American scientists have developed and tested a machine learning model to recognize the characteristics of the crystalline structure of the samples (the type of Bravais lattice and the crystallographic group) for images of the diffraction patterns obtained by the diffraction of reflected electrons. Both of the used neural networks were able to accurately (over 90%) to determine these parameters. The results of a study published in the journal Science.
Crystal structure of a material strongly affects its properties, so the structure determination of proteins, micro – and macromolecules, pharmaceuticals, new materials and geological objects is very important. Most often for the solution of complex problems of determining the parameters of the lattice, symmetry of crystals and phases used or the method of diffraction of x-rays, or electron diffraction convergent beam.
One of the more convenient methods of structure determination of crystalline materials and geological objects becomes the method of diffraction of reflected electrons, combined with a scanning electron microscope. It does not require a complicated sample preparation, as methods using transmission electron microscope, and enables you to analyze large sample areas in less time. This makes it convenient for exploring the orientations with high precision (up to two degrees), resolution angles up to two tenths of a degree and a spatial resolution of about 40 nanometers.
Development of methods of automation of image processing in the nineties accelerated the analysis and made it possible to use the method for more complex and time-consuming tasks, but still data processing takes time. Often used to identify phases and orientations in samples with multiple phases. The researcher selects the phase that presumably is in the sample, and the program searches for the most suitable on the experimentally obtained diffraction pattern. In contrast transmission microscopy or x-ray diffraction method, the diffraction of reflected electrons allows us to see several phases in the spatial resolution, however, the method is limited by the need to specify the phases that are present in the sample, and they are not always known in advance.
Kevin Kaufmann (Kevin Kaufmann) with colleagues from the University of California, San Diego have developed a machine learning algorithm that is able to determine the parameters of the crystalline structure of the sample (Bravais lattice or crystallographic group) for the diffraction pictures, obtained by the method of diffraction of reflected electrons. The authors trained and tested two convolutional neural network. The layers were enrolled as finding algorithm of motifs, which are answered one way or another crystallographic symmetry on the diffraction pattern.