README.md

metadata

dataset_info:
  features:
    - name: nr_atoms
      dtype: int64
    - name: atomic_types_z
      sequence: int32
    - name: atomic_positions
      sequence:
        sequence: float64
    - name: lattice_nsc
      sequence: int32
    - name: lattice_origin
      sequence: float64
    - name: lattice_vectors
      sequence:
        sequence: float64
    - name: boundary_condition
      sequence:
        sequence: int32
    - name: h_matrix
      sequence:
        sequence: float64
    - name: s_matrix
      sequence:
        sequence: float64
  splits:
    - name: nr_atoms_32
      num_bytes: 33281328
      num_examples: 12
    - name: nr_atoms_64
      num_bytes: 2360995092
      num_examples: 213
    - name: nr_atoms_2
      num_bytes: 3644732
      num_examples: 323
    - name: nr_atoms_3
      num_bytes: 4090160
      num_examples: 164
    - name: nr_atoms_8
      num_bytes: 104429660
      num_examples: 599
  download_size: 2290540058
  dataset_size: 2506440972
configs:
  - config_name: default
    data_files:
      - split: nr_atoms_32
        path: data/nr_atoms_32-*
      - split: nr_atoms_64
        path: data/nr_atoms_64-*
      - split: nr_atoms_2
        path: data/nr_atoms_2-*
      - split: nr_atoms_3
        path: data/nr_atoms_3-*
      - split: nr_atoms_8
        path: data/nr_atoms_8-*
license: mit
tags:
  - chemistry
  - materials
pretty_name: aBN

Card Description for Dataset

---

Dataset Name: Atomic Structures and H/S Matrices of aBN Configurations Description: This dataset contains computationally generated atomic structures of amorphous boron nitride (aBN) with various configurations containing 2, 3, 8, 32, and 64 atoms per unit. Each structure is described by its atomic positions, lattice properties, and associated Hamiltonian (H) and overlap (S) matrices, which are commonly used in quantum mechanical simulations and electronic structure calculations.

---

Abstract

---

We introduce HForge, a machine learning (ML) framework for predicting Hamiltonian (H) and Overlap (S) matrices directly from atomic structures, with a focus on amorphous boron nitride (aBN) and hexagonal boron nitride (hBN). Leveraging graph-based descriptors derived from the MACE [1] model and reference Hamiltonians computed via Siesta, HForge enables efficient electronic structure predictions. In this poster, we present how the choice of training structures impacts model performance and demonstrate that incorporating a diverse set of smaller structures significantly enhances the model’s ability to generalize to larger systems—a key strategy, given that training is 3–4 times more expensive than inference. We evaluate both equivariant and non-equivariant ML architectures, showing that equivariant models better preserve the physical symmetries of quantum interactions and outperform their non-equivariant counterparts in extrapolation tasks. Building on recent advancements in equivariant [2] graph-based atomic environment representations and universal message passing, our findings underscore the potential of scalable, ML-driven Hamiltonian prediction to accelerate classical DFT computations and enable quantum simulations of materials like aBN and hBN.

---

Poster:

---

Features:

1. nr_atoms: Number of atoms in the structure (2, 3, 8, 32, 64).
2. atomic_types_z: Atomic types represented by their atomic numbers.
3. atomic_positions: Cartesian coordinates of the atoms.
4. lattice_nsc: Lattice scaling factors.
5. lattice_origin: Origin of the lattice system.
6. lattice_vectors: Lattice vectors defining the cell geometry.
7. boundary_condition: Boundary conditions applied to the system (e.g., periodic).
8. h_matrix: Hamiltonian matrix representing the quantum mechanical interactions.
9. s_matrix: Overlap matrix describing the basis set overlap.

---

Use Cases: This dataset is suitable for research and applications in materials science, quantum chemistry, and machine learning for materials discovery. It can be used for:

• Developing and validating machine learning models for predicting electronic properties.
• Benchmarking models for predicting H and S matrix base on the atomic structure.
• Understanding the structural and electronic properties of amorphous boron nitride at different scales.

---

Splits: The dataset is organized into splits based on the number of atoms per structure:

• nr_atoms_2
• nr_atoms_3
• nr_atoms_8
• nr_atoms_32
• nr_atoms_64

Each split corresponds to structures with the specified number of atoms, enabling targeted analysis or model training.

---

Licensing: Specify the license under MIT license.

Citation: If using this dataset, please cite the source or include the following acknowledgment: " xxxxx"

---