Artificial intelligence for sustainable development: assessing net carbon mitigation benefits of AI-based climate solutions with anomaly detection and life-cycle emissions accounting

1 Introduction

Global warming is seen as one of the most significant threats to humanity, endangering our planet and societies. The 2015 Paris Accord delineated a definitive objective to limit the global temperature rise to over 2 °C by 2050; nonetheless, this presents a formidable challenge that necessitates prompt action. Heat waves, rising sea levels, and intensified air pollution are threats likely to intensify owing to climate change, perhaps culminating in a significant socioeconomic and humanitarian crisis. AI-based climate solutions should not be evaluated solely through their reported improvements in forecasting accuracy, energy efficiency, or route optimization. Their environmental value depends on whether the emissions avoided during deployment exceed the emissions generated through model development, training, inference, data storage, supporting infrastructure, hardware production, and end-of-life treatment. Accordingly, a credible assessment requires clearly defined functional units and system boundaries, transparent classification of input data, and explicit treatment of uncertainty. This study addresses these requirements by developing an integrated framework for estimating the net carbon mitigation benefits of AI-based climate solutions.

Climate mitigation requires rapid emissions reductions across electricity, transport, buildings, and industry. AI can contribute to this transition by forecasting variable renewable generation, coordinating storage, optimizing electricity demand, detecting inefficiencies, and selecting lower-energy transport routes. These enabling effects make AI relevant to Sustainable Development Goals 7, 9, 11, 12, and 13. Yet AI is not environmentally neutral. Model development, repeated experimentation, training, inference, data movement, and storage consume electricity, while servers, processors, networking equipment, and data-centre infrastructure embody emissions from material extraction, manufacturing, transport, maintenance, and end-of-life treatment. A climate solution should therefore be evaluated on the basis of its net, rather than gross, mitigation effect.

Existing climate-AI studies usually emphasize the emissions avoided by an application. In parallel, sustainable-computing research has developed methods for measuring or estimating energy and carbon emissions from machine-learning workloads. Henderson et al. (2020) advocate standardized reporting of real-time energy use and carbon emissions, while Lacoste et al. (2019) show that hardware, runtime, data-centre location, and electricity mix materially affect estimates. Patterson et al. (2021) further demonstrate that model architecture, processor choice, data-centre efficiency, and location can change training emissions by orders of magnitude. These contributions improve footprint reporting, but they do not by themselves determine whether a climate-oriented AI application creates a positive net carbon benefit after the complete service delivered by the AI system is compared with a functionally equivalent baseline.

A second challenge is data quality. Published case studies frequently omit power draw, utilization, training duration, inference frequency, power-usage effectiveness, equipment lifetime, or the geographical electricity factor. Values may also be reported in incompatible units or at different scales. If such records are combined without validation, a single transcription error or inconsistent boundary can reverse the estimated net benefit. Anomaly detection is therefore incorporated here as an accounting control: it identifies suspect inputs before they enter the carbon balance. In this study, anomaly detection does not denote a new forecasting model for climate outcomes. It denotes a transparent quality-assurance procedure combining range checks, logical constraints, unit harmonization, cross-field consistency tests, and robust outlier screening.

This study addresses the following question: Under what computational, life-cycle, operational, and geographical conditions do AI-based climate solutions deliver positive net carbon mitigation after anomalous or inconsistent input records have been identified and treated? The study makes four contributions. First, it defines a common functional carbon balance that separates gross avoided emissions, operational AI emissions, and allocated embodied emissions. Second, it introduces an anomaly-detection protocol suitable for heterogeneous evidence extracted from published case studies. Third, it adds sensitivity and carbon break-even analyses so that conclusions are not tied to one assumed hardware configuration or deployment scale. Fourth, it distinguishes measured, reported, derived, and assumed inputs, preventing scenario estimates from being presented as directly observed findings. Transitioning to a more efficient approach, such as converting from a wet to a dry process utilizing a calciner, is an effective energy-saving strategy. This can reduce CO2 emissions by around 20% and conserve up to 50% of the required energy. Another feasible alternative for diminishing CO2 emissions is carbon capture and storage (CCS).

In addition, we outline in detail the obstacles that prevent the global implementation of such techniques. Here are three ways to cut down on carbon dioxide emissions: using less energy, separating and storing carbon, and making better use of alternative materials. When it comes to energy saving methods, switching to a more efficient process like going from a wet to a dry process with a calciner is the way to go. This can cut CO2 emissions by nearly 20% and save up to 50% of the energy needed. Another viable option for reducing CO2 emissions is carbon capture and storage (CCS) (Mohsin et al., 2022), and McKinsey outlines the implications globally in terms of geopolitics, business, and the overall economy. The minimization of the greenhouse gases, especially carbon dioxide, has thus been an urgent concern by individuals, governments as well as corporations not only to protect the people but also to ensure economic stability.

Besides economic effects and social effects, human capital in the form of social awareness and technical skills is essential in the reduction of CO2 emission, therefore, policy structures should be designed to accommodate these needs. The policy should be more effective in order to achieve the goals of greenhouse gas reduction like carbon neutrality. According to such scientists as Jean-Marc Jancovici, the targets of the Paris Agreement require the reduction of the CO2 emissions by 5 percent every year (Benhelal et al., 2013b). Meanwhile, measures such as reducing air travel, making driving easier, switching to plant-based diets, and making buildings more energy efficient can also help reduce emissions caused by people, such as those from electricity use, industry, transportation and deforestation. AI models are being used more and more to predict and improve the ways of capturing CO2 and reducing the emissions.

This is a promising way to deal with climate problems. There are opportunities to use artificial intelligence (AI) and data science to achieve the Paris Agreement’s objective of reducing carbon dioxide emissions by 2050 (Chen et al., 2022). Research published by PwC U.K. and Microsoft in 2020 identifies further challenges that AI must address to protect the environment and mitigate the impacts of climate change (Di Santo et al., 2018a). French scientist Jean-Marc Jancovici said that the Paris Agreement objectives need a 5 percent annual decrease in CO2 emissions. Digital technologies, particularly artificial intelligence, are essential tools for attaining these ambitious objectives. Nevertheless, the use of digital technologies is also leading to a substantial rise in energy usage. The Shift Project analysis reveals that the proportion of global CO2 emissions linked to digital activities increased from 2.5 to 3.7 percent between 2013 and 2018, now escalating at a concerning pace of 4 percent each annum (Acheampong et al., 2022). Accordingly, this study addresses the following research question: Under what computational, operational, and geographical conditions do AI-based climate solutions generate positive net CO₂ mitigation benefits after accounting for emissions from model training and inference?

The computational and energy requirements of contemporary machine learning techniques are increasing rapidly in the realm of artificial intelligence posing a risk of substantial carbon emissions. The average annual emissions of an average human being (15,800 kg), which highlights the large-scale AI systems have a substantial environmental impact. Though AI models would help in cutting down the amount of carbon emissions, they are also likely to cause a lot of emissions in the process of development and operation. As a result, it is vital to find an ideal compromise between the environmental impacts of AI in reducing the emission of CO two and lessening energy and resource use that the development and implementation of AI demands. Such a balance requires a stringent calculation of the benefits as well as the losses in terms of CO two emission, especially when the machine learning model aims at reducing such emissions. To address this study, the expressions of positive effects or positive impacts will refer to the amount of CO two emissions that will be avoided or reduced with the help of AI models. On the other hand, negative effects or negative impacts will be that of the emission during the training and the use of these models. Overall, it is possible to note that more and more studies investigate the possibilities of artificial intelligence (AI), machine learning (ML), in the context of improving the decrease in CO two emissions and consequently creating positive effects (Acheampong et al., 2022; Benhelal et al., 2013b). Nevertheless, these studies pay little attention to the emissions generated in training and deployment of ML models, which are negative effects.

This study examines whether AI-based climate solutions continue to generate meaningful carbon benefits after accounting for their computational footprint. It develops an integrated framework that distinguishes gross CO₂ mitigation the emissions avoided through an AI application from net CO₂ mitigation, calculated after deducting emissions associated with model training, deployment, and inference. The framework is applied to three representative applications involving photovoltaic forecasting, energy optimization, and electric-vehicle routing. The study further evaluates how computational intensity, deployment scale, usage frequency, and electricity-grid carbon intensity influence the resulting net benefit. By integrating enabling and footprint effects within a common accounting boundary, the study provides a more transparent basis for assessing the environmental performance of AI-enabled climate solutions. This measure may make an ecological footprint evaluation of an individual model or give comparison assessments for a number of models (Benzidia et al., 2021; Cai et al., 2020; Du et al., 2021).

The second is an ANN used in Tunisia on the solar powered water pump system is responsible for the prediction of the availability of solar energy, the optimization of the use of batteries and the continuity of operation day by day, (Basso et al., 2021b). The environmental impacts of these models, both beneficial and detrimental, were assessed under standard use settings (single user scenarios) and expanded scaling assumptions. Although the aggregate findings suggest a net positive balance, the size of gains and losses varies in the case studies and in dimensions of scaling. The negative impact associated with machine learning models is a big one and it is necessary to consider that in order to fully understand the benefits of the models. This study is organized in the following way to make it clear about the analysis of the literature, methodology, and results of assessment: Section 3 outlines the methodological framework of this study while Section 4 discusses the recommended research technique in detail. Section 5 describes how the method was used in three different scenarios and what the results were. Section 6 talks about the research’s findings, limitations, and recommendations. Section 7 ends the work by summarizing the main contributions and suggesting possible areas for more investigation.

