Machine-Learning Assisted Exit-wave Reconstruction for Quantitative Feature Extraction

Reconstruction of the exit wave is a powerful tool to extract the maximal amount of information from High-resolution Transmission Electron Microscopy (HRTEM). In addition to the three-dimensional structure of the nanoparticle, the reconstructed exit waves also contained information about the beam-stimulated vibrations of the atoms nearthe edge of the nanoparticle. We have recently demonstrated that convolutional neural networks are able to reconstruct the exit wave fromafocal serieswith a low number of images. We train the neural networks on simulated images. The simulated images are produced with the multislice algorithm using the abTEM software, both the exit wave function and images produced with three different values of the defocus are saved. The neural network is then trained to reconstruct the exit wave from the images. The network is validated on a different set of simulated images, and if applicable applied to experimentally obtained data. We demonstrated that it is possible to train neural networks to reconstruct the exit wave for a varied set of samples consisting of all structures in the Computational 2D Materials Database (C2DB). For a specialized dataset such asMolybdenum Disulphide (MoS2) supported on graphene, a slightlylower error rate can be obtained(Figure 2), and realistic results can be obtained when the network is applied to experimental data. In this work, we investigate how far the convolutional neural networks can be optimized towards obtaining quantitative information from experimental data, with a particular focus on the kind of data i.e.,reconstructing exit waves with sufficient accuracy to extract the three-dimensional structure and the amplitudes of the atomic vibrations. This can be realized with more flexible training sets than in our previous publicationand by training the network to ignore the support when reconstructing the exit wave.