Programme

P.92 -- Posters, VCTP-51

Date: Thursday, 6 August 2026

Time: 09:30 - 10:30

Machine-Learning Hamiltonian Approach to Electronic Structures and Moiré Physics in Twisted Transition Metal Dichalcogenides

Thi Phuong Thao Nguyen¹ and Emi Minamitani¹

(1) SANKEN, The University of Osaka, 8-1 Mihogaoka, Ibaraki, Osaka, 567-0047, Japan

Twisted transition metal dichalcogenides (TMDs) have emerged as a promising platform for investigating moiré quantum phenomena due to their intrinsic semiconducting band gaps, strong spin-orbit coupling (SOC), and valley-dependent electronic properties [1]. However, first-principles calculations of large moiré supercells remain computationally demanding, particularly when considering SOC, heterostructures, and structural perturbations. In this work, we apply a machine-learning-based Hamiltonian model using graph neural network (GNN) techniques [2] to efficiently investigate the electronic properties of twisted TMD systems with near-density-functional-theory (DFT) accuracy. The Hamiltonian model is trained on a dataset of bilayer structures with varying interlayer translations and separations and subsequently applied to twisted bilayer MoS₂, WSe₂, and their heterobilayers. Benchmark comparisons demonstrate excellent agreement between the GNN Hamiltonian and DFT calculations. By systematically varying the twist angle, we identify the emergence of moiré-induced flat bands in twisted bilayer MoS₂, with ultraflat bands appearing at small twist angles around 3.5°, accompanied by strong spatial localization of electronic states. Furthermore, we investigate the role of spin-orbit coupling and show that SOC-induced valley splitting is substantially larger in W-based systems than in Mo-based systems and becomes strongly suppressed in low-energy moiré bands at small twist angles. In twisted MoS₂/WSe₂ heterobilayers, we find that the electronic structure is qualitatively distinct from that of twisted homobilayers. The intrinsic band offset between MoS₂ and WSe₂ induces strong layer polarization of the low-energy electronic states, while the lattice mismatch introduces strain as an additional tuning parameter for moiré electronic properties. As a result, the heterobilayer exhibits a reduced band gap and enhanced tunability compared with the corresponding homobilayer systems. Our results demonstrate that machine-learning Hamiltonians provide an efficient and accurate framework for exploring the interplay among moiré physics, spin-orbit coupling, layer polarization, and structural degrees of freedom in large-scale twisted TMD systems, opening new opportunities for studying correlated and topological phenomena in moiré quantum materials. [1] F. Wu, T. Lovorn, E. Tutuc, I. Martin, and A. H. Mac-Donald, Phys. Rev. Lett. 122, 086402 (2019). [2] Y. Zhong, H. Yu, M. Su, X. Gong, and H. Xiang, npj Computational Materials 9, 182 (2023).

Presenter: Nguyễn Thị Phương Thảo