Unified deep learning network for enhanced accuracy in predicting thermal conductivity of bilayer graphene, hexagonal boron nitride, and their heterostructures.

In this research, we utilized density functional theory (DFT) computations to perform ab initio molecular dynamics simulations and static calculations on graphene, hexagonal boron nitride, and their heterostructures, subjecting them to strains, perturbations, twist angles, and defects. The gathered energy, force, and virial information informed the creation of a training set comprising 1253 structures. Employing the Neural Evolutionary Potential framework integrated into Graphics Processing Units Molecular Dynamics, we fitted a machine learning potential (MLP) that closely mirrored the DFT potential energy surface. Rigorous validation of lattice constants and phonon dispersion relations confirmed the precision and dependability of the MLP, establishing a solid foundation for subsequent thermal transport investigations. A further analysis of the impact of twist angles uncovered a significant reduction in thermal conductivity, particularly notable in heterostructures with a decline exceeding 35%. The reduction in thermal conductivity primarily stems from the twist angle-induced softening of phonon modes and the accompanying increase in phonon scattering rates, which intensifies anharmonic interactions among phonons. Our study underscores the efficacy of the MLP in delineating the thermal transport attributes of two-dimensional materials and their heterostructures, while also elucidating the micro-mechanisms behind the influence of the twist angle on thermal conductivity, offering fresh perspectives for the design of advanced thermal management materials. [ABSTRACT FROM AUTHOR]