Machine-Learning Interatomic Potentials for Twisted Moiré Materials

Simulating atomic reconstruction in twisted moiré materials poses a major challenge due to their large supercells and complex interlayer relaxations. We present a machine-learning interatomic potential framework designed to bridge the gap between accuracy and scalability in modeling moiré superlattices. By integrating a pretrained potential with minimal fine-tuning data, our approach reproduces key structural and energetic features across twist angles and stacking configurations. A hierarchical training and compression strategy further enables efficient deployment on large supercells, while maintaining sufficient accuracy and transferability. The resulting model enables systematic comparison of atomic-scale reconstruction across twist angles, layer numbers, and material combinations, offering a scalable route for exploring the moiré design space in two-dimensional systems.