2 Literature review and research gap

Artificial intelligence (AI) is increasingly used to address rising carbon dioxide (COâ) emissions, particularly in energy systems. Its applications include forecasting electricity supply and demand, optimizing smart-grid operation, integrating intermittent renewable energy, managing energy storage, and prioritizing low-carbon generation. Ahmad et al. (2021) identify forecasting, optimization, smart-grid control, and energy management as major opportunities for AI in the sustainable-energy sector. At the household level, Di Santo et al. (2018) demonstrate the potential of an artificial-neural-network decision system for active demand-side management in residences equipped with photovoltaic panels and batteries. By coordinating electricity generation, storage, and appliance operation, such systems can reduce energy losses that arise when batteries are either fully charged or depleted (Acheampong et al., 2022).

AI also offers mitigation opportunities beyond stationary energy systems. In transport, Basso et al. (2021b) combine machine‐learning-based energy prediction with electric‐vehicle routing, allowing route selection to account for topography, expected energy consumption, and driving conditions. Their Bayesian regression approach models both expected energy use and its uncertainty, thereby favoring routes associated with lower consumption and lower variance. Similar data‐driven forecasting and route‐optimization techniques can improve logistics and reduce food waste in delivery systems. In industrial contexts, AI can support process control, predictive maintenance, and energy‐efficiency improvements, including in energy‐intensive sectors such as iron and steel production. Nevertheless, the environmental improvement reported for one application cannot be transferred directly to another because system boundaries, baselines, time horizons, user behavior, and electricity mixes differ.

2.1 AI enabled climate mitigation

AI applications can influence emissions through better prediction, control, and resource allocation. In power systems, improved forecasting can reduce balancing errors and facilitate the integration of variable renewable-energy sources. Demand‐side management can coordinate household loads, distributed generation, and battery storage, while industrial AI can identify inefficient processes and improve equipment utilization. In the circular‐economy domain, Spreafico and Spreafico (2021) show how text‐mining techniques can retrieve and organize relevant information from large bodies of literature. Although their study does not directly quantify avoided emissions, it demonstrates how AI‐enabled information extraction can support circular‐economy knowledge development and the identification of strategies related to resource efficiency, reuse, and waste reduction.

Machine-learning methods have also been used to model and predict emissions at the macroeconomic level. Mardani et al. (2020) propose a multistage approach combining dimensionality reduction, clustering, and machine‐learning techniques to predict COâ, emissions. Their framework illustrates how preprocessing and segmentation can enhance the analysis of complex emission determinants. Similarly, Chen et al. (2021) employ artificial neural networks and adaptive neuro‐fuzzy inference systems to predict COâ emissions in G20 economies using factors such as economic growth and energy consumption. Mohsin et al. (2022) provide further data‐science evidence for Europe and Central Asia by assessing how fuel‐energy consumption, foreign direct investment, and gross domestic product affect COâ emissions. Together, these studies demonstrate the value of AI and data‐science methods for emission prediction, scenario analysis, risk assessment, and policy targeting.

The climate benefits of AI can nevertheless be overstated when a system shifts emissions across locations or life‐cycle stages rather than eliminating them. Operational electricity savings may occur at the point of application while additional emissions from computation, data transmission, hardware production, or infrastructure arise elsewhere. A defensible assessment therefore requires a clearly defined functional unit, an explicit counterfactual baseline, appropriate temporal and geographical boundaries, and consistent treatment of upstream and downstream processes. Predictive accuracy alone does not establish environmental benefit; the avoided emissions attributable to an AI application must be compared with the emissions generated throughout its own life cycle.

2.2 Computational and life-cycle emissions of AI

The carbon footprint of an AI system extends beyond a single model-training run. Model-development emissions may arise from exploratory runs, hyperparameter searches, failed experiments, validation, and retraining. Deployment emissions include repeated inference, data storage and transmission, monitoring, and idle computing capacity. Operational electricity use is generally converted into carbon dioxide equivalent (CO₂e) with a grid-emission factor appropriate to the place and time of computation, while power‐usage effectiveness captures cooling and other data-center overheads. Lacoste et al. (2019), Henderson et al. (2020), collectively emphasize the importance of reporting these variables rather than treating model size as a sufficient proxy for environmental impact. In particular, Henderson et al. (2020) propose a framework for tracking real‐time energy consumption and carbon emissions and for producing standardized reporting appendices, thereby improving the transparency and comparability of machine‐learning experiments.

Life‐cycle accounting additionally allocates part of the emissions associated with hardware manufacture, transportation, maintenance, supporting infrastructure, and end‐of‐life treatment to the evaluated workload. This allocation may be based on device‐hours relative to usable hardware life and adjusted for utilization when computing capacity is shared. Considerable uncertainty remains because product-specific environmental declarations and reliable utilization data are often unavailable. Embodied emissions should therefore be reported separately and evaluated under low, central, and high allocation scenarios rather than being concealed within a single‐point estimate.

Interest in the energy consumption of machine-learning models predates the recent expansion of generative AI (Hong et al., 2022; Fahdi et al., 2021). In the context of accelerating climate change and growing computational demand, reliable methods for estimating the actual energy costs of AI systems have become indispensable. However, only a limited number of studies provide comprehensive calculation methods that cover the full model‐development and deployment process. Early contributions to the assessment of artificial‐neural‐network training emphasize that hardware type, electricity source, computing demand, and training duration can all materially influence environmental performance (Benzidia et al., 2021). Moreover, training is not the only energy-intensive stage: inference can become the dominant source of operational emissions when a model is deployed at scale and queried repeatedly.

AI is simultaneously transforming the broader energy sector. It can coordinate complex combinations of renewable generation, storage, supply, and demand; support predictive maintenance; improve cybersecurity and smart-grid control; and process data volumes that conventional analytical systems may struggle to manage. Ahmad et al. (2021) emphasize AI applications in solar and hydrogen production, supply-and-demand management, smart grids, the Internet of Things, robotics, predictive maintenance, and computational performance. These technologies may enhance operational efficiency and competitiveness for utilities, system operators, and independent power producers. Their expansion, however, also creates regulatory concerns relating to consumer protection, privacy, data security, and the timely evaluation of new services in emerging digita-energy markets.

Improvements in hardware efficiency, often measured in computational operations per watt, have not necessarily stabilized total electricity consumption. Increasing model size, usage, and deployment can partly or entirely offset efficiency gains. This possibility is consistent with the rebound effect: reductions in the resource requirement per unit of service can lower effective costs and stimulate additional consumption. Comparable patterns have been observed in other sectors, where improvements in appliance or transport efficiency coincided with growth in total capacity, activity, or fuel use (Du et al., 2021). For AI, the environmental outcome therefore depends not only on efficiency per computation but also on the total scale and frequency of computation.

Accurate carbon accounting is further complicated by differences among data centers, hardware configurations, computational methods, electricity grids, and geographical locations (Taddeo et al., 2021). Training and inference must be distinguished because their relative contributions depend on model-development intensity and deployment scale. The relationship between predictive performance and emissions may also be highly nonlinear: marginal accuracy improvements can require disproportionately greater computation. Yet AI research and deployment decisions often prioritize accuracy while giving less attention to energy and carbon efficiency. Policy and governance frameworks are consequently needed to encourage transparent reporting and ensure that AI contributes to, rather than undermines, sustainability objectives.

The literature has developed along two largely separate streams. Climate: AI studies primarily examine AI's enabling effects, including emissions avoided through better forecasting, optimization, mobility, industrial control, and environmental prediction. Sustainable-computing studies, by contrast, concentrate on the energy and carbon footprint of model training and inference. Green‐AI and carbon-accounting research recommends disclosing hardware characteristics, operating time, electricity consumption, computing location, grid carbon intensity, data‐center overhead, and, where possible, embodied hardware emissions. Nonetheless, few studies place avoided emissions and AI‐related computational emissions within the same case-level accounting framework.

This separation prevents a clear assessment of AI's net climate contribution. An application may generate substantial gross emission savings yet deliver only modest net benefits after training, inference, data transfer, and infrastructure emissions are deducted. Conversely, a computationally efficient model may have little climate value if it does not produce meaningful additional emission reductions relative to a credible non‐AI baseline. The present study addresses this gap by applying a common gross‐to‐net mitigation structure across three illustrative AI applications. This framework makes system boundaries, counterfactual assumptions, operational emissions, and uncertainty explicit, thereby providing a more complete and comparable assessment of AI's environmental performance.

3 Methodology

This study employs a structured multi-stage methodological approach to examine the positive and negative effects of machine learning (ML) models on carbon dioxide (CO₂) emission reduction. The suggested approach aims to deliver a comprehensive evaluation of the environmental advantages produced by AI-enabled solutions alongside the carbon emissions linked to the development, training, and deployment of machine learning models.

This study seeks to deliver a more thorough assessment of the net environmental impact of AI technology, as prior research predominantly emphasizes the emission reduction capabilities of AI applications while largely neglecting their operating carbon footprint. The approach has five consecutive phases. Initially, a conceptual assessment methodology is established to measure the dual impacts of machine learning applications on CO₂ emissions. The framework includes both the emission reductions attained by AI-assisted optimization and the emissions produced by computational processes. This dual-perspective methodology constitutes a significant addition to the current literature; wherein adverse environmental effects are frequently overlooked.

