Machine Learning-Enabled Power Converter Optimization for Next-Generation Data Centers: A Review

Data centers are serving as the backbone for digital infrastructure, with advances in cloud-based services and virtualization enabling the move towards next-generation data centers. However, the rapid growth in the number of newly established as well as expanding data centers, the raising energy demand and associated cost become significantly important. For the same, the effectiveness and performance of data centers and their dependence on power converters become crucial. The adaptability, efficiency, and scalability of conventional power converter designs in data centers are constrained, demanding innovative approaches to improve their performance. This article highlights the limitations of current power converter designs for next-generation data centers, emphasizing the requirement for adaptable and scalable topologies through the utilization of machine learning-based techniques and optimization schemes. We also compare the various approaches, including model-based methods, conventional control schemes, and the utilization of wide-bandgap power electronic devices, highlighting key issues, challenges, existing solutions, implications, and future research directions, based on their strengths and constraints. The evaluation reveals that while conventional designs are rigid and not scalable, redesigning and going through a complex and manual iterative process is incredibly challenging and yet not completely optimized. This paper further discusses the potential and effectiveness of machine learning-based tools and their application to power converter topology design, component section, and dynamic control. We show that ultra-high-efficiency power converter designs can significantly improve the overall power loss of data centers if provided with a robust design, where one possibility is to optimize both the converter topology and control together using machine learning to ensure adaptability to dynamic loads with low power loss and improved overall performance.