Dimensional Reduction by Auto-Encoders in Machine Learning Based Power Integrity Analysis

This study
^1
1
This work is funded by the German federal ministry for economic affairs and climate action of the Federal Republic of Germany under the funding code 19A21006R delves into the application of auto-encoders (AE) to reduce the large dimension of the parameter space for design problems of printed circuit boards (PCB). After dimensional reduction to an adequate sized latent space with controlled information loss, efficient machine learning (ML) methods, such as artificial neural networks (ANN), can support the analysis and the design of PCBs by operating on latent space data. As an example the combination of a trained encoder and a downstream ANN for predicting impedances between various ports of a complex PCB is studied. The decoder of the AE can subsequently be used to remap a latent space representation of PCB data back into the physical space. The efficiency of dimensional reduction due to an AE is compared to low dimensional representation of impedance responses via vector fitting. Finally, SHAP (Shapley Additive Explanations) values are employed to show the significance of individual design parameters on impedance responses. Frameworks linking dimensional reduction by AEs with ANN-based predictive models may provide deeper insights into the complex interactions within PCBs, enable precise predictions of their electrical properties, and, thus, support PCB design.