Exploiting machine learning clustering methods to optimize OPC model

The first and foremost crucial step in generating an effective term-based OPC model is defining an appropriate model structure. This process involves combining diverse types of linear and non-linear terms. With each term having tunable parameters and considerations of cross-terms, a possible number of term combinations can easily grow exponentially and become intractable. Additionally, an optimal term combination is often process-specific which makes a manual term-selection process daunting and laborious. Therefore, we propose a new machine learning (ML)-based feature selection approach to facilitate the term-selection process more systematically and intelligently. Our approach utilizes the high-performance supervised clustering algorithms to effectively reduce dimensionality by removing irrelevant and duplicate terms. In this study, we used three different data types of CD gauge, EP gauge, and hybrid of the two to extract feature vectors and preconfigured term libraries as input to the clustering methods. We applied the new approach to both etch and resist modeling cases and evaluated their result by focusing on model accuracy, degree of overfit, and effect of number terms. Result of etch case was promising by showing equivalent performance to the best performing model created by existing manual process yet containing substantially fewer number of terms. For the case of not satisfactory result such as our resist case, we suggest using the proposed approach as a starting point to build a more advanced model for higher accuracy.