Machine Learning in Carbon Capture, Utilization, Storage, and Transportation: A Review of Applications in Greenhouse Gas Emissions Reduction

Abstract

Carbon Capture, Utilization, and Storage (CCUS) technologies have emerged as indispensable tools in reducing greenhouse gas (GHG) emissions and combating climate change. However, the optimization and scalability of CCUS processes face significant technical and economic challenges that hinder their widespread implementation. Machine Learning (ML) offers innovative solutions by providing faster, more accurate alternatives to traditional methods across the CCUS value chain. Despite the growing body of research in this field, the applications of ML in CCUS remain fragmented, lacking a cohesive synthesis that bridges these advancements to practical implementation. This review addresses this gap by systematically evaluating ML applications across all major CCUS components—CO₂ capture, transport, storage, and utilization. We provide structured representative examples for each CCUS category and critically examine various ML techniques, optimization objectives, and methodological frameworks employed in recent studies. Additionally, we identify key parameters, practical limitations, and future opportunities for applying ML to enhance CCUS systems. Our review thus offers comprehensive insights and practical guidance to CCUS stakeholders, supporting informed decision-making and accelerating ML-driven CCUS commercialization.

1. Introduction

Carbon dioxide (CO₂) is the most significant contributor to global greenhouse gas (GHG) emissions, accounting for approximately 75% of global emissions. Over the past century, CO₂ emissions have risen dramatically, from 1.95 billion metric tons in 1900 to 36.3 billion metric tons in 2021, primarily due to industrialization and rapid societal development. This sharp rise in emissions has led to a corresponding increase in global temperature, with average temperatures rising by 0.6 °C to 1 °C since the late 19th century [1]. The International Panel on Climate Change (IPCC) estimated in 2005 that the average global temperature will increase by around 1.9 °C, and the sea level will rise by 38 m, by 2100 without significant intervention [1]. Therefore, reducing atmospheric CO₂ is crucial to mitigate climate change and its environmental and economic consequences.

As the global community confronts the pressing challenge of climate change, Carbon Capture, Utilization, and Storage (CCUS) has emerged as a critical technology in transitioning towards a low-carbon future. CCUS refers to a suite of processes that capture CO₂ emissions from industrial and power generation sources, and utilize the captured CO₂ in various applications or permanently store it underground, preventing its release into the atmosphere [2]. This technology offers a way to mitigate climate change and enables economic growth. Figure 1 provides a detailed breakdown, highlighting additional steps such as CO₂ source and transportation. The CO₂ source in the CCUS process can originate from various pathways, including atmospheric CO₂, emissions from fossil fuel combustion, human activities, and industrial processes [3]. Carbon capture involves separating CO₂ from other gases produced in industrial processes. It commonly uses post-combustion capture using amine solvents, pre-combustion capture in gasification plants, direct air capture for removing CO₂ directly from ambient air, chemical absorption, membrane separation, or chemical looping [4,5]. According to the IEA’s report [6], the projected CO₂ capture capacity for 2030 increased by 35%, while the announced storage capacity saw a 70% rise. As a result, the total expected CO₂ capture capacity for 2030 reached approximately 435 million tonnes per year, with the announced storage capacity rising to around 615 Mt per year. Storage involves sequestering CO₂ in deep geological formations, depleted oil and gas reservoirs, or saline aquifers to prevent its release into the atmosphere [7,8]. Utilization focuses on converting captured CO₂ into valuable products or services, such as enhanced oil recovery, chemical conversion, concrete curing, or the production of synthetic fuels [9]. Transportation is also a critical part of the CCUS chain. It involves safely and efficiently moving captured CO₂ from its source to storage sites or utilization facilities [10]. CCUS can enhance energy security and economic resilience by supporting the continued use of existing energy infrastructure while integrating renewable energy sources [11]. It offers a pragmatic approach to achieve significant emissions reductions while maintaining economic stability.

Despite its promise, widespread CCUS deployment faces technical and economic barriers, including high energy demands, limited material efficiency, uncertainties in storage integrity, and overall economic viability. Overcoming these barriers requires innovative solutions to optimize performance and reduce operational costs. Machine learning (ML) is a branch of AI that enables computers to learn from data and make predictions or decisions without being explicitly programmed. By analyzing vast amounts of data, ML models can identify patterns, optimize processes, and make data-driven decisions, often surpassing human capabilities in speed and accuracy. It can improve efficiency, reduce cost, provide predictive insights, and handle complex, non-linear systems that are difficult to manage with traditional methods [12,13,14,15]. The benefits of applying ML in CCUS include several aspects: (1) It improves the accuracy of predictive modeling: it allows the real-time optimization of chemical reactions, material selection, and energy efficiency in carbon capture plants during the CO₂ capture process [16,17]; (2) It accelerates material discovery and process design: ML speeds up the discovery of new absorbents, membranes, and catalysts for CO₂ capture by predicting their efficiency before experimental testing. It reduces the time and costs with trial-and-error approaches [18]; (3) It enables fast screening of CO₂ storage sites and monitoring: by analyzing geophysical, geological, and petrophysical data, ML enhances subsurface reservoir characterization and helps predict CO₂ plume migration, assess leakage risks, and improve long-term storage security [19]; (4) It reduces cost and improves economic feasibility: ML-driven economic models assess the financial viability of CCUS projects by predicting carbon credit values and cost reductions over time [20]. By employing ML, CCUS technologies can be more effectively deployed and scaled, contributing to global efforts to mitigate climate change and transition to a low-carbon economy.

Recently, there has been an ongoing interest in assessing the application of ML in the field of CCUS. In this regard, some literature reviews have been published to make reviews/assessments of the state of the art. Rahimi et al. [21], Hosseinzadeh et al. [22], and Hussin et al. [13] conducted systematic review research in the area of carbon capture, primarily focusing on absorption- and adsorption-based processes; however, their studies primarily focus on CO₂ capture and do not comprehensively address the broader CCUS value chain, such as utilization and storage. Li et al. [23] reviewed the geological CO₂ storage, covering trapping mechanisms, storage sites, global projects, and ML in sequestration. Their work focuses on storage aspects without exploring other areas. Other studies such as Yan et al. [12], Gupta and Li [24], and Al-Sakkari et al. [14] have provided a state-of-the-art review of the ML applications across the entire CCUS value chain. However, while these studies provide valuable overviews, they primarily focus on summarizing ML applications rather than offering a detailed examination of specific models, input variables, optimization objectives, and methodological frameworks. As a result, they lack in-depth discussions on the practical implementation and comparative effectiveness of different ML techniques in CCUS.

