Disentangling Structural Confusion through Machine Learning: Structure Prediction and Polymorphism of Equiatomic Ternary Phases ABC

• DOI: 10.1021/jacs.7b08460
• Publication Date: 2017-11-13T06:48:26Z
• Authors: Anton O. Oliynyk; Lawrence A. Adutwum; Brent W. Rudyk; Harshil Pisavadia; Sogol Lotfi; Viktor Hlukhyy; James J. Harynuk; Arthur Mar; Jakoah Brgoch
• Keywords: Confusion

A method to predict the crystal structure of equiatomic ternary compositions based only on constituent elements was developed using cluster resolution feature selection (CR-FS) and support vector machine (SVM) classification. The supervised machine-learning model first trained with 1037 individual compounds that adopt most populated 1:1:1 types (TiNiSi-, ZrNiAl-, PbFCl-, LiGaGe-, YPtAs-, UGeTe-, LaPtSi-type) then validated an additional 519 compounds. CR-FS algorithm improves class discrimination indicates 113 variables including size, electronegativity, number valence electrons, position periodic table (group number) influence preference. final prediction sensitivity, specificity, accuracy were 97.3%, 93.9%, 96.9%, respectively, establishing this is capable reliably predicting given its composition. power SVM classification further demonstrated by segregating polymorphs, specifically examine polymorphism in TiNiSi- ZrNiAl-type structures. Analyzing 19 are experimentally reported both types, correctly identifies, high confidence (>0.7), low-temperature polymorph from high-temperature form. Interestingly, learning also reveals certain cannot be clearly differentiated lie a "confused" region (0.3-0.7 confidence), suggesting polymorphs may observed single sample at experimental conditions. ensuing synthesis characterization TiFeP adopting structures sample, even after long annealing times (3 months), validate occurrence structural uncertainty predicted learning.