The continuous demand for high-performance and low-power semiconductor devices has driven aggressive scaling and the transition toward complex three-dimensionalarchitectures. However, these advancements have significantly increased manufacturing compl...
The continuous demand for high-performance and low-power semiconductor devices has driven aggressive scaling and the transition toward complex three-dimensionalarchitectures. However, these advancements have significantly increased manufacturing complexity and reduced the critical dimension (CD) margin, making precise and robust metrology crucial. Among various methods, optical critical dimension (OCD) metrology has become essential due to its non-destructive, high-throughput capability and its ability to analyze buried structures using optical signals. Nevertheless, as device dimensions shrink, the spectral response weakens and the signal contrast between structural variations decreases, resulting in limited resolution and modeling accuracy when using conventional rigorous coupled-wave analysis (RCWA)-based modeling.
To investigate these limitations, the spectral responses of two representative structures—a trench pattern with CD > 50 nm and a fine pattern with CD < 10 nm—were analyzed using principal component analysis (PCA) feature extraction. The results revealed that the spectra of nanoscale structures exhibit complex nonlinear relationships with CD variations, confirming that spectral interpretation and CD reconstruction become increasingly difficult at smaller dimensions.
To overcome these challenges, this study applies machine learning (ML) to predict nanoscale CD values from measured spectra. Two representative regression models—support vector regression (SVR) and deep neural networks (DNN)—were trained using 228 spectra collected from 12 wafers measured after the bit-line etching process of DRAM, with cross-sectional TEM measurements serving as reference CD values. The trained models were then evaluated on six independent blind-test wafers to verify accuracy and generalization. The results show that the blind-test RMSE was reduced from approximately 0.9 nm (conventional OCD) to 0.6 nm with the ML models, while the R² improved from 0.1 to 0.5, demonstrating a clear enhancement in prediction accuracy.
Furthermore, performance comparison under varying training data amounts revealed that SVR achieved better stability with limited data (< 100 samples), whereas DNN outperformed SVR when sufficient data (> 150 samples) were available. Overall, both models exhibited improved performance as the training data increased, but a saturation trend was observed beyond approximately 70% samples, indicating limited performance gain from additional data.
These findings experimentally validate that ML-based approaches can effectively overcome the limitations of conventional optical metrology, providing a quantitative framework for optimizing data requirements and establishing a practical foundation for next-generation in-process monitoring in advanced semiconductor manufacturing.