DEEP LEARNING IN LITHOGRAPHIC PROCESS OPTIMIZATION

Abstract [English]

The rapid advances in the development of semiconductor devices have placed control, precision and yield optimisation in the printing process under pressures hitherto unseen. In the past few levels of technology, empirical and physics-based models worked well. But in advanced lithography, they have a hard time capturing the complicated non-linear relationships of exposure, focus, resist and etching parameters. New developments in deep learning (DL) have opened up the use of predictive modelling, finding flaws and adaptable optimisation in the printing process in new ways. This research explores how deep learning models such as convolutional neural network (CNN), recurrent neural network (RNN), generative adversarial network (GAN) and transformer can be applied to enhance the lithography process. Data from litho tools and physics-based simulation are used to train models that can be used to predict regularity of critical dimensions (CD), mistakes in overlays, and the accuracy of resist profiles. Model generalisation is enhanced with dimensionality reduction techniques and feature extraction techniques. Mean absolute error (MAE), R2, and process window gain are performance measures and are helpful in evaluation. The study also speaks about new areas of research that require more attention.