This review aims to present a detailed analysis by providing an in-depth analysis of current applications of ML in CCUS, focusing on CO₂ capture, transport, storage, and utilization, as well as the details of ML applications that have been employed. By examining recent research and developments, we aim to highlight how ML can address key challenges in CCUS and unlock its full potential in combating climate change. In comparison with most of the existing review papers, the novelty/originality of our paper is that it consists of several key aspects from critical review and gap analysis perspectives, summarized in the following items:

• This review systematically covers ML applications across all major CCUS processes—capture, transport, storage, and utilization—rather than focusing only on one segment. Unlike previous work, we distinctly provide representative examples for each category, enhancing the comparative analysis and practical applicability of our findings.
• We conduct a critical review of the current CCUS modeling, simulations, and techniques for optimization of the design and operations of CCUS systems up to 2025.
• We identify the critical parameters/objectives and key findings and evaluate different ML techniques applied during the CCUS process.
• We provide guidance and recommendations for the CCUS stakeholders, in which a detailed outlook on the opportunities for ML and related technologies for further applications in the CCUS field is presented.

The paper is structured as follows: Section 2 reviews common ML methods. Section 3 introduces the methodology, and is followed by Section 4, Section 5, Section 6 and Section 7, which examine ML’s role in CO₂ capture, storage, utilization, and transportation, respectively. Section 8 outlines the benefits and limitations of ML and explores future pathways. Section 9 concludes the review.

2. Background on ML Methods

ML has experienced rapid growth over the past decade due to advancements in computational power, the increasing availability of large datasets, and significant progress in algorithm development. ML involves using algorithms and statistical models that enable computers to perform tasks without explicit instructions, relying instead on patterns and inferences drawn from data [25]. ML’s ability to process vast amounts of information, identify patterns, and optimize processes has made it a valuable tool in diverse fields, including healthcare, finance, autonomous systems, and, more recently, environmental sciences and industrial processes such as CCUS.

2.1. Categories of ML

In general, ML methods are classified into four main categories: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning [26,27], as shown in Figure 2.

(1)

Supervised learning: uses labeled datasets to train algorithms, making predictions or decisions based on input–output pairs. This is widely used in CCUS to predict system performance, optimize process parameters, and detect anomalies in real-time data. Examples such as artificial neural networks (ANNs), support vector machines (SVMs), and random forests (RFs) are commonly used to estimate CO₂ capture efficiency and optimize storage conditions.

(2)

Unsupervised learning: employs unlabeled data to discover hidden patterns or structures. Clustering techniques, such as K-means or hierarchical clustering, can be applied in CCUS for site characterization and reservoir classification, where geological or operational data are used to group storage sites based on their CO₂ storage potential or risk profiles.

(3)

Semi-supervised learning: combines labeled and unlabeled data to improve learning accuracy when only limited labeled data is available. This can be especially useful in CCUS applications where comprehensive datasets are complex to acquire, such as in novel material discovery for CO₂ adsorption and storage.

(4)

Reinforcement learning: enables algorithms to make decisions by interacting with an environment and receiving feedback through rewards or penalties. This approach is increasingly being explored for process optimization in CCUS, where it can adapt to dynamic system changes in real time. For instance, reinforcement learning could optimize the injection of CO₂ into reservoirs by continuously adjusting parameters to maximize efficiency while minimizing risks such as leakage.

2.2. Popular ML Algorithms in CCUS

The most widely applied algorithms include artificial neural network (ANN), support vector machine (SVM), K-nearest neighbor, decision tree (DT), random forest (RF), and extreme gradient boosting (XGB), plus several deep learning (DL) methods (e.g., recurrent neural network (RNN), convolutional neural network (CNN), gated recurrent units (GRU), long short-term memory (LSTM)) specifically useful for large datasets, time series, and image analysis [29]. Here, we only provide a short description of these algorithms.

(1)

ANN: ANN is modeled after the human brain’s neural network, consisting of interconnected nodes (neurons) that can process and learn from data. ANNs are widely used in CCUS for predictive modeling, such as estimating CO₂ capture rates, optimizing transport systems, and forecasting reservoir behavior. Deep learning, a more advanced form of ANN, has also been applied to model complex relationships in large datasets, particularly in geological storage and material discovery.

(2)

SVM: SVM is a classification technique that seeks to find the hyperplane that best separates data into different categories. In CCUS, SVM has been used for material selection, process optimization, and real-time monitoring of CO₂ transport systems. For instance, SVM models have been employed to predict the effectiveness of CO₂ sorbents and membranes under varying conditions.

(3)

RF: RF is an ensemble learning technique that combines multiple decision trees to improve prediction accuracy and prevent overfitting. It has been used in CCUS for site selection, reservoir classification, and risk analysis, offering robust predictions in cases where data are incomplete or noisy.

(4)

XGBoost: XGBoost is a scalable decision-tree-based ensemble method applied in material discovery, optimizing capture processes, and predicting CO₂ behavior in storage formations. Its ability to handle missing data and provide essential features makes it a valuable tool for interpreting complex datasets.

(5)

DL Models: Deep learning architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), gated recurrent units (GRUs), and long short-term memory (LSTM) networks are well suited for time-series forecasting, image analysis, and large-scale simulations. DL models have been instrumental in monitoring CO₂ flow, assessing subsurface CO₂ behavior, and processing seismic data in storage sites.

ML offers significant advantages in CCUS by improving predictive accuracy, optimizing complex processes, and enabling data-driven decision-making. Its ability to analyze large datasets and recognize patterns allows for enhanced material discovery, operational efficiency, risk assessment, and system optimization. ML techniques, including supervised, unsupervised, and reinforcement learning, contribute to automation, cost reduction, and improved scalability across CCUS applications. By integrating ML, CCUS technologies can overcome existing challenges, accelerate deployment, and enhance their role in global decarbonization efforts.

3. Methodology

This review was conducted using a structured and systematic approach to identify, evaluate, and synthesize recent studies on the application of ML in CCUS. The review was guided by the following research questions:

• Can ML be applied in CCUS, and how can it be applied?
• What are the common ML methods applied in CCUS processes?
• What are the typical input variables, outputs, and objectives of ML models used in each CCUS domain?
• What are the current challenges and opportunities for further integration of ML into CCUS systems?

Initially, literature was searched using databases including Web of Science, Scopus, and Google Scholar, with keywords such as “machine learning”, “artificial intelligence”, “CCUS”, “carbon capture”, “CO₂ storage”, “CO₂ utilization”, and “CO₂ transportation”. We focused primarily on studies published up to 2025. Papers were selected based on their relevance, recentness, and the significance of their contributions. Representative examples were carefully chosen and structured into specific CCUS categories to avoid redundancy and enhance clarity. Through a detailed examination of relevance, technical depth, and quality, 101 representative studies were selected. Among them, 41 papers were provided with detailed information. This methodological approach ensures comprehensive coverage and effective synthesis of key ML contributions within the CCUS value chain.

4. ML in Carbon Capture

4.1. Overview

Carbon capture reduces the CO₂ released into the atmosphere from industrial processes and power generation. This technology involves separating and capturing CO₂ at its source, preventing it from contributing to the greenhouse effect that drives global warming [30]. Carbon capture can be applied to many emissions-intensive industries, including fossil fuel power plants, cement production, steel manufacturing, and chemical processing.

