Convolution and Long Short-Term Memory Neural Network for PECVD Process Quality Prediction

Abstract

Plasma Enhanced Chemical Vapor Deposition (PECVD) is the most important process in the manufacturing of solar panels, and the deposition effect determines the energy conversion efficiency of solar panels. In this process, multi-sensor data is collected by the system for quality prediction and control. Existing research on quality prediction only pays attention to one of temporal characteristics and spatial characteristics but ignores the other. The multi-sensor data collected in the PECVD process has the characteristics of time series data, and the data collected by different sensors is also relevant, so it is necessary to integrate the temporal characteristics and the spatial characteristics. This paper proposes a hybrid prediction network model and uses Convolutional-LSTM Network to establish a regression model between multi-sensor data and deposition quality, then evaluates the prediction accuracy of the hybrid model and the single model. In the proposed hybrid prediction model, the spatial characteristics of the data are extracted through the convolutional neural network (CNN), and then the temporal characteristics of the data are extracted through the long short-term memory network (LSTM). The results show that compared with existing prediction models, the proposed CNN-LSTM method reduces the prediction error by at least 31% without significantly improving the convergence time.