Nam Le1,Michael Pekala1,Alexander New1,Eddie Gienger1,Janna Domenico1,Christine Piatko1,Elizabeth Pogue1,Tyrel McQueen2,Christopher Stiles1,2
Johns Hopkins University Applied Physics Laboratory1,Johns Hopkins University2
Nam Le1,Michael Pekala1,Alexander New1,Eddie Gienger1,Janna Domenico1,Christine Piatko1,Elizabeth Pogue1,Tyrel McQueen2,Christopher Stiles1,2
Johns Hopkins University Applied Physics Laboratory1,Johns Hopkins University2
High-throughput materials discovery systems require the ability to rapidly screen promising candidates. X-ray diffraction (XRD) provides a useful modality in settings where properties are strongly tied to particular crystalline phases. For example, many A15 phases are Type-II superconductors at relatively high temperatures among metallic alloys; identifying novel A15 phases could therefore lead to a high proportion of novel superconductors. Machine learning (ML) models have been shown effective for automating similar phase identification tasks from XRD patterns. However, previous work has generally been limited to material systems of up to five elements. We report the results of convolutional neural networks trained to classify A15-like phases from XRD patterns from a broad space of binary and ternary material systems spanning 23 elements. High performance can be achieved whether using crystal structures measured experimentally (Inorganic Crystal Structure Database, ICSD) or computed theoretically using density functional theory (Materials Project, MP). Performance decreased significantly when trained and tested on different sources, but can be recovered by augmenting datasets with both experimentally- and theoretically-derived structures. High classification performance can even be maintained on held-out XRD patterns measured experimentally in-house. This work suggests that ML models for XRD screening can be effective for phase identification not only within material systems with already experimentally-measured patterns, but also in novel material systems through careful augmentation with theoretically-computed structures.