As described in Figure 1, the CO₂ capture approaches include post-combustion, pre-combustion, oxy-fuels, and DAC. Post-combustion CO₂ is captured from the flue gases produced after the combustion of fossil fuels. This approach is widely used in power plants and industrial processes [31]. Pre-combustion capture involves removing CO₂ before the fuel is burned. This is typically done by converting fossil fuels into syngas (a mixture of hydrogen and carbon monoxide), followed by a shift reaction to produce CO₂ and hydrogen [32]. Oxy-fuel combustion uses pure oxygen instead of air, producing a flue gas consisting primarily of CO₂ and water vapor. The water vapor can be easily condensed, leaving nearly pure CO₂ for capture and storage [33]. DAC captures CO₂ directly from ambient air, potentially removing CO₂ emissions regardless of their source. This technology uses chemical sorbents or filters to absorb CO₂ from the atmosphere [34]. While post-combustion, pre-combustion, oxyfuel combustion, and DAC represent the primary routes for CO₂ capture, various technological processes—such as adsorption, absorption, chemical looping, membranes, and advanced materials—are employed within these routes to separate and capture CO₂ effectively. These methods are tailored to each route’s requirements, including the composition and concentration of CO₂ in the gas stream [35].

ML is increasingly applied in CO₂ capture to enhance efficiency, optimize processes, and reduce costs. By analyzing vast datasets from carbon capture systems, ML models can predict optimal operational conditions, such as temperature, pressure, and chemical reactions, to maximize CO₂ absorption and minimize energy consumption. ML algorithms can also identify real-time patterns and anomalies, helping operators adjust capture processes dynamically to prevent equipment failures and improve system performance. Additionally, these models are used to design advanced materials, such as adsorbents or membranes, by predicting their CO₂ capture potential and accelerating the discovery of more efficient technologies [12,14]. This integration of ML in CO₂ capture enables more precise control, better decision-making, and improved scalability of capture solutions, driving the overall success of carbon reduction efforts.

4.2. ML Application in Carbon Capture

ML is being applied across several key areas in CO₂ capture, with adsorption and absorption being two prominent domains [12]. It also has an impact on other areas, such as membrane technologies, process optimization and control, and advanced material discovery:

(1)

Adsorption: ML significantly optimizes CO₂ adsorption processes, where solid materials (adsorbents) capture CO₂ from gas streams. By analyzing experimental and simulation data, ML models can predict the performance of different adsorbent materials, such as metal–organic frameworks (MOFs) or zeolites [36], under varying conditions. These models can identify materials with high CO₂ uptake, faster regeneration times, and selectivity, significantly speeding up the discovery and design of new adsorbents. Additionally, ML aids in the optimization of operating conditions, such as pressure and temperature, to enhance the efficiency of CO₂ capture in industrial processes.

(2)

Absorption: In absorption, CO₂ is captured using solvents like amines. Commn amines include monoethanolamine (MEA), diethanolamine (DEA), and methyl diethanolamine (MDEA) [24]. The application of ML models on MEA has been a hot topic for researchers [17,37,38]. ML models can predict the behavior of different solvents under various pressures and temperatures, optimizing the capture and regeneration cycles to reduce energy consumption. These models can also be used to develop new solvent formulations by predicting their CO₂ absorption capacity, efficiency, thermal stability, and reaction kinetics. This accelerates the discovery of more energy-efficient solvents for carbon capture. Hosseinpour et al. [22] has summarized a comprehensive review of the absorption-based post-combustion process.

(3)

Chemical looping: ML is used to predict the performance of oxygen carriers, such as metal oxides and calcium oxide, by analyzing large datasets to identify materials that offer better reactivity and stability. For example, calcium looping uses CaO as a solid sorbent to capture CO₂. It is particularly suitable for large-scale applications like power plants or industrial processes. Readers can reference Song et al. [39] for a comprehensive review of ML applications for chemical looping.

(4)

Membrane technologies: Membranes can be classified into three types based on their material composition: (a) polymeric membranes are made from polyimide, polyetherimide, polysulfone, or polyethylene oxide; (b) inorganic membranes are made from ceramics, zeolites, or metallic materials; (c) mixed-matrix membranes are hybrid membranes combining polymeric and inorganic properties [40]. ML is used to predict the performance of membranes for CO₂ separation, identifying the best materials for selectivity and permeability. This helps researchers design new membranes that enhance CO₂ capture efficiency while minimizing energy costs.

Table 1. Selected examples of ML application in CO₂ capture.

