A Non-Parametric Approach to Behavioral Device Modeling

Dragoljub (Gagi) Drmanac,  Brendon Bolin,  Li-C. Wang
UCSB


Abstract

This work proposes a non-parametric methodology for quick and effective behavioral macromodeling of complex digital and analog devices. Gaussian Process Regression (GPR) learning algorithms are used to generate simple, robust, and widely applicable time-domain models without specifying device equations or parameters. SPICE simulations expose device dynamics to train behavioral models while exhaustive validation ensures accurate and efficient models are generated. Average speedups of 97X are observed over SPICE simulation while maintaining accurate outputs with 95% confidence intervals.