Applications of artificial intelligence in carbon management and carbon footprint research: a bibliometric review

Thematic distribution and research ecology in AI-enabled carbon management

This dataset is derived from a sample of literature in the field of “AI + Carbon” retrieved from the Web of Science. It enables a quantitative analysis of thematic distribution through article counts and proportional representation [Table 2]. Such an approach offers a structured overview of the research ecology and evolutionary dynamics within AI-enabled carbon management. From the perspective of technology-scenario interrelation, research focusing on “AI + carbon emissions” constitutes an absolute core theme. It accounts for 15.38% of the 385 articles examined. This prominence reflects a dual underlying research logic. On the one hand, carbon emissions represent the fundamental “baseline problem” in carbon management. The integration of AI has fundamentally transformed conventional research paradigms. It has shifted from estimation based on statistical models toward ML-driven dynamic prediction. Additionally, it enables optimization of emission reduction pathways through deep reinforcement learning. This technological evolution has catalyzed increased research density and cumulative intellectual advances. On the other hand, the significant proportion of carbon emission-related studies also mirrors the growing imperative. Under the “dual carbon” goals, there is increased demand for precise emission regulation in domains such as industrial decarbonization and energy transition^[35]. This urgency channels AI technological development toward these high-impact scenarios. It accelerates both thematic convergence and methodological penetration. Consequently, a virtuous cycle has emerged. Societal and regulatory demands stimulate technological innovation, which in turn fuels further research and application.

The second tier consists of the carbon cycle, carbon source, and carbon footprint. There are 191 carbon cycle articles accounting for 7.63%, 170 carbon source articles accounting for 6.79%, and 154 carbon footprint articles accounting for 6.15%. Together, these build a “digital twin” foundation for carbon management. Carbon cycle research focuses on the ability of the AI to simulate complex systems. It couples with satellite remote sensing and DL, to analyze the spatiotemporal dynamics of ecosystem carbon sinks/sources. Essentially, it aims to use technology to break through the cognitive boundaries of natural carbon processes. Carbon source recognition relies on the pattern recognition advantages of AI. It uses convolutional neural networks to screen high-emission industrial sources, solving the problem of “missed and misjudged” in traditional monitoring. It also provides precise targeting for emission-reduction strategies. After introducing AI into carbon footprint accounting, the upgrade from static inventory to dynamic tracking is achieved, and the three work together to build a digital foundation of “monitoring diagnosis traceability”, supporting the transformation of carbon management from empirical decision-making to data-driven.

Driven by policy scenarios, research on “AI + carbon neutrality” and carbon peak/carbon trading has become increasingly popular. There are 126 papers on carbon neutrality, accounting for 5.03%, and 87 papers on carbon peak/carbon trading, accounting for 3.48%, reflecting a research shift toward “technology-responsive policies”. The theme of carbon neutrality covers the entire chain of “emission reduction, removal, and offsetting”, and AI’s role has evolved from a single-point tool to a system-level solution. Carbon Peak focuses on “peak prediction, path simulation, and effect evaluation”, with AI providing quantitative support for policy formulation through multi-scenario simulation. Carbon trading extends to market mechanism innovation, connecting the policy implementation chain of “target disassembly, path design, and market synergy”, reflecting the transition of AI from technology application to governance tools.

However, carbon sinks, Carbon Capture, Utilization, and Storage (CCUS), and other niche areas such as carbon capture, utilization, and storage, as well as carbon offset, remain underexplored. Carbon sinks account for 0.88% with 22 articles, CCUS for 0.80% with 20 articles, and carbon offset for 0.48% with 12 articles. Despite the small number of publications, there is a hidden “second curve” of technological breakthroughs. In carbon sink research, AI integrated with remote sensing and the Internet of Things attempts to address the “wide area, dynamic, and accurate” challenges of natural carbon sink monitoring. In the CCUS scenario, AI intervention in carbon capture process optimization and storage site simulation can fill efficiency gaps in engineering technology. The carbon offset mechanism leverages AI for intelligent project certification and precise quantification of carbon credits. These “technology niche scenarios” may become a key pivot to complete the “dual carbon” closed-loop in the future.

