On Optimal Latin Hypercube Design for Yield Analysis of Analog Circuits

Abstract

Combining Monte Carlo (MC) experiment and Response Surface Modeling is an efficient way to analyze parametric yield as an important reliability measure of analog circuits. Latin hypercube (LH) is widely employed to design experiments as a technique that can decrease the simulation time by reducing the number of MC samples (simulation runs) required to achieve certain accuracy. Although traditional LH design shows good projective properties on any single dimension, optimal LH designs (OLHDs) are needed to provide effective performance in high-dimensional MC experiments, which are usually the case in analog circuits. However, constructing optimal LHDs in high-dimensions is generally problematic and computationally expensive. This paper describes a method to build a class of OLHD using a permutation genetic algorithm (PermGA) based on a chromosome-length-expansion (CLE) scheme. Through circuit examples, it is shown that the proposed CLE technique provides a major improvement in terms of computational effort needed in high-dimensional problems as compared to the common PermGA OLHDs.