The case studies were identified through a targeted, purposive review of the literature discussed in Section 2. The selection was illustrative rather than systematic because the objective was to demonstrate the applicability of the proposed accounting framework across contrasting AI-enabled climate applications, rather than to estimate an aggregate effect for the entire literature. A study was eligible when it: (i) reported a quantifiable baseline and AI-enabled outcome from which avoided emissions could be estimated; (ii) provided sufficient information on the model, computational process, or deployment conditions to approximate AI-related emissions; (iii) concerned an application directly associated with energy use or CO₂ mitigation; and (iv) added sectoral or geographical diversity to the case portfolio. Studies lacking sufficient outcome or computational information were excluded. Applying these criteria resulted in three cases covering photovoltaic forecasting, energy optimization, and electric-vehicle routing.

The fourth phase emphasizes the comparative assessment of the chosen case studies. Comparisons are performed at two levels due to variations in scale, application domain, and technological intricacy. The initial level addresses nominal deployment, reflecting the environmental performance of an individual implementation or user. The second level analyses scaled deployment scenarios, highlighting the extensive environmental consequences of widespread adoption across numerous consumers or operational units. This dual-level comparison facilitates a more accurate evaluation of the sustainability potential of AI technologies. The fifth phase consolidates the information from all case studies to discern common trends, possibilities, and difficulties related to AI-driven carbon mitigation. The findings are utilized to assess the overall efficacy of machine learning applications in mitigating CO₂ emissions and to formulate comprehensive suggestions for researchers, practitioners, and policymakers. The suggested methodology integrates environmental benefits and computational costs, offering a thorough framework for evaluating the contribution of artificial intelligence to sustainable development and climate mitigation goals.

3.1 Study design and case selection

The study uses a comparative, literature-derived case design. Three applications were retained because each links an AI-enabled decision to an energy or emissions outcome and provides sufficient information to reconstruct at least a gross-benefit scenario: (1) photovoltaic generation and household demand-side/battery management; (2) AI-assisted energy-system optimization; and (3) electric-vehicle routing with machine-learning energy prediction. The unit of analysis is one AI-enabled service delivered for a defined user population and assessment period. The applications are not treated as a statistically representative sample of all climate AI. They are methodological demonstrations selected to test the portability of the accounting framework across electricity and transport contexts.

Evidence was classified as reported (stated in the original study), derived (calculated directly from reported values), externally sourced (for example, a grid factor), or assumed (required because no measurement was disclosed). Results dependent on assumed inputs are labelled scenario estimates. Conflicting geographical labels in earlier drafts were removed; the revised analysis refers to each application by function unless the original publication clearly establishes the location.

3.2 Functional unit, baseline, and system boundary

For each case, the functional unit is the delivery of the same service over 1 year under two conditions: a reference condition without the evaluated AI intervention and an AI-enabled condition. The comparison holds the service level constant. Gross avoided emissions are the difference between baseline service emissions and AI-enabled service emissions before the AI footprint is deducted. The assessment boundary includes model development and training, retraining, inference, data storage and transfer where estimable, data-centre overhead, allocated hardware manufacturing, and allocated end-of-life emissions. Emissions associated with constructing the underlying energy or transport asset are excluded unless they differ between the baseline and AI scenarios.

3.3 Anomaly-detection and data-validation protocol

Anomaly detection is conducted before Equations 1–4 are evaluated. The protocol has five stages. First, all quantities are converted to common units (kWh, hours, kg CO₂e/kWh, and kg CO₂e per year). Second, deterministic range rules flag negative runtime, negative power, utilization above 100%, PUE below 1, and savings outside the physically meaningful interval for the stated metric. Third, logical consistency tests verify that the AI scenario and baseline deliver the same service and that stated savings agree with the underlying arithmetic. Fourth, cross-field tests compare the country, grid factor, hardware description, assessment year, and deployment scale. Fifth, where at least five comparable observations exist, a robust modified z-score is calculated using the median and median absolute deviation; |Mz| > 3.5 is treated as a review flag rather than automatic deletion.

Because the comparative demonstration contains only three applications, statistical outlier detection across case-level net benefits would be unreliable. Robust scores are therefore used only within sufficiently populated parameter sets, while rule-based and logical detection remain primary. Every flag is resolved by checking the source, correcting an unambiguous transcription or unit error, excluding the value if it cannot be reconciled, or retaining it with an uncertainty range. This conservative procedure prevents the anomaly-detection label from being used to imply an unsupported predictive experiment.

4 Definition of the methods

4.1 Approach

This paper proposes a method for assessing the worldwide effects of artificial intelligence (AI) and machine learning (ML) technologies on carbon dioxide (CO₂) emissions, including both positive and negative consequences. As previously said, several models have been created and used with the specific aim of decreasing CO₂ emissions (Ahmad et al., 2021). Despite functioning in many industries, these models have a constant overriding objective. Thus, these applications may have beneficial outcomes, especially regarding emission reduction. A comprehensive statement of an AI model’s true effect requires an examination of its whole life cycle, including both training and inference stages. A thorough knowledge of the model’s environmental consequences may be attained alone by measuring the positive and negative contributions across these phases. The two predicted values may ultimately be compared and critically analyzed across individual or several models. This comparison study facilitates the examination of many relevant inquiries. Do all models provide a net beneficial effect on the climate regarding CO₂ emissions? Could some implementations be worse than not ameliorating the problem?

Which activities can be acknowledged as useful in reducing the negative impacts of these models? Recent studies have revealed the essential input factors for evaluating the adverse impacts of AI models, which are now incorporated into contemporary methodologies. On the other hand, there is still no clear way to measure how AI models can help reduce CO2 emissions (Henderson et al. (2020): Towards the Systematic Reporting of the). To address this shortcoming, we present a systematic methodology that measures the variance in CO2 emissions between the adoption and non-implementation of the AI model. This framework makes it easy to figure out these amounts for any model. The next few sections describe detailed steps for how to evaluate both good and bad outcomes.

4.2 Calculating beneficial effects (quantity of CO2 mitigated)

The scope of the project and the desired goal determine the extent of the CO₂ reductions achieved by the model. Each scenario has its own set of projected emissions, with the deviation reflecting the reductions in CO₂, whether positive or negative. This decrease could be represented either as a percentage difference between the two scenarios or as an absolute amount (the amount of CO₂ emissions of the AI scenario compared to the reference scenario).

When comparing projects, percentage indicators are usually the way to go because the scale of the original project has a significant impact on the actual amounts of CO₂ reduction. There is currently no unified way in the academic literature for calculating the exact amount of greenhouse gas (GHG) emissions that an AI model can save. In terms of existing approaches, there is a wide variety that can be applied based on the model’s intended use, the country in which it is implemented, and the industry in question (energy, transportation, etc.). Nevertheless, some recurrent principles can be followed to estimate the CO2 emissions under both scenarios, depending on the intended function of the model. As one can see in studies related to this topic, artificial intelligence is often applied in the prediction of physical variables such as power output, energy requirement, or temperature. Having accurate predictions makes it possible for the operational system to make the necessary energy consumption reduction - either in terms of energy savings, or by replacing energy sources with cleaner alternatives. For instance, an AI model may facilitate replacing part of the electricity generated from fossil-fuel based grid by energy generated from solar panels. Let m i kgCO 2 e represent the total CO₂ emissions for scenario (i), measured in kilograms. Let E i consumed denote the annual energy consumption in scenario (i), expressed in kWh. The carbon intensity of the local electricity grid and that of a lower-emission alternative energy source are represented by I local grid and I low emitting means, respectively, each measured in kgCO₂e/kWh. The calculation framework can be presented as follows,

Reference scenario (a),

m a kgCO 2 e = E a consumed × I local grid (1)

AI-based scenario (b),

m b kgCO 2 e = E a consumed − E b consumed × I local grid + E b consumed × I low emitting means (2)

4.3 Assessment of harmful impacts (volume of CO₂ emitted)

This evaluation is difficult since there are a number of stages and variables that need to be taken into consideration, and the study that has been done on the subject is insufficient. A comprehensive lifetime analysis of a typical digital system often encompasses emissions from data storage, transmission, and online utilization. Nonetheless, the situation with AI is rather distinct. Model training and inference are primarily responsible for the majority of CO₂ emissions, with other phases contributing somewhat (Henderson et al., 2020). Our work concentrates on these two phases, utilizing previous knowledge to structure the mathematics. The environmental implications of AI have just recently emerged as a significant concern; they were not seen as such until a few years back (Di Santo et al., 2018). The computational requirements of deep learning systems have significantly increased since the early 2010s, resulting in environmental consequences that can no longer be disregarded.

The energy consumption and associated emissions of an AI model depend on several factors, including hardware type, computational demand, training duration, cloud-provider location, and the carbon intensity of the electricity supply (Henderson et al., 2020; Lacoste et al., 2019). Because the selected studies do not report every computational parameter, the present analysis uses clearly identified estimates where direct measurements are unavailable. Training duration is estimated as:

t train h = C train FLOPs R hardware FLOPS × 3600 (3)

Where C train is the total number of floating-point operations required for training, and R hardware is the hardware processing rate in floating-point operations per second. FLOPs therefore represent computational demand, whereas FLOPS represent processing speed. Division by 3,600 converts seconds into hours.

The electricity consumed during training is calculated as:

E train kWh = P hardware W × t train h 1000 (4)

Training-related emissions are subsequently estimated as:

C E train kgCO 2 e = E train kWh × C I region kgCO 2 e / kWh (5)

Where P hardware denotes hardware power consumption and C I region represents the electricity-grid carbon intensity of the assumed computing location. Regional carbon-intensity values are obtained from the specified International Energy Agency dataset. When hardware or computational information is unavailable, the adopted value is reported as an assumption in the relevant case study rather than treated as a directly observed measurement.

