Predicting Phonon Transport in Disordered Alloys from a Highly Accurate Machine Learning Interatomic Potential

The II₂IV family of materials, such as Mg₂Si, Mg₂Sn, Sr₂Si, and Sr₂Ge, among others, are highly regarded as promising high-performance thermoelectric materials. In our previous research, we calculated the maximum figure of merit ZT for the promising II-IV family thermoelectric compounds Sr₂Si and Sr₂Ge, yielding values of 1.15 and 1.44 at 900 K through first-principles calculations. To improve thermoelectric performance, the common practice involves alloying to reduce lattice thermal conductivity and enhance the Seebeck coefficient. Nevertheless, determining the optimal alloy ratios through first-principles calculations can be quite challenging because disordered effects require a large supercell in the computations. Here, we introduce a highly accurate machine learning interatomic potential (MLIP) for Sr₂Si_1-xGe_x disordered alloys. This MLIP is created through a machine learning technique trained on first-principles density functional theory (DFT) data, and it attains accuracy levels comparable to those achieved with DFT. This approach empowers us to carry out efficient molecular dynamics simulations for entire alloy concentration 0 ≤ x ≤ 1 in Sr₂Si_1-xGe_x and make accurate thermal property predictions. Our work provides a solution to explore compositions that offer the most potential for high-performance thermoelectric disordered alloys while assessing the contributions of phonon modes to phonon transport.