Power converter design is a complicated, time-consuming process that includes many steps, including the development and applications of steady-state equations, parameter solutions, component selection, printed circuit board design and fabrication, system assembly and packaging, and experimental validation. A common objective in the design of switched-mode power converters is to meet specifications while minimizing losses, size, cost, or a combination of these performance metrics. The incorporation of data-driven design automation is a relatively new way of thinking about the converter design process. This work presents a machine learning (ML)-based, data-driven approach to the component selection step of power converter design. Large quantities of component data scraped from distributor sites and component datasheets are the foundation of the approach. The data is used to train machine learning models and then uses these models within an optimization algorithm. The desired outcome of this work is to develop a systematic way of selecting components to optimize an objective function while meeting design constraints, ultimately assisting power converter designers by significantly reducing the design time and costs by providing automated component recommendations to generate high-performing converters. The approach is developed into an easy-to-use tool that allows for rapid iteration of varying design specifications.

This work covers five major topics involved in the approach. First, the acquisition of large quantities of machine-readable component data from real, commercially available components. Second, the development of datasheet-based, design-oriented parameters for use in the power loss model of a converter topology. The goal is to develop an accurate power loss model that can be computed using only datasheet data and known quantities about the design. Third, the development of machine-learning models trained on component data. These models are used within the approach to quickly come up with proxy component parameter sets for every component in the design at each iteration of the optimization algorithm. Fourth, the development of an approach which is based on an optimization algorithm in order to solve the presented multivariate combinatorial optimization problem of minimizing an objective function subject to constraints and bounds. These optimized parameters are used to reduce the component databases such that they can be used to return sets of specific manufacturer part numbers.

The entire approach is implemented via a generalized, easy-to-use Python-based tool. Once the designer inputs their power loss model and design specifications, the tool quickly determines an optimal component combination and returns a set of manufacturer part numbers available on commercial distributor sites. Finally, the approach is validated using hardware prototypes, to compare the theoretical and experimental performance given the returned parameters. Multiple converter designs are selected for experimental validation. One design is of 48-to-12V, 5A synchronous buck converters, built with the objective of minimizing total power loss subject to constraints on the board area and a variety of component cost constraints, using either Si or GaN FETs for the switches. Additional design examples include a 12-to-48V, 2A output boost converter and a 430W photovoltaic microinverter. The experimental results show that the presented approach can be used to successfully select components for high-performance converters that achieve design constraints.