Machine learning for discrimination of phase-change chalcogenide glasses

Chalcogenides, despite their versatile functionality, share a notably similar local structure in their amorphous states. Particularly in electronic phase-change memory applications, distinguishing these glasses from neighboring compositions that do not possess memory capabilities is inherently difficult when employing traditional analytical methods. This has led to a dilemma in materials design since an atomistic view of the arrangement in the amorphous state is the key to understanding and optimizing the functionality of these glasses. To tackle this challenge, we present a machine learning (ML) approach to separate electronic phase-change materials (ePCMs) from other chalcogenides, based upon subtle differences in the short-range order inside the glassy phase. Leveraging the established structure–property relations in chalcogenide glasses, we select suitable features to train accurate machine learning models, even with a modestly sized dataset. The trained model accurately discerns the critical transition point between glass compositions suitable for use as ePCMs and those that are not, particularly for both GeTe–GeSe and Sb₂Te₃–Sb₂Se₃ materials, in line with experiments. Furthermore, by extracting the physical knowledge that the ML model has offered, we pinpoint three pivotal structural features of amorphous chalcogenides, that is, the bond angle, packing efficiency, and the length of the fourth bond, which provide a map for materials design with the ability to “predict” and “explain”. All three of the above features point to the smaller Peierls-like distortion and more well-defined octahedral clusters in amorphous ePCMs than non-ePCMs. Our study delves into the mechanisms shaping these structural attributes in amorphous ePCMs, yielding valuable insights for the AI-powered discovery of novel materials.