PDN impedance prediction using cross-stage densely connected network based on attention augmentation
Zhonghong Ou, Xinyuan Xiang, Shufang Li, Xiaohe Chen · 2024
Traditional full-wave electromagnetic simulation is inefficient in modeling and simulating the impedance of power distribution network (PDN) for printed circuit board (PCB) with multiple layers of stacking. This paper utilizes convolutional neural networks (CNN) to model the PDN networks of different PCBs and proposes an attention-enhanced cross-stage densely connected network (CSDNet) for PDN impedance prediction. Compared to DenseNet, CSDNet improves the accuracy of PDN impedance prediction by 41.5% and speeds up training by 12.2%. This paper innovatively introduces transfer learning (TL) into PDN impedance prediction, performing transfer training on a small amount of PDN impedance data from different PCBs. Compared to direct training, the accuracy of TL is improved by $\mathbf{6 6. 7 \%}$, and the amount of labeled data required for training is reduced by 80%. Additionally, CSDNet takes 2.3 seconds on average to predict the PDN impedance of 100 sets of 6-layer stacked PCBs, which is $\mathbf{1 3 2 0}$ times faster than traditional full-wave electromagnetic simulation.