication.

Author Response

We are grateful for your valuable comments and your approval.

Reviewer 3 Report

1. Please cite the following paper about using ML to detect abnormalities:

Zheng, D., Ozbayoglu, E., Miska, S. Z., Liu, Y., & Li, Y. (2022, April). Cement sheath fatigue failure prediction by ANN-based model. In Offshore Technology Conference (p. D011S013R009). OTC.

2. Do you characterize your model's output as a probability? Is there a reference for this?

3. What are the details of input parameters? You need to have a list of physical meanings of An.

4. What is your input dimension? Based on the confusion matrix in Fig. 7, you did not have much test data. There were only 18 cases for testing, which means that your full dataset only has 90 records. You don't have enough input for solid model verification.

5. Your writing on page 13 after table 8 is fussy and complex. Please make sure your writing is brief and clear.

6. In the appendix, Young's Modulus should have a unit.

Comments for author File: Comments.pdf

The paper needs further/major revision.

Author Response

We are grateful for your valuable comments and suggestions. The manuscript has been revised carefully, and the revised parts were highlighted in the revised version. The detailed responses are listed below.

1. Please cite the following paper about using ML to detect abnormalities.

Zheng, D., Ozbayoglu, EMiska, S.Z., Liu, Y., & Li, Y (2022, April). Cement sheath fatigue failure prediction by ANN-based model. In Offshore Technology Conference (p. D011S013R009) OTC.

A: Thanks for your advice. We have added a citation about the suggested paper in the Section 1 Introduction.

[24]Zheng D, Ozbayoglu E, Miska S Z, et al. Cement sheath fatigue failure prediction by ANN-based model[C]//Offshore Technology Conference. OTC, 2022: D011S013R009.

2. Do you characterize your model's output as a probability? ls there a reference for this?

A: Thanks for your comments. It is of reference significance to characterize the model's output as a probability. The reason is that the probability of a model's output representation is affected by the quality of its training data. If the training data is of high quality and well representative, then the output characterization of the model as probability is credible.

3. What are the details of input parameters? You need to have a list of physical meanings of An.

A: Thanks for your comments. We apologize for the incomplete statement of , which complete statement is . In addition, to avoid misunderstanding, we remove the Table A1, the physical meanings of , from Appendix into section 2, marked as Table 1.

4. What is your input dimension? Based on the confusion matrix in Fig. 7, you did not have much test data. There were only 18 cases for testing, which means that your full dataset only has 90 records You don't have enough input for solid model verification.

A: We would be really grateful for your suggestions. First, the input dimension of methods in Section 2 is 17, corresponding to the all factors in Table 1, while the input dimension of methods in Section 3 is 6, corresponding to the main controlling factors in Table 6. The related statement has been represented in Page 5 and Subsection 3.1 in Page 12. Moreover, to make the experimental results more reliable, we adjust our experimental apparatus. In the revised manuscript, the full dataset now has 538 records while the test data has 107 records. The modified experimental results are shown in Figure 7.

5. Your writing on page 13 after table 8 is fussy and complex. Please make sure your writing is brief and clear.

A: Thanks for your suggestion. We have reorganized the related statement in Page 13 carefully to make it brief and clear, which details has been presented in Page 13 and 14.

6. In the appendix, Young's Modulus should have a unit.

A: Thanks for your suggestion. We have proved the corresponding unit for Young's Modulus, and now it is shown in Table 1.

Reviewer 4 Report

Minor revision is suggested at this stage, detailed comments are listed below.

1.     Given the application of a surrogate model for predicting abnormal conditions in hydraulic fracturing, how does the model perform in terms of specificity and generalization across different geological formations and operational conditions? Are there validation datasets from diverse geological backgrounds to support the model's robustness?

2.     Machine learning models, especially in complex geological contexts, are subject to uncertainties. How do the authors quantify and mitigate the uncertainties inherent in the prediction of abnormal conditions, particularly given the stochastic nature of factors like rock properties and fracture behaviors?

3.     The manuscript mentions using field data for model training and validation. Could the authors elaborate on the diversity and volume of this data, how it was curated, and any challenges encountered in its collection and preprocessing?

4.     Some texts in the figures are not clear enough, please enhance the clarity.

5.     For the background introduction, kindly enhance the diversity including some chemical tracers development for horizontal well. The article below is suggested to be consulted as a starting point. Status and outlook of oil field chemistry-assisted analysis during the energy transition period. Energy Fuels 2022.

