CNNs for orientation mapping of LNO electron diffraction patterns together with test-data.

Content: Classification and Regression neural networks for predicting the orientation of LiNiO2 transmission electron diffraction images. Also included are two test datasets. One a set of simulated images of orientations not in the training datasets of the networks and one of a experimentally recorded LiNiO2 sample. Training data: The CNNs were trained on synthetic diffraction patterns of LiNiO2 simulatad with the Bloch-wave algorithm of the py4DStem python package. Three different training datasets were created and the number at the end of the networks' name indicate on which they were trained. The training datasets differ in the manner they cover the orientation space of the fundamental zone of the material. For the first datatset orientations were randomly selected. For the second dataset, the three Euler angles were each varied in steps of 2.5 degrees and for the third dataset equi-spaced orientations were simulated. The training diffraction patters were simulated with a acceleration voltage of 300kV, a semi-convergence angle of 1.5mrad and a maximum field of view of 8mrad. The images were augmented to include shifting centers, different camera lengths and amorphous material on the sample. Networks: The networks follow a conventional CNN architecture. They take a 256x256 image as input and encode it using a number of convolutional blocks. The latent space consists of two fully connected layers each containing 3,200 nodes. The classification networks end in a fully connected layer with a softmax activation. The number of nodes ranges from 70,000 to 500,000 depending on the training dataset used. The mapping from these nodes to the sets of Euler angles is given in .xlsx files inside each network folder. The regression network ends in a FC-layer with three nodes, one for each Euler angle. Test datasets: The simulated test-datasets consists of 580,000 unaugmented images, following the same covering of the fundamental zone as the third training dataset, but with a higher resolution in the orientation space. The experimental test-datset of a LiNiO2 sample was recorded with a JEOL JEM-3010 operating at 300kV, using a convergence semi-angle of 1.6 mrad. A NanoMegas P2010 system is used to precess and scan the electron beam across the sample with a precession angle set to 0.6 degrees. The experimental images were labeled with the ASTAR pattern matching software.