Recent Advancements in Applying Machine Learning in Power-to-X Processes: A Literature Review

Comment 1: In the abstract, the authors mention different types of PtX systems, but it is recommended to explain the specific ML techniques discussed for clarity and to give readers a better idea of the technical focus. I recommend the authors to include one or two key insights on how ML has successfully influenced PtX processes, as this would make the abstract more engaging.

Response 1: Thank you for pointing this out. We agree with the comment. Therefore, we have updated the abstract to include specific insights on how deep reinforcement learning and data-driven optimization have successfully influenced Power-to-X (PtX) processes. Specifically, we now mention how deep reinforcement learning has improved real-time decision-making in Power-to-Gas (PtG) systems by optimizing energy dispatch and reducing operational costs, and how predictive diagnostics have enhanced system reliability by identifying early failures in key components such as proton exchange membrane fuel cells (PEMFCs). The changes can be found in the revised manuscript on the abstract page.

Comment 2: In the introduction, the authors discuss the global shift from fossil fuels to renewable energy and the challenges associated with integrating intermittent renewable energy sources. However, the central role of machine learning (ML) in these systems is missing from the outset. I recommend that the authors highlight how ML specifically addresses the technical challenges of PtX to provide the reader with a clearer understanding of the paper.

Response 2: Thank you for your insightful comment. We understand the importance of highlighting the role of ML in addressing the challenges posed by intermittent renewable energy. However, we chose to first introduce energy storage systems, such as PtX, as a core solution to intermittency, since these systems fundamentally manage the supply-demand mismatch caused by renewable variability. We believe ML plays a critical, yet supportive, role in optimising PtX operations and forecasting renewable energy outputs rather than directly solving the intermittency issue. Therefore, in the introduction, we deliberately framed ML as an optimization tool within the context of PtX systems later in the paper, where its role is fully explained. We feel this maintains the flow of the narrative and reflects the true nature of ML’s role in these systems.

Comment 3: A short paragraph at the end of the introduction outlining the structure of the paper is missing, which would improve readability and guide the reader through the subsequent sections.

Response 3: Thank you for your comment. We would like to note that the structure of the paper was already outlined in the last paragraph of the introduction. However, we agree that it could be enhanced for better clarity, as it may not have effectively captured the reader's attention. To address this, we have revised the paragraph to provide a more explicit and detailed outline of the paper’s structure. This update can be found in the last paragraph of the Introduction.

Comment 4: In the methodology review, the authors review literature published between 2008 and 2024, but there is no clear rationale provided for this time frame. I recommend that the authors explain why this particular range was chosen and how it aligns with the development of ML and PtX technologies.

Response 4: Thank you for your comment. We understand the need for clarity regarding the time frame. In fact, the rationale for selecting 2008–2024 is already addressed in the methodology section. Specifically, we mention that the application of machine learning methods in Power-to-X processes “has emerged as a young but growing research area since almost 2008”. To ensure clarity, we have refined the wording in the methodology section to make this rationale more explicit.

Comment 5: In the third section, the authors outline the main challenges associated with PtX systems, such as economic barriers, technical scalability, and environmental concerns. However, I recommend that the authors include a more in-depth explanation of the technical processes behind different PtX systems and their efficiencies.

Response 5: Thank you for this suggestion. We agree that providing more in-depth explanations of the technical processes behind different PtX systems would enhance the reader's understanding of the paper. We have added more technical detail about key PtX processes, including Power-to-Gas, Power-to-Liquid, Power-to-fuel, and Power-to-Heat systems, as well as a more comprehensive description of electrolysis methods, such as Proton Exchange Membrane (PEM), Alkaline Electrolyzers, and Solid Oxide Electrolysis (SOE). Also, a quick note on the evolving and fluid terminology of PtX is added to the context. These updates can be found in Section 3.

Comment 6: I recommend that the authors add a more detailed discussion on the environmental impacts of PtX technologies, especially in terms of resource use and carbon emissions.

Response 6: Thank you for this valuable suggestion. We agree that further detail on the environmental impacts of PtX technologies would enhance the clarity and depth of the discussion. In response to this comment, we have added more information regarding the resource consumption of PtX systems, such as the significant water demand for electrolysis and the indirect emissions from material and energy inputs. Additionally, we expanded the discussion to include the challenges of sourcing low-carbon or carbon-neutral CO₂ and the critical role of lifecycle assessments (LCAs) in determining the overall sustainability of PtX systems. These changes can be found in the last parts of the Section 3.

