A Review on AI-Driven Optimization of Data Center Energy Efficiency and Thermal Management

As data centers grow and face energy challenges, traditional thermal management struggles with dynamic loads, multi-scale coupling, and heterogeneous control. This review examines AI-driven solutions for energy efficiency, focusing on integrating deep learning and reinforcement learning. Key innovations include physics-data hybrid models and constrained RL controllers, achieving PUE<1.2, a 55.7% reduction in fan energy, and enhanced thermal stability. Challenges remain in explainable decision-making, hardware compatibility, and the complexity of multi-physics simulation. Our evaluation framework emphasizes PUE and energy savings, advocating for future advancements in digital twins, edge AI deployment, and renewable cooling integration. Policy-supported AI implementation could increase annual energy savings to 8-12%, promoting sustainable digital infrastructure. Future research should explore multi-scale optimization, reliable AI mechanisms, and renewable-cooling coordination to meet dynamic demand and support carbon neutrality goals.