Characterizing material structure using X-ray or neutron scattering and Pair Distribution Function (PDF) analysis often involves refining a structure model against experimental data. Identifying a suitable model is challenging. This paper introduces a Machine Learning based Motif Extractor (ML-MotEx) that trains a machine learning algorithm on thousands of fits and uses SHAP (SHapley Additive exPlanation) values to identify important model features for fit quality. The method is demonstrated on four chemical systems, including disordered nanomaterials and clusters, assigning importance values to each model feature based on explainable machine learning.

Andy S. Anker, Emil T. S. Kjær, Mikkel Juelsholt, Troels Lindahl Christiansen, Susanne Linn Skjærvø, Mads Ry Vogel Jørgensen, Innokenty Kantor, Daniel Risskov Sørensen, Simon J. L. Billinge, Raghavendra Selvan, Kirsten M. Ø. Jensen