Methods to overcome limited labeled data sets in machine learning-based optical critical dimension metrology

With the aggressive scaling of semiconductor devices, the increasing complexity of device structure coupled with tighter metrology error budget has driven up Optical Critical Dimension (OCD) time to solution to a critical point. Machine Learning (ML), thanks to its extremely fast turnaround, has been successfully applied in OCD metrology as an alternative solution to the conventional physical modeling. However, expensive and limited reference data or labeled data set necessary for ML to learn from often leads to under- or overlearning, limiting its wide adoption. In this paper, we explore techniques that utilize process information to supplement reference data and synergizing physical modeling with ML to prevent under- or overlearning. These techniques have been demonstrated to help overcome the constraint of limited reference data with use cases in challenging OCD metrology for advanced semiconductor nodes.