Using Machine Learning Methods to Predict the Magnitude and the Direction of Mask Fragments Displacement in Optical Proximity Correction (OPC)

The paper studies the effectiveness of machine learning methods in computational photolithography. The first task is to determine the direction of displacement of the mask contour fragment. The second task is to determine the amount of displacement of the mask contour fragment. The machine learning models were trained on the data generated with Calibre WORKbench CAD in the form of radiation intensity vectors around the center of the segment. Comparisons were made between linear regression, random forest, gradient boosting, and feedforward convolutional neural network models. The most accurate results were demonstrated by the random forest model. With its help, it is possible to achieve an absolute error of 2 nm and an accuracy of displacement’s direction prediction of 97.9%.