Unsupervised Machine Learning Prediction of a Novel 1:3 Intermetallic Phase with the Synthesis of TbIr ₃ (PuNi ₃ -type) as Experimental Validation

Crystal structure classification of binary intermetallic structures with 1:3 stoichiometry was done with machine learning algorithms. The data set included 97 features and a total of 2366 reported compounds adopting six different structure types. An unsupervised learning method based on principal component analysis (PCA) followed by clustering using the K-means method was applied to cluster compounds belonging to different structure types. With the recommendation engine, we predicted the expansion of the clusters and then identified the cluster/structure-type overlap. The PuNi₃-type was among the clearly segregated structure types according to the unsupervised model, and a novel representative, TbIr₃, was selected for experimental validation, adopting this structure. The final supervised machine learning predictions were done with partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), and XGBoost, confidently predicting that the novel TbIr₃ belongs to the PuNi₃-type with accuracies of 96.6, 99.8, and 99.9% respectively. Successful crystal structure segregation was attributed to a novel set of descriptors comprising both compositional and structural features. Given that the predicted PuNi₃-type of the TbIr₃ phase could be controversial due to the extensive study of the Tb-Ir phase diagram and the reports of the TbIr₃ in two different structure types, we conducted two independent experimental structural validations to confirm the existence of TbIr₃ in the PuNi₃-type structure. Subsequent theoretical validation explained that Ir-Ir contacts are the primary stability factor of TbIr₃ in the PuNi₃-type structure compared to other structure types.