Implementation of Virtual Sensor for Semiconductor Process Verification Using Machine Learning

Abstract

The facility that performs the exposure process during semiconductor process verification includes more than 2500 sensors, and if an abnormal value is detected in any of these sensors, the wafer is treated as an exception. Wafers treated as exceptions undergo and additional process of determining wherein it is determined through physical measurement whether they are normal or abnormal. In this study, to reduce the time and cost of physical measurement, machine learning is used to learn sensor values to predict measurement results and to implement virtual sensors based on predicted values. The method for determining whether an exposure process is normal or abnormal is suggested using the implemented virtual sensor. The virtual sensor implements a single virtual sensor and multiple virtual sensors, and the algorithms used are extreme gradient boosting (XGBoost) and multi-output classifier. As a result of the experiment, the single virtual sensor detected an average of 93% of defects, furthermore, when measuring the detected defective wafers, an average of 99% of the actual defects were detected. Multiple virtual sensors also exhibited an average performance of 93%. Based on the findings of this study, 2500 real sensors were implemented as 250 virtual sensors, enabling the reduction of the time required to verify semiconductor exposure process results.