An Approach to DC-DC Converter Optimization using Machine Learning-based Component Models

This paper presents a novel approach to power converter design and optimization, where the optimal component characteristics for a DC-DC converter architecture are selected given a specified objective function and various design constraints. Extensive amounts of component data available on commercial distributor sites are used to train supervised regression Machine Learning (ML) models used throughout the optimization algorithm. Using ML-based techniques allows the data-based optimization task to become tractable over a large design space. The developed tool can be used to compare performance measures across converter topologies or component technologies and to predict performance contributions of indi-vidual components. As an example, optimized designs of 48-to-12 V, 5 A buck converters based on GaN and Si MOSFETs are compared in terms of losses, size and cost.