Deep reinforcement learning for energy-efficient thermal management in 2U air-cooled server systems

Reinforcement learning offers a promising path for reducing the energy footprint of server cooling systems. This study develops a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework for the thermal management of a 2U air-cooled server. By assigning an independent agent to each fan and employing a centralized critic, the framework learns cooperative control strategies that eliminate redundant cooling. The agents’ learning is guided by a novel physics-informed reward function that divides the server’s thermal headroom into distinct operational zones, adding penalties to mitigate fan vibrations while dynamically balancing energy efficiency and thermal safety. To validate generalization, the MADDPG algorithm is trained in a simulation environment and subsequently deployed on experimental mock-up servers. A total of five configurations and power maps are used for validation. Each fan agent relies solely on local temperatures of its state space, while the centralized critic receives the global state of the server during training to penalize redundant cooling actions. The MADDPG controller reduced fan energy consumption by an average of 31.4 % compared to a conventional fan-table controller, while maintaining all component temperatures below their critical thresholds. The results also revealed that performance is highly dependent on server layout, with energy savings ranging from 43.8 % in centrally-located CPU configurations to 20.5 % when CPUs are at the chassis extremes, highlighting the importance of hardware-aware control policies.