Machine Learning Tight-Binding Hamiltonians for Quantum Transport Simulation in Monolayer and Multilayer 2D Materials

Abstract

Quantum transport simulation of two-dimensional (2D) semiconductor materials typically relies on a pipeline of density functional theory (DFT), Wannierization, and the non-equilibrium Green's function (NEGF) formalism. The Wannierization step, which transforms the DFT band structure into a real-space tight-binding Hamiltonian, presents a practical bottleneck due to its high computational cost. This thesis adopts the machine learning tight-binding (MLTB) method as a replacement for Wannierization, and extends it to bilayer and multilayer 2D semiconductor systems through a three-stage Hamiltonian construction procedure.

The MLTB methodology is first showcased by applying it to monolayer NiN2, a recently synthesized pentagonal 2D semiconductor, providing the first characterization of its quantum transport properties. NiN2 is a direct bandgap semiconductor with pronounced in-plane transport anisotropy arising from its four-fold lattice symmetry. The anisotropic effective mass structure along the 45-degree transport direction is shown to enhance source-to-drain tunneling in p-type devices at short channel lengths, while n-type devices remain largely insensitive to transport orientation.

The MLTB framework is then extended to bilayer systems through a three-stage procedure in which independent monolayer Hamiltonians are first constructed for each layer, and interlayer coupling matrices are subsequently fitted to the bilayer DFT band structure with the monolayer blocks held fixed. This approach preserves the layer-resolved structure of the Hamiltonian required for NEGF simulation and avoids the need for a full bilayer Wannierization. The procedure is validated on bilayer MoS2, where the simulated transport characteristics show close agreement with those obtained from a Wannier-derived Hamiltonian. The method is subsequently applied to bilayer Janus MoSSe in the 2AA' and 3AA' stacking configurations. The two configurations are found to exhibit a tradeoff in device performance: the 2AA' configuration shows better DIBL but worse subthreshold swing than 3AA', attributed respectively to its smaller total channel thickness and the more centrally located electron density relative to the gate electrodes.

The results demonstrate that the MLTB-based simulation pipeline is applicable to both monolayer and multilayer 2D semiconductor systems, and provides a more automated and scalable alternative to Wannierization for quantum transport simulation of novel materials.