AI-Driven Optimization of Thin Film and CMP Processes in 3-D NAND Manufacturing

Abstract

Artificial intelligence (AI) and machine learning (ML) demonstrate transformative potential in optimizing complex semiconductor manufacturing. This study addresses two critical areas: process control for 3D NAND fabrication and predictive maintenance (PdM) of fab equipment. For optimizing Thin Film deposition and Chemical Mechanical Planarization (CMP) processes, Artificial Neural Networks (ANNs) capture complex, non-linear dynamics, achieving prediction accuracies up to ~91.4% for controlling surface topography and roll-off. This significantly outperforms Lasso regression (~63.8%), Transformer models (~86.5%), and other customized ML models. In PdM applications, gradient boosting frameworks like eXtreme Gradient Boosting (XGBoost) deliver superior classification F1 score (exceeding 0.995) for chamber failure prediction on structured tabular data, outperforming Random Forest (0.991) and Convolutional Neural Networks (CNNs) (0.984). The enhanced performance of XGBoost stems from its sequential error-correcting tree building and effective handling of feature interactions. This work underscores the critical principle of selecting AI models aligned with data characteristics: ANNs for complex process optimization with rich feature dynamics, and tree-based ensembles like XGBoost for tabular PdM data. This tailored approach enables enhanced yield, reduced downtime, and higher operational efficiency, advocating for hybrid AI modeling in smart semiconductor Fabs.