hermanhugging/qmof_project

QMOF & ODAC23 MOF Datasets

Pre-processed LMDB datasets for training graph neural networks on Metal-Organic Framework (MOF) property prediction. Covers two source datasets — QMOF (electronic band gap) and ODAC23 IS2R (CO₂ / H₂O adsorption energy) — prepared for three model architectures: CGCNN, MGT, and PMT.

Note: The Hugging Face Dataset Viewer is not available for this repository because the archives are not in WebDataset format. Download and extract locally.

---

Source Datasets

Dataset	Task	Target	Structures	Source
QMOF	Electronic band gap regression	outputs.pbe.bandgap (eV)	~20,372 MOFs	Rosen et al., Matter 2021
ODAC23 IS2R — CO₂	Adsorption energy regression	y_relaxed (eV)	~90k entries	Sriram et al., arXiv 2307.01547
ODAC23 IS2R — H₂O	Adsorption energy regression	y_relaxed (eV)	~70k entries	Sriram et al., arXiv 2307.01547

QMOF provides DFT-computed band gaps (PBE level) for ~20k MOFs. ODAC23 provides relaxed adsorption energies for MOF–adsorbate systems, targeting direct air CO₂ capture. Both datasets share the same MOF framework structures (ODAC23 is sourced from QMOF).

---

Repository Structure

qmof_project/
├── data/
│   ├── QMOF_Full.tar                          # Full QMOF dataset (CIF + metadata), 376 MB
│   ├── ODAC_init.tar                          # ODAC23 initial (unrelaxed) structures, 505 MB
│   │
│   ├── original_data/                         # Raw source files (unmodified from upstream)
│   │   ├── QMOF/
│   │   │   └── qmof_database.zip              # Full QMOF database ZIP, 392 MB
│   │   └── ODAC23/
│   │       ├── odac23_is2r.tar.gz             # Full IS2R relaxed dataset, 848 MB
│   │       ├── pristine_CO2.tar.gz            # Pristine MOF + CO₂ structures, 93 MB
│   │       └── pristine_H2O.tar.gz            # Pristine MOF + H₂O structures, 72 MB
│   │
│   ├── lmdb/                                  # Pre-processed LMDB files, ready for training
│   │   ├── CGCNN/                             # Crystal Graph CNN format
│   │   │   ├── qmof_cgcnn_lmdb.tar            # QMOF bandgap, 5.69 GB
│   │   │   ├── odac_co2_cgcnn_lmdb.tar        # ODAC CO₂ adsorption energy, 15.1 GB
│   │   │   ├── odac_h2o_cgcnn_lmdb.tar        # ODAC H₂O adsorption energy, 9.39 GB
│   │   │   └── preprocess/
│   │   │       ├── preprocess_qmof_to_lmdb.py
│   │   │       └── preprocess_is2r_full_to_lmdb.py
│   │   │
│   │   ├── MGT/                               # Materials Graph Transformer format
│   │   │   ├── qmof_mgt_lmdb.tar              # QMOF bandgap, 17 GB
│   │   │   ├── odac_co2_mgt_lmdb.tar          # ODAC CO₂ adsorption energy, 38.5 GB
│   │   │   ├── odac_h2o_mgt_lmdb.tar          # ODAC H₂O adsorption energy, 23.7 GB
│   │   │   └── preprocess/
│   │   │       ├── preprocess_qmof_to_lmdb.py
│   │   │       └── preprocess_is2r_full_to_lmdb.py
│   │   │
│   │   └── PMT/                               # PMT model format
│   │       ├── qmof_pmt_lmdb.tar              # QMOF bandgap, 1.42 GB
│   │       └── preprocess/
│   │           └── qmof_preprocessor.py
│   │
│   ├── util/                                  # Shared utilities and target CSV files
│   │   ├── atom_init.json                     # 92-element atom feature embedding (CGCNN standard)
│   │   ├── id_prop_qmof.csv                   # QMOF structure IDs + bandgap targets
│   │   ├── id_prop_init_co2_adsorption_energy.csv     # CO₂ targets from initial (unrelaxed) structures
│   │   ├── id_prop_init_h2o_adsorption_energy.csv     # H₂O targets from initial (unrelaxed) structures
│   │   ├── id_prop_relaxed_co2_adsorption_energy.csv  # CO₂ targets from relaxed structures
│   │   ├── id_prop_relaxed_h2o_adsorption_energy.csv  # H₂O targets from relaxed structures
│   │   ├── make_small_data.py                 # Script to create a small subset for testing
│   │   ├── prepare_dataset_co2.py             # Dataset preparation helper for CO₂
│   │   └── prepare_dataset_h2o.py             # Dataset preparation helper for H₂O
│   │
│   └── cif_models/                            # ⚠️ DEPRECATED — see DEPRECATED.md inside
│       ├── DEPRECATED.md
│       ├── cgcnn_mgt/
│       └── mof_transformed/
│
└── README.md

