Machine Learning-Based Predictions of Power Factor for Half-Heusler Phases

Abstract

A support vector regression model for predictions of the thermoelectric power factor of half-Heusler phases was implemented based on elemental features of ions. The training subset was composed of 53 hH phases with 18 valence electrons. The target values were calculated within the density functional theory and Boltzmann equation. The best predictors out of over 2000 combinations regarded for the p-type power factor at room temperature are: electronegativity, the first ionization energy, and the valence electron count of constituent ions. The final results of support vector regression for 70 hH phases are compared with data available in the literature, revealing good ability to determine favorable thermoelectric materials, i.e., VRhGe, TaRhGe, VRuSb, NbRuAs, NbRuBi, LuNiAs, LuNiBi, TaFeBi, YNiAs, YNiBi, TaRuSb and NbFeSb. The results and discussion presented in this work should encourage further fusion of ab initio investigations and machine learning support, in which the elemental features of ions may be a sufficient input for reasonable predictions of intermetallics with promising thermoelectric performance.