Process	Material	References	Methods	Inputs	Output(s)	Key Findings
Adsorption	Activated carbon	Subraveti et al. [41]	ANN-based NSGA-II	Adsorption time, low pressure, feed velocity	CO2 recovery and purity during PSA	The ANN with NSGA-II optimization provides a systematic trade-off analysis and lower computational cost.
	Zeolite 13X	Pai et al. [42]	DT, RF, SVM, GPR, ANN	Adsorption time, low pressure, intermediate pressure, feed velocity	Prediction of CO2 purity, recovery, energy consumption, productivity, and cyclic steady state profiles.	GPR-based models provided highly accurate forecasts with an adjusted R2 > 0.98.
	MOF	Burns et al. [18]	RF, GBDT, GA	Adsorption metrics, geometric properties, process parameters	CO2 purity, recovery, parasitic, productivity	ML can classify MOFs with an accuracy of 91%; N2 adsorption behavior is the strongest predictor.
	Biomass waste-derived porous carbon	Yuan et al. [43]	GBDT, LGB, XGBoost	Surface area, total pore volume, micropore volume, carbon/hydrogen/nitrogen/oxygen content, temperature, pressure	CO2 adsorption capacity	GBDT-based models provided highly accurate predictions with R2 > 0.9. Micropore volume was identified as the most critical factor influencing CO2 adsorption capacity.
	Coalbed	Alanazi et al. [44]	DT, RF, GBR, KNN, ANN, FN, ANFIS	Coal composition and operation conditions	CO2 adsorption capacity	RF provided the highest prediction accuracy, with R2 ≈ 0.99.
	Solid amine-functionalized porous adsorbent	Zhang et al. [45]	RF	Amine loading, amine type, pore volume, pore size, specific surface area	CO2 adsorption capacity	RF provided accurate predictions with R2 = 0.99; amine loading was the most critical factor affecting CO2 adsorption performance, followed by pore volume.
Absorption	Piperazine	Yarveicy et al. [46]	ANN, ANFIS, LSSVM, AdaBoost-CART	Temperature, CO2 partial pressure, piperazine concentration	CO2 loading capacity	AdaBoost-CART was the best-performing model with R2 = 0.9934.
	MEA	Sipocz et al. (2011) [47]	ANN	Mass% of CO2, mass flow, temperature, reboiler duty, solvent circulation rate, solvent lean load, CO2 removal efficiency	Mass flow of CO2 captured, rich load, specific duty	The ANN models accurately replicate the CO2 capture process modeled by the process simulator CO2SIM, with errors below 0.2% for the closed-loop system.
	MDEA and Piperazine	Zhan et al. [48]	ANN	MDEA and PZ concentration, liquid volumetric flow rate, high gravity factor, temperature	H2S and CO2 absorption efficiency, mass-transfer coefficient	The addition of PZ significantly enhances CO2 absorption. The ANN model has accurate prediction with a deviation within ±10%.
	MEA, MDEA, TEA, DGA, DEA	Nuchitprasittichai and Cremaschi [49]	ANN	Solvent circulation rate, concentration of primary/secondary/tertiary amine, reboiler duty, temperature, number of stages in the absorber/stripper column	CO2 capture cost	RSM and ANN can optimize CO2 capture, but ANN is better for handling complex relationships. Blended amines performed better than single amines in cost reduction and efficiency.
	K2CO3 (with various promoters)	Khan et al. [50]	GPR, SVM, RF	pH, contact time, pressure, temperature, concentration, type and concentration of promoter, CO2 loading rate and concentration	CO2 absorption rate and efficiency	GPR performed best with R2 = 0.9958; Adding promoters (e.g., Piperazine, MEA) significantly enhanced CO2 absorption.
Chemical looping	Calcium oxide	Hanak and Manovic [51]	ANN	Process variables, economic variables, scale and size variables, operating parameters	Levelized cost of electricity, specific total capital requirement	ANN-based stochastic economic assessment provides a more reliable feasibility evaluation than traditional deterministic methods.
	Oxygen carrier (Mn-, Ni-, Ca-)	Tahir et al. [52]	ANN (with LM, BR, SCG)	Gasifier temperature, gasifying agent, steam-to-biomass ratio, oxygen carriers used, equivalence ratio, biomass type, moisture content of biomass, pressure	Syngas composition, O2 transport capacity of O2 carrier, H2 production efficiency, CO2 capture efficiency	ANN-BR performed best with R2 = 0.999; Ca-based oxygen carriers were the best for hydrogen production and CO2 capture.
	Oxygen carrier (Fe2O3, Al2O₃, SiO2	Wang et al. [53]	ETR, RF, ANN	Biomass feedstock properties, oxygen carrier composition, operational conditions	H2/CO ratio, gas yield, carbon conversion efficiency, CO yield, H2 yield, CH₄ yield	ETR outperformed RF and ANN, achieving R2 > 0.88 for all target variables; BCLG with Fe-based oxygen carriers is effective for CO2 capture and hydrogen production.
Membrane	Metal–organic framework	Situ et al. [54]	SVM, KNN, DT, RF, GBDT, LGBM, XGBoost	Largest cavity diameter, pore-limiting diameter, volumetric surface area, fractional free volume, density, pore size distribution percentage	Permeability of CO2, N2, and O2 in MOF membranes, selectivity for CO2/N2 and CO2/O2 gas pairs	XGBoost was selected as the best-performing model. Gas permeability is strongly influenced by MOF structure.
	Polymeric membrane	Ahmad et al. [55]	SVR	Fractional free volume, glass transition temperature	Plasticization pressure during membrane separation	The SVR model successfully predicted plasticization pressure with R2 = 0.8837 for training data and R2 = 0.9433 for testing.
	Mixed-matrix	Guan et al. [56]	RF	MOF structure properties, polymer properties, membrane properties, operating conditions	Relative CO2 permeability, relative CO2/CH₄ selectivity, relative CO2/N2 selectivity, Young’s modulus	MOF pore size and surface area were the most influential factors affecting gas permeability and selectivity.

provides selected representative examples of various CO₂ capture materials and their applications in ML-driven carbon capture technologies. It highlights different CO₂ capture processes, listing key materials used (e.g., activated carbon, zeolite, MOFs, amine-based solvents, and membranes), the diverse ML methods (such as ANN, RF, SVM) applied, input and output variables, and the key findings of these studies. By summarizing key findings from recent research, the table serves as a comparative reference, showcasing how different ML techniques contribute to optimizing CO₂ capture efficiency, selectivity, and material performance.

4.3. Summary

ML has been applied to nearly all primary CO₂ capture materials, including adsorbents, solvents, chemical looping materials, and membranes. ML models such as ANN and RF have been used in adsorption-based capture to predict adsorption capacity and optimize sorbent materials like activated carbon, zeolites, and MOFs. Similarly, ML has aided in solvent selection and process optimization in absorption-based capture, improving the efficiency of amines and carbonate-based solutions for CO₂ removal. ML has been leveraged for chemical looping to identify high-performance oxygen carriers, such as metal oxides and calcium-based materials, optimizing reaction conditions for improved CO₂ capture efficiency. In membrane-based separation, ML has been extensively used to enhance the design of polymeric, inorganic, and mixed-matrix membranes, predicting gas permeability and selectivity with high accuracy. In conclusion, ML has become a transformative tool in advancing carbon capture technologies. As the field continues to evolve, the integration of ML will drive further innovation, making carbon capture technologies more efficient and accessible for large-scale deployment in industries critical to global decarbonization efforts.

5. ML in CO₂ Storage

5.1. Overview

CO₂ storage refers to the process of securely and permanently storing CO₂ in geological formations deep underground to prevent it from entering the atmosphere [23]. The primary goal of CO₂ storage is to provide a long-term solution for managing CO₂ emissions from industrial processes, power generation, and other sources. CO₂ storage is essential for mitigating climate change, as it provides a way to safely and permanently remove large quantities of CO₂ from the atmosphere. The primary geological storage options include:

(1)

Deep saline aquifers are ideal for CO₂ storage due to their vast capacity, deep location, and the multiple natural trapping mechanisms they offer. Found at depths greater than 800 m, these aquifers consist of porous rock formations filled with highly saline water, providing ample space to store CO₂ under high pressure, where it becomes a supercritical fluid [57]. Impermeable caprock layers prevent CO₂ from escaping, while residual trapping, solubility trapping, and mineral trapping ensure long-term containment. Additionally, saline aquifers are widely distributed globally and do not compete with freshwater resources, making them a key option for large-scale carbon sequestration.

(2)

Depleted oil reservoirs typically have well-mapped porous rock formations, which provide the necessary pore space to store CO₂ after oil production has ceased. The presence of a natural caprock, which initially trapped the oil, also prevents the injected CO₂ from escaping [58]. The existing infrastructure from oil operations, such as wells and pipelines, further reduces costs, making depleted oil reservoirs an economically and technically viable option for long-term CO₂ storage.

(3)

Unmineable coal seams present a viable option for CO₂ storage due to their ability to adsorb CO₂ onto the surface of the coal, effectively trapping the gas within the rock matrix. In these seams, the CO₂ molecules adhere to the micropores of the coal, a process known as adsorption, which securely holds the gas and prevents its migration. While these coal seams are too deep or uneconomical to mine for coal extraction, they can serve a dual purpose by enhancing methane recovery; injecting CO₂ can displace methane, allowing for coalbed methane (CBM) extraction as an additional energy resource [44]. The impermeable nature of the coal seams and the adsorption process provide a robust mechanism for long-term CO₂ storage, making them a unique and promising geological storage option in areas with abundant unmineable coal deposits. Figure 3 illustrates the different geological formations suitable for underground CO₂ storage. A CO₂ injection well is shown on the left side of the schematic, while an oil production well is depicted on the right. The diagram includes representations of saline aquifers, depleted oil and gas reservoirs, and unmineable coal seams.