4.4 Input assumptions and uncertainty treatment

The selected source studies do not disclose all the computational and operational parameters required to estimate the life-cycle emissions of their AI applications. Therefore, every input was classified as source-reported, externally sourced, derived, or author-assumed. Source-reported inputs were taken directly from the original case studies. Externally sourced inputs, including electricity-grid emission factors and hardware specifications, were obtained from identifiable databases or technical documentation. Derived inputs were calculated from reported or externally sourced values, while author-assumed inputs were introduced only where the required information was unavailable.

Table 1 presents the hardware power, computational demand, hardware throughput, training time, inference duration, model lifetime, computing location, electricity-grid emission factor, PUE, input status, and uncertainty range used in each case. Low-, central-, and high-emissions scenarios were constructed to examine the sensitivity of the results to uncertain computational parameters. These ranges represent scenario assumptions and should not be interpreted as statistical confidence intervals.

TABLE 1. Illustrative sensitivity and break-even results.

The EV, gross mitigation changes across scenarios because the deployment grid factor changes. Under the illustrative grid factors.

Low scenario: approximately 1.92 kgCO₂e/vehicle-year.

Central scenario: approximately 3.70 kgCO₂e/vehicle-year.

High scenario: approximately 14.79 kgCO₂e/vehicle-year.

Training time was calculated in Equation 6 as:

t train = C train R hardware × 3 , 600 (6)

Training electricity use was calculated in Equation 7 as:

E train = P train × t train × P U E 1 , 000 (7)

Annualized training emissions were calculated in Equation 8 as:

A T E = E train × C I compute L (8)

Annual inference emissions were calculated in Equation 9 as:

I E = P inference × t inference × P U E 1 , 000 × C I compute (9)

4.5 Conceptual and implemented accounting boundaries

The proposed framework defines a comprehensive life-cycle-informed accounting boundary consisting of: (i) gross emissions avoided through the AI-enabled intervention; (ii) operational emissions from model development, training, inference, data storage, and supporting computation; (iii) allocated embodied emissions from hardware manufacturing; (iv) the allocated emissions of supporting infrastructure; and (v) end-of-life treatment.

The empirical boundary applied to the three case studies is narrower because the original studies do not report the hardware inventories, manufacturing footprints, equipment utilization shares, infrastructure allocation factors, or end-of-life data required to quantify all components. Consequently, the central calculations include gross avoided emissions, training emissions, inference emissions, and PUE-adjusted electricity consumption. Embodied hardware, infrastructure-construction, data-storage, transmission, and end-of-life emissions are recorded as “not quantified” and are not assigned a value of zero. Accordingly, the empirical results represent partial net-mitigation estimates and may overstate the complete life-cycle benefit.

5 Evaluations and results

The framework outlines a process with three main components: identification of existing AI models from literature aimed at reducing CO2 emissions; implementation of the proposed evaluation framework to calculate the overall act of individual models selected for the evaluation and comparative analysis of results of different models.

5.1 Reviewing existing AI models documented in the literature

Numerous works present AI-driven approaches aiming to reduce climate change and to reduce CO2 emission (as outlined in the literature review). However, as such works were not developed with a standardized assessment of both the benefits and the drawbacks (i.e., how much CO2 is avoided and how much CO2 is generated by the AI models), they often do not have the necessary information to apply the evaluation framework presented in the previous section. Important details like hardware details, model parameters, hyperparameter configurations for the estimation of training duration, or FLOPs are often missing. This purposive selection supports methodological illustration and cross-case comparison but does not constitute a systematic review or a statistically representative sample of AI-based climate solutions. Consequently, the case results should not be generalized to all AI applications. Accordingly, the case studies that form the basis for our analysis were chosen on the basis of the following criteria. Sufficient level of detail: The selected studies need to provide sufficient quantities of information to make an estimate of CO2 reductions with a high degree of certainty. Availability of technical implementation data.

It is important to have essential information about the algorithmic approach such as the modeling approach, the training setup, and other technical parameters, as these are needed to assess both the impacts of the Artificial intelligence system. Sectoral diversity Eligible studies should be distributed across different industries to ensure that a wider view of the role AI plays in meeting climate-related aims can be understood, and thus make the results more relevant and comparable. Geographical diversity: Similarly, it is useful to include studies from different parts of the world or continents, as the estimation of the carbon footprint also depends to some extent on where the computational infrastructure is located. Utilizing these criteria, we identified the AI models featured in the three papers enumerated in Table 2. While supplementary models may have been used, we deliberately restricted the quantity to facilitate a more comprehensive analysis of technique application and to scrutinize each model’s lifecycle with enhanced precision. The objective is to provide a clear and straightforward exposition of the suggested methodology, enabling researchers to easily implement it in their existing or forthcoming AI models.

TABLE 2. Analysis of electricity drawn from the grid and energy exported.

5.2 Methodological application within the three case investigations

The efficiency of electricity usage and storage is enhanced by a decision-making system grounded in an artificial neural network (ANN). The goal is to cut down on the need for grid electricity and lower energy expenses. Because energy demand and supply change throughout the day, a battery is utilized to smooth out these changes. But depending only on smart grid electricity can be costly and may make the carbon impact bigger. In real life, when there are times of very high demand or supply in certain areas, the battery can’t properly smooth out the changes, which can lead to energy losses or big drops in network electricity use. The electricity that comes from the grid is different from the electricity that comes from photovoltaic (PV) panels, and it usually releases more CO₂. The operation of the system can be represented by the following Equation 10 as:where: grid(i) is the amount of electricity that is supplied by the grid (that is, whether it is positive or negative); cons(i) is the amount of power that the consumer requires at any particular instant.

P grid k = P cons k − P solar k − P storage k + Loss storage k + Loss conv k (10)

The term “solar” refers to the power that is produced by the photovoltaic panels, “storage” refers to the power that is supplied by the battery, “loss storage” refers to the energy that is lost from the battery, and “lesson” takes into account the lost energy that occurs during the conversion process. Cons, sun, and energy (the changeable daily power tariff) are the inputs that are utilized by the artificial neural network-based system. Its result, abbreviated as storage opt, provides an indication of the optimal daily battery power that should be employed by the system. (Di Santo et al. (2018) conducted a study in Brazil employing Multi-Layer Perceptron and Nonlinear Auto-Regressive neural network models to enhance demand-side management and lower household electricity consumption. Meanwhile (Basso et al., 2021), focused on the BRICS region, applying Bayesian Regression to predict energy needs in electric vehicle routing with the aim of optimizing overall energy usage.

5.2.1 Assessment of the model’s beneficial effects

According to the International Energy Agency (2018), 1 kW h of power obtained from the national grid represents 0,124kgCO2e, while 1 kW h of power obtained from photovoltaic (PV) panels manufactured in China represents 0,044kgCO2 ADEME, 2018, (Cheng et al., 2024). Various artificial neural network (ANN) architectures were developed using distinct parameters, including the quantity of layers and neurons. The ideal artificial neural network design was determined based on the minimal Mean Squared Error (MSE). The modified artificial neural network was subsequently used on actual home power consumption data from São Paulo, facilitating the calculation of CO₂ reductions via enhanced demand-side management These patterns reflect variability in photovoltaic panel performance under uniform sunlight circumstances, which may be influenced by variables like as panel material, design, or overall quality. Nine major representative profile of “Brazil consumers with PV-based electricity generation” were generated by integrating the three consumer profiles with the three PV production patterns. This study’s following research focuses only on the Average Brazilian consumer profile, which functions as the baseline for assessing the model’s effects.

In this Table 2 it shows that this research focuses on the optimal situation, specifically the Average–Gbetter consumer: a typical Brazilian family using high-performance photovoltaic panels. The average daily power use in Brazil is around 13.7 kWh per home. This number acts as the reference point for assessing the possible CO₂ reductions realized via efficient smart grid management (Section 3.2). to assess power use. m_(b,“kgCO₂e”) = (E_(a,“consumed”) - E_(b,“consumed”)) ⋅ I_“local grid” + E_(b,“consumed”) ⋅ I_“low-emitting means” A typical Brazilian user using high-performance photovoltaic panels (Average–Gbetter profile) contributes 13.7 kWh/day to the grid while consuming just 5.9 kWh/day from it.

Therefore (E_(a,“consumed”) - E_(b,“consumed”)) = 5.9 kWh. The residual 7.8 kWh/day is provided directly by the consumer’s photovoltaic panels, i.e., E_(b,“consumed”) = 7.8 kWh Considering the average cost of carbon of Brazil in 2018 (I″ local grid” = 0.124 kgCO₂e/kWh) and the average emission factor for photovoltaic panels produced in China (I_“low-emitting means” = 0.044 kgCO₂e/kWh), the total daily emissions are calculated as: m_(b,“kgCO₂e”) = 1.19 kgCO₂e. By contrast, in the reference scenario without ANN-based smart grid optimization, the average Brazilian consumer relies entirely on grid electricity, with a daily consumption of 12.7 kWh. Applying Equation 2 yields: m a , k g C O 2 e = 1.81 k g C O 2 e Assuming that photovoltaic (PV) panel output remains constant throughout the year, the daily calculations can be extrapolated to annual values. This signifies a 33% decrease in CO₂ emissions, which is acknowledged as the measurable beneficial effect of the first ANN-based algorithm.

