Test pattern clustering for fast and accurate lithography modeling

Abstract

Lithography model is a basis of simulating lithography processes including light exposure and photoresist development. It is an empirical model, so test patterns are used to calibrate a number of model parameters. A key problem in calibration process is test pattern clustering, which we address in this paper: test patterns are clustered and a minimum number of representative patterns are chosen while model accuracy is kept as high as possible. Our method consists of two components: (1) each pattern is represented by image parameter set (IPS) values to reflect light exposure effect, and a few principal component analysis (PCA) components out of a series of convolution values between Gaussian kernels and aerial image to reflect photoresist development, and (2) each pattern is associated with two gauges (in horizontal and vertical direction) where CD measurement is performed

a unique problem of gauge clustering with a goal of minimizing the number of chosen patterns is identified and efficient clustering algorithm is presented. Experiments with 10nm contact patterns demonstrate that the proposed method yields 31% smaller number of test patterns yet CD errors are reduced by 55%, compared to the popular method of using IPS values for clustering.