(4)

Mineral storage, also known as mineral carbonation, involves the long-term sequestration of CO₂ by chemically reacting it with certain minerals to form stable carbonate compounds. This process occurs naturally in rocks rich in minerals like calcium, magnesium, or iron, which react with CO₂ to form solid carbonates such as calcite, magnesite, and siderite. Once the CO₂ is transformed into these stable mineral forms, it becomes permanently trapped, making this one of the most secure and permanent methods of carbon storage [59]. Although the process occurs naturally over geological timescales, research is focused on accelerating the reaction in suitable formations, like basalt or peridotite, to enhance storage potential. While mineral storage is slower than other methods, it offers a highly secure solution with virtually no risk of CO₂ leakage, making it ideal for regions with appropriate geological conditions.

5.2. ML Application in Storage

ML is crucial in enhancing the efficiency, safety, and scalability of CO₂ storage in geological formations. One key application is in reservoir characterization and site selection [60]. ML models can analyze vast datasets, such as seismic data, well logs, and geological surveys, to identify optimal storage sites. By processing this data quickly and accurately, ML helps predict different formations’ porosity, permeability, and CO₂ storage capacity, reducing uncertainty in selecting the best locations for long-term storage. These models can also quantify risks by predicting the behavior of faults or fractures that could lead to CO₂ leakage.

Another significant use of ML in CO₂ storage is in monitoring and optimizing CO₂ injection [61]. Once CO₂ is injected into a reservoir, ML algorithms can analyze real-time data, such as pressure, temperature, and geophysical measurements, to monitor the migration of CO₂ and detect early signs of leakage or anomalies. Additionally, ML models help optimize the rate and pressure of injection to prevent formation damage and improve the distribution of CO₂ in the subsurface. By integrating data from multiple sources—seismic surveys, surface sensors, and boreholes—ML creates predictive models that track the long-term behavior of CO₂ in storage formations, ensuring its secure containment while minimizing costs.

Table 2. Selected examples of ML application in CO₂ storage.

Storage Method	References	Methods	Inputs	Objectives	Key Findings
Saline aquifer	Song et al. [57]	ANN	Geological and reservoir parameters (depth, permeability, porosity, salinity, etc.), synthetic datasets from simulation, and real field data.	Residual trapping index, solubility trapping index, total trapping index	ANN model achieved high predictive accuracy with R2 = 0.9847 for up to 300 years post-injection.
	Amar & Ghahfarokhi [62]	GMDH, GEP, RF, DT	Temperature, pressure, viscosity of the solvent.	CO2 diffusivity coefficients in brine	The GEP model performed best with R2 = 0.9979; temperature is the most influential factor.
	Wang et al. [63]	LinearSVM, RNN, SVMr, DNN	Synthetic seismic features, pore pressure, TDS.	CO2 saturation levels at three depths	For different depths, the model performs differently; seismic features were the most important for detecting CO2 saturation.
Depleted oil & gas fields	Liu et al. [64]	BPNN-ILSO, BP-GWO, BP-SFLA	Vertical-horizontal ratio, dimensionless dip parameter group, CO2–oil mobility ratio, buoyancy effect, initial oil saturation.	CO2 storage coefficient	BPNN-ILSO model significantly outperformed traditional BP networks and other optimization-based models.
	Selveindran et al. [19]	AdaBoost, RF, ANN, ridge regression	Geological and well properties, injection parameters, production parameters.	Incremental oil production, CO2 storage capacity, revenue metrics	The stacked learner outperformed individual models; time-of-flight is a critical feature
	Hassan et al. [65]	MLR, RF, regression tree, GBM, bagging	Reservoir parameters, operational parameters.	Cumulative CO2 injected	RF provided the best predictive accuracy; operational parameters have a greater impact.
Unmineable coal seam	Alanazi et al. [44]	DT, RF, GBR, KNN, ANN, FN, ANFIS	Moisture content, ash content, volatile matter, fixed carbon content, vitrinite reflectance, pressure, temperature.	CO2 adsorption capacity in coal formations	RF, GBR, and KNN were the most reliable models, with RF achieving the best accuracy.
	Xue et al. [66]	SVM-RFE-LSTA-ELM	Time, gas injection volume, inlet pressure, CH4 concentration, mixed gas flow rate.	CH4 diffusion rates, CH4 cumulative production, and CO2 sequestration	The developed SVM-RFE-LTSA-ELM model outperformed all baseline models
Ocean	Zemskova et al. [67]	CNNs, RNNs	Sea surface temperature, flow velocity at the surface, near-surface wind velocity, surface particle pressure of CO2.	Dissolved inorganic carbon (DIC) concentration in the Southern Ocean up to 4 km depth	DIC concentration decreased in the ocean interior during the 1990s–2000s; DIC increased near the surface during the 2010s.
Mineral	Tariq et al. [59]	FFNN	Temperature, pressure, mineral type.	Advancing/receding contact angles of minerals/CO2/brine system	FFNN successfully predicted wettability behavior with R2 = 0.981; Pressure had the most significant influence on the contact angle.

presents selected examples of ML applications in CO₂ storage. Instead of listing all available studies, we have chosen representative examples showcasing various ML techniques in different storage settings. These examples highlight how various ML methods optimize CO₂ trapping mechanisms, predict storage capacity, and enhance monitoring accuracy.

5.3. Summary

The application of ML in CO₂ storage has become increasingly prevalent, offering enhanced accuracy in site selection, injection optimization, and long-term monitoring. This section illustrates how various ML techniques have been successfully applied across different storage sites, including saline aquifers, depleted oil and gas reservoirs, unmineable coal seams, ocean storage, and mineralization. These applications leverage ML’s ability to process complex geological, geophysical, and operational datasets, ultimately improving predictions of CO₂ trapping efficiency, storage capacity, and leakage risks. While different ML methods have been explored, there is no single best approach that applies universally across all storage sites. Each ML technique has strengths and limitations depending on the specific geological conditions, data availability, and prediction objectives. However, these studies collectively demonstrate that ML is a feasible and valuable tool for addressing various challenges in CO₂ storage. The continued development and integration of ML models will further enhance the reliability, efficiency, and scalability of CO₂ sequestration, supporting its role in long-term climate mitigation strategies.

6. ML in CO₂ Utilization

6.1. Overview

CO₂ utilization includes a range of processes that convert captured carbon dioxide into commercially valuable products, thereby mitigating CO₂ emissions, offsetting operational costs, and generating revenue streams. Various CO₂ utilization pathways differ in terms of technological maturity, economic feasibility, scalability, and market potential. The primary methods include Enhanced Oil Recovery (EOR), chemical conversion, mineralization, biological conversion, and industrial utilization. Among these, EOR remains the predominant and most commercially established method, responsible for the vast majority of CO₂ utilization worldwide [12]. EOR has been extensively employed in the oil and gas industry for several decades as a mature technology that enhances hydrocarbon recovery while sequestering significant volumes of CO₂. On a global scale, approximately 30–40 million metric tons of CO₂ are injected annually for EOR operations [68]. The potential for CO₂ utilization in EOR is expected to grow substantially with advancements in CO₂ capture, transportation infrastructure, and storage technologies. Projections indicate that with widespread implementation, EOR could utilize between 140 and 370 million metric tons of CO₂ per year by 2050, contingent on oil price dynamics, regulatory frameworks, and technological advancements [69]. Beyond EOR, emerging CO₂ utilization strategies—such as chemical conversion into synthetic fuels, polymer production, and mineral carbonation—are gaining attention for their potential to create sustainable value chains. However, challenges such as high energy input requirements, economic constraints, and limited scalability continue to hinder their widespread adoption. Future developments in catalysis, renewable energy integration, and policy incentives will play a crucial role in expanding the viability of these alternative utilization pathways.

