Heusler compounds form a diverse group of intermetallic materials encompassing many compositions and structures derived from cubic prototypes, and exhibiting complicated types of disorder phenomena. In particular, preparing solid solutions between half-Heusler ABC and full-Heusler compounds AB2C offers a means to control physical properties. However, as is typical in materials discovery, they represent only a small fraction of possible intermetallic compounds. To address this problem of unbalanced data sets, a machine-learning model was developed using an ensemble approach involving the synthetic minority oversampling technique to predict new compounds likely to adopt half-Heusler structures. The training set was based on experimental crystal structures, including those of nonstoichiometric compounds. The model achieved an accuracy of 98% on the validation set and gave excellent performance in terms of balanced statistical measures. A subset of compounds predicted to adopt half-Heusler structures having existing full-Heusler counterparts was then targeted for preparation. Six of seven of these candidates were successfully synthesized and confirmed to be half-Heusler compounds.

The site preferences within the structures of half-Heusler compounds have been evaluated through a machinelearning approach. A support-vector machine algorithm was applied to develop a model which was trained on 179 experimentally reported structures and 23 descriptors based solely on the chemical composition. The model gave excellent performance, with sensitivity of 93%, selectivity of 96%, and accuracy of 95%. As an illustration of data sanitization, two compounds (GdPtSb, HoPdBi) flagged by the model to have potentially incorrect site assignments were resynthesized and structurally characterized. The predictions of the correct site assignments from the machine-learning model were confirmed by single-crystal and powder X-ray diffraction analysis. These site assignments also corresponded to the lowest total energy configurations as revealed from first-principles calculations.

Machine learning attempts to find underlying trends in data and offer predictions of outcomes. When machine learning is applied to materials science, in a discipline called materials informatics, the complex relationships between composition, structure, and properties can be unraveled even when the quantity of data is limited. To illustrate this application, the large class of materials known as Heusler compounds are modeled through machine learning, enabling new candidates to be predicted or existing compounds to be screened for potentially interesting properties. Data, algorithms, and preprocessing techniques are important components of a successful machine‐learning model. Efforts to predict structures and properties of Heusler compounds are reviewed, and other machine‐learning approaches to discover materials in general are discussed. Ultimately, a machine‐learning model is only valuable if its predictions are validated by experimental results. Thus, perspectives are offered to guide experimentalists on how machine learning can be useful for targeting new Heusler compounds.

The site preferences within the structures of half-Heusler compounds have been evaluated through a machine-learning approach. A support-vector machine algorithm was applied to develop a model which was trained on 179 experimentally reported structures and 23 descriptors based solely on the chemical composition. The model gave excellent performance, with sensitivity of 93%, selectivity of 96%, and accuracy of 95%. As an illustration of data sanitization, two compounds (GdPtSb, HoPdBi) flagged by the model to have potentially incorrect site assignments were resynthesized and structurally characterized. The predictions of the correct site assignments from the machine-learning model were confirmed by single-crystal and powder X-ray diffraction analysis. These site assignments also corresponded to the lowest total energy configurations as revealed from first-principles calculations.

No abstract