Performance modeling of analog integrated circuits using least-squares support vector machines

Abstract

This paper describes the application of least-squares support vector machine (LS-SVM) training to analog circuit performance modeling as needed for accelerated or hierarchical analog circuit synthesis. The training is a type of regression, where a function of a special form is fit to experimental performance data derived from analog circuit simulations. The method is contrasted with a feasibility model approach based on the more traditional use of SVMs, namely classification. A design of experiments (DOE) strategy is reviewed which forms the basis of an efficient simulation sampling scheme. The results of our functional regression are then compared to two other DOE-based fitting schemes: a simple linear least-squares regression and a regression using posynomial models. The LS-SVM fitting has advantages over these approaches in terms of accuracy of fit to measured data, prediction of intermediate data points and reduction of free model tuning parameters.