Compressed SPICE-Compliant IC Models via Machine Learning Kernel Regression

This paper introduces a fully behavioral machine learning methodology for generating compact and accurate models of IC buffers. The proposed approach leverages a vector-valued implementation of the kernel Ridge regression to construct models based on observations of device transient responses recorded during normal operation. A key focus is placed on developing an efficient compression scheme to minimize model complexity (i.e., the number of regression coefficients), resulting in a compact mathematical representation that can be efficiently integrated into any SPICE-based solver.