Item

Abstract

In this thesis, we perform high-throughput screening calculations for the thermal conductivity (κ) at 300 K of 152 materials using first-principles methods. Due to the various approximations involved in these calculations, especially the low-order treatment of anharmonicity, a first set of 49 calculations is used to validate the computational approach. In general, this shows that the calculations agree relative well with experiments, but also that stronger anharmonic effects indeed lead to increased prediction errors for thermal insulators. In a second step, 103 additional materials are investigated, 83 of which are found to be potential thermal insulators with κ < 10W/mK. Since the degree of anharmonicity affects the accuracy of our computational predictions, we further investigate to which extent anharmonicity can be quantitatively measured. To do this, we compare between three metrics for anharmonicity (γ, σ^A and σ^A_os) and discuss their correlation with thermal conductivity. While σ^A and σ^A_os show a relatively promising correlation with κ, the Grüneisen parameters γ do not. To refine this finding and to find even better descriptors for the thermal conductivity, we eventually use a machine-learning based symbolic regression approach, i.e., the SISSO (sure-independence screening and sparsifying operator) method. By applying it to our data, we can show that the models identified by the SISSO approach show promising predictive accuracy and thus have the potential to facilitate the discovery of thermal insulators in future.