Machine Learning-Enhanced Model-Based Optical Proximity Correction Framework With Convolutional Neural Network-Based Variable Threshold Method Near the Diffraction Limit

As semiconductor manufacturing approaches the diffraction limit of lithography, traditional model-based optical proximity correction (MBOPC) techniques face accuracy bottlenecks due to physical modeling errors. This study proposes a Machine Learning (ML)-enhanced MBOPC framework that employs a convolutional neural network (CNN) to predict mask edge imaging thresholds, thereby mitigating modeling deviations caused by complex lithographic conditions in the 28 nm technology node under immersion lithography. Compared to the conventional polynomial-based variable threshold method, the CNN-based variable threshold strategy significantly improves threshold prediction accuracy and critical dimension (CD) simulation consistency by integrating geometric features of local patterns and slice’s optical intensity distribution parameters. Specifically, the CNN module reduces the root mean square error (RMSE) by approximately 78% on training datasets and 69% on testing datasets versus the constant threshold strategy. In CD simulations for typical patterns, the hybrid model reduces error medians and confines the statistical upper and lower limits of the distribution ranges to ±5 nm. The simulation results demonstrate tighter error distributions across typical 1D and 2D patterns in the metal layer of 28 nm node, highlighting the method’s robustness. This work presents a combined approach that integrates physical models with data-driven methods for the OPC model, providing a new paradigm for next-generation MBOPC technologies that balance physical interpretability and high precision in advanced nodes.