Developing accurate and efficient machine-learning potentials to bridge the gap between theory and experiments for understanding short-range order in complex semiconductor alloys

Group-IV and III-nitride semiconductor alloys hold great promise for various applications, including electronics, optoelectronics, sensors, topological quantum technologies, solid-state lighting, solar cells, and ferroelectric devices. However, understanding the complex behavior of these alloys, particularly their short-range order (SRO), presents significant challenges. Accurate modeling using density functional theory (DFT) is often constrained by spatiotemporal limitations, leading to a gap between theoretical models and experimental characterizations. To bridge this gap, we develop machine-learning interatomic potentials based on a neuroevolution potential (NEP) approach combined with high-quality DFT datasets. The developed NEPs achieve near-DFT accuracy and EAM-like efficiency in atomistic simulations, enabling side-by-side comparisons with advanced characterization techniques such as atom probe tomography (APT) and four-dimensional scanning transmission electron microscopy (4D-STEM), providing deeper insights into the nature of SRO in these semiconductor alloys. Our findings indicate the potential of atomic-order-based engineering as a third degree of freedom beyond composition and strain.