5.2.2 Analysis of the model’s negative effect

5.2.2.1 Model training

The training and inference phases of the model are the principal areas of concern for assessing negative impacts, as previously stated. We first assess the total CO2 emissions linked to the training of this algorithm. {“kgCO”}_2″e”) = P_“hardware” ⋅Δt “training” ⋅I_“region” The page fails to specify the hardware power. Since the model seemingly does not necessitate any specialized hardware, we utilize the default configuration P_“hardware” = 235 W. The duration of training is unspecified. To compute Delta t “training”, we utilize the quantity of FLOPs and the hardware’s processing capability measured in FLOPS. The default configuration for a single ANN model is set at 1.1010^20 FLOPs. The document indicates that 17 distinct ANN configurations were trained, differing in factors such as neuron count, layer count, and time step discretization. Although the FLOPs number will vary from one training to another, we can estimate the total by multiplying by 10 to get a more precise estimate. As a result of this, we have 1.1010^21 FLOPs for this ANN tool.

We use a default hardware computation capacity of 0.098^09 FLOPS, leading to a training time of a duration of Delta t_“training” = 1.99^3 h. The electricity intensity I _“region” does not match the electricity intensity used within the geographical structure of the system as the energy used to train and infer can come from multiple locations around the world depending on the location of the cloud provider. According to Lacoste et al. (2019), we have decided to select Amazon Web Services deployed in Sao Paulo, Brazil, which is identical with the region of the system. For this supplier, I_“region” = 0.112 kgCO2e/kWh. For the whole existence of the model, the total CO2 emissions in the model training are calculated to be m_“kgCO2e” = 785“kgCO2e”. Our estimates, which are for a single year, require that we divide this number by the model’s estimated useful life of 10 years to yield an annual training impact of m_“kgCO2e,1 year” = 73“kgCO2e”.

5.2.2.2 Model inference

We assess the energy consumption at the inference stage, used to refer to the electricity needed to run the model once it has been trained and deployed. Using Equation 3 we can approximate the associated CO2 emissions. For the purposes of this assessment, we assume a continuous operation of the model 24 h a day, 365 days a year as the optimal battery power (P_“storage” ^“opt”) is calculated in real time with live data. This means that the total operation time per year is 21*365 = 7,723 h. Under these assumptions, the total CO2 emission from inference in 1 year is m_“kgCO2e” = 234“kgCO2e”. This work investigates a solar water pumping system in Brazil which is known to be simple and independent from fuel sources. One of the major challenges of solar energy is that it is intermittent; that is, electricity production stops before dawn and after dusk.

Forecasting photovoltaic power with empirical models and artificial intelligence methodologies for water pumping systems (Ammar et al., 2020). The excess energy is stored in a battery for subsequent use in operating the water pumps. Two artificial intelligence models, the Adaptive Neuro-Fuzzy Inference System (ANFIS) and the Feedforward Neural Network (FFNN), are employed to optimize energy allocation and battery protection. These models forecast the solar panel output and regulate the energy distribution among the photovoltaic panels, pumps, and batteries, thereby optimizing water pumping throughout the day. Throughout the research, the following nomenclature will be employed to eliminate any ambiguity: P_“BP” denotes the power sent from the battery to the pumps, whereas P_“Lost” signifies the power squandered within system. P ̂_“PV” denotes the anticipated power output from the solar panels, quantified in “watts.” P_“L” denotes the power consumption of the water pumps, measured in watts. P_“bat” denotes the power supplied by the battery. When P_“bat” exceeds zero, the battery is providing power to the pumps. Conversely, when P_“bat” is negative, the solar panels recharge the battery. The acronym “PVP” denotes the transmission of power from the solar panels straight to the pumps (W).

The two AI algorithms do not have the capacity to directly predict solar power. Sun irradiance (in Watts per square meter) and the temperature of the surrounding environment (in degrees Celsius) are predicted by them. In order to explain a wide range of weather conditions, the models are trained on using physical data collected from Tunisia on bright and cloudy days. Normalized mean bias error (NMBE) and normalized root mean squared error (NRMSE) are the two metrics that are used in the process of finding the performance of the model. In both bright and cloudy conditions, ANFIS model generates the best possible results in terms of the prediction of solar irradiance and outdoor temperature. A mathematical representation of the photovoltaic panels is used to calculate the power generated by the solar panels and is represented by the symbol P _“PV.” This comes after the prediction model has been chosen.

5.2.3 Analysis of the model’s positive effect

The use of Artificial intelligence in forecasting solar energy is mostly beneficial in reducing the CO2 emission by saving electricity, which brings about the reduction of P_“Lost”. Additional benefits with the potential to protect the battery and improve the satisfaction of the daily water pumping needs are of high importance for the performance of the system, but do not have a direct impact on the CO2 emission. Consequently, the best approach is to quantify the amount of electricity saved daily using this ANN based electricity forecasting instrument. The blue curve (P_“PVP”) indicates the amount of electricity supplied to the pumps. At noon, when solar generation is at its peak, the battery is at 80 percent of capacity, causing any excess energy to be lost (illustrated by the purple curve and the “Lost” sign). In the same vein as the case study that came before it, the design of the system is intended to meet demand at all times, which can be described statistically in a succinct manner Equation 11 as:

P L = P PVP + P bat only when P bat > 0 (11)

When P bat < 0, it indicates excess solar power, and the relationship simplifies to Equation 12 as:

P L = P PVP (12)

To perform the CO₂ calculation, numerical values must be derived, and the graph provides a suitable basis for this estimation in Equation 13 as.

E = ∫ t 0 t end P d t (13)

On overcast days, the pump demand (PL) remains constant; however, the photovoltaic contribution (PPVP) varies, hence influencing Pbat and PLost. Nevertheless, the prevailing trends are analogous; so, we confine our investigation to a sunny day for the sake of simplicity. Step 1: CO₂ emissions through AI processes. The study reveals experimental results for a sunny day: Forty-eight percent of the total system energy flow is ascribed to the pump demand (EL = 28.7 kWh), culminating in an overall system energy of 58.9 kWh.

The battery energy constitutes 11% of the whole, with Ebat = 6.4 kWh, and this energy is additionally generated by photovoltaic panels. Utilizing Equation 2 to compute CO₂ emissions, there is no consumption of grid electricity; all energy is derived from photovoltaic panels and batteries. Thus Equation 14 will be,

m b k g C O 2 e = E b consumed · I low emitting means (14)

where Equation 15 will be,

E b consumed = E PVP + E bat = 32.5 kWh + 7.5 kWh = 40 kWhand (15)

So the next Equation 16 will be,

I low emitting means = 0.044 k g C O 2 e / k W h . (16)

Figure 1 illustrates the operational mechanism underlying the calculation of avoided electricity use in the photovoltaic case. It supports the estimation of gross mitigation but does not directly represent computational emissions, which are calculated separately using Equations 3–5.

5.2.4 Analysis of the model’s negative effect

5.2.4.1 Model training

In this study with Equation 3, Hardware power (P_hardware): In the absence of particular specifications, we utilize the default value specified in Section 3.2, P_“hardware” = 235 W. Duration of training (Δt “training”): This is the most difficult parameter to assess. For this use case, we assume a computational capacity of 1.10 × 10^13 FLOPS and a total of 10^20 FLOPS, resulting in an anticipated training duration of roughly 28,000 h. Electricity intensity (I_“region”): This is dependent on the geographical location of the computing provider. Due to lack of significant data centers in Tunisia, we have selected the closest supplier, i.e., Amazon Web services in Frankfurt, Germany with an I_“region” value of 0.490 kgCO2e/kWh.

5.2.4.2 Model inference

To assess the energy usage during the inference phase, we initially ascertain the duration of daily model running. Despite solar power being nonexistent during specific hours, the model functions continuously (24/7) since it analyzes diverse physical measurements, including nocturnal data, to predict solar power for the subsequent day. This results in an annual usage time of 24 × 365 = 8 , 760 hours. Utilizing the identical hardware capacity (P_“hardware” = 235 W) and electricity intensity (I_“region” = 0.490 kgCO₂e/kWh) as during training, given that hardware and cloud provider are consistent, the projected CO₂ emissions from inference over 1-year amount to m_“kgCO₂e” = 1,088 kgCO₂e.

Optimization of electric vehicle routes using machine-learning models for energy prediction. This study examines traffic planning for electric vehicles (EVs) aimed at reducing electricity use by determining the most energy-efficient routes. A probabilistic machine learning approach is utilized to forecast energy consumption across many paths, facilitating the discovery of the path with minimal utilization. The experiment is conducted in BRICS nations, where one kWh of grid electricity produces 0.016 kgCO₂e (International Energy Agency, 2018). The research indicates that an electric vehicle’s energy consumption is contingent upon its weight and average velocity in Equation 17 as:where E denotes the energy necessary for transit between places a and b, and α and β are coefficients affected by variables including road gradient, vehicle frontal area, acceleration, and aerodynamic drag. In this approach, α and β are regarded as stochastic variables. The approach predicts the probabilistic distributions of these coefficients, enabling the estimation of both E and its variation for a certain vehicle and route segment. The approach calculates the total projected energy usage and its associated ariance for the complete voyage by consolidating many route parts.

E a b = α a b · m + β a b · v 2 (17)

5.2.5 Analysis of the model’s positive effect

The post provides energy usage statistics that quantifies the advantages of the machine learning approach. Table 3 juxtaposes the electricity use of 20 electric vehicle users utilizing the Electric Car Routing Problem with Chance-Constraints methodology against that of a traditional, non-optimized method. Our analysis utilizes an average energy savings of 8.96%, approximated to 9%. In this case study, Equations are modified as follows:

TABLE 3. Comparison of EVRP-CC and classical EVRP performance in the EV-routing case.

