Reinforcement Learning for Data Center Cooling Performance Optimization: A Systematic Review and Gap Analysis for Deployment in Tropical Wet Climates: Article No. 118163

Tropical wet climates represent a rapidly expanding share of global data center capacity, despite these regions being among the most thermodynamically challenging environments for data center operation, characterized by year-round high ambient temperatures and humidity. The applicability of existing research transferability to these climate contexts remains unestablished for data center cooling system performance optimization. This systematic review synthesizes 64 empirical studies published between 2021 and March 2026 on the application of reinforcement learning (RL) to data center cooling optimization, conducted in accordance with the PRISMA 2020 framework. Three research questions are addressed: the RL methods and efficiency outcomes reported across the literature; the climatic representativeness of the evidence base; and the methodological gaps limiting transferability to tropical deployment. Actor-critic off-policy approaches constitute the dominant algorithmic category, with reported energy savings spanning 2-52.7%. Power Usage Effectiveness (PUE) is reported in only 18 of 64 studies, Water Usage Effectiveness (WUE) in only 3, and Carbon Usage Effectiveness (CUE) in only 1, indicating substantive cross-study comparability limitations on energy, water, and carbon efficiency dimensions. A four-tier climate classification of all 64 included studies reveals that only 2 studies confirmed physical deployment under tropical wet ambient conditions consistent with ASHRAE Climate Zone 0A, while 50% of the evidence base derives from demonstrably non-tropical contexts. Five structured research gaps are identified: the absence of real tropical testbed evidence; the under representation of emerging cooling architectures; the absence of humidity control in RL formulations; the predominance of single-objective energy reward functions; and the absence of renewable energy integration in RL cooling control frameworks. These findings indicate that the existing literature does not yet provide a sufficient evidence base for confident RL-based cooling deployment in tropical data center facilities.