Comment 7: In the fourth section, the authors categorise different ML methods as supervised, unsupervised, and reinforcement learning. I recommend that the authors add more technical depth on the specific ML techniques that are most relevant to PtX systems. For example, techniques like deep reinforcement learning, neural networks, or genetic algorithms are highly relevant to optimising energy systems and could be discussed in more detail.

Response 7: Thank you for the suggestion. We agree with the comment, and we have added a more detailed explanation of specific ML techniques, such as DRL, NNs, and GAs, which are most relevant to optimising PtX systems. These techniques have been explored in terms of their underlying mechanisms and their applications in improving energy conversion, storage, and resource management within PtX systems. The updated section can be found in Section 4.

Comment 1: The author did not explain clearly why it is necessary to analyze machine learning techniques to drive current goals, rather than more advanced methods such as deep learning.

Response 1: Thank you for your comment. We understand that deep learning (DL) is a subset of machine learning (ML), and therefore, when we discuss ML techniques, it inherently includes DL methods where relevant. However, given that by "machine learning techniques" you may be referring to traditional ML methods, such as reinforcement learning, genetic algorithms, and data-driven optimization techniques, we agree that it would be helpful to provide a clearer distinction between traditional ML methods and DL. In the revised manuscript, we have added an explanation to clarify this distinction and to explain why a broader focus on ML, including DL where relevant, was chosen for this study. Traditional ML methods are highly effective in optimizing PtX processes, especially when interpretability and computational efficiency are crucial.

While DL is powerful for handling high-dimensional data (such as images or complex sequences), simpler ML methods can be more appropriate for many PtX tasks involving process control, resource management, and predictive maintenance. Introducing DL unnecessarily could increase the complexity of the solution without adding substantial value. Traditional ML techniques, including regression, clustering, and reinforcement learning, are already proving to be impactful in PtX processes. There's merit in starting with these approaches before exploring more advanced methods like DL, especially if simpler methods meet the required accuracy or performance benchmarks. We have updated the manuscript to reflect this reasoning, and this change can be found on Page 14, Section 4.

Comment 2: Please explain the meaning of "Finally, the article concludes by synthesising the findings and highlighting future research directions." This sentence is not coherent enough with the previous research content.

Response 2: Thank you for your comment. The sentence in the comment means: the article concludes by bringing together the findings and pointing out potential directions for future research. However, to improve clarity and avoid any potential misunderstanding, we have simplified and replaced the sentence. This change can be found in the Introduction section.

Comment 3: The expression in Figure 2 is prone to ambiguity for readers, and Figure 2 cannot clearly observe the corresponding numerical results for each year, requiring subtraction to obtain.

Response 3: Thank you for your comment. We respectfully disagree with the suggestion that Figure 2 is ambiguous. The purpose of this figure is to show the overall trend in publication growth related to ML applications in PtX fields, while also presenting the cumulative total in the final bar, which is clearly distinguished with a different color. We specifically chose a waterfall bar chart to highlight both the yearly progression and the total publications in a visual way. While the exact number of publications for each year is not the focus of this figure, it’s available to those who wish to calculate it, but it’s not essential for understanding the trend. We believe the current format effectively communicates the increasing interest in this field, which is the key takeaway here, rather than the precise yearly numbers.

Comment 4: Please explain the meaning of Figure 3 (a), why there are circles and ellipses, and how their sizes are defined.

Response 4: Thank you for raising this point. We appreciate the need for further clarification. Figure 3(a) is a Venn diagram designed to represent the distribution and overlap of various machine learning tasks applied in the context of PtX systems, based on keyword analysis of the identified literature. The diagram makes a visualised approximation of the extent to which different tasks (such as prediction, forecasting, optimisation, etc.) intersect, reflecting their interchangeable use in the literature or their combination in specific research studies. The choice of circles and ellipses in this diagram is not tied to any specific geometrical meaning but rather to provide flexibility in representing both overlaps and area among the tasks, which is common practice when dealing with Venn diagrams involving a higher number of categories. The size of each shape reflects the relative frequency of publications related to each task, while the overlaps show the degree to which certain tasks are used together in the literature. For example, prediction and forecasting are shown as having significant overlap due to their frequent interchangeable use, while other tasks, such as process control, exhibit more distinct boundaries but still intersect with other tasks like optimisation. We think although this representation is an approximation, it offers a meaningful way to map the scope and relationships and intersections among key ML tasks in the literature. We have updated the manuscript to provide a more detailed explanation of this figure. This revision can be found on page 8, Section 2.

