Zernike-aware neural networks for accurate wafer map prediction in chemical vapor deposition processes

초록

This study proposes a physics-aware predictive neural network for efficiently optimizing semiconductor fabrication processes. Recognizing Chemical Vapor Deposition (CVD) as a critical process in semiconductor fabrication, accurately and rapidly predicting deviations within CVD is essential for maximizing production yield and device reliability. Traditional multiscale computational fluid dynamics (CFD) models, while detailed, are computationally demanding due to iterative calculations between macro-and microscale interactions, making them unsuitable for real-time optimization. To address this challenge, we introduce a hybrid prediction modeling approach combining a physics-aware initialization module based on Zernike polynomial, which provides primary low-order predictions, with predictive neural networks for high-order refinement. Our physics-aware neural networks achieve meaningful improvements over conventional methods by accurately predicting wafer map profiles across entire wafers with minimal data requirements, effectively capturing complex process conditions. Validated under diverse and challenging process conditions, the proposed neural network model consistently outperforms conventional approaches in both accuracy and training stability, particularly when applied to CVD datasets characterized by limited data availability.

키워드