Accuracy improvement of 3D-profiling for HAR features using deep learning

We applied deep learning techniques to improve the accuracy of 3D-profiling for high aspect ratio (HAR) holes. As deep learning requires big data for training, we developed a method for generating a large amount of BSE line-profiles by a numerical calculation in which the aperture angle and the aberration effects of the electron beam are considered. We then utilized these numerically calculated datasets to train the deep learning model to learn the mapping from the BSE line-profiles to the target cross-sectional profiles of the HAR holes. Two different one-dimensional neural network architectures: convolutional neural network (CNN) and multi-scale convolutional neural network (MS-CNN) were trained, and different loss functions were investigated to optimize the networks. The test results show that the MS-CNN model with a defined loss function of weighted mean square error (WMSE) provided higher accuracy than the others. The mean absolute percentage error (MAPE) distribution was narrow and the typical MAPE was 4% over 2810 items of test data. This model enables us to predict the cross-section of the HAR holes with different sidewall profiles more accurately than our previously proposed exponential model. These results demonstrate the effectiveness of the learning approach for improving the accuracy of 3D-profiling of the HAR features.