Learning Self-Supervised Representations of Powder-Diffraction Patterns

Abstract

The potential of machine learning (ML) models for predicting crystallographic symmetry information from single-phase powder X-ray diffraction (XRD) patterns is investigated. Given the scarcity of large, labeled experimental datasets, we train our models using simulated XRD patterns generated from crystallographic databases. A key challenge in developing reliable diffraction-based structure-solution tools lies in the limited availability of training data and the presence of natural adversarial examples, which hinder model generalization. To address these issues, we explore multiple training pipelines and testing strategies, including evaluations on experimental XRD data. We introduce a contrastive representation learning approach that significantly outperforms previous supervised learning models in terms of robustness and generalizability, demonstrating improved invariance to experimental effects. These results highlight the potential of self-supervised learning in advancing ML-driven crystallographic analysis.