Scale-invariant machine-learning model accelerates the discovery of quaternary chalcogenides with ultralow lattice thermal conductivity

This paper presents the development of a scale-invariant machine-learning (ML) model to discover novel quaternary chalcogenides (AMM′Q3) with ultralow lattice thermal conductivity (κ). The model, based on a crystal graph convolutional neural network, is insensitive to the input crystal structures' volumes. Screening approximately 1 million compounds for thermodynamic stability using iterative ML and DFT calculations identified 99 DFT-validated stable compounds. Calculations using the Peierls–Boltzmann transport equation revealed ultralow κ (<2 Wm⁻¹K⁻¹ at room temperature) in these compounds, attributed to soft elasticity and strong phonon anharmonicity. The study highlights the efficiency of the scale-invariant ML model in predicting novel compounds.

Koushik Pal, Cheol Woo Park, Yi Xia, Jiahong Shen, Chris Wolverton