A Review of Artificial Intelligence for Data Center Energy Efficiency: Toward Bridging the IT–Cooling Divide
Energies · 11.11.2026 · 19 (22) · 5417
DOI: 10.3390/en19225417 ↗Kurz gefasst
Fachartikel in der Zeitschrift Energies, online erschienen am 11.11.2026. Verfasst von Mouna Nassim und 1 weiteren Autor:innen. Die Arbeit nennt 83 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Data centers consume vast amounts of electricity, and cooling still accounts for a large share of it. Artificial intelligence is increasingly used to reduce this demand, but computing resources and cooling are usually optimized as two separate problems, a separation we call the IT–cooling divide. This review examines how AI is applied to cooling, to Information and Communication Technology (ICT) resources, and to the data center as a whole (Full-DC). We searched Scopus and Web of Science (last searched on 3 August 2026) for journal articles published in 2021–2025 that report quantitative results. We supplemented the database search with a targeted search of ScienceDirect and reference checking. We report the process following PRISMA 2020. As no formal risk-of-bias tool exists for simulation-based engineering studies, the validation level and baseline of each study were recorded, and results were synthesized narratively. The first author selected 45 studies (16 cooling, 21 ICT, 8 Full-DC) on the basis of titles and abstracts and read all of them in full for data extraction. Because selection was made by a single reviewer, the reported shares describe the reviewed sample rather than the whole literature; a reproducible re-screening of all 933 database records, checked double-blind by both authors, indicates that reinforcement learning is over-represented, and metaheuristics under-represented, in this sample. Reinforcement learning is the most frequent primary approach for control and deep learning for prediction. Energy results are heterogeneous: they use different baselines (PID control, heuristics, competing algorithms), are reported at different levels (cooling plant, IT servers, network, or facility), and come almost entirely from simulation, trace-driven, or offline evaluations; no included study reports closed-loop deployment in production. Among the studies reporting an energy reduction in percent, the median value is about 9% for cooling (n = 7) and 13% for ICT (n = 7), and the two Full-DC studies report 16–17%. Only two included studies explicitly co-optimize computing and cooling decisions. Joint IT–cooling optimization therefore remains a promising but thinly tested direction, and progress depends on facility-level reporting, real-world validation, and open datasets. The review was not registered and received no external funding.
Zitieren
APA
Nassim, M., & Bezza, M. (2026). A Review of Artificial Intelligence for Data Center Energy Efficiency: Toward Bridging the IT–Cooling Divide. Energies, 19(22), 5417. https://doi.org/10.3390/en19225417
Vancouver
Nassim M, Bezza M. A Review of Artificial Intelligence for Data Center Energy Efficiency: Toward Bridging the IT–Cooling Divide. Energies. 2026;19(22):5417. doi:10.3390/en19225417
BibTeX
@article{Nassim2026,
author = {Nassim, Mouna and Bezza, Mohamed},
title = {A Review of Artificial Intelligence for Data Center Energy Efficiency: Toward Bridging the IT–Cooling Divide},
journal = {Energies},
year = {2026},
volume = {19},
number = {22},
pages = {5417},
doi = {10.3390/en19225417}
}
EUVIMED