PaperlayerView exact revision on arXiv

Source-linked AI summary

A Systematic Review of Green AI

Roberto Verdecchia, June Sallou, Luís Cruz

arXiv:2301.11047v3cs.AI

TL;DR

AI’s growing carbon footprint has prompted Green AI research, but a comprehensive characterization of the field was missing. This systematic review analyzes 98 peer-reviewed publications and finds substantial growth since 2020, promising energy savings, and a need to connect laboratory findings with industrial practice.

  • Problem

    AI’s carbon footprint has become significant, motivating Green AI research, while a comprehensive characterization of this research body was missing.

  • Method

    The paper systematically reviews Green AI literature using selection criteria and recursive bidirectional snowballing to identify and characterize primary studies.

  • Results

    98 peer-reviewed publications show significant growth since 2020, while most Green AI publications report energy savings, reaching up to 115%.

  • Takeaways & Limitations

    The review indicates that Green AI is promising and that more field experiments are needed to support practitioners’ adoption of green strategies.

  • Takeaways & Limitations

    The review follows an a priori protocol and standard systematic-review guidelines, with data documented for reproducibility and replicability.

Abstract

from arXiv · show

With the ever-growing adoption of AI-based systems, the carbon footprint of AI is no longer negligible. AI researchers and practitioners are therefore urged to hold themselves accountable for the carbon emissions of the AI models they design and use. This led in recent years to the appearance of researches tackling AI environmental sustainability, a field referred to as Green AI. Despite the rapid growth of interest in the topic, a comprehensive overview of Green AI research is to date still missing. To address this gap, in this paper, we present a systematic review of the Green AI literature. From the analysis of 98 primary studies, different patterns emerge. The topic experienced a considerable growth from 2020 onward. Most studies consider monitoring AI model footprint, tuning hyperparameters to improve model sustainability, or benchmarking models. A mix of position papers, observational studies, and solution papers are present. Most papers focus on the training phase, are algorithm-agnostic or study neural networks, and use image data. Laboratory experiments are the most common research strategy. Reported Green AI energy savings go up to 115%, with savings over 50% being rather common. Industrial parties are involved in Green AI studies, albeit most target academic readers. Green AI tool provisioning is scarce. As a conclusion, the Green AI research field results to have reached a considerable level of maturity. Therefore, from this review emerges that the time is suitable to adopt other Green AI research strategies, and port the numerous promising academic results to industrial practice.

1 Introduction

Green AI emerged in response to concerns about AI’s environmental impact, but the field lacked a broad overview. This review systematically characterizes its research and finds substantial growth and promising energy savings.

  • Green AI addresses the carbon footprint of developing and running AI models, motivating research into computationally sustainable AI.
  • AI systems span multiple lifecycle stages and artifacts, requiring sustainability research to address the system lifecycle rather than isolated components.
  • The graphical abstract characterizes Green AI research as solution-oriented, often context- and algorithm-independent, and ready for transfer from academia to industry.
  • The review analyzes Green AI literature to characterize its evolution, topics, approaches, and artifacts.
  • 76% of reviewed papers were published since 2020, while reported energy savings range from 13% to 115%.

2 Methodology

The study follows a systematic literature-review process to identify, select, snowball, and analyze software-centric Green AI research. It uses predefined criteria and a structured data-extraction framework.

  • The review follows established software-engineering systematic literature-review guidelines.
  • The review frames its goal as characterizing Green AI literature from researchers’ and practitioners’ viewpoints in environmental sustainability.
  • The search combines Google Scholar, Scopus, and Web of Science with iterative bidirectional snowballing until theoretical saturation.
  • Studies had to concern AI, environmental sustainability, the environmental sustainability of AI, and the software level.
  • The selection excludes non-English, unavailable, duplicate, secondary or tertiary, non-scientific, and non-study-format publications.
  • Three authors independently screened candidates, reconciled decisions, and extracted study characteristics through exploration and framework-based analysis.
  • The extraction framework records Green AI definition, study type, topic, domain, data type, intended reader, and tool availability.

3 Results

The 98-study review shows rapid growth and a concentration of Green AI research on monitoring, hyperparameter tuning, benchmarking, and deployment. Studies report substantial energy savings across several software-focused interventions.