Comment 5: Although Figure 8 cites corresponding references, was this figure drawn by the author themselves? Please pay attention to the corresponding copyright issue.

Response 5: Thank you for pointing this out. We agree with your observation. We have now corrected the reference in Figure 8 to properly credit the source and avoid any potential copyright issues. The updated reference can be found in the caption of Figure 8.

Comment 6: Figure 9 has the same issue as Figure 8 and requires attention to capitalization of the first letter.

Response 6: Thank you for your comment. We have now corrected the capitalization in the caption of Figure 9. Also we confirm that the permission for the referenced figure has been granted.

Comment 7: The author should indicate more in the conclusion the research direction that the author needs to focus on in future studies.

Response 7: Thank you for pointing this out. We agree with this comment. Therefore, we have revised the conclusion to clearly emphasise specific future research directions. These include the need to focus on the scalability of machine learning algorithms for industrial-scale PtX systems, development of more robust real-time data management frameworks, and integration of machine learning with comprehensive life cycle assessments to ensure sustainability.

Technical Comments:

Comment 1: The manuscript discusses data-driven models for optimising PtX processes but lacks technical details about how these models were trained, validated, and benchmarked. More information on datasets, validation methods (e.g., cross-validation), and performance metrics would strengthen the reliability of the models.

Response 1: Thank you for your valuable comment. While we understand the importance of providing detailed technical insights into the training, validation, and benchmarking of machine learning models, we respectfully disagree with the necessity of including this level of detail within the scope of the current manuscript. The primary focus of this review is to provide a holistic overview of how machine learning is applied across different PtX systems. This includes a discussion of key trends, challenges, and opportunities rather than technical case studies on the implementation of specific models. The studies referenced in the manuscript cover a wide range of applications, and in-depth technical details (such as datasets, cross-validation techniques, or performance metrics) for each individual study would be better suited for specialised technical papers rather than a comprehensive review like this one. Additionally, the core purpose of the review is to provide insights into how ML can enhance PtX processes and offer directions for future research. Readers interested in the technical specifics of model validation and benchmarking can refer to the original papers cited in the manuscript (which are numerous), where those details are thoroughly presented.

Nevertheless, we understand the value of technical detail and have included references to important studies that handle model validation comprehensively. The intention is to keep the manuscript broad and accessible, while still pointing readers toward the more technical resources.

Comment 2: The manuscript does not clearly specify the performance metrics used to assess machine learning algorithms in PtX processes. Including quantitative metrics such as accuracy, F1 score, or root mean square error (RMSE) would provide clearer comparisons and a more solid evaluation of the models' effectiveness.

Response 2: Thank you for the thoughtful comment. While we agree that including quantitative metrics such as accuracy, F1 score, or RMSE would provide a more solid evaluation of the machine learning models, the primary aim of this paper is to offer a broad overview of the emerging applications of ML in PtX processes, rather than delving deeply into specific technical advancements or benchmarks. The scope of this review is meant to address a wide audience, many of whom are researchers seeking insights into how ML is being applied to PtX systems in the context of sustainability and energy systems, rather than ML specialists focused on technical performance comparisons. We recognize that the inclusion of detailed metrics and benchmarks could be highly valuable in a future paper with a more specific focus on individual ML techniques or PtX applications. However, expanding the scope of this current paper to include such technical details would have been challenging given the tight revision timeline and the broader aims of this review. Therefore, while we agree that these metrics are important, they fall outside the scope of the present study. This paper is intended to serve as an introduction to the field and a starting point for further, more technically focused research.

Comment 3: While the manuscript acknowledges uncertainties in renewable energy inputs, it does not describe how ML models handle these uncertainties. Including techniques such as Bayesian methods or Monte Carlo simulations to quantify and manage uncertainty would improve the technical robustness of the analysis.

Response 3: Thank you for this valuable observation. We agree that Bayesian methods and Monte Carlo simulations are important techniques for handling uncertainty, and they are mentioned briefly where relevant studies utilised them. However, in this review, we have prioritised the discussion of methods that are particularly well-suited for real-time operational optimization and uncertainty management in PtX systems, such as distributionally robust optimization (DRO), deep reinforcement learning (DRL), and predictive diagnostics. As discussed in Section 5.1.5 (Handling Uncertainty with Data-Driven Robust Optimisation, p. 19), these methods have demonstrated substantial improvements in system flexibility and reliability in handling uncertainty in renewable energy inputs, energy flow optimisation, and multi-energy system integration. We believe these approaches provide more relevant insights into managing the specific complexities and real-time demands of PtX systems. While Bayesian methods and Monte Carlo simulations are indeed valuable, the review focuses on those models that align best with the practical challenges of PtX systems. Nonetheless, we agree that exploring additional uncertainty quantification techniques may offer further robustness in future research with a more specific focus on ML techniques and PtX processes.

