Currently machine learning tools are not capable to provide analysis solutions for complex printed circuit boards. It is unknown how to prepare the data and how to determine the optimal architecture of the machine learning process. We show that both collaborative and extended artiﬁcial neural networks (ANNs) are capable to compensate drops in accuracies for predicting target impedance violations in an extended design space. It is proven that the extended ANN has the advantage of requiring less samples during the training process compared with the collaborative approach. The necessity of either approach is highly depending on the design space and the inﬂuence of the variation on the power delivery network.
Submitted to: SPI 2020 - 24th IEEE Workshop on Signal and Power Integrity