Learning the Electronic Hamiltonian of Large Atomic Structures

Chen Hao Xia, Manasa Kaniselvan, Alexandros Nikolaos Ziogas, Marko Mladenović, Rayen Mahjoub, Alexander Maeder, Mathieu Luisier

Abstract

Graph neural networks (GNNs) have shown promise in learning the ground-state electronic properties of materials, subverting ab initio density functional theory (DFT) calculations when the underlying lattices can be represented as small and/or repeatable unit cells. Realistic systems are, however, non-ideal and generally characterized by higher structural complexity, requiring large unit cells and thousands of atoms. At these scales, DFT becomes computationally prohibitive. In this work, we present a strictly local equivariant GNN capable of learning the electronic Hamiltonian (H) of realistically extended materials. It incorporates an augmented partitioning approach that enables training on arbitrarily large structures while preserving local atomic environments beyond boundaries. We demonstrate its capabilities by predicting the electronic Hamiltonian of various systems with up to 3,000 nodes, 500,000+ edges, and $\leq$0.53% error in the eigenvalue spectra.

Introduction

Understanding a material's electronic properties, such as conductivity or bandgap, requires calculating its ground-state Hamiltonian matrix. This matrix encodes quantum mechanical couplings between atomic orbitals. Traditionally, Density Functional Theory (DFT) is the gold standard for these calculations. However, DFT suffers from cubic scaling ($O(N^3)$) with the number of atoms, making it impractical for large, realistic systems containing defects, interfaces, or amorphous structures.

While GNNs have emerged as a faster alternative, they have primarily been applied to small molecules or periodic crystals. Scaling GNNs to massive, disordered structures presents a memory bottleneck. Standard graph partitioning methods often sever atomic bonds at boundaries, losing critical local physical context.

Methodology: Equivariant GNNs and Augmented Partitioning

The authors propose a strictly local, equivariant GNN that respects physical symmetries. If a 3D atomic structure is rotated, the predicted Hamiltonian matrix transforms accordingly. The architecture utilizes efficient SO(2) convolutions and multi-headed attention to model inter-atomic interactions.

To address the memory constraints of large-scale graphs, the researchers introduced an augmented partitioning approach. This method involves slicing the massive graph into sub-graphs suitable for GPU memory. To maintain physical accuracy at the boundaries, each slice is padded with "virtual" nodes and edges. These virtual components provide the necessary local geometry for boundary atoms without contributing to the loss function during training. This allows for independent, parallel processing across multiple GPUs.

Overview of the equivariant GNN architecture and the augmented partitioning scheme.

A key physical insight driving this approach is quantum "near-sightedness": Hamiltonian matrix elements decay exponentially with distance. By focusing on a local radius (8–12 Angstroms), the GNN captures essential physics without the need for computationally expensive long-range message passing.

Experimental Evaluation

The model was evaluated on custom datasets of large, amorphous materials generated via molecular dynamics simulations, including a-HfO2 (up to 3,000 atoms), a-GST (1,008 atoms), and a-PtGe (2,688 atoms). It also demonstrated zero-shot compositional generalization on sub-stoichiometric HfOx, accurately predicting properties for varying oxygen vacancy concentrations without retraining.

Large-scale amorphous structures used for evaluation: a-HfO2, a-GST, and a-PtGe.

Key Results

  • Scalability: Successfully predicted Hamiltonians for systems with ~28 million orbital interactions.
  • Accuracy: Achieved prediction errors between 2.17–2.58 meV. For a 3,000-atom HfO2 structure, the relative error in the eigenvalue spectrum was $\le 0.53\%$.
  • Efficiency: The augmented partitioning method provided a 6.5x speedup and 7.2x memory reduction per GPU compared to full-graph training. Inference is orders of magnitude faster than DFT.
  • Application: The GNN-predicted Hamiltonian was used in quantum transport simulations of a memory device, successfully tracking electrical current changes across five orders of magnitude.

Conclusion and Future Work

This work expands the applicability of GNNs to complex, large-scale materials science problems. While the current spatial slicing can lead to load imbalances across GPUs, future iterations could integrate more advanced graph-partitioning algorithms. The framework's ability to learn rotationally-equivariant properties also opens doors for predicting other physical characteristics, such as phonon dispersions.

Summary updated

Abstract

Graph neural networks (GNNs) have shown promise in learning the ground-state electronic properties of materials, subverting ab initio density functional theory (DFT) calculations when the underlying lattices can be represented as small and/or repeatable unit cells (i.e., molecules and periodic crystals). Realistic systems are, however, non-ideal and generally characterized by higher structural complexity. As such, they require large (10+ Angstroms) unit cells and thousands of atoms to be accurately described. At these scales, DFT becomes computationally prohibitive, making GNNs especially attractive. In this work, we present a strictly local equivariant GNN capable of learning the electronic Hamiltonian (H) of realistically extended materials. It incorporates an augmented partitioning approach that enables training on arbitrarily large structures while preserving local atomic environments beyond boundaries. We demonstrate its capabilities by predicting the electronic Hamiltonian of various systems with up to 3,000 nodes (atoms), 500,000+ edges, ~28 million orbital interactions (nonzero entries of H), and $\leq$0.53% error in the eigenvalue spectra. Our work expands the applicability of current electronic property prediction methods to some of the most challenging cases encountered in computational materials science, namely systems with disorder, interfaces, and defects.
BibTeX
@inproceedings{xia2025learning,
  author = {Chen Hao Xia and Manasa Kaniselvan and Alexandros Nikolaos Ziogas and Marko Mladenović and Rayen Mahjoub and Alexander Maeder and Mathieu Luisier},
  title = {{Learning the Electronic Hamiltonian of Large Atomic Structures}},
  booktitle = {International Conference on Machine Learning},
  year = {2025},
  url = {https://openreview.net/forum?id=WGejWCgrpD},
}
Figure