Bi-LSTM-Based Dynamic Prediction Model for Pulling Speed of Czochralski Single-Crystal Furnace

Abstract

Abstract With the rapid development of microelectronics science and technology, the quality of IC-grade silicon single crystal directly affects the yield and stability of the performance of semiconductor device production. As the main equipment for the preparation of such materials, the monitoring and maintenance of the working condition of the single crystal furnace are crucial. Bi-directional long short-term memory (Bi-LSTM) is an innovative neural network paradigm that is used to predict future occurrences by learning the bi-directional long-term dependencies of time-steps and serial data. This paper built a Bi-LSTM based model that can dynamically predict the pulling speed of a Czochralski (Cz) single-crystal furnace by modeling the time series of operational parameters. The Bi-LSTM model is validated using real data from a silicon single-crystal factory. It is proven that the model achieved higher accuracy than LSTM, ANN, SVR, and XGBOOST. The experimental results verify the validity of modeling the pulling speed of single-crystal furnace devices through the Bi-LSTM model by using the time series of multi-dimensional parameters. Therefore, the Bi-LSTM model can serve as a reference for modeling the parameters of such devices.