Shape-aware dissimilarity measures for process TCAD model calibration

Abstract Traditional calibration of semiconductor TCAD models for topography modifying processes often relies on matching critical dimensions (CDs). This approach is prone to subjective measurement and fails to capture overall profile shape differences, thereby limiting the accuracy and effectiveness of model optimization. In this work, we address this limitation by adapting computer vision techniques for TCAD model calibration. We comprehensively evaluate several shape dissimilarity measures (SDM): the established Chamfer Distance (CHD) and two novel level-set-based measures, the Area Difference (AD) and the Sparse Field Distance (SFD). Through a rigorous computational case study, testing for sensitivity, robustness, and efficiency, we demonstrate that SFD and CHD significantly outperform traditional CD matching and area-based methods for semiconductor profile comparison derived from a representative scanning electron microscopy (SEM) image. These standardized SDMs are crucial for leveraging emerging large-scale electron microscopy datasets and enabling robust, machine learning-driven model development.