Machine learning in X-ray diffraction for materials discovery and characterization

Machine learning (ML) is a promising analytical method for large high-throughput, in situ , and operando X-ray diffraction (XRD) datasets. However, ML methods are, by default, physics agnostic and must therefore be interpreted carefully. In this review, we survey how supervised ML methods are used to predict symmetries and phases in pure and mixed-composition materials, and we highlight challenges related to experimental artifacts and model interpretation. We also review recent uses of unsupervised ML methods in the extraction of patterns hidden in high-dimensional data, such as in in situ and microscopic studies. Finally, we discuss the importance of problem formulation, data transferability, and reporting, leveraging examples from the literature, and we provide various resources throughout to expedite the learning curve for readers new to XRD or ML. We advocate for greater scrutiny of ML methods and how they are reported in the literature, and we explain how to conduct data-driven research responsibly. X-ray diffraction (XRD) is an immediate and powerful characterization technique that provides detailed information on the lattice structure and long-range order in crystalline materials. In recent decades, the quality and quantity of available crystal structure data have exploded, in large part due to the advent of high-throughput materials synthesis and processing, online crystal structure databases, increased use of in situ and operando methodologies, and user-accessible beamlines. The new wealth of data has also spawned an increasing use of machine learning (ML) to either construct high-throughput surrogates of established analysis or extract patterns from large datasets. However, XRD analysis has for decades been solved via Rietveld refinement, while most ML techniques are simply complex statistical evaluation methods that are physics agnostic. The discrepancy between data analysis and the underlying physics can lead to incorrect conclusions and/or limit the widespread adoption of ML techniques. In this review, we begin to bridge the gap between ML and XRD spectroscopy with introductions both for new data scientists interested in XRD and for experimentalists interested in applying ML to their existing data. We also advocate for greater collaboration in the sharing of experimental data and appropriate material metadata, enabling cross-study meta-analysis and training of predictive ML models from multiple sources. Machine learning has been demonstrated as an effective method to extract structural information from X-ray diffraction data, promising future autonomous and semiautonomous experimental workflows. The accurate classification of material symmetry and phase identity and the visualization of high-dimensional in situ and operando experiments may increase the throughput of materials discovery and characterization in conjunction with recent advances in high-throughput and combinatorial synthesis.