Model aggregation for virtual metrology in high-volume manufacturing

Virtual metrology (VM) plays a pivotal role in enhancing productivity, improving quality, and reducing maintenance costs in semiconductor manufacturing by replacing traditional physical metrology. Many implementations of VM rely on the predictive modeling approach that leverages equipment sensor data to estimate crucial process outcome variables. Despite its inherent advantages, VM has not been deployed widely in actual manufacturing processes due to the lack of accuracy and scalability. Most machine learning algorithms struggle to effectively address practical challenges such as data drift, data shift, sparse ground-truth data, and inconsistent data quality prevalent in semiconductor manufacturing processes. To address these limitations, we introduce the aggregated adaptive online model (AggAOM), a novel approach that effectively tackles the challenge of data scarcity in VM. By leveraging the hierarchical structure of manufacturing equipment, AggAOM captures and utilizes the underlying commonalities among equipment chambers within the same hierarchy besides their individual variations. This methodology enables more efficient use of limited data and substantially improves the prediction accuracy of VM running in mega-fabs for high-volume manufacturing. We present the experimental results by utilizing the datasets collected from SK hynix over nine months and demonstrate that AggAOM outperforms existing models significantly in accuracy. This progress marks a significant step forward in optimizing VM for semiconductor manufacturing.