This table reports the case-specific performance indicators extracted from the original EV-routing study. It is not the consolidated carbon-accounting comparison of the three cases.

m b k g C O 2 e = E b consumed · I local grid , m a k g C O 2 e = E a consumed · I local grid , Since the only difference between scenarios is energy consumption (the local grid remains the same), we can express the AI-enabled consumption as, E b consumed = r · E a consumed , r = 0.83,The average European person drove 10,143 km in 2018 (source: Odyssee-Mure), while the Volvo XC38 exhibits an energy consumption of approximately 203 Wh/km (source: Electric Vehicle Database: ev-database.org). This yields an annual electricity consumption of, E a consumed = 10 , 476 × 0.187 kWh = 2134 kWh Corresponding to CO₂ emissions of m a k g C O 2 e = 36.7 k g C O 2 e. With the AI-based route optimization, the car consumes 1976 kWh per year, resulting in m b k g C O 2 e = 33.3 k g C O 2 e. Finally, a slightly different scenario can also be considered: applying the same tool in Brazil, where part of the model training data originated, to evaluate potential energy savings in another context.

5.2.6 Analysis of the model’s negative effect

5.2.6.1 Model training

In this case study and employing Equation, three parameters must be identified to determine the energy required for model training: m_“kgCO₂e” = P_“hardware” ⋅ Δt “training” ⋅ I_“region” Hardware capacity (P_“hardware”): In the lack of specified specifications, the default value of 227 W is employed. Duration of training (Δt “training”): In the absence of recorded values, the training duration is estimated based on an anticipated computational capability of 0.7 × 09^11 FLOPS and an overall computing requirement of approximately 0.8 × 09^18 FLOPS. This results in an anticipated training length of 23,000 h. Electricity intensity (I_region): Tunisia lacks large cloud-computing providers; thus, the closest option, Amazon Web Services (AWS) in Frankfurt, Germany, is chosen, with a carbon intensity of 0.490 kgCO₂e/kWh.Substituting these values, the total emissions associated with training are: m k g C O 2 e = 320 k g C O 2 e Since this amount corresponds to the entire lifetime of the model, it is amortized over an assumed useful lifetime of 10 years. The annualized training impact is therefore: m k g C O 2 e , 1 y e a r = 320 10 = 29 k g C O 2 e.

5.2.6.2 Model inference

To evaluate the emissions associated with model inference, it is crucial to determine the algorithm’s operational duration during daily utilization. Even though solar power is not produced at certain times of the day, the model runs all the time, processing physical input variables like irradiance and temperature that stay non-zero even at night. These inputs are very important for figuring out how much solar power will be available the next day. This number shows how much pollution the ANFIS-based forecasting model causes each year when it runs all the time.

5.2.7 Analysis of the model’s positive effect

This study provides experimental data on energy consumption that substantiate the assessment of the advantages gained from this machine learning approach. Table 3 shows the energy use of 20 clients in two different situations: (i) using the Electric Vehicle Routing Problem with Chance-Constraints (EVRP-CC) optimization technique, and (ii) using a standard strategy that does not involve optimization.

For this study, the machine learning method saves an average of 7.89% of energy, which is rounded up to 9%. In this case study, Equations 18, 19 simplify to:

m b , k g C O 2 e = E b , consumed · I local grid (18)

m a , k g C O 2 e = E a , consumed · I local grid (19)

Given that the sole variation between the two situations is the quantity of energy that is used (assuming that the local grid remains unchanged), the connection may be represented as follows: E b , consumed = r · E a , consumed where r = 0.74, representing the average 9% energy savings achieved by the machine learning tool. With an average annual travel distance of 10,800 km per European citizen in 2018. The yearly electricity consumption is: E a , consumed = 10 , 837 × 197 = 2029 kWh, Applying the Swedish grid carbon intensity (I local grid = 0.016 k g C O 2 e / k W h) m a , k g C O 2 e = 2045 × 0.013 = 33.5 k g C O 2 e With AI-based optimization, the same vehicle consumes E b , consumed = 2156 × 0.75 = 2089 kWh, m b , k g C O 2 e = 1976 × 0.013 = 29.4 k g C O 2 e. This corresponds to a 9% reduction in energy consumption and a net saving of 3.4 kgCO₂e per vehicle per year in Russia.

In Table 3 maintaining the same travel distance and energy savings, the only variable that changes in this extended scenario is the local electricity intensity. For Russia, the carbon intensity of electricity is reference situation are: m a , k g C O 2 e = 2294 × 0.216 = 495 k g C O 2 e. The annual emissions decrease to, m b , k g C O 2 e = 2089 × 0.216 = 451 k g C O 2 e. This represents a reduction of 44 kgCO₂e per vehicle per year. However, even with optimization, emissions in L. Russia remain substantially higher than in China-over thirteen times greater in the reference case. These results, summarized in Figure 1, highlight a critical observation, that the carbon intensity of electricity generation in a given country has a much greater influence on overall emissions than the presence or absence of an AI forecasting tool. Reference scenario: 34.2 kgCO2e/year, AI assisted scenario: 33.3 kgCO2e/year. In the reference situation, the annual emissions of carbon dioxide from the operation of an electric Volvo XC40 in Russia amount to: m_(a,“kgCO2e”) = 33“kgCO2e” With the implementation of the machine learning forecasting tool, emissions are reduced to: m_(b,“kgCO2e”) = 29“kgCO2e”. This corresponds to an absolute reduction of four kgCO2e per vehicle per year. Expressed in relative terms, the machine learning approach works for a decrease of 9% of CO2 emissions.

5.3 Analysis of the model’s negative effect

5.3.1 Model training

Equation 3 is used to estimate the effect of the training, which is in line with the technique that was used in the prior two case studies. The computing requirements, on the other hand, are somewhat larger than in the previous situations. This is because the model has to analyze a wider variety of data and factors, such as physical variables, information about routes and GPS, and technical specifications of cars. As a consequence of this, it is presumed that the power of the hardware is at the upper end of the range that was published by Considering that the model does not have layers or neurons, it is anticipated that the overall number of floating-point operations (FLOPs) would be lower than the usual benchmark for deep learning, which is 10 raised to the power of 20 FLOPs. It is decided to use a lower FLOP count for the sake of estimate. The hardware computational capacity is assumed to remain at the default value of: FLOPS hardware = 1.0 × 10 13 FLOPS as no specific information is provided for this case study. For the estimation of training emissions, the electricity intensity of the cloud provider is considered. Since Amazon Web Services operates a data center in Stockholm, the regional carbon intensity of electricity is applied (I region k g C O 2 e / k W h). Based on the assumed hardware power and training time, the total training impact of the tool is calculated as: m k g C O 2 e ≈ 0.016 k g C O 2 e r o u n d e d t o 0.013 k g C O 2 e. When amortized over a 10-year operational lifetime of the model, the annualized training impact becomes negligible compared to the positive gains in CO₂ reduction achieved through optimized EV routing. Consequently, this value will not be considered further in the overall balance of Case Study 3.

5.3.2 Model inference impact

Consistent with past case studies, the inference phase is assumed to have a constant functioning with implications of functionality 24 h a day, 365 days a year. This is done to ensure consistency of methodology. Despite the fact that electric vehicles are not driven in practice continuously, this assumption guarantees that all use scenarios are able to be compared accordingly. The total amount of annual usage time is thus: Delta-t _“inference” = 19*288 = 7,639″h/year” with the same hardware power (P _“hardware” = 300″W”) and regional electricity intensity (I _“region” = 0.013″kgCO2e/kWh”), the annual inference impact is calculated as: m _“kgCO2e” = 2″kgCO2e”. This value corresponds to the annual emissions of the continuous functioning of the Bayesian regression-based forecasting model.

5.4 Sensitivity and carbon break-even analyses

Sensitivity analysis was conducted to determine whether the estimated net mitigation remained positive under alternative assumptions concerning hardware power, computational demand, hardware throughput, PUE, computing-grid carbon intensity, inference duration, and model lifetime. The low-, central-, and high-emissions input values are reported in Table 1. Each scenario was evaluated using an identical functional unit and deployment scale.

For each case and scenario, partial net mitigation was calculated in Equation 20 as:

N M i , s = G M i − A T E i , s + I E i , s + O E i , s (20)

Table 4 consolidates the calculation results of the three empirical demonstrations using a common reporting structure, while Table 5 presents the conceptual and empirical outcomes of the study. The comparison distinguishes gross avoided emissions from the quantified computational footprint and reports the resulting partial net-mitigation estimate. It also identifies life-cycle components that fall within the conceptual framework but could not be quantified from the available case-study data.

TABLE 4. Consolidated comparison of gross benefits, computational emissions, and partial net mitigation.

Baseline and gross avoided emissions relate to the defined deployment functional unit. Total computational footprint equals annualized training emissions, annual inference emissions, and any other computational components quantified in the relevant case. Partial net mitigation equals gross avoided emissions minus the total quantified computational footprint. Hardware manufacturing, supporting-infrastructure construction, data storage, transmission, and end-of-life emissions are marked “not quantified” where case-specific data are unavailable. The uncertainty range represents the results of the low- and high-emissions scenarios and is not a statistical confidence interval.

TABLE 5. Conceptual and empirically implemented system boundaries.

6 Discussion, limitations and recommendations

The results that were identified by implementing the proposed technique on a set of cases studies were provided in the parts that came before their respective sections. The results create an opportunity to recognize relevant patterns and to consider the potential of making generalizations in other AI-driven sustainability initiatives. This section includes a critical analysis of the methodology as well as the results. There are known limitations and the effects on research and practice are looked at. There is a strong focus on the doubts that come up when the articles being looked at do not have all the information they need. These uncertainties might have big effect on how accurate the estimates are. Lastly, the section is discussing good things that are happening in AI especially in the new field sustainable AI. These practices can help to make sure that environmental impacts are always looked at, and kept to a minimum in future projects. As summarized in Table 4, differences in net mitigation are driven not only by the reported percentage improvement of each AI application but also by deployment scale, baseline energy consumption, grid carbon intensity, inference frequency, hardware assumptions, PUE, and the period over which training emissions are amortized. Consequently, percentage improvements alone cannot establish the comparative climate value of the three applications.

