Deep Learning Models to Identify Common Phases across Material Systems from X-ray Diffraction

X-ray diffraction (XRD) is an indispensable tool for materials discovery and development, allowing researchers to confirm the synthesis of expected phases and determine the structures of new compounds in unexplored material systems. A common bottleneck is the identification of relevant phases from the XRD patterns, which even using current software requires human intervention and is thus not suitable for automated workflows. As a case study, we take the example of screening for A15-type phases based on XRD measurements of samples from systems spanning a broad space of 23 elements. Many A15 phases are type-II superconductors at comparatively high temperatures relative to other metallic alloys, motivating a hypothesis that seeking novel A15 phases will lead to a high proportion of novel superconductors. We report the results of convolutional neural networks trained to classify A15-like phases based on databases of synthetic data with simulated noise. Models exhibit high performance even when training, validation, and test sets are constructed such that they do not include examples from any material systems in common. Furthermore, performance is equally high on two data sets constructed from databases of crystal structures measured experimentally (Inorganic Crystal Structure Database, ICSD) and those computed using density functional theory (Materials Project). However, model performance decreased markedly when trained and tested across the two different sources, demonstrating a need for careful consideration when models are transitioned from large density functional theory (DFT)-computed databases to experimental data. A strategy of augmenting DFT-based patterns with experiment-based patterns is shown to be simple yet effective in improving performance across both MP- and ICSD-based data. Good performance is also observed on held-out XRD patterns measured experimentally on in-house samples. The approach demonstrated here can extend readily to other materials discovery settings where properties are strongly tied to particular phases that are distinguishable based on XRD patterns.

Automated summary

X-ray diffraction (XRD) is a crucial tool for materials discovery and development, allowing researchers to confirm synthesis and determine structures of new compounds. The identification of relevant phases from XRD patterns remains a challenge for automated workflows. This study presents convolutional neural networks trained on synthetic data to classify A15-like phases based on XRD measurements. The models show high performance even without common material systems in training, validation, and testing. However, when transitioning from density functional theory (DFT)-computed databases to experimental data, the model's performance decreases significantly. Augmenting DFT-based patterns with experiment-based patterns improves performance across different data sources. The approach is also effective on experimentally measured XRD patterns. The method can be applied to other materials discovery settings where properties are tied to specific phases distinguishable by XRD patterns.