2D-diffractogram analysis: Kinematic-diffraction simulator for neural-network training-data generation

Abstract

To exploit the information contained in 2D X-ray diffractograms fully, quantitatively, automatically, and with high throughput, e.g. for analyzing video sequences from in-situ experiments, we can train deep-learning NNs (neural networks) with simulated diffractograms. Realistic models of materials microstructures require “ground truth” training datasets of high cardinality. To produce these, we developed a “kinematic-diffraction simulator,” implemented in the Wolfram Language and executed within a high-performance computing environment. The simulator can rapidly generate Fraunhofer diffractograms for diverse crystal- and microstructure models over a significant multi-dimensional space of parameters. We conclude that simulated diffractograms can enable suitable training of deep-learning NNs – in spite of not including some “real-world” features that occur in experimental diffractograms – and that high-performance computing achieves training data generation rates that support modeling of microstructures with a realistically large number of parameters.