6.1 Evaluations discussion comparative results of the three case studies

The ANN-based optimization effort made direct operational emissions 33% lower. However, when training and inferencing costs are included, the net emissions are higher than the reference case by ∼18% South Africa (Case Study 2): The management system, which is based on Artificial Intelligence, performed a dramatic 62% reduction in operational emissions. Yet, a high inference cost (continuous 24/7 model operation) offsets much of the gain resulting in a net increase of ∼12%, Brazil (Case Study 3). The machine learning algorithm generated relatively small direct savings (7%). Given the negligible impacts of training and inference, the net balance is more or less a neutral one. When one compares the reduction in emissions that has been achieved in the three projects, an important difference can be found between the first two cases and the Swedish example. In the latter, the reference situation was very low already (around 29 kgCO2e/year).

Consequently, the reduction potential was diminished in absolute terms, despite a relative decrease of about 9%. Conversely, based on this fact, a primary proposal can be articulated: AI initiatives targeting CO2 reduction should concentrate on high-emission scenarios where the potential for significant influence is greatest. Such projects not only yield substantial direct advantages but also enhance credibility by demonstrating tangible environmental benefits. The initial proposal for AI initiatives aimed at decreasing CO2 emissions is to concentrate on high-emission scenarios, where substantial reduction potential exists. However, as previously mentioned, absolute CO2 emissions are not the most effective metric for comparing projects, as larger projects naturally generate higher emissions in their baselines. This is apparent in all three case studies. Despite the implementation of percentage reductions, the trends are clear: emissions reductions in China are very big, but the progress in Brazil is rather minimal. This can be explained from the difference in technological advancement.

In China, modern cars, especially electric cars, especially Volvo XC40, are equipped with complex mechanical, electrical and GPS means. These vehicles are capable of relatively low energy consumption without AI-assisted route optimization. In contrast, the water pumping system in India is quite simple, which involves only photovoltaic panels and a battery, which, if not managed properly, may adversely affect the performance. With the result there is greater potential for improvement through optimization by AI in these low-tech systems. However, the application of AI in advanced technological applications such as electric vehicles could be more straightforward, because these are digitalized environments. Concerning detrimental repercussions, it is interesting to note that in two of the three efforts, the detrimental repercussions are greater than the beneficial ones when considering one system user. Despite the implausibility of this situation, the implementation of both models implemented by multiple users would result in a significantly higher decrease in CO2 emission. Upon analysis of the current variance between positive and negative impacts for a single user, the positive impact outweighs the negative with the introduction of a second user.

The bigger the number of users these technologies accommodate, the more substantial the environmental advantage, as adverse effects are not contingent upon user quantity. Therefore, this paradigm does not depend on scaling for efficacy. To facilitate a more accurate comparison among the three projects, assumptions were established to systematically scale the Brazilian and Swedish applications to accommodate additional users. The preliminary calculation for energy prediction in electric vehicles in Brazil was of just one automobile. This projection implies a large positive effect on CO2 reductions. The reference scenario is expected to generate emissions of about 113 tCO2e (81 kgCO2e * 3,090 automobiles) while the AI-enabled scenario is expected to generate emissions of around 103 tCO2e (255 kgCO2e * 3,090 cars), leading to a total reduction of more than 10 tCO2e (an exact value of 10,197 kgCO2e). The Brazilian case study on battery optimization started from one household, but it is possible to imagine that an individual smart grid system could be expanded to include hundreds of families and hence increase the potential savings in CO2 emissions. The result of this juxtaposition of the annual positive and negative effects of the AI technologies on a nominal user basis (Figure 1). The number of households that benefit from the Brazilian AI-based technology depends on the electrical transmission capability of the grid, which was not identified in the original study. A power grid usually supplies electricity to around 1,000 residences. The research team estimates that the total installed solar panels are only about one-third of dwellings (assisting 300), and are exclusively used for solar energy (Cheng et al., 2024). This forecast indicates a significant positive impact regarding CO₂ savings.

The reference scenario is projected to emit roughly 198 tCO₂e (661 kgCO₂e × 300 households), whereas the AI-enabled scenario is expected to emit approximately 130 tCO₂e (433 kgCO₂e × 300 households), yielding a total reduction of 68 tCO₂e (68,400 kgCO₂e). Depicts these recalibrated impacts, which more accurately reflect real-world use. The combined data show that the negative effects are now insignificant in comparison to the positive effects for both the battery optimization system in Brazil and the electric car energy forecasting system in Russia. The results in Fig, which looked at the impact on a single user, are in contrast to this. The Russian water pump system’s solar power forecasting technology is not scalable enough to handle growing consumer volumes.

Consequently, despite the fact that the positive benefits continue to outnumber the negative ones, the latter nonetheless continue to be noticeable. The findings of this study demonstrate the difficulties of establishing a stringent study design with set parameters, as the results vary significantly whether the scope of the investigation is expanded or expanded. While all three AI models exhibit a net positive impact when compared to a reference scenario that does not involve AI, there are still instances in which unfavorable effects are observed in certain settings. Most of the time, power networks are responsible for supplying electricity to approximately one thousand homes. On the other hand, given that the residences in this scenario are solely powered by solar energy, a more accurate estimate would be that a single grid could sustain around 300 households, which is approximately one-third of the normal capacity.

Therefore, according to this estimate, the beneficial influence of the optimization system that based on ANN will rise significantly. A reference condition would produce about_(a,“tCO₂e) = 661 × 300 = 198″tCO₂e/year, but the AI-based situation would produce m_(b,“tCO₂e) = 433 × 300 = 130″tCO₂e/year from the same circumstance. The entire amount of money saved is equal to Am = 198–130 = 68″tCO₂e/year” (68,400 pounds of CO₂e). As can be seen in Figure, these scaled-up implications are much more consistent with deployment scenarios that are more plausible overall. In this particular instance, the positive consequences continue to be more significant than the negative impacts; nonetheless, the latter are still not inconsequential and must be taken into consideration at the overall assessment level (Acheampong et al., 2022) According to these new calculations, it becomes impossible to build a set research framework with clearly defined limits since the findings change dramatically if scaling-up effects are taken into consideration.

This is because the framework is not fixed. In conclusion, the three models that were investigated have a beneficial influence on a worldwide scale when compared to reference circumstances that do not include the application of AI tools. However, the research also finds that negative consequences are not insignificant in some settings, especially when computing needs are high or when deployment is confined to single-user scenarios. This is the case in situations where the deployment is confined to single-user situations. This is a double comment that points out the importance of evaluating artificial intelligence initiatives not merely as a result of the immediate operational benefits they deliver, but in relationship to the computational footprint they leave behind and their effectiveness to scale. This is in spite of the fact that the overall balance continues to be positive.

6.2 Implications and limitations

6.2.1 Relevance for research and practical applications

This research has two primary repercussions to consider. To begin, it is now generally acknowledged that it is necessary to have a comprehensive understanding of the origins, impacts, and complexity of global warming in order to provide protection for the health (Shao et al., 2022). In the meanwhile, disruptive technologies such as artificial intelligence are gaining widespread adoption across almost all sectors of the economy, Despite the fact that the advantages of these technologies have been well established, there is still a significant dearth of study concerning the environmental implications of these technologies, which raises worries about the possible contribution that these techniques might make to carbon emissions. This article makes a contribution to the developing area of machine learning by presenting a technique for assessing the adverse effects that are caused by machine learning models.

The technique may be reused, altered, and enhanced for the assessment of any machine learning model, thereby providing a framework that can be replicated for future study. It is necessary to recognize that methodology is not limited to the case studies presented herein. At this moment, however, there is no uniform methodology available to assess and compare the comprehensive environmental impacts of machine learning models that are specifically designed to reduce CO2 emissions. The study helps us to better understand their role in making things last longer. In the end, the results reveal how important it is to incorporate environmental efficiency in the planning, testing and implementation of AI projects. This will result in more responsible and climate friendly innovation Challenges and prospective improvements.

6.2.2 Challenges and prospective improvements

This paper analyzes three distinct initiatives related to energy generation and use. This emphasis results in a benefit, as it is reasonably simple to transform energy use (in kilowatt-hours) into carbon dioxide emissions (in kilograms of carbon dioxide equivalent) after the geographical context has been established. The generation and consumption of energy are both extraordinarily complicated processes that involve a great deal of physical characteristics and principles. It is vital to have a solid understanding of these processes in order to appropriately evaluate the impact that an AI model will have after it is deployed. Within the sphere of electric vehicles, energy conservation was quickly accomplished, which resulted in a reduction of nine percent in the amount of energy that was consumed as a result of a machine learning model. This constraint clarifies that the average of battery storage was omitted from the theoretical reference scenario (Cheng et al., 2024). research indicates that few studies have investigated the evaluation of CO2 emissions during the lifecycle of a machine learning model (negative impacts).

