A Physics-Informed Hybrid Neural Network for High-Precision Temperature Prediction in Semiconductor Process Equipment.

High-precision thermal regulation in semiconductor process equipment is critical for product quality, yet it is challenged by actuator transport delays, limited actuator bandwidth due to hardware dynamics, and broadband inlet disturbances in temperature-controlled process fluids. This paper presents a systematic solution integrating architecture optimization with a physics-informed hybrid prediction model to enable effective feedforward compensation. Frequency-domain analysis justifies placing the temperature fluctuation attenuator (TFA) upstream of the heater to filter mid-to-high-frequency disturbances without compromising feedback stability. To address actuation delays, a Physics-CNN-LSTM predictor is developed using a residual learning strategy. This framework employs a mechanism model for baseline estimation and a deep learning network to correct persistent low-frequency residuals caused by unmodeled dynamics. Comparative experiments on industrial data demonstrate that the model achieves a Root Mean Square Error (RMSE) of 3.56×10-5 K under low-to-mid-frequency inlet disturbances, reducing error by approximately 51.8% compared to a standard LSTM. The model also exhibits strong robustness against disturbance frequency shifts (R2>0.996 on unseen data). Furthermore, closed-loop simulations confirm that the proposed feedforward compensation enhances temperature stability in high-precision thermal control.