Machine learning prediction of thermal and elastic properties of double half-Heusler alloys.

Double half-Heusler alloys are promising materials for applications as magnetocaloric materials, topological insulators, but especially thermoelectric materials. Four different elements in their composition provide a wide range of possible compositions, which, on the other hand, is difficult to study directly by applying traditional first-principles approaches to large number of compositions. In this work, based on the gradient boosting method, regression models are constructed that allow rapid prediction of the lattice thermal conductivity, as well as a number of other thermal and elastic properties, based on the composition and crystal structure of a compound. This made it possible for the first time to calculate the lattice thermal conductivity, as well as Grüneisen parameter, Debye temperature, and elastic moduli for a number of double half-Heusler compounds. We observe that the predicted thermal conductivity is in better agreement with the experimental data than the results of density functional theory calculations available in the literature. Half-Heusler compounds with thermal conductivity values lower than those previously known have been found. In addition, we have analyzed the importance of various features for predicting each of the studied properties, and the effect of the crystallographic symmetry of the compound on the prediction accuracy. • Regression models predicting thermal conductivity and elastic moduli are built. • Thermal conductivity of 90 novel double half-Heusler alloys is predicted. • Quantitative agreement with available experimental data is obtained. • The double half-Heuslers with thermal conductivity lowest in the class are revealed. • Half Heusler alloys with the lowest thermal conductivity include Mg, Pd or Pt. [ABSTRACT FROM AUTHOR]