Crystal Structure Reconstruction from X-Ray Diffraction Patterns Based on Supervised Contrastive Learning

Techniques based on X-ray diffraction (XRD) are the most related to the identification of crystal structures. The latest generation of tools for high-throughput diffraction experiments and simulations has led to collection of a large volume of XRD data. The handling of these large-sized dataset calls for big data and machine learning (ML)-based approaches. Several recent studies were reported, where powder XRD 1D curves were used to classify crystal symmetry such as crystal systems and space groups. However, these studies are limited to identifying crystal symmetry, and not yet extended to the full reconstruction of three dimensional (3D) crystal structure.<br/>In this work, we report a deep learning (DL)-based approach to reconstruct 3D crystal structures from XRD patterns, based on supervised contrastive learning (SupCon). This protocol largely consists of crystallographic prototype classifier and element position assigner. The former process attempts to classify crystallographic prototype, which is one of crystal structure symmetry groups. According to AFLOW, 1,100 prototypes are available today, and each has a unique combination of material composition (stoichiometry), crystal symmetry (Pearson symbol and space group), and Wyckoff position. Using SupCon, we achieved the classification accuracy around 90%, based on the datasets involving 28 prototypes. The latter process is element position assigner, which finds the position of elements as well as lattice parameters for a given prototype. Through our novel protocols leveraging prototype classification, we demonstrate several examples where 3D crystal structures were fully reconstructed from XRD patterns.<br/>We additionally emphasize the importance of SupCon, which is much better in contrasting the subtle differences of similar crystal prototypes than traditional supervised classification methods, which was key to achieving classification accuracy around 90%. This work offers a method to reconstruct 3D crystal structure directly from XRD patterns, which will be useful in understanding the structural information more intuitively.