---

LMDB Graph Format

All LMDB archives contain pre-computed crystal graphs. Each entry is a pickle-serialised dict with the following fields:

Field	Shape	Description
atom_fea	(N, 92)	Per-atom feature vectors from atom_init.json
nbr_fea	(N, 12, 41)	Gaussian-expanded pairwise distances (8 Å cutoff)
nbr_fea_idx	(N, 12)	Indices of 12 nearest neighbours per atom
target	scalar	Property value (band gap in eV, or adsorption energy in eV)
cif_id	string	Structure identifier

Graph parameters (same across all CGCNN/MGT datasets):

Parameter	Value
Cutoff radius	8.0 Å
Max neighbours	12
Atom feature length	92
Edge feature length	41 (Gaussian expansion, step 0.2 Å)

ODAC23 IS2R note: Only MOF framework atoms (tags == 0) are included in the graph. Adsorbate atoms (tags == 1) are excluded. The adsorbate type (CO₂ or H₂O) is encoded via the separate id_prop_*.csv files.

---

Quick Start

1. Download and extract a dataset

# QMOF for CGCNN
wget https://huggingface.co/datasets/hermanhugging/qmof_project/resolve/main/data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar
tar -xf qmof_cgcnn_lmdb.tar

# ODAC CO₂ for MGT
wget https://huggingface.co/datasets/hermanhugging/qmof_project/resolve/main/data/lmdb/MGT/odac_co2_mgt_lmdb.tar
tar -xf odac_co2_mgt_lmdb.tar

Or using the HuggingFace Hub Python client:

from huggingface_hub import hf_hub_download

# Download QMOF CGCNN LMDB
path = hf_hub_download(
    repo_id="hermanhugging/qmof_project",
    filename="data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar",
    repo_type="dataset",
)

2. Read an LMDB file

import lmdb, pickle
import numpy as np

env = lmdb.open("path/to/train.lmdb", readonly=True, lock=False, subdir=False)
with env.begin() as txn:
    n = int(txn.get(b"length").decode())   # total entries
    entry = pickle.loads(txn.get(b"0"))    # first entry

print(entry.keys())        # ['cif_id', 'target', 'atom_fea', 'nbr_fea', 'nbr_fea_idx']
print(entry["target"])     # band gap or adsorption energy in eV
print(entry["atom_fea"].shape)   # (N_atoms, 92)

3. Create a small subset for local testing

python data/util/make_small_data.py \
  --source-dir ./qmof_cgcnn_lmdb/train.lmdb \
  --output-dir ./qmof_small \
  --num-samples 500

---

Dataset Details

QMOF — Band Gap

Property	Value
Total structures	20,372
Train / Val / Test split	16,297 / 2,037 / 2,038 (80/10/10, seed=42)
Target	PBE band gap (eV)
Target mean	~2.09 eV
Target median	~1.58 eV
Target range	0 – 10+ eV
LMDB size (CGCNN)	5.69 GB
LMDB size (MGT)	17 GB
LMDB size (PMT)	1.42 GB

