Speaker
Description
With recent improvements to the efficiency of collecting single crystal diffuse scattering at synchrotron x-ray sources, large contiguous scattering volumes of diffraction data comprising 100GB can be collected in 20 minutes with high dynamic range and low backgrounds. The Python package, NXRefine, implements a complete data reduction workflow, from ingesting the data, orienting the single crystals, transforming the data into reciprocal
space coordinates, and generating 3D-ΔPDF maps, i.e., maps of real space interatomic vector probabilities. Data from the Advanced Photon Source are now streamed to Argonne’s Leadership Computing Facility for on-demand processing, enabling diffuse scattering to be tracked as a function of parametric variables such as temperature in a few hours as fast as it is collected. With data volumes of several TB a day, it is imperative to have advanced methods of interrogating the data in real time. We have implemented two complementary machine learning approaches. In the first, billions of voxels collected at multiple temperatures are grouped into a finite set of clusters using the Gaussian Mixture Model in order to identify automatically distinctive temperature dependences resulting, for example, from the growth of superlattice peaks at a structural phase transition. This has been implemented in a Python package called X-TEC. In the second, 3D-ΔPDF maps are modeled by generating all the symmetry modes in the entire space group tree using online crystallographic databases. Convolutional neural networks then identify those subgroups that are compatible with the experimental data, allowing the interatomic displacements to be optimized. I will also discuss how the collection of such large datasets enable different contributions of the diffuse scattering to be separated using Independent Component Analysis.
Supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division.
