Fingerprints of low lattice thermal conductivity compounds: half-Heusler case study and beyond

Compounds with rattling atoms can have low lattice thermal conductivity compared to structurally similar materials. The rattling effect can arise because of loose bonding of filler atoms. Identifying rattlers without in-depth analysis of vibrational modes, is a challenge. In fact, such low thermal conductivity compounds can exist in material groups where one might not suspect rattling behavior to begin with, notably the half-Heusler group [Phys. Rev. B 101, 064301 (2020)] where a few of the compounds contain rattlers. In this case study, we first show that the known rattling half-Heuslers can be identified from their structural and chemical properties through machine learning using random forest regression. Careful feature selection is used to ensure high reliability in the prediction and improve feature interpretability. Finally, we illustrate how the machine-learning based intuition can aid in the identification of low lattice thermal conductivity compounds in broader classes of systems.