Data-Driven Virtual Metrology and Retraining Systems for Color Filter Processes of TFT-LCD Manufacturing

Abstract

With the innovations in commercial-electronics products and fierce competition in the global panel industry, most panel manufacturers have adopted the small batch production mode to deal with a wide range of customization. With small batch sizes, productive yield and troubleshooting are considered the top priority, so monitoring capability conducted via virtual metrology (VM) is essentially necessary to satisfy the requirement of process quality. In this paper, the real-world dataset involving the photoresist (PR) spacer heights of color filter (CF) in the array sector of TFT-LCD manufacturing is investigated. In practice, the PR spacer heights can only be measured in an infrequent manner due to the scheduling restriction. Without taking additional sample measurements, how to design a high-accuracy VM model based on small batch sizes warrants urgent research for the TFT-LCD industry. The proposed framework can be divided into two parts. First, two novel distance-measuring methods are proposed, Direct (fully-connected) Neural Network (DNN) and K-means, to allocate the most affiliated positions so that the VM system can utilize these positions to execute metrology prediction on the same product. Next, a modified random forests (RFs) regression model is integrated into the VM system to create an ensemble VM predictor that can handle multiple products of different small batch sizes. Lastly, the VM model is repeatedly evaluated, to check the performance of the re-training procedure when the model needs to be renewed between production maintenance periods.