Improvement of critical dimension uniformity using machine learning

This paper introduces a method for improving photomask CDU (Critical Dimension Uniformity) using machine learning. The CDU of a photomask is one of the key factors affecting device quality. To improve the CDU of photomask, local dosage corrections are performed on the writing tool. However, it is difficult for human to predict the amount of correction for entire writing area in advance. In particular, non-critical layer masks for memory device, which are often used in peripheral circuits, significant degradation in CD distribution has been observed. In this study, we evaluated the performance of several machine learning models, called GBDT (Gradient Boost Decision Tree), in predicting the CD distribution of non-critical layer masks. The motivation behind this research is that the consistent production of highly accurate photomasks leads to reduce costs in photomasks and device development. Methods for increasing the accuracy of the model are also presented. Not only numerical data but also categorical data were used to generate the features used in the model. To avoid leakage problem of time series data, the data were divided into training, validation, and test data along the time sequence. Multiple models were used in ensemble to construct a highly accurate and stable model. In the product test, simulation results with CD correction using machine learning predictions showed a 20% improvement in CDU at the median compared to the conventional method.