English fine

Author Response

We are grateful for your valuable comments and suggestions. The manuscript has been revised carefully, and the revised parts were highlighted in the revised version. The detailed responses are listed below.

1. Given the application of a surrogate model for predicting abnormal conditions in hydraulic fracturing, how does the model perform in terms of specificity and generalization across different geological formations and operational conditions? Are there validation datasets from diverse geological backgrounds to support the model's robustness?

A: Thanks for your comments. Our research refers to the deep shale gas reservoir in China, but limited to the challenge in collecting dataset, there is not enough data in terms of geological backgrounds. However, our experimental result of abnormal condition prediction in hydraulic fracturing demonstrates the superior generalization and robustness of our method.

2.Machine learning models, especially in complex geological contexts, are subject to uncertainties. How do the authors quantify and mitigate the uncertainties inherent in the prediction of abnormal conditions, particularly given the stochastic nature of factors like rock properties and fracture behaviors?

A: Thanks for your comments. In this paper, we have not particularly considered the geological contexts impacting on machine learning. However, we have indeed quantified and mitigated the impact of uncertainties on machine learning by utilizing the controlling factors analysis in section 2, including 5 methods, i.e., correlation analysis, grey correlation analysis, main control factor identification based on mutual information, feature importance ranking based on embedded method, and apriori correlation analysis. Meanwhile, in optimization process, the predicted target is not specific abnormal condition, but the probability of abnormal condition occurring mapped by machine learning. The related statement and mapping result is shown in Subsection 3.2 Table 8 and the contexts.

3. The manuscript mentions using field data for model training and validation. Could the authors elaborate on the diversity and volume of this data, how it was curated, and any challenges encountered in its collection and preprocessing?

A: Thanks for your suggestions. We have reanalyzed and researched the backgrounds and characteristics of the dataset, and added relevant descriptions as well as the challenges in preprocessing in Section 2. The target gas filed refers to a typical deep shale gas field in China. We conducted the experiments on its real-world oilfield dataset sampled from the shale gas reservoir in Sichuan Province, China. Notably, the total dimensions and records were 18 and 538, respectively. There were some outliers in dataset, and we have adopted the related algorithms to identify and remove the outliers.

4. Some texts in the figures are not clear enough, please enhance the clarity.

A: Thanks for your suggestions. We have modified the figures, including Figure 3-5, to enhance the clarity.

5. For the background introduction, kindly enhance the diversity including some chemical tracers development for horizontal well. The article below is suggested to be consulted as a starting point. Status and outlook of oil field chemistry-assisted analysis during the energy transition period. Energy Fuels 2022.

A: Thanks for your comments. We have studied some literatures further in terms of chemical tracers development for horizontal well, and then add a literature [13] to improve quality of manuscript in Section 1 Introduction.

[15] Jia B, Xian C, Tsau J S, et al. Status and outlook of oil field chemistry-assisted analysis during the energy transition period[J]. Energy & Fuels, 2022, 36(21): 12917-12945.

Round 2

Reviewer 1 Report

Thank you for revising the manuscript. however, two comments can be answered in detail:

Comment 4: Adding the results of the recommended works in the previous review to the manuscript may improve the widespread use of RSM.

comment 6: the selected algorithm can be described in detail and added to the manuscript.

Author Response

1. Adding the results of the recommended works in the previous review to the manuscript may improve the widespread use of RSM.

A: Thanks for your comments. We studied the RSM optimization algorithm, to conform to the style of the research paper, we have added some relevant literature on RSM optimization algorithm [36,37] into the Section 3 Optimization of the fracturing scheme base on surrogate model.

[36] Veza I, Spraggon M, Fattah I M R, et al. Response surface methodology (RSM) for optimizing engine performance and emissions fueled with biofuel: Review of RSM for sustainability energy transition[J]. Results in Engineering, 2023: 101213

[37] Nakkeeran, G., and L. Krishnaraj. "Prediction of cement mortar strength by replacement of hydrated lime using RSM and ANN." Asian Journal of Civil Engineering 24.5 (2023): 1401-1410.

2. The selected algorithm can be described in detail and added to the manuscript.

A: Thanks for your comments. To support the selected optimization algorithms including GO, PSO, and DE algorithms, we have added a brief detail of the selected algorithms in the Section 3.

Reviewer 3 Report

Thank you for clarification.

Author Response

Thanks for your comments.