Diffusion-Limited Crystal Growth of Gallium Nitride Using Active Machine Learning

Abstract

Gallium nitride (GaN) is an important semiconductor with properties that make it particularly suitable for high-temperature applications in power systems. Unfortunately, its crystalline synthesis involves an energy-intensive process requiring high temperatures and pressures. A new additive manufacturing process offers a lowertemperature alternative, but little is understood of the molecular-scale mechanisms that drive its crystallization from the melt. Traditional semiempirical force fields within a molecular dynamics (MD) simulation are typically unable to capture bond-making and -breaking, whereas ab initio MD, while more accurate, suffers from high computational expense. This paper uses a machine-learned force field based on the FLARE++ framework to mimic the liquid-phase epitaxy of GaN, simulating the diffusion of nitrogen atoms through liquid gallium to form GaN. We show that nitrogen diffusion through Ga is slow, and relatedly, there is a pronounced tendency for nitrogen to phase-segregate within liquid Ga. This leads to nitrogen being less available to react and form crystalline GaN. As a result, the predicted crystal growth at the melt/crystal interface is extremely slow, as seen experimentally. In this work, we demonstrate the potential of a MLFF to describe complex, multiphase behavior under different conditions. We also uncover the key atomistic-level mechanism of diffusion-limited GaN crystal growth, which is an important step toward further control of the additive manufacturing process.