CNN-ALSTM Soft-Sensing Model for Resistivity Gradient in CZ Monocrystalline Silicon Wafers

Abstract

This study addresses the challenge of real-time resistivity gradient measurement in the Czochralski (CZ) silicon production process. Due to the inability to directly measure this parameter, we propose a CNN-ALSTM soft-sensing model that enhances traditional LSTM by integrating convolutional neural networks (CNN) and an attention mechanism to overcome time lag variations during silicon pulling. The CNN module extracts spatial features from multi-source sensor data, while the attention-enhanced LSTM (ALSTM) dynamically adjusts historical parameter weights, enabling accurate resistivity gradient prediction. Experiments with real production data show that CNN-ALSTM outperforms SVR, FNN, and RNN, improving prediction accuracy by 11.76%, 16.67%, and 21.05%, respectively. This soft-sensing approach enhances real-time monitoring and optimization of monocrystalline silicon growth.