Target CSV: data/util/id_prop_qmof.csv

ODAC23 IS2R — CO₂ Adsorption Energy

Property	Value
Entries	~90,000
Target	Relaxed adsorption energy y_relaxed (eV)
Target range	~−5 to +1 eV
Adsorbate	CO₂
LMDB size (CGCNN)	15.1 GB
LMDB size (MGT)	38.5 GB

Target CSVs:

• data/util/id_prop_relaxed_co2_adsorption_energy.csv — targets from relaxed structures
• data/util/id_prop_init_co2_adsorption_energy.csv — targets from initial (unrelaxed) structures

ODAC23 IS2R — H₂O Adsorption Energy

Property	Value
Entries	~70,000
Target	Relaxed adsorption energy y_relaxed (eV)
Target range	~−4 to +1 eV
Adsorbate	H₂O
LMDB size (CGCNN)	9.39 GB
LMDB size (MGT)	23.7 GB

Target CSVs:

• data/util/id_prop_relaxed_h2o_adsorption_energy.csv — targets from relaxed structures
• data/util/id_prop_init_h2o_adsorption_energy.csv — targets from initial (unrelaxed) structures

---

Preprocessing Scripts

Each lmdb/{MODEL}/preprocess/ folder contains the scripts used to generate the LMDB files from the raw source data. These are provided for full reproducibility.

Script	Input	Output
preprocess_qmof_to_lmdb.py	original_data/QMOF/qmof_database.zip	{MODEL}/qmof_*_lmdb.tar
preprocess_is2r_full_to_lmdb.py	original_data/ODAC23/odac23_is2r.tar.gz	{MODEL}/odac_co2_*_lmdb.tar, odac_h2o_*_lmdb.tar
data/util/qmof_preprocessor.py	QMOF CIF files	PMT-format LMDB

---

Utility Files

File	Description
data/util/atom_init.json	92-element atom feature embedding used by CGCNN and MGT. Maps atomic number → 92-dim vector.
data/util/id_prop_qmof.csv	(qmof_id, bandgap) pairs for all 20,372 QMOF structures.
data/util/id_prop_relaxed_co2_adsorption_energy.csv	(sid, y_relaxed) for CO₂, relaxed structures.
data/util/id_prop_relaxed_h2o_adsorption_energy.csv	(sid, y_relaxed) for H₂O, relaxed structures.
data/util/id_prop_init_co2_adsorption_energy.csv	(sid, y_relaxed) for CO₂, initial structures.
data/util/id_prop_init_h2o_adsorption_energy.csv	(sid, y_relaxed) for H₂O, initial structures.
data/util/make_small_data.py	Creates a small LMDB subset for local development.
data/util/prepare_dataset_co2.py	Dataset preparation helper for CO₂ training.
data/util/prepare_dataset_h2o.py	Dataset preparation helper for H₂O training.

---

Deprecated

data/cif_models/ is deprecated and kept only for backward compatibility. Use data/lmdb/ for all current workflows. See data/cif_models/DEPRECATED.md for migration notes.

---

Citations

If you use these datasets, please cite the original sources:

QMOF Database:

@article{rosen2021machine,
  title   = {Machine learning the quantum-chemical properties of metal-organic frameworks
             for accelerated materials discovery},
  author  = {Rosen, Andrew S. and others},
  journal = {Matter},
  volume  = {4},
  pages   = {1578--1597},
  year    = {2021},
  doi     = {10.1016/j.matt.2021.02.015}
}

ODAC23:

@article{sriram2023open,
  title   = {The Open DAC 2023 Dataset and Challenges for Sorbent Discovery
             in Direct Air Capture},
  author  = {Sriram, Anuroop and others},
  journal = {arXiv preprint arXiv:2307.01547},
  year    = {2023}
}

---

License

MIT — see LICENSE for details.