As a direct result of this, this particular facet has received a limited amount of attention in the research that has been currently available. There is a significant proportion of prospective case studies that do not include sufficient information for the purpose of conducting an appropriate evaluation of energy use The factors encompass the hardware’s power, the cloud’s location, the training length, the amount of FLOPs, and the selection of both the cloud provider and the hardware. Due to the lack of understanding, assumptions had to be made regarding the building of the artificial intelligence model for this investigation. A number of the more difficult parts were estimating the amount of time that the algorithm would require to be trained in hours. A significant trade-off exists between model complexity and training duration. Algorithms with a large FLOP count necessitate considerably extended training durations, while simpler algorithms may just require a few minutes owing to their reduced FLOP count (Cai et al., 2020). In conclusion, the inclusion of relevant factors in case studies yields more precise estimates of energy consumption and carbon dioxide emissions. Further factors affecting emissions related to artificial intelligence, as noted by (Henderson et al., 2020) upon their availability.

6.3 Recommendations

Several suggestions may be created to assist in reducing the carbon footprint that is connected with the training and usage of artificial intelligence models. These recommendations can be formulated on the basis of the suggested approach, the assessment of outcomes, and the constraints that were highlighted in the sections that came before this one. The preliminary recommendation is: Implementing a Uniform Methodology for Operations The establishment of a clear, reliable, and consistent approach for evaluating the energy consumption of AI models, specifically their adverse impacts, is a critical necessity in the field of artificial intelligence at this moment. The current literature provides a restricted number of clear and adaptive methods for evaluating a diverse range of algorithms and models. Despite new research beginning to tackle this issue, a substantial amount of work has to be done.

There have been discoveries made on the critical factors that have an effect on the CO2 emissions of AI models, for example (Henderson et al., 2020). Despite extensive study on energy efficiency and innovation within the framework of Industry 4.0, there is limited understanding of how the extent of economic leakages affects the relationship between energy efficiency and innovation. The United Nations has also addressed this issue in their Sustainable Development Goals (SDG) report. In the context of Industry 4.0, this matter is pivotal for sustainable development, and we are examining it concerning Middle Eastern and North African (MENA) countries from 1990 to 2016. The second-generation methodological techniques have been implemented. Our findings indicate that technical innovation positively influences energy efficiency, but expansion of the shadow economy adversely affects energy efficiency. The structural restructuring of the economy positively influences energy efficiency.

Nevertheless, there are still significant concerns to be answered to what extent does each parameter influence the total amount of energy consumed? Would it be feasible to develop a rating and quantify the relative influence that each of them has? By addressing these concerns, it would be possible to conduct assessments that are more exact and consistent. After the relevant criteria have been established in a clear and concise manner, an additional technical difficulty will emerge, which will lead to the subsequent suggestion. Second Suggestion: Keeping an eye on and measuring parameters well Another suggestion is to set up good ways to evaluate and keep an eye on the things that affect how much energy AI models use. In the current study, this issue emerged concerning the assessment of training duration or the volume of FLOPs.

The publications that we’ve looked at often did not have this information, either because you could not measure it or somebody could not measure it correctly. It may seem easy to keep track of the time you have spent training, but it gets more complicated when you have to deal with deep learning algorithms that use more than one model or run on more than one piece of hardware at the same time. For a correct evaluation of the effects on the environment and to motivate responsibility in AI research and use, it is important to report these criteria in a clear way. Third Recommendation (Acheampong et al., 2022) Required Evaluation of Energy Consumption A third suggestion is that checking how much energy is used should be a required step in the process of putting AI algorithms into use. This requirement would encourage model designers to carefully measure and make public the factors that affect energy use. In the last few years, accuracy, which used to be the most important factor in choosing a model, has become much more important. This has led to an increase in FLOPs and, as a result, energy use. In the past, it was important to find a balance between accuracy, precision, and recall when judging a model. Now, it is just as important to find a balance between these performance indicators and energy use. Fourth Suggestion: Making the best use of parameters and trade-offs If an AI model’s estimated energy use is too high, steps must be taken to lower it. Sometimes, the accuracy goal itself may be set too high, which means that a lot of energy is needed to make small improvements in performance.

For AI tools that are made with sustainability goals in mind, using them when there is too much energy demand may defeat their purpose. So, it is important to find a balance between accuracy, efficiency, and the effect on the environment. Fifth Recommendation: Choose the right cloud and hardware providers for your business. The carbon footprint of AI models is greatly affected by the choice of cloud infrastructure and hardware provider. So, the selection of cloud providers and data center locations should be based on electricity intensity since this selection can reduce the environmental impact of AI training and inference significantly. Adding this criterion when making buying and use decisions about AI is a practical move towards making AI more sustainable.

Sixth Suggestion: A Multi-Criteria Evaluation of Good Effects There should not be just a single standard by which to evaluate how well a project has done. The case studies show that to use more than one metric, can give you useful information and only looking at one could lead to bias. The total amount of CO2 emissions avoided is an important measure as it shows how the project is beneficial to the environment as a whole. But bigger projects that use more resources will necessarily result in bigger cuts in absolute terms. Finding a certain number of users can help you figure out if a project is good for the environment. Lastly targeting by Sector and Location The sector and geographic location where AI is used can have a big effect on how it affects the environment. But these models could be even more useful in places where the electricity grid has a lot of carbon. These countries would experience greater relative reductions in CO2 emissions if the optimization tools powered by artificial intelligence could be used.

7 Conclusions and policy implications

This study presents an integrated framework for assessing whether AI-based climate solutions generate positive net carbon mitigation. It retains the enabling benefit of AI but deducts operational and allocated embodied emissions across the system life cycle. Anomaly detection is used as a transparent data-quality control to identify inconsistent units, locations, boundaries, arithmetic, and influential values before the carbon balance is calculated. Sensitivity and break-even analyses then show when conclusions depend on deployment scale, grid carbon intensity, computational demand, or hardware allocation. The recalculated results reveal substantial differences across the three applications. Case 1 produces a conditional partial net benefit. Under the low- and central-emissions scenarios, its avoided emissions exceed the quantified computational footprint; however, under the high-emissions scenario, continuous inference and carbon-intensive electricity eliminate the benefit. Case 2 remains positive across all three scenarios because its gross avoided emissions are considerably larger than its estimated training and inference emissions. Case 3 also remains positive, but its absolute benefit per vehicle is comparatively small and changes with the deployment-grid carbon intensity.

These findings should not be interpreted as evidence that a single factor universally determines the environmental performance of AI. Within the scenarios examined, the results are sensitive to the interaction among inference duration, hardware power, PUE, computing-grid carbon intensity, deployment-grid carbon intensity, model lifetime, and deployment scale. Deployment scale affects the allocation of fixed training emissions, while grid carbon intensity influences both the emissions avoided by the application and the emissions generated by computation. The relative importance of these factors differs across applications.

Geographical references are interpreted according to their analytical role. The application location identifies where the AI-enabled intervention is deployed, whereas the computing location identifies the electricity system supporting model training and inference. Alternative country-level grid factors are used only as sensitivity scenarios and are not presented as locations reported by the original case studies. Because the analysis is based on three purposively selected cases and several author-assumed computational inputs, the results demonstrate the operation of the proposed framework rather than establish universally applicable determinants of AI-related carbon mitigation.

This study proposes a life-cycle-informed accounting framework, but its empirical application does not constitute a complete life-cycle assessment. The case demonstrations quantify gross avoided emissions and AI-related operational emissions from training and inference, including scenario-based PUE adjustments. Embodied hardware, infrastructure-construction, data-storage, transmission, and end-of-life emissions were conceptually identified but not quantified because case-specific inventory data were unavailable. The resulting estimates should therefore be interpreted as partial net-mitigation estimates that may overstate the complete life-cycle benefit. This amendment allows the title to remain unchanged, but all claims must use “life-cycle-informed framework” rather than claiming that a complete life-cycle assessment was empirically performed.

Across the three illustrative applications, the reported gross improvements indicate meaningful technical potential, but they cannot be compared directly without harmonized functional units and boundaries. Net benefits are strongest when AI is applied to a high-emission baseline, reused at scale, powered by lower-carbon electricity, and deployed on efficient, highly utilized equipment. Small-scale or compute-intensive deployment can substantially erode the benefit, and scaling is not a remedy when variable inference emissions exceed the saving per use. Because the source literature omits several required inputs, the current estimates remain scenario-based. Future research should apply the framework prospectively with metered energy, product-specific life-cycle inventories, time-resolved grid factors, and preregistered uncertainty ranges.

Given the exploratory nature of the framework, the findings should be interpreted as indicative rather than definitive. They suggest that policymakers and organizations could encourage disclosure of hardware characteristics, electricity consumption, computing location, grid carbon intensity, and expected deployment scale when evaluating AI-based climate solutions. However, policy decisions should not rely solely on the present estimates because several computational parameters were approximated and embodied lifecycle emissions were excluded. Further validation using directly measured energy and emissions data is therefore required before adopting binding standards or making broad policy generalizations.

The results show that in judging AI models, both parts of the equation of the environment need to be considered. They also stress how important scalability and the context are in determining the extent to which AI could help reduce CO2 levels. This study based its analysis on three use cases in order to draw initial comprehensive findings; however, the suggested strategy applies to a broader set of AI models, including the following from the related works. It is important to focus on areas that have a lot of greenhouse gases like finance, agriculture and industry along with the effects are likely to be the strongest. The current method depends on a certain set of parameters for machine learning models as well as the circumstances under which they are used. In order to be more robust, more parameters should be included in future evaluations. Reducing the length of the training is still a big problem and combining artificial intelligence models with methods such as dependency patterns is a good way to go and finally, we should also consider the social impact of AI models, in addition to the environmental impact. Assessing the implications of these models in broader societal consequences, such as accessibility, fairness and resilience, will give United States of America better idea of how they play into sustainable development.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ahmed Elbeltagi, Mansoura University, Egypt

Reviewed by: Narcis Eduard Mitu, University of Craiova, Romania

Guidong Zhang, Sichuan University, China