Abstract
This paper describes a methodology to build a virtual metrology (VM) model for semiconductor chemical mechanical polishing (CMP) process control. The VM model predicts the polishing rate based on equipment-derived data as soon as allowed, and immediately applies the results to advanced process control (APC). The proposed methodology uses Markov chain Monte Carlo (MCMC) methods to build an analytical model with many parameters for individual consumed materials from historical data in small quantities. The mutual interference of two kinds of consumed materials: dresser and pad are modeled in a form of multilevel predictive model. The methodology uses MCMC methods again to identify the multilevel predictive model taking into account the assumed operation of an actual manufacturing line, for instance, using preliminary test result, learning a model parameter online, and being affected by metrology lag as disturbance. The simulation results show the APC with the proposed VM model is low sensitivity to metrology lag and high precision on polishing amount control.
