Abstract

In this paper, the applicability of transfer learning (TL) combined with artificial neural networks (ANNs) for power integrity analysis of printed circuit boards (PCBs) is investigated. Reusing already existing data samples from a database enables to reduce the amount of data samples required for a new problem setting. Here, more than 30 000 electromagnetic numerical simulations are evaluated of different PCB shapes, geometries, and used materials. The format and processing of the data is adapted at hand, e.g. the plane capacitance is used as one additional input feature. If less than 50 training samples are available the error is reduced by a factor 2 using TL.