Ordinary Kriging metamodel-assisted Ant Colony algorithm for fast analog design optimization

Abstract

This paper explores an ordinary Kriging based metamodeling technique that allows designers to create a model of a circuit with very good accuracy, while greatly reducing the time required for simulations. Regression and interpolation based methods have been researched extensively and are a commonly used technique for creating metamodels. However, they do not take into account the effect of correlation between design and process parameters, which are critical in the nanoscale regime. Kriging provides an improved metamodeling technique which takes into effect correlation effects during the metamodel generation phase. The ordinary Kriging metamodels are subjected to an Ant Colony Optimization (ACO) algorithm that enables fast optimization of the circuit. This design methodology is evaluated on a sense amplifier circuit as a case study. The results show that the Kriging based metamodels are very accurate and the ACO based algorithm optimizes the sense amplifier precharge time with power consumption as a design constraint in an average time of 3.7 minutes (optimization on the metamodel), compared to 72 hours (optimization on the SPICE netlist).