Modular hybrid machine learning and physics-based potentials for scalable modeling of van der Waals heterostructures

Abstract

Accurately modeling the structural reconstruction and thermodynamic behavior of van der Waals (vdW) heterostructures remains a significant challenge due to the limitations of conventional force fields in capturing their complex mechanical, thermal, electronic, and tribological properties. To address these limitations, we develop a hybrid framework that combines single-layer machine-learned potential ($s$MLP) with physics-based anisotropic interlayer potential (ILP), effectively decoupling intralayer and interlayer interactions. This $s$MLP+ILP approach modularizes the modeling of vdW heterostructures like assembling LEGOs, reducing the required training configurations by at least an order of magnitude compared to the pure MLP approach, while retaining predictive accuracy and computational efficiency. We validate our framework by accurately reproducing the mechanical and thermal transport properties of graphite and bulk hexagonal boron nitride ($h$-BN), and by resolving intricate Moir\'e patterns in graphene/$h$-BN bilayer and graphene/graphene/$h$-BN trilayer heterostructures, achieving excellent agreement with experimental observations. Leveraging the developed $s$MLP+ILP approach, we reveal the stacking order-dependent formation of Moir{\'e} superlattice in trilayer graphene/$h$-BN/MoS$_2$ heterostructures, demonstrating its ability to accurately model large-scale vdW systems comprising hundreds of thousands of atoms with near $ab$ $initio$ precision. These findings demonstrate that the hybrid $s$MLP+ILP framework remarkably outperforms existing pure machine-learned or empirical potentials, offering a scalable and transferable solution for accurately and extensively modeling complex vdW materials across diverse applications, including sliding ferroelectricity, thermal management, resistive switching, and superlubric nanodevices.