hermanhugging/qmof_project

MOF Property Prediction Datasets

Pre-processed datasets for training graph neural networks on Metal-Organic Framework (MOF) property prediction. Covers five source databases — QMOF, ODAC23, hMOF, CoREMOF, and MOSAEC — prepared for three model architectures: CGCNN, MGT, and PMT.

Each database is available in up to three forms depending on coverage:

• data/cif_plus_idprop/ — CIF structures bundled with id_prop.csv target files, ready for direct CGCNN/MGT input
• data/lmdb/ — pre-computed LMDB graph databases, ready for training
• data/original_data/ — raw upstream source archives (unmodified)

The target variable for training is not fixed at preprocessing time — it is selected when launching a model. All multi-property databases retain the full set of available targets in their id_prop.csv files.

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

---

Source Databases

Database	Structures	Source
QMOF	~20,372 MOFs	Rosen et al., Matter 2021
ODAC23 IS2R	~160k MOF–adsorbate pairs (CO₂ + H₂O)	Sriram et al., arXiv 2307.01547
hMOF	~137k hypothetical MOFs	Boyd & Woo, Nature Chemistry 2019 (MOFDB)
CoREMOF	~14k experimental MOFs	Chung et al., J. Chem. Eng. Data 2019
MOSAEC	TBD	TBD

---

LMDB Coverage by Model

Pre-computed LMDB archives are currently available as follows. All databases are compatible with all three model architectures and can be preprocessed into LMDB format using the provided scripts.

Database	CGCNN LMDB	MGT LMDB	PMT LMDB
QMOF	✅ ready	✅ ready	✅ ready
ODAC23 CO₂	✅ ready	✅ ready	—
ODAC23 H₂O	✅ ready	✅ ready	—
hMOF	✅ ready	—	—
CoREMOF ASR	✅ ready	—	—
MOSAEC	—	—	—

Databases without a pre-built LMDB can be preprocessed from data/cif_plus_idprop/ using the scripts in data/lmdb/{MODEL}/preprocess/.

---

Repository Structure

qmof_project/
│
├── data/
│   │
│   ├── cif_plus_idprop/                     # CIF archives + id_prop.csv — direct CGCNN/MGT input
│   │   ├── QMOF.tar                         # QMOF CIFs + targets
│   │   ├── ODAC_init.tar                    # ODAC23 initial (unrelaxed) MOF structures
│   │   ├── HMOF.tar                         # hMOF CIFs + targets
│   │   ├── CoreMof_SI_NOT_FULL_ASR.tar      # CoREMOF ASR subset (not full)
│   │   ├── CoreMOF_CR_Full.tar              # CoREMOF CR full dataset
│   │   └── MOSAEC_full_and_partial.tar      # MOSAEC full + partial structures
│   │
│   ├── lmdb/                                # Pre-processed LMDB files, ready for training
│   │   │
│   │   ├── CGCNN/                           # Crystal Graph CNN format
│   │   │   ├── qmof_cgcnn_lmdb.tar          # QMOF — 5.69 GB
│   │   │   ├── odac_co2_cgcnn_lmdb.tar      # ODAC CO₂ — 15.1 GB
│   │   │   ├── odac_h2o_cgcnn_lmdb.tar      # ODAC H₂O — 9.39 GB
│   │   │   ├── hmof_cgcnn_lmdb.tar          # hMOF
│   │   │   ├── asr_cgcnn_lmdb.tar           # CoREMOF ASR
│   │   │   └── preprocess/
│   │   │       ├── preprocess_qmof_to_lmdb.py
│   │   │       └── preprocess_is2r_full_to_lmdb.py
│   │   │
│   │   ├── MGT/                             # Materials Graph Transformer format
│   │   │   ├── qmof_mgt_lmdb.tar            # QMOF — 17 GB
│   │   │   ├── odac_co2_mgt_lmdb.tar        # ODAC CO₂ — 38.5 GB
│   │   │   ├── odac_h2o_mgt_lmdb.tar        # ODAC H₂O — 23.7 GB
│   │   │   └── preprocess/
│   │   │       ├── preprocess_qmof_to_lmdb.py
│   │   │       └── preprocess_is2r_full_to_lmdb.py
│   │   │
│   │   └── PMT/                             # PMT model format
│   │       ├── qmof_pmt_lmdb.tar            # QMOF — 1.42 GB
│   │       └── preprocess/
│   │           └── qmof_preprocessor.py
│   │
│   ├── original_data/                       # Raw upstream source files (unmodified)
│   │   ├── QMOF/
│   │   │   └── qmof_database.zip                              # Full QMOF DB, 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
│   │   ├── hMof/
│   │   │   └── bulk-dl-mofdb-version-dc8a0295db.zip           # hMOF bulk download from MOFDB
│   │   └── CoreMof/
│   │       ├── NOT_FULL_FROM_ZENODO_CoREMOF2024DB_SI_20250204.zip
│   │       └── Core_MOF_CSD-modified.zip
│   │
│   ├── util/                                # Shared utilities and preprocessing scripts
│   │   ├── atom_init.json                   # 92-element atom feature embedding (CGCNN standard)
│   │   ├── prepare_dataset_co2.py           # Dataset prep helper for ODAC CO₂
│   │   ├── prepare_dataset_h2o.py           # Dataset prep helper for ODAC H₂O
│   │   ├── prepare_hmof.py                  # Dataset prep helper for hMOF
│   │   └── prepare_asr.py                   # Dataset prep helper for CoREMOF ASR
│   │
│   └── 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:

