Data-Driven Component-Level Design of Switching Power Converters Using Predictive Modeling

This article shows how the design of switching power converters at the component level can be assisted by data-driven, automated, systematic methods. This work uses regression machine learning (ML) techniques, where predictive component models are trained using large amounts of component data and design-oriented parameters derived from the data. For a given converter topology and specification set, the ML-based component models are used to predict converter power loss, bill-of-material (BOM) cost, and footprint area. This allows for an algorithm to iterate over design parameters to optimize the objective function in terms of efficiency, size, cost, or a combination of these performance metrics while also meeting all design constraints. Once the algorithm converges, the optimized parameters are applied to filter the original dataset, identifying components that are most likely to be chosen for the specific design. The developed design automation approach allows for comparisons between topologies, device technologies, and operating modes, and reduces the design time by quickly providing component recommendations. To illustrate an application of the ML-based design approach, the tool is applied to the design of 48-to-12 V, 5 A synchronous buck dc–dc converters using silicon or gallium nitride power semiconductors, where the design objective is to minimize loss while meeting BOM cost and area constraints.