We present a unified machine-learning-based framework to investigate the electronic structure and plasmonic response of graphene/hexagonal boron nitride (hBN) moiré heterostructures. This framework integrates machine-learning-derived interatomic potentials and a data-driven Hamiltonian model, facilitating a fully ab initio-trained and parameter-free strategy that effectively links atomic relaxation processes to electronic property predictions. By applying this methodology, we conduct a systematic exploration of the graphene/hBN moiré heterostructures at four representative twist angles, revealing distinct out-of-plane corrugations in the relaxed structures predicted by the machine-learned potential. Our findings uncover a twist-angle-dependent reduction in the Dirac-point gap within the electronic bands. Additionally, the calculated plasmon spectra exhibit pronounced electron-hole asymmetry due to the presence of the secondary Dirac-point gap, consistent with previous theoretical expectations. Importantly, our approach eliminates the necessity for empirical parameter fitting, thus enhancing its applicability to a broader range of two-dimensional heterostructures with complex interlayer geometries. This work paves the way for a generalized framework that enables automated, high-accuracy modeling of moiré systems, effectively integrating ab initio data, machine learning, and the study of collective excitations.

Close

PrevNext

View figure in article

• Plasmons
• Graphene
• Hexagonal boron nitride
• Twisted heterostructures
• Density functional calculations
• Machine learning