Source-linked AI summary
Ab initio Modeling of MoS2/Oxide Device Interfaces with Machine Learned Electronic Structures
Manasa Kaniselvan, Mauro Dossena, Denghui Lu, Alexander Maeder, Nicolas Vetsch, Alexandros Nikolaos Ziogas, Mathieu Luisier
TL;DR
The paper addresses the cost and realism limitations of ab initio quantum-transport simulations for atomically detailed semiconductor devices. It combines an ML Hamiltonian model with a QT solver to simulate large MoS2/oxide systems, finding that realistic oxides and undercoordinated interface metals substantially affect current magnitude and propagation.
Problem
DFT-based Hamiltonian construction limits device-scale QT simulations, while simplified or periodic oxide representations do not capture atomistic interface variability.
Method
The framework uses MALOQ to predict DFT-level Hamiltonians for atomically resolved MoS2/oxide structures and QuaTrEx to compute NEGF quantum-transport observables.
Results
>10,000× speedup over DFT enables Hamiltonian generation for >20,000-atom devices, while MALOQ reproduces DFT electronic structure and transmission functions and explicit oxides reduce ON-state current by factors of 2–3.
Takeaways & Limitations
Undercoordinated Hf or Al atoms near MoS2/oxide interfaces are linked to leakage points and influence potential valleys, current magnitude, and current flow through MoS2.
Abstract
from arXiv · showhide
We introduce a new ab initio approach to simulate semiconductor devices that integrates scalable machine-learned (ML) electronic structure models with an advanced quantum transport (QT) solver. The developed framework enables 10,000X speedups over density functional theory to produce the Hamiltonian matrix of devices made of >20,000 atoms, while offering high prediction accuracy. We use its unique features to investigate MoS2/oxide samples and single-layer MoS2 field-effect transistors, where the surrounding oxide layers, here, HfO2 or Al2O3, are explicitly included into the QT domain. In particular, we reveal that the presence of undercoordinated metal atoms (Hf or Al) close to the semiconductor-oxide interface significantly affects the magnitude of the electronic current and its propagation through MoS2.
I. INTRODUCTION
MoS2 FET experiments have advanced substantially but remain below ab initio transport predictions, while realistic oxide structure and device-scale Hamiltonian construction remain central challenges.
- Experiments reported MoS2 FET milestones including EOT=0.8 nm, Lg=30 nm, W=25 nm, and ION=1,135 μA/μm at Vds=1.5 V.
- Ab initio QT simulations predict MoS2 ON-state currents exceeding 2,000 μA/μm, with one study anticipating ION >3,000 μA/μm at Vds=0.64 V including contact resistances.
- Electron–phonon interactions, interface scattering, and surface optical phonons are generally cited as explanations for the experiment–simulation gap, but modeling them requires well-parameterized scattering rates.
- Atomistic construction of surrounding amorphous dielectrics is necessary because their atomic configuration can strongly influence TMDC behavior.
- DFT preparation of device Hamiltonians scales as O(N3), while tiling small cells creates periodic interface motifs and uncontrollable oxide defect densities.
- The paper introduces an ML Hamiltonian model with O(N) complexity for atomically resolved MoS2/HfO2 and MoS2/Al2O3 stacks exceeding 20,000 atoms.
II. METHODS
MALOQ learns DFT-like Hamiltonian matrices from atomistic structures using an SO(2)-equivariant graph neural network, which then feeds a NEGF quantum-transport solver.
- MALOQ is an SO(2)-equivariant graph neural network operating in spherical tensor representations aligned with the Hamiltonian matrix.
- Its graph convolution layers predict diagonal and off-diagonal Hamiltonian blocks while preserving rotational symmetries of the atomic lattice.
- The workflow constructs MoS2 and oxide domains atom by atom before MALOQ predicts a DFT-level Hamiltonian matrix.
- QuaTrEx uses the predicted Hamiltonian in the NEGF formalism to return charge densities, electrostatic potentials, I-V characteristics, and spatial current distributions.
1) Dataset generation & training:
The training dataset combines atomistic MoS2/oxide structures with CP2K Hamiltonian and overlap labels, and the resulting model reproduces key electronic-structure and transport quantities accurately.
- Training structures are extracted from high-temperature ML-potential MD trajectories of 3.2×3.2×2.8 nm3 MoS2/oxide samples containing Mo, S, Al, Hf, and O.
- CP2K computes H_DFT and S_DFT labels using DZVP basis functions for MoS2 and SZV basis functions for the oxides.
- 38 meV average error across 16,608 eigenvalues leaves nearly exact MoS2/HfO2 eigenvalues near the conduction-band edge.
- MALOQ reproduces DFT conduction bandstructures and transmission functions for the systems of interest.
3) Generalization at scale:
MALOQ scales to large unseen MoS2/oxide structures while retaining close agreement with DFT, enabling efficient Hamiltonian generation beyond conventional DFT sizes.
- 200 node hours for a 16,704-atom DFT structure versus <0.1 node hours with MALOQ yields a >10,000× speedup for Hamiltonian generation.
- DFT generation time scales as O(N2.82), close to its O(N3) algorithmic complexity, whereas MALOQ inference has linear scaling.
- A single 25k-atom inference uses 0.1 instead of 200 node hours and amortizes dataset-generation and training costs.
- For unseen 12.8×3.2×2.8 nm3 MoS2/HfO2 and MoS2/Al2O3 structures, MALOQ preserves prediction accuracy relative to DFT.
- QuaTrEx transmission functions computed from H_ML and H_DFT agree very well across a large energy range for both oxide types.
4) Device simulations:
Explicitly modeling amorphous oxide layers in large MoS2 FET transport domains reduces current and reveals atomistic interface mechanisms. Undercoordinated Hf or Al atoms create localized effects that reshape current propagation through MoS2.
- More than 20,000-atom MoS2/oxide FET simulations explicitly included HfO2 or Al2O3 in the quantum-transport domain.The learned Hamiltonian exceeded 180,000 in size.
- A physical oxide reduces ON-state current by a factor of 2 for Al2O3 and 3 for HfO2 relative to an ideal dielectric.
- Oxide-induced electrostatic hills and valleys alter current trajectories and narrow the effective channel width compared with an ideal dielectric.Current crowds in potential valleys.
- Interface leakage points trace to undercoordinated Hf or Al atoms, including localized leakage into the oxide near an undercoordinated Hf atom.The same region shows current crowding within the MoS2 channel below the atom.
IV. CONCLUSION AND OUTLOOK
The paper presents a scalable machine-learned electronic-structure and quantum-transport workflow for atomistic MoS2/oxide devices. It combines learned Hamiltonians with quantum transport to study large systems and resolve how oxide interfaces affect electronic current.
- MALOQ predicts device Hamiltonian matrices from atomic structures, and QuaTrEx uses them with overlap matrices to compute quantum-transport observables.The workflow outputs I-V characteristics and bond-resolved current.
- The training dataset contains 50 approximately 2,000-atom MoS2/oxide structures generated with amorphous-oxide melt-quench simulations and DFT electronic structures.
- MALOQ validation shows an average Hamiltonian matrix-element error of 39 μeV and an average eigenvalue error of 38 meV, concentrated largely at high energies.The high-energy states are outside the range of interest.
- MALOQ reduces Hamiltonian-generation scaling from measured O(N2.82) for CP2K to O(N1.01), with single-shot inference instead of approximately 100 DFT SCF iterations.
- The workflow supports atomistic current-flow analysis in MoS2 devices containing more than 20,000 atoms and exposes interface effects from undercoordinated oxide metal atoms.