Next-generation virtual metrology for semiconductor manufacturing: A feature-based framework

In semiconductor manufacturing, virtual metrology (VM), also known as soft sensor, is the prediction of wafer properties using process variables and other information available for the process and/or the product without physically conducting property measurement. VM has been utilized in semiconductor manufacturing for process monitoring and control for the last decades. In this work, we demonstrate the shortcomings of some of the commonly used VM methods and propose a feature-based VM (FVM) framework. Unlike existing VM approaches where the original process variables are correlated to metrology measurements, FVM correlates batch features to metrology measurements. We argue that batch features can better capture semiconductor batch process characteristics and dynamic behaviors. As a result, they can be used to build better predictive models for predicting metrology measurements. FVM naturally addresses some common challenges that cannot be readily handled by existing VM approaches, such as unequal batch lengths and/or unsynchronized batch trajectories. A simulated and an industrial case studies are used to demonstrate the effectiveness of the proposed FVM method. We discuss how to generate and select features systematically, and demonstrate how feature selection affects FVM performance using a case study. Finally, the capabilities of FVM in addressing process nonlinearity is investigated in great details for the first time, which helps establish the theoretical foundations of the proposed framework for the semiconductor industry. (C) 2019 Elsevier Ltd. All rights reserved.