Unveiling bonding heterogeneity-driven anharmonicity and ultralow lattice thermal conductivity in NbSe2Br2: a machine learning accelerated discovery

Transition metal chalcogenide halides (TM-Ch-X) compounds with significant heterogeneity in their chemical bonding have immense potential for thermoelectric applications. Their mixed ionic-covalent bonding nature, combined with intrinsic low lattice symmetry, provides a favorable platform for achieving strong lattice anharmonicity and ultralow lattice thermal conductivity. In this work, we developed a temperature-included crystal graph convolutional neural network (CGCNN) to accurately predict mode-resolved Grüneisen parameters, a key descriptor of lattice anharmonicity. Using this approach, two-dimensional NbSe2Br2 is identified as a thermoelectric candidate with strong anharmonicity and ultralow lattice thermal conductivity. First-principles results reveal that the strong anharmonic lattice dynamics originate from its weak and heterogeneous chemical bonding, further leading to ultralow lattice thermal conductivity. NbSe2Br2 also exhibits favorable electronic transport behavior, resulting in a maximum ZT of 1.63. Our work provides a theoretical understanding on the origin of low lattice thermal conductivity in TM-Ch-X compounds with bonding heterogeneity and should encourage further exploration of potential thermoelectric materials.