A Voxelized Transformer-Based Neural Network for 3D Reconstruction From Multi-Energy SEM Backscattered Electrons

Three-dimensional (3D) data analysis, especially 3D reconstruction, is critical within integrated circuits for quality control and inspection. The traditional 3D reconstruction, like multi-energy deconvolution scanning electron microscopy (MED-SEM), can achieved resolution of 5 nm in the horizontal direction and 10 nm in the vertical direction from backscattered electron images of objects at different depth layers. However, the traditional methods are difficult to obtain high-precision reconstruction results in the longitudinal direction, and the reconstruction results of most methods contain a large number of artifacts. Thus, in this work, a hybrid CNN-Transformer network, called BSE-VoxNets, is proposed for 3D high precision reconstruction. The proposed BSE-VoxNets consists three parts: transformer feature extractor, convolutional upsampling block and feature fusion block. 10000 BSE samples were used for training BSE-VoxNets, with 200 samples for testing, and the optimal cross-union ratio (IoU) of the test dataset reached 0.953. Two numerical cases are used to respectively verify the inversion performance of the proposed method for different depths and different feature scales. The reconstruction results show that the proposed BSE-VoxNets can reach a resolution of 2 nm in both the horizontal and vertical dimensions.