Deep Learning-Based Quantum Transport Simulations in Two-Dimensional Materials

Jian‐Ping Zou, Zhanghao Zhouyin, Qiangqiang Gu, Shishir Kumar PandeyarXiv 2025 · arXiv

TL;DR

This paper introduces DeePTB-NEGF, a novel framework that leverages deep learning to predict quantum transport properties of 2D materials with unprecedented speed and accuracy. By learning tight-binding Hamiltonians directly from first-principles data, it achieves orders-of-magnitude speedup over traditional methods, paving the way for large-scale exploration and design of next-generation electronic devices.

What Problem Does This Paper Address?

Accurately simulating the quantum transport behavior of materials is crucial for designing advanced nanoelectronic devices. While established methods like Density Functional Theory (DFT) coupled with the Non-Equilibrium Green's Function (NEGF) formalism provide high accuracy, they are computationally very expensive. This high cost severely limits their application to large-scale systems or high-throughput screening of new materials. The challenge, therefore, is to develop methods that can achieve first-principles accuracy at a significantly reduced computational cost, enabling broader exploration and faster design cycles.

Methods

The core of the DeePTB-NEGF framework lies in its two main components:

  1. DeePTB (Deep Learning Tight-Binding): This deep learning model learns to predict tight-binding (TB) Hamiltonians. It takes local atomic environments as input and outputs corrections to Slater-Koster integrals, which are then used to construct the TB Hamiltonian. This approach aims to capture the physics of the system with high fidelity while maintaining the sparsity of the Hamiltonian, crucial for efficiency. The neural networks are trained to map atomic environments to Hamiltonian matrix elements.
  2. DPNEGF (Deep Learning-accelerated Quantum Transport Package): This is an open-source package that implements the NEGF formalism for quantum transport simulations. The DeePTB-predicted TB Hamiltonians are integrated into DPNEGF. The NEGF formalism is used to compute the retarded Green's function, from which the transmission spectrum (a key indicator of transport properties) is derived.

The workflow involves training DeePTB on first-principles data and then using the trained model to generate Hamiltonians for transport simulations within DPNEGF.

Datasets

The paper uses first-principles calculations as the source of "ground truth" data for training and validation. Specifically:

  • Graphene and h-BN: DFT calculations were performed using the SIESTA package.
  • MoS₂: DFT calculations were performed using the ABACUS package.

The exchange-correlation effects were described using the Perdew-Burke-Ernzerhof (PBE) functional within the generalized gradient approximation (GGA). A vacuum spacing of over 15 Å was used to avoid spurious interlayer interactions. The DFT band structures served as training labels for DeePTB. Transmission spectra were computed using DPNEGF and compared against conventional DFT-NEGF results.

Key Insights and Ideas

  • Deep Learning as a Hamiltonian Generator: The paper demonstrates that deep learning can effectively learn the complex mapping from atomic structure to the parameters of a tight-binding Hamiltonian, bypassing the need for computationally intensive self-consistent DFT calculations for every simulation.
  • Bridging the Accuracy-Efficiency Gap: DeePTB-NEGF successfully bridges the gap between the high accuracy of DFT-NEGF and the need for computational efficiency, making large-scale quantum transport simulations feasible.
  • Generality Across Material Types: The framework's ability to accurately model graphene (semimetal), h-BN (insulator), and MoS₂ (semiconductor) highlights its broad applicability across the spectrum of 2D material electronic behaviors.
  • Decoupling Electronic Structure from Transport: The framework allows for the use of DeePTB-predicted Hamiltonians from DFT codes that lack native NEGF capabilities (like ABACUS for MoS₂), significantly expanding the range of materials that can be studied for transport properties.

Key Results

  • Accuracy: DeePTB-NEGF shows excellent agreement with conventional DFT-NEGF results for band structures and transmission spectra for graphene, h-BN, and MoS₂.
  • Efficiency: The framework achieves an order-of-magnitude speedup in simulation time. For example, for graphene, DFT-NEGF takes 3024 seconds, while DeePTB-NEGF completes the task in 56 seconds. For h-BN, the times are 3521 seconds and 59 seconds, respectively.
  • Reproducibility: The method accurately reproduces the characteristic features of each material, such as the Dirac cone in graphene, the large band gap in h-BN, and the direct band gap in MoS₂.

Limitations and Future Work

  • Training Data Preparation: While the inference step is fast, the initial preparation of training data and the training process for DeePTB itself require significant computational resources.
  • Scalability for Very Large Systems: The paper notes that the efficiency gains become even more pronounced for larger systems, as the computational cost of traditional self-consistent calculations increases drastically with system size. However, the exact scaling limits and performance for extremely large systems are not exhaustively detailed.
  • Future Work: The authors suggest that DeePTB-NEGF opens avenues for large-scale exploration and device design in 2D materials, implying future work will focus on applying this framework to discover new materials and optimize device architectures. The ability to decouple electronic structure calculations from transport simulations also suggests potential for integration with a wider array of DFT codes.

Summary updated

Abstract

Two-dimensional (2D) materials exhibit a wide range of electronic properties that make them promising candidates for next-generation nanoelectronic devices. Accurate prediction of their quantum transport behavior is therefore of both fundamental and technological importance. While density functional theory (DFT) combined with the non-equilibrium Green$'$s function (NEGF) formalism provides reliable insights, its high computational cost limits applications to large-scale or high-throughput studies. Here we present DeePTB-NEGF, a framework that combines a deep learning-based tight-binding Hamiltonians derived learned directly from first-principles calculations (DeePTB) with efficient quantum transport simulations implemented in the DPNEGF package. To validate the method, we apply it to three prototypical 2D materials: graphene, hexagonal boron nitride (h-BN), and MoS$_2$. The resulting band structures and transmission spectra show excellent agreement with conventional DFT-NEGF results, while achieving orders-of-magnitude improvement in efficiency. These results highlight the capability of DeePTB-NEGF to enable accurate and efficient quantum transport simulations, thereby opening avenues for large-scale exploration and device design in 2D materials.
BibTeX
@misc{zou2025deep,
  author = {Jian‐Ping Zou and Zhanghao Zhouyin and Qiangqiang Gu and Shishir Kumar Pandey},
  title = {{Deep Learning-Based Quantum Transport Simulations in Two-Dimensional Materials}},
  year = {2025},
  eprint = {2512.11291},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.11291},
  howpublished = {arXiv},
}
Figure