If an electron beam passes through a material using a transmission microscope, it is scattered at the potentials of the atoms in the sample. The complex patterns formed by the electrons behind the material can provide information on the properties of the sample under investigation. Interfering dynamic effects can be reduced by precessing the electron beam [Vincent et al. 1994].
With our work we want to investigate how well neural networks are suitable for evaluating electron scattering images. We examine position averaged convergent beam electron diffraction images (PACBED) to determine the sample thickness and the tilt angle of the sample to the beam axis. Furthermore, we consider scattering images with small convergence angles to investigate the orientation of the sample.
Scattering images of a STEM are simulated using a multi-slice algorithm [Oelerich et al. 2017]. Four materials (Si, GaP, Ge, GaAs) up to a thickness of 300nm and a tilt angle of 10mrad are created. These are used to train convolution neural networks (CNN). The networks are then used to determine the thickness and tilt on experimental images. To analyse the sample orientation, CNNs were trained with images from Bloch-wave simulations (Si, LiNiO2). The experimental images were acquired using a JEOL JEM 2200FS with a nanomegas ASTAR PED system and a pnCCD pixelated detector.
Our networks are highly competent in predicting the tilt angle of the sample. We achieve accuracies of up to 100% with tolerances of 0.3mrad. When determining the sample thickness, our networks show very good results for thin samples (i.e. up to about 60nm). At higher thicknesses, the accuracy of the predictions decreases steadily. It also shows that the networks are able to predict images of crystal directions that were not included in the training data.
When determining the orientation, our networks show a high agreement between the input images and simulated images of the predicted classes.
In conclusion, neural networks are a promising method for analysing scattering data from electron microscopes. The biggest hurdles are samples with very high thickness and the resource-intensive generation and storage of sufficient training data.