Feature Construction for Dense Inline Data in Semiconductor Manufacturing Processes

Management and analytics for inline manufacturing process data has become a critical but increasingly complex task in the semiconductor fabrication industry. The importance of new methods for construction of informative features has been accentuated by advancements in the data collection technology employed in this industry, which has recently increased sampling rates for inline data from values below 1 Hz to frequencies in excess of 10Hz. In this paper, a new feature construction method is proposed that aims to extract as much of the information accessible with these increased sampling rates as possible, while simultaneously minimizing redundancy and user involvement. The proposed method results in a meaningful dynamics inspired feature set, which provides insight into the underlying process and equipment dynamics. The advantages offered by this feature set are established using data collected at several modern 300mm fabs, for chamber and tool matching tasks, as well as wafer defect level prediction.