Quality Prediction of PECVD Process with Random Forest and Long Short-Term Memory Neural Network.

Plasma enhanced chemical vapor deposition (PECVD) is a key process for producing the surface thin film of high-energy batteries for that it could enhance the absorption and energy conversion rates of solar cell films. While the PECVD coating process is complex enough to be realized and requires various sensors to collect data as well as monitor the quality throughout the process. This paper first performs correlation analysis on the collected data, examines the relationship between the multi-sensor data, and proposes a hybrid algorithm of random forest and long short-term memory network (LSTM) to predict the relationship between the process parameters and the quality parameters data collected by multiple sensors in PECVD process. In this hybrid algorithm model, the importance of multi-sensor data is extracted by random forest module, and the time series features of the data are analyzed by LSTM. The proposed novel method overcomes the limitations of the existing algorithm models that only consider either temporal features or important features, which are not comprehensive enough for practical applications. The result reveals that the RF-LSTM improves the prediction accuracy by 40% without significantly increasing the iteration time.