  • Publication venues: 47 papers appeared in conferences, 39 in journals, and 12 in workshops.
  • Definitions and study types: Energy efficiency is the dominant Green AI definition, appearing in 81 papers, compared with 20 for carbon footprint and 9 for ecological footprint.
  • Definitions and study types: Solution papers are most common, with 51 studies, followed by 35 observational and 12 position papers.
  • Topics: Monitoring, hyperparameter tuning, model benchmarking, deployment, and model comparison are the leading topics, and the top four cover 61% of papers.
  • Precision and energy: Removing redundant neurons can reduce energy without significant accuracy loss, but further removal can substantially reduce accuracy.
  • Approaches: Green AI studies include monitoring tools, energy estimation, data-centric reduction, adaptive inference, algorithm changes, framework comparisons, and emissions analysis.

3.6 Green AI Topics by Study Type

Study types vary by Green AI topic: solution studies generally dominate, while benchmarking and library research are mainly observational. Position papers cluster in the least represented topics.

  • Most topics are dominated by solution studies, followed by observational and position papers.
  • Model benchmarking and libraries differ from the general pattern because they are mostly observational and compare models, libraries, or design decisions.
  • Ethics, policy, and emissions are least represented and are mainly covered by position papers; among ten studies, only one is observational and none is solution-based.
  • Among the ten largest topics, only six papers are position studies, whereas the bottom four topics contain ten position papers.

3.7 Domains

Green AI studies usually address AI energy efficiency without targeting a specific application domain. Among domain-specific studies, edge computing is the most recurrent.

  • 3.7 Domains: 58 of 98 papers study AI energy efficiency in a general context rather than a specific domain.
  • 3.7 Domains: 24 of 98 papers address edge computing, making it the most recurrent specific domain.
  • 3.7 Domains: Computer vision appears in 6 papers, while cloud and mobile domains appear in 5 and 4 papers, respectively.
  • 3.7 Domains: Health, autonomous driving, smart cities, human activity, wearables, and embedded systems each appear only once.
  • 3.8 AI Pipeline Phases: The AI pipeline distinguishes training, inference, and all-phase studies, with 49, 17, and 32 papers respectively.

3.9 Considered Artifacts

Green AI research predominantly studies models and neural networks, often using image data and laboratory experiments. Dataset sizes vary widely among studies that report them.

  • 3.9 Considered Artifacts: 63 of 98 papers address the model or associated algorithm, while 24 treat AI systems generally and fewer study data, pipelines, or other artifacts.
  • 3.9 Considered Artifacts: 51 of 98 papers focus on a specific algorithm, including 41 on neural networks, 5 on decision trees, 5 on logistic regression, and 1 on genetic algorithms.
  • 3.9 Considered Artifacts: Among deep neural network studies, 8 focus on convolutional neural networks, while transformers and spiking neural networks each appear once.
  • 3.9 Considered Artifacts: Image data appears in 42 of 98 papers, followed by textual data in 22, numeric data in 10, video in 4, and audio in 2.
  • 3.9 Considered Artifacts: 32 of 98 papers specify no data type, a pattern attributed primarily to position and theoretical papers.
  • 3.9 Considered Artifacts: Among 48 studies reporting dataset size, counts range from 1k to 40M data points, with 25 using thousands and 23 using at least one million.
  • Research Strategies: Laboratory experiments are used by 73 of 98 papers, compared with 6 field experiments and 5 computer simulations.

3.14 Energy savings

Reported Green AI energy savings range from 13% to 115%, although only about one-third of reviewed studies quantify savings. Tool availability and industrial authorship remain limited.

  • 3.14 Energy savings: 115% energy savings is reported for structure simplification in deep neural networks, the highest percentage among reported Green AI strategies.
  • 3.14 Energy savings: 27 of 98 studies explicitly report energy savings, and 17 of those 27 report savings of at least 50%.
  • 3.14 Energy savings: 97% savings are reported for quantized decision-tree inputs, 92% for data-centric techniques, and 91% for virtualized cloud-fog deployment.
  • Intended readers: 85 of 98 studies target academic readers, while 8 target both academic and industrial readers and 5 also address the general public.
  • Green AI Tool Provision: Only 15 of 98 studies make Green AI tools available, despite numerous studies proposing solutions.

4 Discussion

Green AI research has consolidated rapidly, but its evidence base remains concentrated in energy efficiency, training, generic settings, and laboratory experiments. The review therefore identifies clear opportunities to broaden environmental coverage and move promising results toward industrial practice.

