Beyond the Virtual Crystal Approximation: Machine Learning Potentials and Hydrodynamic Theory for Lattice Thermal Conductivity in Ga_xIn_1-xAs

Ga_xIn_1-xAs is a promising material used in transistors and optoelectronic devices due to its high electron mobility and modulation of the direct bandgap. Inherent to its disordered nature, this alloy is also a low lattice thermal conductivity material, which makes it a compelling candidate for thermoelectric applications.  However, predictive modeling of thermal transport in alloys remains a challenge. The virtual crystal approximation, which successfully combines density functional theory and lattice dynamics for alloys like Si_1-xGe_x, breaks down due to the differences in the chemical bonding between Ga and In with As in this system.

In this work, we overcome this limitation by using machine learning potentials, including one that is specifically designed to treat III-V semiconductors, to compute the phonon properties and heat transport in Ga_xIn_1-xAs alloy. We first establish the fidelity of our approach by benchmarking against the phonons and elastic properties of the binary end-members (GaAs and InAs). For the random alloy, we go beyond the Quasi-Harmonic Green-Kubo approach by incorporating hydrodynamic corrections to accurately capture the physics of low-frequency modes. This effectively corrects for the spurious finite-size effects in the lattice dynamics of disordered systems and enables accurate thermal conductivity predictions.