From a holistic perspective, the data reveal a three-layer structure of “dense sedimentation at the foundation layer, accelerated response at the policy layer, and forward-looking layout at the technical layer”, forming a logical loop from “micro diagnosis of carbon source identification” to “macro design of the carbon neutrality system”. This provides empirical support for summarizing the current status of quantitative research and anchoring core themes. Combined with tools such as CiteSpace, it can also reveal theme intersections and the “technology spillover effects” of AI in carbon management, showing how technological achievements from basic research transfer into policy scenarios and niche fields. This promotes the systematic evolution of carbon research and offers a deep analytical perspective for mapping the technology diffusion path and innovation ecology in the “AI + carbon” field.

Major research methods

Currently, various research methods are used in AI applications of carbon. Among these, econometric and policy evaluation methods are one of the core tools^[36]. This type of method is primarily used to evaluate the actual impacts of AI-related policies or technological applications on carbon reduction. It focuses particularly on identifying causal relationships^[16]. For example, research on the policies of the AI Innovation and Development Pilot Zone (AIPZ) of China widely uses the dual difference method (DID) and the multi-period DID model^[17]. Pilot cities are set as treatment groups and non-pilot cities as control groups. The net effect of policies on urban carbon emissions or corporate carbon performance is then quantified^[37]. To ensure the robustness of the results, studies also employ parallel trend tests, placebo tests, and propensity score matching (PSM-DID). These methods eliminate the interference of other confounding factors^[38]. In addition, instrumental variable methods and the Heckman two-step approach are used to address endogeneity issues, tackling reverse causality and sample selection bias. For example, the AI application level of other companies in the same industry and city can serve as instrumental variables to enhance the reliability of conclusions^[39].

ML and data modeling methods play an important role in AI carbon application research. They are especially suitable for prediction, optimization, and small sample scenarios^[40]. On the one hand, various ML models are used for predicting carbon-related indicators. These include artificial neural networks (ANN), K-nearest neighbors (KNN), and linear regression (LR). For example, they are used to predict the compression strength of concrete with carbon nanotube additions^[41]. The accuracy of the models is evaluated through statistical indicators such as Coefficient of Determination (R²), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE)^[42]. On the other hand, data augmentation techniques are used to solve the problem of insufficient sample size. These include using conditional generative adversarial networks (CGAN), Copula generative adversarial networks and Gaussian Copula. These generative modeling methods synthesize virtual samples that are consistent with the distribution of real data. Combined with semi-supervised learning, they improve the precision and generalization ability of prediction models^[43]. Automated ML (AutoML) is also applied to the process of automatically designing data preprocessing, model selection, and hyperparameter tuning, improving modeling efficiency^[44].

Multi-objective optimization and decision-making methods are commonly used in AI-driven low-carbon material or process design, with a focus on achieving multiple goals such as carbon reduction and cost reduction while meeting performance requirements. For example, in the research and development of ultra-high-performance concrete (UHPC), researchers integrate multi-objective evolutionary algorithms, such as the Adaptive Grid-based Evolutionary Multi-objective Algorithm (AGE-MOEA), for multi-objective optimization. They optimize indicators including compressive strength, flexural strength, carbon footprint, and cost to generate non-dominated solution sets. Decision-making methods, such as the technique for order preference by similarity to ideal solution (TOPSIS), are then applied to select the optimal low-carbon and efficient material formula from the candidate solutions. This approach balances technical performance with low-carbon goals by quantifying the weights and constraints of each objective and provides actionable optimization solutions for engineering practice.

Mechanism and heterogeneity analysis methods are used to explore in depth the pathways and boundary conditions of AI’s influence on carbon performance, revealing the inherent logic of the “technology-carbon reduction” relationship. For mechanism analysis, this review constructed a mediation effect model to examine the specific ways AI affects corporate carbon performance through channels such as enhancing total factor productivity, optimizing supply-demand matching efficiency, and promoting green technology research and development. At the urban level, the work analyzes the carbon reduction mechanism through mediating variables including government expenditure structure adjustment, consumer behavior transformation, and industrial structure upgrading. Heterogeneity analysis focuses on differences between characteristic subjects, using methods such as group regression or generalized random forest (GRF) to investigate how factors such as enterprise ownership, financing constraints, pollution levels, and urban size or regional location influence AI-driven carbon emission reductions. This analysis clarifies the applicable scenarios and optimization directions for technology applications.

