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

Area of Science:

• Semiconductor Manufacturing
• Process Control Engineering
• Thermal Management

Background:

• High-precision thermal regulation is crucial for semiconductor product quality.
• Challenges include actuator delays, limited bandwidth, and fluid temperature disturbances.
• Existing methods struggle with broadband disturbances and unmodeled dynamics.

Purpose of the Study:

• To develop an effective feedforward compensation strategy for high-precision thermal control.
• To integrate architecture optimization with a physics-informed hybrid prediction model.
• To address actuator transport delays and broadband inlet disturbances.

Main Methods:

• Frequency-domain analysis to optimize temperature fluctuation attenuator (TFA) placement.
• Development of a Physics-CNN-LSTM predictor utilizing residual learning.
• Hybrid approach combining a mechanism model with a deep learning network for residual correction.

Main Results:

• The hybrid model achieved a Root Mean Square Error (RMSE) of 3.56×10-5 K, a 51.8% reduction compared to standard LSTM.
• Demonstrated strong robustness against disturbance frequency shifts (R2>0.996 on unseen data).
• TFA placement upstream of the heater effectively filtered mid-to-high-frequency disturbances.

Conclusions:

• The proposed physics-informed hybrid prediction model significantly enhances temperature stability in semiconductor process equipment.
• The integrated solution effectively compensates for actuation delays and inlet disturbances.
• This approach contributes to improved product quality through advanced thermal control.