Green AI: A Study of Optimization Approaches, Sustainability Metrics, Applications and Emerging Challenges

Abstract

The recent development of artificial intelligence (AI) has brought spectacular improvements in a great variety of spheres but has also caused significant issues related to the high energy usage, the cost of the calculations and the carbon emissions. In this respect, Green AI has become a significant research focus whereby efforts are undertaken to come up with AI systems that are not only accurate and powerful, but also energy-efficient, environmentally sustainable and economically viable. The study begins by defining what Green AI entails and why it is relevant in mitigating the environmental footprint of the current AI systems. It then introduces a taxonomy of Green AI techniques, which are model-level optimization, trainingtime optimization, system and infrastructure optimization and deployment/runtime optimization. Moreover, the survey examines the feasibility of deploying Green AI in various significant fields such as buildings, agriculture, manufacturing, data centers, and supply chains and how sustainability-oriented AI can be used to enhance the operational efficiency of facilities and minimize environmental impact. The paper also talks about the primary evaluation measures and instruments applied in Green AI such as energy usage, carbon footprint, training time and inference latency. Lastly, the survey also lists the main open challenges, including the compromise between accuracy and efficiency, the unstandardized benchmarks and the growing sustainability cost of large-scale AI models, as well as offers promising future directions, including sustainable scheduling, efficient foundation models and adaptive real-time Green AI systems.