A Meta-Learning Reinforcement Training Method for Machine Learning Image-To-Image Optical Proximity Correction

Abstract

As the scaling down of semiconductor manufacturing nodes, optical proximity correction (OPC) has become more and more crucial, where the OPC using machine learning to construct models, or so-called machine learning optical proximity correction (MLOPC), is flourishing. The forward and reverse networks in MLOPC are two deciding parts. Training the image-to-image OPC dataset using various advanced ML models can lead to less fitted results. The tests on diverse mainstream image-to-image ML models, including Unet, Vnet, ResUnet, U2net, and attention Unet, are conducted in our studies. We propose a meta-learning method based on reinforcement learning (RL) with the forward and reverse networks cascaded to boost the training efficiency and accuracy for MLOPC. The meta-learning is performed so that the selected setting of the next training step is based on the Q-model in RL, and the baseline training method is the direct utilization of adaptive moment estimation (Adam). In the setting of the cascaded network with six different models and three types of datasets, the model using the meta-learning approach takes advantage of the cascade and RL-controlled training path to boost the model accuracy, and the results using meta-learning surpass the baseline results in most of the cases, especially the instances with highly complex data or network architecture. Among them, the networks trained via the proposed meta-learning RL method receive the averaged increments of 19.5%, 1.2%, and 4.8% in the forward models and the averaged increments of 12%, 3.1%, and 13% in the reverse models.