Machine Learning-Enhanced Model-Based Optical Proximity Correction by Using Convolutional Neural Network-Based Variable Threshold Method

As the lithography process continues to become more rigorous in advanced technology nodes, the model-based optical proximity correction (MBOPC), as a core component within computational lithography, necessitates the development of highly precise techniques. In this paper, we propose an approach to enhance MBOPC through the integration of machine learning (ML), utilizing convolutional neural network (CNN)-based variable threshold method. This OPC framework is characterized by retaining the physical lithography model while integrating the mapping capability of the neural network model to rectify errors encountered in MBOPC. We validate the CNN-based MBOPC model's feasibility at implant layers in advanced nodes. The results demonstrate an improvement in the accuracy of threshold value regression compared to conventional variable threshold methods, and confirm the positive impact of ML integration in simulation accuracy across various patterns. The enhanced MBOPC model effectively compensates for lithography differences in both one-dimension (1D) and two-dimension (2D) regions. This research aims to enhance the simulation precision of MBOPC, thereby ultimately contributing to the ongoing advancement of computational lithography technology.