Deep learning with slow feature analysis for silicon single crystal growth state identification in Czochralski process

Abstract

The preparation of semiconductor silicon single crystal for advanced chip process requirements is the top priority of the current integrated circuit industry development. In this process, real-time and accurate identification of the growth state of Czochralski silicon single crystals (Cz-SSC) is the key to ensuring stable and improved crystal quality. Due to the switching of operating conditions, environmental disturbances and high dependence on operator experience, the process dynamics are highly complex, and minor anomalies are often masked by normal operating conditions, posing a serious challenge to real-time crystal growth state identification. To solve this problem, this study proposes a Cz-SSC growth state identification method that integrates slow feature analysis (SFA) and deep learning. Firstly, SFA extracts slow features reflecting the nature of process evolution at the time scale to reduce the masking effect of anomalies on minor anomalies from the source; subsequently, a multi-scale one-dimensional convolutional neural network (MS-1DCNN) is designed and a cross-attention mechanism is introduced for features extracted from convolutional kernels of various scales to achieve the weighted fusion of cross-scalar information, thus comprehensively capturing the discriminative patterns at different time scales. Finally, the experimental results demonstrate that the proposed method achieves superior performance in crystal growth state recognition, outperforming other approaches in terms of overall accuracy, recall, and F1-score. This Cz-SSC growth state identification method provides an effective solution for fine control of the semiconductor SSC manufacturing process.