Accelerated discovery of extreme lattice thermal conductivity by crystal graph attention networks and chemical bonding

Designing materials with targeted lattice thermal conductivity (LTC) demands electronic-level insight into chemical bonding. We introduce two bonding descriptors, namely normalized negative integrated COHP (-ICOHP) and normalized integrated COBI, that correlate strongly with LTC and rattling (mean-squared displacement), surpassing empirical rules and the unnormalized −ICOHP across >4500 inorganic crystals by first-principles. We train a crystal attention graph neural network (CATGNN) to predict these descriptors and screen ~200,000 database structures for extreme LTCs. From 367 (533) candidates with low (high) normalized -ICOHP and normalized ICOBI, first-principles validation identifies 106 dynamically stable compounds with LTC < 5 W m − 1 K − 1 (68% <2 W m − 1 K − 1 ) and 13 stable compounds with LTC > 100 W m −1 K − 1 . The descriptors’ low cost and clear physical meaning provide a rapid, reliable route to high-throughput discovery and inverse design of crystalline materials with ultralow or ultrahigh LTC for applications in thermal insulation, thermoelectrics, and electronics cooling.