Efficient Parameter Reduction for Statistical Behavioral Modeling

Statistical verification is crucial for high-quality circuits in the presence of variations. To reduce the computational effort for large-scale circuits, individual subblocks are often abstracted by behavioral models. The computational cost to build and evaluate these models, as well as their accuracy in predicting the behavior of the physical circuit, depends strongly on the number of statistical parameters that are considered. To cope with this problem, we propose an approach to minimize the number of parameters without sacrificing statistical accuracy and inter-block correlations. To this end, we perform feature selection based on Spearman correlation in combination with the DBSCAN algorithm.