Screening New Quaternary Semiconductor Heusler Compounds By Machine-Learning Methods

Abstract

The following work presents a theoretical workflow to accelerate the discovery of new quaternary Heusler compounds for thermoelectric applications. The process consists of several steps: (i) construction of a consistent home-made DFT learning database with a limited set of compounds mainly involving unary, binary, and ternary configurations; (ii) machine-learning regression to estimate the heat of formation for all possible arrangements of atoms in the 4 crystallographic sites of the Heusler phase; (iii) classification learning to predict the semiconductor (SC) feature; (iv) verification of stable SC compounds through a convex hull analysis of the ground state for each quaternary system; and (v) phonon calculation to check the mechanical stabilities of the final candidates. From a selection of 24 chemical elements, 13 272 unique DFT calculations were performed among all 244 = 331 776 configurations. The learning and screening process led to predict the properties of all quaternaries (24 × 23 × 22 × 21 = 255 024) along with the discovery of 8 new stable semiconducting compounds (TaAlCoMn, TaSiFeMn, NbAlCoMn, NbSiFeMn, ···), promising potential interest in thermoelectric properties.