IC Modeling via Machine Learning Regressions: A Data-Driven Approach to SPICE Integration

This article presents a method for generating accurate and efficient macromodels of high-speed input/output (I/O) buffers. Extending existing techniques, the proposed approach enables a modular and scalable model generation tool based on machine learning. Given the limitations of traditional methods, this work leverages kernel regression to develop SPICE-compliant models. Two compression schemes, random selection and Nyström approximation, are used and thoroughly compared to reduce the number of expansion terms, with beneficial effects in terms of compactness of the SPICE implementation. The effectiveness of the method in terms of model accuracy and efficiency is stressed through real devices and typical signal and power integrity (SIPI) cosimulations.