Transferable dispersion-aware machine learning interatomic potentials for multilayer transition metal dichalcogenide heterostructures

Abstract

Stacking atomically thin transition metal dichalcogenides (TMDs) into heterostructures enables exploration of exotic quantum phases, particularly through twist-angle-controlled moiré superlattices. These structures exhibit novel electronic and optical behaviors driven by atomic-scale structural reconstruction. However, studying such systems with DFT is computationally demanding due to their large unit cells and van der Waals (vdW) interactions between layers. To address this, we develop a transferable neural network potential (NNP) that includes long-range vdW corrections up to 12Å with minimal overhead. Trained on vdW-corrected DFT data for Mo- and W-based TMDs with S, Se, and Te, the NNP accurately models monolayers, bilayers, heterostructures, and their interaction with h-BN substrates. It reproduces equilibrium structures, energy landscapes, phonon dispersions, and matches experimental atomic reconstructions in twisted WS₂ and MoS₂/WSe₂ systems. We demonstrate that our NNP achieves DFT-level accuracy and high computational efficiency, enabling large-scale simulations of TMD-based moiré superlattices both with and without substrates.