Accelerating Numerical Simulations of CO₂ Geological Storage in Deep Saline Aquifers via Machine-Learning-Driven Grid Block Classification

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

 

This manuscript investigates the use of machine learning (ML) to accelerate numerical simulations of CO₂ geological storage in saline aquifers. The proposed hybrid model, combining machine learning and iterative methods, effectively reduces computational costs, making it particularly applicable for large-scale CO₂ sequestration projects. This approach, which integrates ML with traditional numerical simulation techniques, presents a degree of innovation and offers a more efficient and scalable solution for modeling CO₂ storage in saline aquifers.

My suggestion includes：

(1) The introduction could further elaborate on the shortcomings of existing literature regarding the use of machine learning for accelerating reservoir simulations. It should clearly explain how this study overcomes those issues, offering a stronger comparison.

(2) The methodology section introduces the machine learning model and IQR classification method, but it is recommended to provide more details on the machine learning model's training process. How was overfitting prevented? Was cross-validation used to enhance the model's generalization capability? Additionally, the limitations of the model should be discussed. How scalable is the approach when applied to larger datasets or more complex reservoir structures?

(3) In the discussion section, it is suggested to strengthen the quantitative comparison with traditional reservoir simulation methods, particularly in terms of improvements in computational efficiency. Can the paper specify how much time or resources the new method saves? Moreover, it would be valuable to discuss the broader significance of this research. How does this method improve the feasibility of CO₂ sequestration projects in practical applications?

(4) Please pay attention to the language style and accuracy of word choice. The paper should use clear and concise language, avoiding long sentences and obscure vocabulary. Some sentences, especially longer ones, are not very fluent and contain grammatical issues. For instance, the last two sentences of the abstract are too long and could benefit from being split into shorter sentences to improve readability.

(5) Please check the punctuation throughout the text and ensure that the reference format complies with the journal's guidelines. Some references are missing complete information, so it is recommended to carefully review the reference formatting.

Author Response

Please see the attachment. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents a methodology to improve the computational efficiency of CO₂ storage simulations by classifying grid blocks into fast- and slow-varying regions using machine learning (ML). The authors’ focus on grid-based classification offers a practical pathway for CO₂ injection simulations that could scale with increased storage demand.

The authors employ a machine learning-driven framework to classify grid blocks based on dynamic characteristics, particularly pressure and saturation variations. 

The method involves two main stages: (1) a classification model using an interquartile range (IQR)-based statistical approach to detect outliers in prediction errors and designate grid blocks as fast-varying, and (2) an adaptive application of ML proxies to handle slow-varying blocks. The use of feedforward neural networks trained with temporal-spatial data enables the authors to capture reservoir behavior and streamline the computational load effectively. While the neural network model used is relatively simple, its integration within this selective approach enhances both speed and generalizability.

The results indicate that the proposed classification-based methodology provides notable improvements in computational efficiency without significant accuracy loss. By isolating fast-varying blocks and assigning them to iterative solvers, the method ensures that the high-variance regions receive sufficient computational resources, while low-variance regions benefit from the speed of ML predictions. This selective approach demonstrates potential for optimizing the computational demands of large-scale CO₂ storage projects, a critical factor for scalability in climate mitigation efforts.

Comparing these results with literature reveals the significance of proxy models in simulating complex reservoir conditions. The authors' approach aligns with findings from Mohaghegh's categorization of proxy models and furthers it by integrating machine learning classifications directly into simulation workflows, marking a potential improvement in reservoir engineering practices.

Expanding the literature review to highlight environmental and climatic aspects could strengthen the paper’s relevance to carbon capture and storage (CCS) strategies as part of global climate action. The environmental impact of CO₂ injection, such as potential effects on groundwater quality and induced seismicity, requires careful modeling and monitoring. Research addressing these factors, especially with machine learning applications, could provide broader insight into the environmental safety of CO₂ storage. For instance, studies have examined how CO₂ plume behavior influences brine displacement and pressure buildup, which are directly related to geomechanical stability and potential environmental risks. Including references to these findings would deepen the study's connection to environmental sustainability.

Additionally, discussing the relevance of the 2015 Paris Agreement could contextualize the importance of enhancing CO₂ sequestration methods. Climate models indicate that achieving net-zero emissions requires substantial sequestration capacity, which could benefit from the computationally efficient solutions proposed by the authors.

The paper provides a robust methodology for accelerating CO₂ storage simulations, with promising implications for efficient and scalable CCS implementations. Further contextualization within environmental and climate frameworks would strengthen its impact, addressing both computational and ecological challenges in the pursuit of sustainable carbon sequestration practices.

Author Response

Please see the attachment. 

Author Response File: Author Response.pdf