Distilling universal machine-learning potentials for moiré lattices across one million atoms

Atomic reconstruction reshapes moiré materials across multiple scales, from local structure and polarization textures to global electronic topology, yet direct $\mathit{\text{ab}}\mathit{\text{initio}}$ modeling becomes prohibitive for large superstructures such as marginal-twist-angle moirés and moiré-of-moirés. We develop MoiréMLIP by fine-tuning a universal atomistic model on the density functional theory labeled Moiré Kaleidoscope dataset, which spans transition metal dichalcogenide compositions, symmetries, stackings, and twist angles. MoiréMLIP reproduces $\mathit{\text{ab}}\mathit{\text{initio}}$ reconstruction with force errors of 6–8,meV/Å and transfers to smaller twist angles and unseen structures. Knowledge distillation yields MoiréMLIP-mini, which retains this accuracy while extending single-GPU inference to one million atoms. Applied to an alternate-twist MoTe${}_2$ trilayer, it reveals a hierarchical polarization network spanning tens of nanometers arising from large moiré-cell distortions sharply localized along the moiré-of-moiré domain walls. These results overcome key accuracy, transferability, and scaling bottlenecks of existing atomistic models and enable predictive simulations across emergent moiré length scales.