Multi-selection method for physical design verification applications

Abstract

In this paper we present a modular approach which combines model based verification, pattern matching and machine learning methods in order to achieve a high accuracy over computing time ratio. We utilize pattern recognition technique using a supervised machine learning system (as opposed to pattern matching) to classify the patterns either as failures (hotspots) or non-failures, and we use pattern matching to detect all the outlier misses and false detections in each of the regions (based on the calibration set), which will be added or removed from the set of hotspots later on. Doing so allows us to do two things: Reduce the number of patterns that need to be pattern matched since only the outliers of the machine learning system need to be considered and more importantly it allows us to add trained predictability to new configurations that were not in the training set but that can be interpolated from the system. The results indicate that indeed it is possible to successfully combine Machine learning with pattern matching methods in order to achieve better predictability of errors of previously unseen data, while being exact in the treatment of previously observed data. We also explore possible avenues to further speed up the computation of the layout characterization process by inserting a global density grid, and assess the impact of model quality and aliasing under real detection conditions.