Carbon-Aware AI: Reducing Energy Consumption of Machine Learning Models

This paper has presented a comprehensive framework for carbon-aware AI, grounding the discussion in mathematical formulations of energy and carbon metrics, a structured review of efficiency techniques, and simulated experimental results comparing ResNet-50, MobileNetV2, and EfficientNet-B0 on CIFAR-10. The results demonstrate that lightweight architectures can reduce training energy consumption by 72–78% relative to the conventional ResNet-50 baseline while incurring accuracy penalties of less than 2 percentage points—a trade-off that is strongly favorable in the vast majority of practical deployment scenarios. The four-stage Carbon-Aware AI Pipeline introduced in this paper provides a practical operational framework for integrating environmental considerations into the full ML development lifecycle, from initial architecture selection through production scheduling. The mathematical formulations E = P × T and CO₂ = E × CI × PUE provide a standardized basis for reproducible environmental reporting. Several important research directions remain to be pursued, including extending experimental comparisons to NLP and object detection tasks, integrating automated carbon tracking tools into training pipelines, and systematically investigating the cumulative efficiency gains achievable through technique stacking. As generative AI continues to scale, the inference energy of autoregressive large language models deserves targeted study. The trajectory of AI development makes the pursuit of carbon-aware design principles an increasingly urgent responsibility for the research community.

1.  E. Strubell, A. Ganesh, and A. McCallum, "Energy and Policy Considerations for Deep Learning in NLP," in Proc. 57th Annual Meeting of the ACL, Florence, Italy, 2019, pp. 3645–3650. 2.  International Energy Agency, "Data Centres and Data Transmission Networks," IEA, Paris, France, Tech. Rep., 2022. 3.  R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, "Green AI," Commun. ACM, vol. 63, no. 12, pp. 54–63, Dec. 2020. 4.  D. Patterson et al., "Carbon Emissions and Large Neural Network Training," arXiv preprint arXiv:2104.10350, 2021. 5.  A. Canziani, A. Paszke, and E. Culurciello, "An Analysis of Deep Neural Network Models for Practical Applications," arXiv preprint arXiv:1605.07678, 2017. 6.  A. G. Howard et al., "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications," arXiv preprint arXiv:1704.04861, 2017. 7.  M. Tan and Q. V. Le, "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," in Proc. ICML, Long Beach, CA, USA, 2019, pp. 6105–6114. 8.  P. Warden and D. Situnayake, TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers. Sebastopol, CA, USA: O'Reilly Media, 2020. 9.  S. Han, J. Pool, J. Tran, and W. J. Dally, "Learning Both Weights and Connections for Efficient Neural Networks," in Advances in Neural Information Processing Systems (NeurIPS), vol. 28, Montreal, Canada, 2015. 10.  G. Hinton, O. Vinyals, and J. Dean, "Distilling the Knowledge in a Neural Network," arXiv preprint arXiv:1503.02531, 2015.