Density Functional Theory-Derived General-Purpose Machine Learning Interatomic Potential for Monolayer and Bilayer Transition-Metal Dichalcogenides

Small-twist angle bilayer graphene and transition metal dichalcogenides (TMDs) lead to the formation of large-area moiré superlattices. The flat electronic bands and novel moiré excitons in TMD moiré superlattices are closely linked to the structural rearrangement of the constituent atoms.  However, due to the large number of atoms in small-twist angle unit cells, atomic relaxations using density functional theory (DFT) are computationally prohibitive.  Recently, Kolmogorov−Crespi (KC) interatomic potentials have been developed that can efficiently distinguish different stacking interlayer interactions. However, this approach requires refitting KC parameters for each distinct bilayer TMD. Here, we develop a robust general-purpose neural network potential for a wide range of monolayer and bilayer TMDs derived from DFT calculations. This NNP reaches the DFT accuracy for atomic relaxations of large-scale moiré superlattices, and we further use it to compute vibrational and thermal properties of a variety of twisted bilayer TMDs.

* This work is supported by DOE, and computational resources are provided by NERSC