6.2. ML Application Utilization

CO₂-EOR is a widely used technique that injects CO₂ into a reservoir after waterflooding to repressurize rock formations and release trapped hydrocarbons. The injected CO₂ can reduce oil viscosity, increase oil mobility, and facilitate oil flow to the surface. This technology is relatively mature and has over 50 years of history [70]. By increasing oil production while mitigating emissions, CO₂-EOR provides a financial incentive for industries to adopt CCUS, reducing the economic burden of CO₂ capture and transportation.

The applications of ML-based approaches primarily seek to reduce the computational overhead required by calling for the original high-fidelity numerical model, hence shortening the running time and eventually enabling optimization and uncertainty assessment. This application type often generates a proxy or surrogate model coupling with various ML approaches. Du et al. [28] have provided a comprehensive review of the application of ML in CO₂-EOR with 101 papers reviewed, which includes minimum miscibility pressure (MMP) estimation [71,72], water-alternating-gas (WAG) [73,74], CO₂ solubility [75,76], well location optimization [19,77], oil production/recovery factor [78,79], and CO₂ storage [80,81].

Chemical conversion is another promising and innovative approach. During the process, CO₂ is used as a raw material and transformed into valuable chemicals, fuels, or polymers through catalytic reactions. In chemical conversion, CO₂ can be converted into synthetic hydrocarbons, such as methanol, methane, or other liquid fuels (e.g., gasoline or diesel) through catalytic processes like hydrogenation [82,83]. Other valuable chemicals, including ethanol, urea (used in fertilizers), and polymers, are alternative chemical conversion products. This involves catalytic reactions in which CO₂ is combined with other compounds to form new products. CO₂ can be reduced into valuable chemicals using electricity (electrochemical) or light (photochemical) processes. For example, CO₂ can be electrochemically converted into syngas, which can then be used to produce fuels or chemicals (Figure 4). This approach enables the integration of CCUS with the chemical industry, creating a market-driven solution that reduces dependency on fossil-derived feedstocks and creates economic value by producing commercially viable products.

In addition to CO₂-EOR and chemical conversion, CO₂ can be utilized through biological processes like algae cultivation, where microalgae use CO₂ during photosynthesis to produce biomass for biofuels, animal feed, and other products, which provide a sustainable alternative to fossil-based products. For example, Onsree and Tippayawong [85] employed several ML algorithms to predict the solid yield of biomass during torrefaction and found that gradient tree boosting provides the best performance. Mineralization is another key CO₂ utilization method, in which CO₂ reacts with minerals or industrial byproducts like steel slag to form stable carbonates. This process can be applied in the construction industry to produce cement, concrete, and aggregates, providing permanent CO₂ storage and more substantial building materials. The economic advantage lies in reducing material costs, lowering carbon taxes, and creating a sustainable market for CO₂-based construction materials. Song et al. [86] applied ANN to predict the mass fraction of fly ash’s amorphous phase, where fly ash is a byproduct of coal combustion and can be used as supplementary cementitious material to replace ordinary Portland cement in concrete.

Table 3. Selected examples of different applications in ML of CO₂ utilization.

Category	References	Methods	Inputs	Output(s)	Key Findings
CO2-EOR	Huang et al. [71]	ANN	Pure CO2 (TR, xvol, MWC5+, xint), impure CO2 (yH2S, yN2, yCH4, ySO2, Fimp).	MMP of CO2-oil	ANN showed high accuracy in predicting MMP for pure and impure CO2 and is better than other statistical models.
	You et al. [87]	Gaussian SVR—PSO	FOPR*2, gas cycle*5, water cycle*5.	Hydrocarbon recovery + CO2 sequestration volume + NPV	The proposed method can optimize the WAG process with high accuracy.
	Nwachukwu et al. [88]	XGBoost	Well-to-well pairwise connectivity, injector block k and ϕ, initial injector block saturations.	Total profit, cumulative oil/gas produced, net CO2 stored.	Quick evaluation of well placement with 1000 simulation runs and R2 = 0.92.
Chemical conversion	Sedighi et al. [89]	CFNN, GRNN, MLP, RBF	Temperature, space velocity (gas hourly space velocity).	CO2 conversion efficiency, CO2 selectivity, light olefin selectivity.	GRNN was best for CO2 conversion, CFNN was best for CO selectivity.
	Zhong et al. [90]	GPR with DFT	Material dataset, surface adsorption data, catalyst properties.	Predicted CO adsorption energy, optimized catalyst candidates.	GPR and DFT calculations can rapidly screen and optimize electrocatalyst materials.
	Iwama et al. [91]	GPR, RF, Bayesian optimization	Metal oxide compositions, material descriptors, experimental conditions, oxygen vacancy formation energy.	CO2 conversion efficiency, H2 conversion efficiency, predicted optimal metal oxides.	ML models can effectively predict CO2 conversion efficiency; Cu- and Ga-based metal oxides are the most effective for CO2 reduction.
Biomass	Onsree and Tippayawong [85]	GTB, KRR, LR, DT, RF, KNN, SVM	Biomass properties, torrefaction conditions.	Yield of torrefied biomass.	GTB achieved the highest prediction accuracy; torrefaction temperature and residence time were the most critical factors.
	Phromphithak et al. [92]	RF, SVM, GB	Biomass properties, pretreatment conditions, ionic liquid solvent features, catalyst loading.	Cellulose enrichment factor (CEF), solid recovery (SR).	RF outperformed other models, achieving an R2 of 0.94 for CEF and 0.84 for SR; biomass composition and pretreatment conditions are key factors.
New materials	Song et al. [86]	ANN	CaO, Al2O₃, SiO2, Fe2O₃, MgO, total alkali content.	Mass fraction of amorphous phase, network topology parameter, fly ash reactivity classification.	ANN enables rapid and accurate fly ash screening; fly ash reactivity is strongly influenced by chemical composition.

presents selected representative examples of ML applications in CO₂ utilization rather than an exhaustive list of all studies in the field. The table highlights diverse applications, including CO₂-EOR, chemical conversion, biomass utilization, and new material development, detailing the ML methods used, input and output parameters, and key findings for each case.

