neurips-dataset/QHMat_tar

QHMat Dataset

QHMat is a large-scale quantum Hamiltonian dataset for materials comprising 105,260 crystal structures spanning 75 elements. It contains crystal structure features together with Hamiltonian/overlap tensor blocks, serialized as torch_geometric.data.Data objects.

The current release provides a ~50 GiB subset suitable for model development and integration tests. The full dataset (~4.1 TB) is planned to be uploaded after peer review.

Dataset Structure

Files in this release

File / pattern	Description
train-000000.tar, train-000001.tar, ...	WebDataset tar shards containing sample payloads.
XXXXXXXX.pkl	Pickled torch_geometric.data.Data object (payload from LMDB value).
XXXXXXXX.json	Sidecar JSON with lmdb_key and value_bytes.
train-manifest.csv	Sample index table: shard path + member names + source LMDB key.

Splits

• Single split under train-* shards.

Example record fields

Typical decoded Data keys include (exact keys can vary by record):

• Geometry/graph: pos, atoms, lattice, multi_edge_index, multi_edge_vec, lattice_translation_vector
• Hamiltonian/overlap blocks: diagonal_hamiltonian, off_diagonal_hamiltonian, optional overlap tensors and masks
• Metadata: database_idx, mp_id, raw_basename, fermi_level, elapsed_time

Minimal load example (single sample)

Use the same Python environment as training (any env where torch and torch_geometric import correctly).

import json
import pickle
import tarfile

import torch  # noqa: F401

tar_path = "train-000000.tar"
with tarfile.open(tar_path, "r") as tar:
    pkl_info = tar.getmember("00000000.pkl")
    pkl_bytes = tar.extractfile(pkl_info).read()
    meta = json.loads(tar.extractfile(tar.getmember("00000000.json")).read())

data = pickle.loads(pkl_bytes)
print(meta)
print(data)

Uses

Direct Use

• Training and evaluating graph neural networks for Hamiltonian / electronic-structure prediction.
• Data loader validation and distributed training pipeline setup.
• Reproducibility checks against LMDB-originated records.

Out-of-Scope Use

• Safety-critical or high-stakes decision workflows without domain validation.
• Claims of complete physical/materials coverage from this subset alone.

Dataset Creation

Data Collection and Processing

1. A large LMDB (~4.1 TB) is produced from electronic-structure workflows of OpenMX.
2. A ~50 GiB LMDB subset is constructed by ID-list-driven extraction.
3. train-manifest.csv is emitted for deterministic mapping and audit.

Personal and Sensitive Information

No personal data is expected. IDs such as mp_id represent materials-system identifiers.

Bias, Risks, and Limitations

• This is a subset release, not the full corpus yet.
• Data reflects simulation settings and source pipeline assumptions.
• Pickle-based payloads require environment compatibility (torch, torch_geometric).

Recommendations

• Use the same runtime stack as training/preprocessing.
• Pin package versions for reproducibility.
• Validate model behavior on broader/full releases once available.