Challenges and Opportunities of Machine Learning on Neutron and X-ray Scattering

Abstract

Machine learning has been highly successful in boosting the re-search for neutron and X-ray scattering in the past few years [1, 2]. Fordiffraction, machine learning has shown great promise in phase map-ping [3, 4] and crystallographic information determination [5, 6]. Insmall-angle scattering, machine learning shows the power in reachingsuper-resolution [7, 8], reconstructing structures for macromolecules[9], and building structure-property relations [10]. As for absorptionspectroscopy, machine learning has enabled the rapid inverse searchfor optimized structures [11, 12] with improved spectral interpretability[13, 14]. Overall, as a data-driven approach, the success of the machine-learning-based scattering analysis depends on a few criteria, including:• Quantity of available experimental data, and feasibility to extractcertain data labels;• Quality of experimental data that can separate the intrinsic effect(e.g., materials properties) from extrinsic influence (e.g., instru-mental or data artifacts);• Feasibility to generate high volume of computational data;• Accuracy of computational data that can simulate the experimen-tal data.