Virtual Metrology Modeling for CVD Film Thickness With Lasso-Gaussian Process Regression

In order to achieve statistical process control in high volume semiconductor manufacturing, it is of great significance to detect the film thickness of wafers and further achieve early warning of abnormal conditions. Virtual metrology (VM), based on the production parameters, has been utilized for process monitoring and control for the last decades. It can realize the real-time online prediction of measurement value, with the existing sampling system, to reducing the sampling rate and ultimately compress cycle time. However, most of existing VM strategies for CVD usually use global nonlinear models, which are difficult to catch sudden changes of the fluctuating production process. Moreover, the machine learning methods used in existing VM can only obtain the predicted values, but cannot judge the confidence of the results. This also results in a low accuracy of traditional approaches when dealt with the high-dimensional data with small sizes. Based on the just-in-time learning framework, this paper proposes the Lasso-Gaussian Process Regression(LGPR) and develops an intelligent VM system, which can realize the real-time prediction of film thickness in CVD process. The experiment results show that the LGPR is more stable and higher accurate.