Machine-learning-assisted discovery of 212-Zintl-phase compounds with ultra-low lattice thermal conductivity

Abstract

Zintl-phase compounds hold immense potential for thermoelectric applications owing to their intrinsically low lattice thermal conductivity (κ_L). However, numerous 212-Zintl-phase compounds remain largely unexplored due to the challenges in assessing their thermal and electrical transport properties via traditional trial-and-error approaches. Here, we present a gradient boosting regressor (GBR) machine-learning (ML) model to predict and discover 5 unexplored and promising 212-Zintl-phase compounds with κ_L lower than 2 W (mK)⁻¹ at 300 K. The model demonstrated excellent predictive capability with a coefficient of determination (R²) of 0.988 and root mean square error (RMSE) of 0.083 W (mK)⁻¹ on the test set using tenfold cross-validation. Notably, the top-ranked compound Ba₂ZnBi₂ exhibited an ultra-low κ_L of approximately 1 W (mK)⁻¹ at 300 K, substantially lower than those of other types of Zintl-phase compounds like 122-, and 111-types. Our theoretical calculations further validated the ultra-low κ_L of Ba₂ZnBi₂, and revealed that it originates from the large three-phonon scattering rates and the low group velocities due to the weak atomic interaction in the system. Therefore, our study demonstrates the power of combining ML and first-principles calculations to rapidly identify promising candidates for thermoelectric applications.