Learning Compressed Parametric Dynamic Models for High-Speed Electronic Systems via Nyström Kernel Ridge Regression

This paper proposes a compressed parametric dynamic modeling framework for high-speed electronic systems based on Kernel Ridge Regression. The approach combines a NARX representation with a separable kernel that accounts for both dynamic input-output histories and external parameter variations. A Nystrom-compressed formulation is introduced to handle the large training datasets generated by transient simulations while limiting the number of model coefficients. The method is assessed on three signal-integrity examples, ranging from IC-buffer modeling to high-speed interconnects with linear and non-linear terminations with multiple geometrical parameters. The results demonstrate accurate transient and eye-diagram prediction with reduced training time and model complexity.