A novel quality prognostics scheme (QPS) for plasma sputtering in TFT-LCD manufacturing processes is proposed. The QPS consists of a conjecture model and a prediction model. The conjecture model can use processing parameters and sensor data to estimate the processing quality (sputtering thickness) in real time. This conjecture function is also called virtual metrology. On the other hand, the prediction model is capable of predicting the processing quality of the next-lot glasses. Neural networks and weighted moving average algorithms are applied to construct the QPS. In particular, a reliance index is developed such that online evaluation of whether the conjecture results of the QPS are trustworthy or not is possible. For increasing the accuracy of the QPS, a self-searching mechanism is designed to automatically search the best set of parameters and functions used by the conjecture and prediction algorithms for cases that the processing properties vary or the recipe changes. Also, an auto-adjusting mechanism is developed for tuning the system parameters of the QPS and bringing the conjecture accuracy within an acceptable bound. Thorough tests using normal and abnormal processing data on one set of plasma sputtering equipment in a TFT-LCD plant show that the values of the mean absolute percentage error of both the conjecture and the prediction results are less than 2%, which validates the effectiveness of the proposed QPS.
All Science Journal Classification (ASJC) codes
- Electronic, Optical and Magnetic Materials
- Condensed Matter Physics
- Industrial and Manufacturing Engineering
- Electrical and Electronic Engineering