6.3. Summary

In summary, CO₂ utilization offers a wide range of approaches, including EOR, chemical conversion, biomass utilization, and new material development, each contributing to reducing atmospheric CO₂ and generating valuable products. EOR remains the most widely applied method, where ML enhances process optimization, well placement, and CO₂ injection strategies, ultimately improving oil recovery and CO₂ sequestration efficiency. Chemical conversion holds great promise for transforming CO₂ into fuels, chemicals, and polymers, with ML playing a critical role in catalyst discovery, reaction optimization, and process efficiency improvements. Biomass utilization and mineralization provide additional pathways for CO₂ utilization, with ML assisting in predicting material properties, optimizing reaction conditions, and enhancing yield efficiency. Across these domains, ML has demonstrated significant advantages in accelerating material discovery, optimizing process conditions, and improving economic feasibility. Researchers can rapidly screen catalysts, optimize operating parameters, and enhance predictive modeling by leveraging ML techniques, making CO₂ utilization technologies more effective and scalable. As advancements in ML and CO₂ utilization continue to evolve, integrating data-driven approaches will be essential for maximizing efficiency and economic viability, further contributing to the global decarbonization effort.

7. ML in CO₂ Transportation

7.1. Overview

CO₂ transportation is a critical component in the success of CCUS projects, as it serves as the crucial link between CO₂ capture sites and storage or utilization locations. The efficiency and reliability of CO₂ transportation directly impact the overall feasibility and cost-effectiveness of the entire CCUS process. Given the significant distances between emission sources and storage reservoirs, robust transportation networks—whether through pipelines, ships, rail, or other means—are essential to ensuring that captured CO₂ can be safely and economically delivered to its final destination. Furthermore, effective transportation systems help mitigate the risks associated with CO₂ leakage or accidental release, which is vital for maintaining environmental integrity and public trust in CCUS initiatives [10,93]. Without reliable and well-designed transportation infrastructure, the potential benefits of CCUS in reducing greenhouse gas emissions and combating climate change would be severely compromised.

7.2. ML Application in Transportation

ML plays a pivotal role in optimizing the transportation of CO₂ in CCUS systems. By utilizing large datasets and advanced algorithms, ML can predict the most efficient routes, modes of transportation, and operational conditions to minimize costs and risks. It helps in real-time monitoring and management of pipelines, identifying potential leakages or corrosion before they become critical issues, and optimizing or forecasting CO₂ flow rates. Additionally, ML models can be used to forecast demand and optimize scheduling, ensuring that CO₂ is transported at optimal times and in the right quantities [12,14]. This integration of ML enhances the reliability and safety of CO₂ transportation, making the entire CCUS process more efficient and cost-effective.

Unlike oil, water, and natural gas, CO₂ is usually transported near the supercritical point. In this condition, CO₂ is dense like a liquid, which allows it to be transported efficiently through pipelines, yet it also flows easily like a gas. It can maximize the amount of CO₂ that can be transported while minimizing the energy required to move it over long distances. To achieve this supercritical state, CO₂ is typically compressed and maintained at pressure above 7.39 MPa and temperature above 31.1 °C [94,95]. Therefore, a small change in line temperature and pressure may lead to a significant change in the phase of CO₂, resulting in gas–liquid two-phase CO₂ flow. Impurities produced using different capture methods or the existence of water may also affect the phase behaviors of CO₂ flow.

A major issue in transportation is pipeline corrosion, usually due to CO₂ dissolved in water, which then forms carbonic acid (H₂CO₃) and reacts with the steel in pipes. Monitoring the pipeline pressure and flow rate is crucial for a cost-effective and safe CCUS system. Kim et al. [96] employed three deep learning models (MLP, LSTM, and CISM-LSTM) to predict critical alarm events by detecting the changes in inlet and outlet pressures and mass flow rates. Wang et al. [97] applied a Coriolis mass flowmeter incorporating LSSVM models to measure the mass flow rate of gas–liquid two-phase CO₂ flow in both horizontal and vertical pipelines. Su et al. [98] used deep neural networks (DNNs) to predict pressure failure in oil and gas pipelines based on construction material and dimensions.

Table 4. Selected examples of ML application in CO₂ transportation.

References	Methods	Inputs	Output(s)	Key Findings
Kim et al. [96]	MLP, LSTM, CISM-LSTM	Mass flow rate, pressure, temperature, time series data, operational conditions.	Predicted pipeline conditions (normal vs. abnormal), detection of anomalies, confidence intervals for normal operation ranges, critical alarm event detection	CISM–LSTM outperformed MLP and standard LSTM in detecting leaks and hydrates. Application to the East Sea gas field CO2 pipeline shows effectiveness.
Yang et al. [99]	LightGBM, MLPNN	Flow velocity, pH, CO2 concentration, pipe inner diameter, pipe bend angle, pipe bend radius, temperature.	Yearly maximum CO2 corrosion rate.	The LightGBM model achieved an R2 of 0.9985, while MLPNN achieved an R2 of 0.9931.
Wang et al. [97]	LSSVM	Apparent mass flow rate, observed fluid density, damping factor, differential pressure sensor readings, flow regime classification.	Corrected total CO2 mass flow rate, CO2 gas volume fraction.	LSSVM models improve CO2 mass flow rate measurement accuracy; flow pattern-based LSSVM further enhances performance.
Shao et al. [100]	BPNN, RBFNN, LSSVM	Capacitance readings from 12 electrodes inside pipeline, capacitance normalization values, electrode configurations, flow conditions.	Gas volume fraction (GVF).	RBFNN provides the best performance for GVF estimation.
Su et al. [98]	DNN	Pipeline material, defect length/depth/width, pipeline outer diameter, wall thickness.	Failure pressure of defective pipelines.	The DNN model achieves an average prediction error of 2.38%, significantly lower than traditional empirical formulas.

lists the details of selected examples of how ML is applied to different topics of transportation.

7.3. Summary

Despite its potential, ML in CO₂ transportation for CCUS comes with notable challenges. One key disadvantage is the reliance on high-quality and extensive datasets. In many cases, gathering sufficient real-time data on pipeline conditions, environmental factors, and CO₂ flow rates is difficult, especially for new or remote projects [101,102]. ML models can generate unreliable predictions without accurate data, leading to suboptimal decisions that impact efficiency and safety. Additionally, CO₂ transportation systems may encounter highly dynamic operational conditions, including fluctuating pressures, temperatures, or unexpected changes in pipeline integrity. This uncontrolled circumstance will cause supercritical phase CO₂ to switch to a two-phase flow. ML models need constant updates and retraining to adapt to these changes, which can be challenging and time-consuming. Moreover, the current application of ML in transportation primarily focuses on pipeline corrosion or leakage detection; applications in other aspects require further and continuous efforts.

