High-performing virtual metrology with group-wise feature transformation and advanced data aggregation techniques

Data quality and quantity are two of the essential qualifications of a dataset for building a powerful machine learning model. In semiconductor manufacturing, these qualifications are luxurious, making the technical implementation of virtual metrology difficult in practice. Recent studies have proposed advanced data preprocessing techniques to enhance model updatability and adaptability. However, these methods often lack scalability as they depend on specialized domain expertise. This paper introduces advanced data aggregation techniques to collect essential data at scale together with a simple yet powerful data-driven group-wise feature transformation module (GroFT). GroFT dynamically groups feature that share similar data characteristics using a distance-based clustering technique and then applies specifically optimized data transformation for each feature group. Our framework achieves superior performance in publicly available datasets and a high-volume dataset from a major semiconductor manufacturer.