  • Publication trend: Green AI has consolidated quickly within AI research communities, especially since 2020.The topic began in 2015 and is increasingly represented in conferences and journals.
  • Definition of Green AI: The review defines Green AI as practices using AI to mitigate human impacts on natural resources and AI’s impacts on the natural environment.This definition encompasses energy consumption, carbon emissions, and broader environmental effects.
  • Transdisciplinary topics: Green AI spans 13 topics, including monitoring, hyperparameter tuning, deployment, libraries, and estimation, requiring transdisciplinary involvement beyond training.Only a few Green AI papers come from software engineering venues.
  • Research coverage: Training dominates the literature, while image data and laboratory experiments remain common, limiting coverage of other lifecycle phases, data types, and real-world settings.The review calls for more heterogeneous data and field experiments or case studies.
  • Industrial practice: More than half of the reviewed papers report energy savings of 50% or more, but the extent of industrial adoption remains unresolved.The review studies the state of the art rather than the state of practice, and industry-targeted studies are relatively scarce.

5 Threats to Validity

The review addresses validity threats through predefined protocols, broad database searches, structured selection criteria, and coordinated researcher checks. These measures target representativeness, internal consistency, construct validity, and conclusion validity.

  • Validity safeguards: The review followed an a priori protocol and systematic-literature-review guidelines to support study quality.The protocol governed data collection and analysis throughout the review.
  • External validity: Three literature indexers and an unbounded publication year were used to reduce the risk of missing relevant studies.The search covered Google Scholar, Scopus, and Web of Science.
  • Internal validity: Weekly meetings among three researchers helped align study selection and mitigate subjective interpretation.The researchers jointly discussed examples and doubts during selection.
  • Construct validity: A priori inclusion and exclusion criteria, supplemented by bidirectional snowballing, were used to broaden and control the study set.These procedures addressed whether the selected studies answered the research questions.
  • Conclusion validity: Data extraction and analysis followed a predefined protocol, with documentation intended to support reproducibility and replicability.The review also followed established systematic-review practices.

6 Related Work

This review positions itself as the first comprehensive characterization of Green AI research, extending beyond earlier reviews focused on intersections, subdomains, or narrow applications. It covers the broader field rather than a single AI or software-engineering context.

  • Review scope: Earlier reviews considered Green AI only marginally, whereas this study presents itself as the first comprehensive review of the field.The authors distinguish their scope from reviews focused on specific intersections or subdomains.
  • Review scope: The review covers 98 primary studies, compared with 41 in a related review restricted to consumer products and services.The broader count reflects inclusion of non-commercial applications and the whole Green AI field.
  • Review scope: Unlike prior reviews limited to deep learning, information retrieval, or embedded systems, this study targets Green AI across AI and software-engineering subdomains.Its scope is not restricted by a specific application or technical subdomain.
  • Review scope: The study examines more Green AI characteristics than a deep-learning review, rather than mapping only the deep-learning lifecycle and artifacts.This distinction supports its broader characterization of the literature.

7 Conclusion

The review analyzes 98 peer-reviewed Green AI publications and finds rapid growth, concentrated topics, substantial reported energy savings, limited field validation, and low industry participation. It concludes that the field is sufficiently mature to prioritize reproducibility and transfer promising academic results into industrial practice.

  • Conclusion: 98 peer-reviewed publications show significant growth in Green AI research since 2020.The review identifies 2020 onward as the period of substantial expansion.
  • Conclusion: Monitoring, hyperparameter tuning, model benchmarking, and deployment are the most prominent among 13 identified Green AI topics.Data-centric, estimation, and emissions topics are less frequent and warrant further research.
  • Conclusion: Energy savings reach up to 115% with little or no accuracy cost across the reviewed publications.The review characterizes these findings as evidence of Green AI’s potential.
  • Conclusion: Only 23% of publications involve industry partners, while most studies remain laboratory-based and require more field experiments.Field validation is framed as necessary for strategies to become effective, feasible, and measurable in practice.
  • Conclusion: The field appears considerably mature, but only a small fraction of solution papers provides reusable tools or software packages.The authors call for reproducible research and transfer of promising academic results into industrial practice.
  • Conclusion: The review establishes a foundation for follow-up grey-literature and interview studies examining how AI professionals address environmental impact.These studies are proposed as future research directions.
Loading 2301.11047v3…