Composite verifiable ML potentials for moiré materials

Atomic relaxation in large-scale moiré materials presents a unique challenge due to the intricate balance between strong intralayer covalent bonds and weak interlayer van der Waals forces. This work presents a dataset-driven approach to training machine-learned interatomic potentials (MLIPs) by explicitly separating the dataset and training of intralayer and interlayer potentials. Our composite MLIPs trained on this split dataset approach outperform those trained on a unified dataset approach by a factor of ~10x in capturing the interlayer interactions that drive moiré reconstructions. Our dataset-generation and model-splitting methodology is generalizable to classical and machine-learned force fields, offering a robust framework for reconstructing small-angle moiré materials. We demonstrate the validity of this approach in MoS2-WSe2 heterobilayers and predict small angle reconstruction in novel 2D moiré structures.