8. Discussion

8.1. Evaluation of Key Findings

This review provides a comprehensive synthesis of how ML is applied across all major components of CCUS. The structured representative examples provided in Table 1, Table 2 and Table 3 demonstrate that ML significantly improves predictive accuracy, operational efficiency, and optimization of CCUS processes. Our analysis reveals that ML consistently outperforms traditional methods, especially in terms of computational efficiency, accuracy of predictions, and real-time decision-making capabilities. The key findings are:

• In CO₂ capture, ANN and RF are among the most frequently applied ML methods due to their strong predictive capabilities. These models are widely used to predict CO₂ adsorption/absorption performance, optimize process conditions, and screen advanced materials. Surface area, pore volume, solvent concentration, and temperature were consistently ranked as influential among input parameters.
• In storage, ML was primarily used for site characterization, storage capacity estimation, and leakage detection. RF and ensemble models outperformed single regressors in complex reservoir settings. Notably, synthetic datasets derived from simulations were often used to train models when field data were limited.
• In utilization, CO₂-EOR emerged as the most mature area for ML integration. Surrogate modeling with ANN, SVR, and XGBoost accelerated optimization of injection strategies and economic forecasting. In chemical conversion, ML has shown promise in catalyst discovery and reaction optimization, particularly when coupled with DFT or high-throughput screening datasets.
• In transportation, pipeline monitoring, anomaly detection, and flow rate estimation are popular topics. Time-series models such as LSTM and DNN were well suited for predicting critical pressure and flow changes in supercritical CO₂ pipelines.

8.2. Benefits and Limitations of ML in CCUS

The integration of ML in CCUS has revolutionized the field by offering advanced predictive and optimization tools that enhance process efficiency and scalability. One of the key benefits of ML in CCUS is its ability to optimize various stages of the process, from CO₂ capture to storage, transportation, and utilization. By analyzing vast datasets, ML models can predict optimal operational conditions, such as temperature and pressure, which are crucial for maximizing CO₂ adsorption and absorption efficiency. Moreover, ML accelerates the discovery of advanced materials like adsorbents, membranes, and catalysts, essential in making CO₂ capture more cost-effective and scalable. These benefits underscore the potential of ML to drive innovation and reduce the costs associated with CCUS operations.

Another significant advantage of ML lies in its capacity for real-time predictive analytics and decision support. Through the analysis of large datasets, ML provides operators with critical insights into the behavior of CCUS systems, enabling real-time adjustments that enhance safety and efficiency. This is particularly important in monitoring CO₂ storage sites and pipelines, where ML can detect potential risks, such as leakage, before they become significant issues. Additionally, ML-based surrogate processes can dramatically reduce computational overhead, as can be seen in EOR models. By generating proxy models that replace high-fidelity numerical simulations, ML minimizes the time required for optimization, making it feasible to perform complex analyses in a shorter time frame. The scalability offered by ML tools allows CCUS processes to be implemented on a larger scale, thereby contributing to global efforts to reduce greenhouse gas emissions.

However, despite these benefits, there are several limitations to the application of ML in CCUS. One of the main challenges is the heavy reliance on large, high-quality datasets. ML models require extensive data for training and validation, and in many CCUS projects, particularly those in remote or emerging regions, data scarcity can limit the model’s accuracy and utility. Another limitation of ML in CCUS is its adaptability to dynamic operational conditions. While ML models are powerful in predicting and optimizing stable systems, they need continuous retraining to account for fluctuating conditions such as changing temperatures and pressures in pipelines or unexpected failures in storage sites. This retraining can be resource-intensive and time-consuming, limiting the flexibility of ML in environments where conditions are highly variable. Furthermore, regulatory and environmental compliance adds another layer of complexity. CCUS operations must adhere to strict regulations, and the integration of ML-based solutions often requires additional validation and oversight to meet these legal and environmental standards.

In summary, while ML presents immense opportunities to advance CCUS technologies, there are also significant challenges that must be addressed. The ability of ML to optimize processes, reduce costs, and scale operations is promising, but data dependency, integration complexity, and regulatory hurdles must be overcome for these technologies to reach their full potential.

8.3. Further Pathways of ML in CCUS

The future of ML in CCUS is poised to be transformative as technological advancements and computational capabilities continue to grow. Key pathways include the following:

(1)

Hybrid physics-informed ML models: Models will be developed that integrate first-principles equations with data-driven approaches to improve interpretability and generalizability across CCUS applications.

(2)

Advanced materials discovery: ML techniques such as deep learning and reinforcement learning will play a crucial role in discovering new materials for CO₂ capture (e.g., sorbents, membranes, and catalysts). These models can accelerate the screening and design of materials with higher efficiency, lower cost, and improved stability.

(3)

Autonomous CCUS operations: The development of autonomous systems for real-time monitoring and optimization will enable more efficient and cost-effective CCUS operations, especially in CO₂ pipeline leakage detection. These systems could dynamically adjust operational parameters based on changing environmental and process conditions, minimizing human intervention and enhancing system resilience.

(4)

Integration with the Internet of Things (IoT) and big data: The integration of IoT sensors and big data analytics with ML models will provide unprecedented levels of process control and efficiency. This will enable the collection of real-time data from pipelines, storage sites, and capture facilities, feeding directly into ML algorithms to enhance predictive capabilities and reduce operational risks.

(5)

Uncertainty quantification and risk management: Future ML models will focus on uncertainty quantification and risk assessment in CCUS operations, particularly for storage and transportation. By accounting for uncertainties in geological formations and pipeline conditions, ML models can improve the reliability and safety of CO₂ storage.

9. Conclusions

This review has systematically examined the applications of machine learning (ML) in CCUS, highlighting how ML enhances efficiency, scalability, and cost-effectiveness across the CCUS value chain. We analyzed various ML techniques, including supervised, unsupervised, and reinforcement learning, applied to CO₂ capture process optimization, subsurface storage monitoring, chemical conversion and EOR utilization, and safe transportation through pipeline integrity management. By synthesizing recent studies, we have provided an informative overview of ML’s transformative potential in overcoming technical challenges and accelerating the commercialization of CCUS.

Our findings underscore that ML significantly improves process optimization, predictive modeling, and decision-making across CCUS applications. In CO₂ capture, ML accelerates material discovery, enhances solvent efficiency, and reduces computational costs in process simulations. ML-based models improve site selection, leakage detection, and long-term reservoir behavior prediction in storage. CO₂ utilization applications—including enhanced oil recovery, chemical conversion, biomass utilization, and mineralization—benefit from ML-driven catalyst optimization, reaction condition tuning, and process efficiency enhancement. In CO₂ transportation, ML enhances pipeline monitoring, corrosion prediction, flow rate estimation, and anomaly detection, ensuring the safe and efficient movement of CO₂.

Despite these advancements, several key gaps and challenges remain. Data availability and quality pose major limitations, as many CCUS processes rely on high-fidelity simulations or experimental datasets that are expensive and time-consuming to obtain. Model interpretability is another challenge, as deep learning models often function as black boxes, making it difficult to derive physical insights or gain regulatory approval. Furthermore, the dynamic nature of CCUS operations—including fluctuating CO₂ flow conditions, evolving storage site behaviors, and variable industrial processes—necessitates continuous model retraining and adaptation, which increases computational and operational complexity.

As ML technologies evolve, their role in CCUS will expand, enabling more intelligent, data-driven approaches to CO₂ reduction and climate mitigation. By addressing current limitations and focusing on interdisciplinary advancements, ML can further accelerate CCUS commercialization and integration into global decarbonization strategies.