ANN performance for the prediction of high-speed digital interconnects over multiple PCBs

Article Number
9231490
Citation
IEEE 29th Conference on Electrical Performance of Electronic Packaging and Systems (EPEPS 2020)
Contribution to Conference
IEEE 29th Conference on Electrical Performance of Electronic Packaging and Systems, EPEPS 2020 
Publisher DOI
10.1109/EPEPS48591.2020.9231490
Scopus ID
2-s2.0-85096977961
In this paper the performance and the accuracy of artificial neural networks for the prediction of high-speed digital interconnects up to 100 GHz on printed circuit boards are analyzed and evaluated. The prediction accuracy is evaluated both for scattering parameters in frequency domain as well as weighted power sums thereof. The interconnects considered all contain a backplane connected to a daughtercard, showing two via arrays each. Several parameter variations of the basic setup lead to a wide range of possible transmission and crosstalk parameters. Training data sets are obtained using physics-based via modeling up to 100 GHz. Approximately 7000 data sets were made available in total for this study. Neural networks are able to predict the overall link behavior.
Schlagworte
High-Speed Links
Machine Learning
Physics-Based Modeling
Signal Integrity
MLE@TUHH