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 (selected at model launch via id_prop.csv)
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 note: Only framework atoms (tags == 0) are included in the graph. Adsorbate atoms (tags == 1) are excluded.

---

Quick Start

Download a pre-built LMDB (wget)

# 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

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

# QMOF CIF + id_prop (for raw CGCNN/MGT input or re-preprocessing)
wget https://huggingface.co/datasets/hermanhugging/qmof_project/resolve/main/data/cif_plus_idprop/QMOF.tar
tar -xf QMOF.tar

Download via Python

from huggingface_hub import hf_hub_download

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

Read an LMDB file

import lmdb, pickle

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())
    entry = pickle.loads(txn.get(b"0"))

print(entry.keys())             # ['cif_id', 'target', 'atom_fea', 'nbr_fea', 'nbr_fea_idx']
print(entry["atom_fea"].shape)  # (N_atoms, 92)

---

Dataset Sizes

Database	Archive	Size
QMOF	CGCNN/qmof_cgcnn_lmdb.tar	5.69 GB
QMOF	MGT/qmof_mgt_lmdb.tar	17 GB
QMOF	PMT/qmof_pmt_lmdb.tar	1.42 GB
ODAC23 CO₂	CGCNN/odac_co2_cgcnn_lmdb.tar	15.1 GB
ODAC23 CO₂	MGT/odac_co2_mgt_lmdb.tar	38.5 GB
ODAC23 H₂O	CGCNN/odac_h2o_cgcnn_lmdb.tar	9.39 GB
ODAC23 H₂O	MGT/odac_h2o_mgt_lmdb.tar	23.7 GB
ODAC23 (init structures)	cif_plus_idprop/ODAC_init.tar	505 MB
QMOF (raw)	original_data/QMOF/qmof_database.zip	392 MB
ODAC23 IS2R (raw)	original_data/ODAC23/odac23_is2r.tar.gz	848 MB

---

Preprocessing Scripts

All scripts to regenerate LMDB files from source data:

Script	Input	Output
data/lmdb/CGCNN/preprocess/preprocess_qmof_to_lmdb.py	original_data/QMOF/	CGCNN/qmof_cgcnn_lmdb.tar
data/lmdb/CGCNN/preprocess/preprocess_is2r_full_to_lmdb.py	original_data/ODAC23/	CGCNN/odac_*_cgcnn_lmdb.tar
data/lmdb/MGT/preprocess/preprocess_qmof_to_lmdb.py	original_data/QMOF/	MGT/qmof_mgt_lmdb.tar
data/lmdb/MGT/preprocess/preprocess_is2r_full_to_lmdb.py	original_data/ODAC23/	MGT/odac_*_mgt_lmdb.tar
data/lmdb/PMT/preprocess/qmof_preprocessor.py	QMOF CIF files	PMT/qmof_pmt_lmdb.tar
data/util/prepare_dataset_co2.py	ODAC23 structures	ODAC CO₂ training data
data/util/prepare_dataset_h2o.py	ODAC23 structures	ODAC H₂O training data
data/util/prepare_hmof.py	original_data/hMof/	hMOF training data
data/util/prepare_asr.py	original_data/CoreMof/	CoREMOF ASR training data

---

Deprecated

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

---

Citations

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}
}

hMOF (MOFDB):

@article{boyd2019data,
  title   = {Data-driven design of metal-organic frameworks for wet flue gas CO2 capture},
  author  = {Boyd, Peter G. and Woo, Tom K.},
  journal = {Nature Chemistry},
  volume  = {11},
  pages   = {1026--1034},
  year    = {2019},
  doi     = {10.1038/s41557-019-0327-5}
}

CoREMOF:

@article{chung2019advances,
  title   = {Advances, Updates, and Analytics for the Computation-Ready,
             Experimental Metal-Organic Framework Database: CoRE MOF 2019},
  author  = {Chung, Yongchul G. and others},
  journal = {Journal of Chemical \& Engineering Data},
  volume  = {64},
  pages   = {5985--5998},
  year    = {2019},
  doi     = {10.1021/acs.jced.9b00835}
}

---

License

MIT — see LICENSE for details.