Analyzing the global research ecosystem in AI for carbon management

This dataset presents the global geographical landscape of research in the “AI + Carbon” field. Behind this lies the underlying logic of each country or region regarding energy structure, emission reduction pressure, and technological pathway selection. The top ten records are shown in Table 3. China dominates with 1,291 published articles, accounting for 51.58% of the total. This dominance is driven not only by policy-driven research investment under the “dual carbon” target but also by China’s practical needs as a manufacturing- and energy-intensive country. Research largely focuses on application scenarios such as industrial carbon emission monitoring and AI optimization in the coal-fired power industry, forming a closed loop of “policy guidance → technology implementation → industry adaptation”. This loop begins with policy: China’s “dual carbon” policy clearly defines emission reduction targets for key industries, directly guiding the technical direction for AI in carbon management. AI technology is then oriented toward these policy demands to achieve implementation. The United States ranks second with 297 articles, or 11.87%, with research focusing more on fundamental theories and cutting-edge technologies, including algorithmic innovations in ML for carbon cycle simulation and the cross-integration of AI with carbon capture technologies. This reflects the advantage of technological powerhouses in advancing innovation at the source. China’s leading position in research publications is driven by domestic “dual carbon” policies, actual emission reduction needs, and easier access to local data. This demonstrates that policy guidance can effectively stimulate technology application research. In the future, closer coordination between policy and scientific research can enable AI to better serve carbon accounting practices.

India has 233 articles, accounting for 9.31%, and Saudi Arabia has 211 articles, accounting for 8.43%. The research focus in these countries is closely related to their energy structures. As a major coal-consuming country, India emphasizes AI solutions for carbon emission control in the thermal power industry. Saudi Arabia focuses on the low-carbon transformation of the oil and gas sector, exploring AI applications for carbon footprint accounting in oil and gas extraction. This reflects the practical considerations of resource-dependent economies seeking to balance emissions reduction with development.

South Korea has 203 studies, accounting for 8.11%, exhibiting distinct “technology-industry” linkages, particularly in carbon footprint tracing for new energy vehicles and carbon optimization of smart grids. These align with its industrial upgrading and carbon neutrality goals.

England has 142 studies (5.67%) and Australia has 129 studies (5.15%), both emphasizing global climate governance. England excels in integrating AI with carbon market mechanisms, exploring intelligent regulatory technologies for cross-border carbon trading. Australia focuses on monitoring ecological carbon sinks, such as AI-based assessment of forest carbon storage, leveraging its abundant natural carbon sink resources.

Iran has 94 studies, Malaysia 93, and Canada 89. While smaller in proportion, each country shows targeted research: Iran emphasizes AI-based emission reduction in the oil and gas industry, Malaysia focuses on intelligent carbon footprint accounting in the palm oil sector, and Canada explores AI applications in simulating carbon cycles in cold regions. These examples illustrate “local problem-driven research”.

Evaluating research output intensity per capita highlights two factors: concentration of research resources and per-capita research investment. The per-capita output values of the United States (7.19), Saudi Arabia (5.98), and Australia (4.74) are significantly higher than those of China (0.92) and India (0.16). This indicates that although total publications in these countries are lower than China’s, their per-capita research intensity is higher, likely due to denser resource allocation, higher talent density, or more focused research groups. This difference reinforces distinct research patterns: technologically advanced countries tend to promote cutting-edge innovation efficiently and intensively, whereas developing countries rely more on broad policies and industrial drivers to achieve wide coverage.

Overall, this geographical distribution is not simply a difference in quantity. It is a microcosm of the role positioning of various countries in the global carbon governance pattern. China represents the practical path of “reducing emissions in transition” for developing countries. The United States embodies the idea of “innovation driven emission reduction” for technology-leading countries. Resource-based countries explore characteristic solutions in “balancing energy utilization and emission reduction”. These differentiated research directions collectively form the global ecosystem of research in the “AI + carbon” field. They provide multiple perspectives for international technological cooperation and knowledge sharing. They also offer crucial empirical evidence for reviewing and analyzing global research cooperation and differentiated paths.

A network analysis of technical foundations, application scenarios, and regional specializations

The keyword network in the field of “AI + carbon” has established the core framework for research. The deep coupling between technical methods and application scenarios constitutes its main feature [Figures 1 and 2]. AI, with ML as a key subset, relies on underlying algorithms as its computational foundation; together, these form the technological backbone that permeates the entire carbon research process. Among these, the high-frequency use of ML algorithms in carbon data modeling and emission-reduction scenario optimization has become a key link between technological capabilities and practical needs. This technological orientation is particularly prominent in research conducted by technology-leading countries such as the United States, reflecting ongoing investment in basic theoretical innovation.

Figure 1. Keyword co-occurrence network analysis.

Figure 2. Keyword density analysis.

