2026년 4월 8일
Domain knowledge-driven fusion machine learning for overlay prediction enhancement
SPIE Advanced Lithography + Patterning
논문 보기저자
Sungbin Son, Jungsuk Kim, Jeng-Hun Suh, Taekwon Jee
초록
In advanced technology nodes, overlay control can no longer satisfy required precision using conventional feedback-only schemes due to physical metrology limits and process noise. This study proposes a domain knowledge-driven Fusion machine learning (ML) framework that combines scanner logs with process-domain expertise. To sufficiently learn rare and extreme defect signatures observed in manufacturing, we adopt a Virtual Fab-based synthetic pretraining strategy. Under intentionally induced bimodal deformation conditions, we compare Linear Regression (LR), XGBoost, and the proposed model. The proposed model achieves an average root-mean-square error (RMSE) of 0.26 nm, representing a 78% reduction compared to LR (1.19 nm) and a 46% reduction compared to XGBoost (0.48 nm), while maintaining Pearson correlation r above 0.88 in both axes. These results show that a physics- and process-aware model is effective for achieving metrology-grade prediction accuracy in the sub-1nm era, beyond what generic ML typically provides.