DIF-LSTM: A Dual Information Filtering LSTM Network for V/G Value Prediction in Czochralski Silicon Growth

In Czochralski (CZ) silicon growth, controlling the ratio of crystal growth velocity to axial temperature gradient (V/G) is critical for defect management. However, the V/G value is difficult to measure in real-time. Furthermore, it exhibits strong multivariate coupling and extreme non-stationarity under complex thermal fields. While standard deep learning models like LSTM are used for soft sensing, they often misidentify high-frequency hardware noise as true process dynamics, causing severe error amplification in multi-step predictions. To address this, we propose a Dual Information Filtering LSTM (DIF-LSTM). It utilizes an external context-aware mechanism to screen long-term steady-state redundant information and an internal denoising gate coupled with the LSTM input to explicitly block transient high-frequency noise. Furthermore, a confidence evaluation branch and residual decay fusion ensure stable multi-step forecasting. Experimental results an industrial-scale experimental silicon single crystal furnace show DIF-LSTM achieves superior accuracy, obtaining an R2 of 0.9935 and a Mean Squared Error of 1.60×10−6 at a 3-step horizon. Even at a 9-step horizon, it maintains an R2 of 0.9422, significantly outperforming the baseline IF-LSTM (0.8498).