The core cluster of scenario keywords comprises “carbon emissions”, “carbon neutrality”, “carbon cycle”, and “carbon footprint”. The strong correlation between “carbon emissions” and technology keywords reflects the core value of AI in emission monitoring, prediction, and control. This correlation is particularly significant in Chinese research and directly aligns with the demand for industrial emission reduction driven by the “dual carbon” goal. As a macro target keyword, “carbon neutrality” connects sub-directions such as “emission reduction” and “carbon removal”. Its research popularity reflects both global consensus and the differentiated exploration of different countries in the path of target implementation. For example, resource-based countries pay more attention to AI applications in the low-carbon transformation of the oil and gas industry. On the other side, countries with abundant ecological resources focus on developing carbon sink-related technologies. The close correlation between core technology keywords and carbon-related scenarios essentially means that technological capabilities precisely meet the demand for emission reduction. This indicates that AI has shifted from scattered use to system assistance. This suggests that in the future, AI can be used more in areas that are currently less focused on, such as carbon sinks.

The keywords “energy”, “CO₂”, and “climate change” have expanded the boundaries of their association. The high-frequency co-occurrence of “energy” with AI and carbon emissions reveals the central role of low-carbon transformation in energy system research. This direction is prominent in research in countries such as India, where the pressure for energy structure transformation is high. The research on “CO₂” and “climate change” is placed in a global ecological context. The application of AI in greenhouse gas monitoring and climate model simulation reflects the research characteristics of countries such as Australia and the United Kingdom. This reflects a global governance perspective.

Functional keywords such as “prediction”, “optimization”, “monitoring”, and “challenge” reflect the complete chain of research from technological application to practical implementation. “Prediction” and “optimization” reflect the technological capabilities of AI in solving complex carbon problems. Meanwhile, “challenge” and “application” point to practical bottlenecks in technology implementation, such as multi-system integration and data barriers. This practical orientation is reflected in research in various countries, but the focus varies depending on the industry foundation and technological maturity.

Overall, the keyword network forms a multi-layered structure encompassing “technical methods, core scenarios, related fields, and practical dimensions”. The research characteristics of countries/regions are indirectly presented through keyword preferences. Technology-leading countries focus on basic algorithm innovation. Transforming major countries focus on industrial emission reduction scenarios, and resource-based countries focus on adapting to characteristic fields. This pattern of differences and commonalities not only reveals the global collaborative trend of field research. It also provides key clues for sorting out the path of technology diffusion and regional cooperation directions.

Tracing knowledge structure from algorithm development to multi-scenario carbon management

Figure 3 shows a literature co-citation network generated by CiteSpace, with clusters in different colors to distinguish research groups. Yellow nodes represent literature. Size reflects citation frequency. Lines reflect co-citation relationships. In Figure 3, central nodes, such as the publications by Dwivedi YK (2022) and Acemoglu D (2020), represent highly influential studies in the field. The literature within different clusters is closely related. This illustrates the core literature inheritance and academic context of research in the "AI + carbon" field. Clusters, such as green and purple, correspond to sub-research directions, while recent studies by authors such as Ghasemi A (2024) continue to advance these areas.

Figure 3. Literature clustering analysis.

Figure 4 is the time-zone evolution diagram of CiteSpace in the “AI + carbon” field. Using time as the axis, colored arcs connect literature nodes or yellow dots, presenting the dynamic evolution of research. From a temporal perspective, early literature focused on the preliminary combination of ML algorithms and carbon footprint accounting. As time progresses, research is expanding in multiple directions. In the mid-term, AI deepens its application, covering characteristic scenarios such as porous media and system integration. Recently, new themes such as synergistic effects and energy digitalization have emerged, reflecting the deep correlation between AI, dual carbon goals, and energy systems. Different colored nodes correspond to time zones, and node sizes reflect citation frequency. Cross-time-zone connections reveal theme inheritance: early carbon emission algorithms laid the foundation for mid-term carbon cycle simulation, while mid-term model optimization supports the construction of recent carbon neutrality systems. This clearly demonstrates the evolution trajectory of the field from technical trials to system applications, from single-point breakthroughs to multi-scenario collaboration, helping to trace the origin of research and identify key development stages and innovation directions. The changes in research from early simple integration to current multi-scenario applications result from the combined effects of global emission reduction demand upgrading and AI technology progress, reflecting the law of “demand-driven technology, technology-supported emission reduction”. In the future, AI needs to pay more attention to cross-domain integration in carbon accounting and provide more comprehensive solutions.

Figure 4. Evolution of time zones in literature.