Comment 4: The manuscript lacks detailed case studies demonstrating real-world applications of ML in PtX processes. Adding specific examples where ML was successfully integrated, such as in ammonia production or energy storage, would provide more practical insights and strengthen the manuscript's relevance to industry applications.

Response 4: Thank you for this observation. While we partly agree with this comment, it’s essential to acknowledge that both PtX and ML are emerging fields, and fully scaled industrial applications integrating ML into PtX processes are indeed still rare. Most of the existing studies tend to propose solutions or demonstrate implementations through digital twins, simulations, or controlled pilot environments, rather than in large-scale industrial settings. This is especially the case with PtX, as challenges such as scaling and system integration are still significant obstacles, making widespread industrial adoption difficult. However, there might be studies in the literature that present applications of ML in technologies related to PtX, but they may not explicitly frame them as part of the PtX umbrella. Therefore, while there are close-to-practical case studies in our review (with data from industry or real-world applications), these are not yet widespread in industrial-scale PtX systems. It would indeed be valuable to pursue another study specifically focused on real-world industrial applications of ML in PtX processes, once such cases become more prominent. This would require a comprehensive and targeted research effort to capture the evolution of these technologies as they scale up. We have noted this as a valuable direction for future research.

General Comments:

Comment 1: The manuscript would benefit from more consistent use of key terms. For instance, the terms "Power-to-X" and "PtX" are used interchangeably, which could cause confusion. Clarifying and standardising terminology would improve readability.

Response 1: Thank you for your suggestion. We agree that consistency in terminology is important for readability. We have revised the manuscript to ensure the consistent use of key terms and abbreviations. Wherever possible, we have used the full term when first introduced and then the corresponding abbreviation consistently throughout the text. Additionally, any interchangeable terms are clarified at their first introduction (or in titles) to avoid confusion.

Comment 2: Since the manuscript includes numerous technical terms, it is strongly recommended to add a nomenclature section for clarity and ease of understanding for the readers.

Response 2: Thank you for pointing this out. We agree with this suggestion. We have added a nomenclature table in the revised manuscript. The table could be found immediately after keywords, before the Introduction section.

Comment 3: The following references are highly relevant and should be cited to strengthen the manuscript:

a)       A. M. Sadeq et al., "Hydrogen energy systems: Technologies, trends, and future prospects" (2024) Justification: The manuscript discusses hydrogen as a key component of PtX technologies, making this citation necessary. This work offers an extensive review of hydrogen energy systems, technologies, and future trends, aligning directly with the hydrogen production and storage sections of the manuscript.

b)      A. M. Sadeq et al., "Development of Machine Learning Models for Studying the Premixed Turbulent Combustion of Gas-To-Liquids (GTL) Fuel Blends" (2023)

Justification: This citation is essential in the sections covering machine learning applications in PtX systems. The paper presents the use of machine learning models to optimize combustion processes, specifically for GTL fuel blends, which directly supports the manuscript’s discussion on improving energy conversion efficiency through advanced algorithms.

Response 3: Thank you for your suggestion to include the two references. Upon reviewing both, we have made the following decisions:

A. M. Sadeq et al., "Hydrogen energy systems: Technologies, trends, and future prospects" (2024): While this paper provides a comprehensive review of hydrogen energy systems, it does not specifically discuss the role of machine learning or data-driven methods in PtX technologies, which is a core focus of our manuscript. Therefore, we believe including this reference would not add significant value to our discussion on ML's role in PtX systems. For this reason, we have decided not to cite it.

A. M. Sadeq et al., "Development of Machine Learning Models for Studying the Premixed Turbulent Combustion of Gas-To-Liquids (GTL) Fuel Blends" (2023): This paper is highly relevant as it covers the application of ML in optimising combustion processes, specifically for Gas-to-Liquids (GTL) fuel blends. We agree that this strengthens our discussion on energy conversion efficiency through advanced algorithms, and we have now included this reference in the relevant sections of the manuscript. The update could be found under subsection 5.3.