A high-quality, DFT-computed dataset of high-entropy CoOOH surface structures for the oxygen evolution reaction (OER), together with the fine-tuned EquiformerV2 model checkpoints used in the paper:

Decoding active sites in high-entropy catalysts via attention-enhanced model Science Advances (2026). DOI: 10.1126/sciadv.aea1170

This repository backs a data-driven workflow — machine-learning prediction → active-site identification → mechanistic analysis → automated synthesis validation — for discovering high-entropy electrocatalysts. The model is an equivariant graph neural network (EquiformerV2) augmented with a Post-Attention Adapter (Post-Att Adapter) and a multi-target transfer-learning strategy that jointly predicts two key descriptors:

OER overpotential (η, V)
Doping formation energy (E_form, eV)

Dataset Summary

4,822 DFT-computed high-entropy CoOOH structures (transition metals randomly occupying Co sites), each associated with OER and/or doping descriptors.
Stored in LMDB format compatible with the fairchem / Open Catalyst Project (OCP) data pipeline used by EquiformerV2 (atomic graphs with positions, atomic numbers, cell, and targets).
Two prediction targets in two separate databases (overpotential.lmdb, doping_energy.lmdb).
10 fine-tuned EquiformerV2 checkpoints: a 5-fold cross-validation ensemble (folds 0–4) for each of the two tasks (over and doping).

⚠️ Gated dataset. Access requires accepting the contact-information / usage agreement on the Hugging Face page.

Repository Structure

yinliang22/oer_dataset/
├── oer_data/
│   ├── overpotential.lmdb            # OER overpotential dataset (LMDB; ~10 GB map size)
│   ├── overpotential.lmdb-lock
│   ├── doping_energy.lmdb            # Doping formation-energy dataset (LMDB; ~1 GB map size)
│   └── doping_energy.lmdb-lock
├── equiformerV2-0-0-over.pt          # OER overpotential head — CV fold 0   (~469 MB)
├── equiformerV2-0-1-over.pt          # OER overpotential head — CV fold 1
├── equiformerV2-0-2-over.pt          # OER overpotential head — CV fold 2
├── equiformerV2-0-3-over.pt          # OER overpotential head — CV fold 3
├── equiformerV2-0-4-over.pt          # OER overpotential head — CV fold 4
├── equiformerV2-0-0-doping.pt        # Doping formation-energy head — CV fold 0 (~469 MB)
├── equiformerV2-0-1-doping.pt        # Doping formation-energy head — CV fold 1
├── equiformerV2-0-2-doping.pt        # Doping formation-energy head — CV fold 2
├── equiformerV2-0-3-doping.pt        # Doping formation-energy head — CV fold 3
└── equiformerV2-0-4-doping.pt        # Doping formation-energy head — CV fold 4

Total size ≈ 16.7 GB. The .lmdb file sizes reflect the allocated LMDB map size, not necessarily the on-disk data volume. The *.lmdb-lock files are LMDB lock files.

Data Fields

Each LMDB entry is a serialized atomic-graph object (PyTorch-Geometric Data, OCP/fairchem style) with, at minimum:

Field	Description
`pos`	Cartesian atomic coordinates (Å)
`atomic_numbers`	Atomic numbers / species of each atom
`cell`	Periodic unit-cell lattice vectors
`natoms`	Number of atoms in the structure
`y` (target)	`overpotential` (V) in `overpotential.lmdb`; `doping_energy` (eV) in `doping_energy.lmdb`

The two databases use the same structural representation but carry different regression labels. Field names follow the fairchem/OCP convention; load and inspect a sample to confirm exact keys for your fairchem version.

Prediction (Screening) Dataset — Baidu Netdisk

The 17,500 candidate structures used for high-throughput virtual screening (from which the 8 high-activity, structurally stable systems — including the top performer TiFeNiZn-CoOOH — were selected) are distributed separately due to size:

File: predict_dataset.zip
Baidu Netdisk: https://pan.baidu.com/s/1EA_nXZGOLLEU6_9WN9yW-A?pwd=btg4
Extraction code (提取码): btg4

Usage

Download

from huggingface_hub import snapshot_download

# Requires accepting the gated-access agreement and `huggingface-cli login`
local_dir = snapshot_download(
    repo_id="yinliang22/oer_dataset",
    repo_type="dataset",
    local_dir="oer_dataset",
)

Read an LMDB database

import lmdb, pickle

env = lmdb.open(
    "oer_dataset/oer_data/overpotential.lmdb",
    subdir=False, readonly=True, lock=False, readahead=False, meminit=False,
)
with env.begin() as txn:
    n = pickle.loads(txn.get(b"length"))          # number of entries (key convention may vary)
    sample = pickle.loads(txn.get(f"{0}".encode()))  # a PyG Data object
print(n, sample)

In practice these LMDBs are intended to be consumed through fairchem's LmdbDataset (the EquiformerV2 data loader). Point the loader's src at the .lmdb file.

Load a fine-tuned checkpoint

import torch

ckpt = torch.load("oer_dataset/equiformerV2-0-0-over.pt", map_location="cpu")
# Use within the fairchem/EquiformerV2 trainer config with the Post-Att Adapter enabled.

For ensemble predictions, average the outputs of the five folds (0–4) of the relevant head.

Dataset Creation

System: high-entropy CoOOH, a representative OER catalyst, with multiple transition metals randomly substituting Co sites — yielding a vast space of local coordination environments.
Labels: computed by DFT — OER overpotential (via the adsorption free energies of the *OH, *O, *OOH intermediates along the conventional 4-electron OER pathway) and doping formation energy (thermodynamic stability of substitution).
Scale: 4,822 structures form the training/validation set; the model was then applied to 17,500 candidates for screening, and feature-importance/statistics were extended to a prediction space of >5 million structures, revealing that Zn has the highest active-site occupation probability and that the [CoNiZn] coordination consistently yields the lowest overpotential.

Experimental Validation (context)

The screened TiFeNiZn-CoOOH was synthesized and characterized end-to-end by an automated laboratory. In 1 M KOH it reaches an OER overpotential of 263 mV at 100 mA cm⁻² (≈93 mV lower than undoped CoOOH), 338 mV at 1000 mA cm⁻², a Tafel slope of 39.2 mV dec⁻¹, and retains 97.5 % of its performance after 120 h of continuous operation at 100 mA cm⁻².

Considerations & Limitations

Targets are DFT-derived approximations of catalytic activity/stability, not direct experimental measurements; absolute values carry the usual DFT/functional uncertainties.
The dataset focuses on the CoOOH OER chemistry; transferability to other host lattices or reactions is not guaranteed.
Field names and key conventions inside the LMDBs follow the fairchem/OCP pipeline of the time of release; verify against your fairchem version before training.

License

Released under CC-BY-4.0. You are free to share and adapt the material with appropriate credit. Please cite the paper below.

Citation

@article{yin2026decoding,
  title   = {Decoding active sites in high-entropy catalysts via attention-enhanced model},
  author  = {Yin, Liang and Ma, Tiantian and Zhu, Zibo and Ran, Nian and Zhou, Wei and Liu, Jianjun},
  journal = {Science Advances},
  year    = {2026},
  doi     = {10.1126/sciadv.aea1170},
  publisher = {American Association for the Advancement of Science}
}

Authors & Contact

Co-first authors: Liang Yin (SICCAS), Tiantian Ma (Beihang University), Zibo Zhu (SICCAS)
Co-corresponding authors: Nian Ran, Wei Zhou (Beihang University), Jianjun Liu (SICCAS)
Affiliations: Shanghai Institute of Ceramics, Chinese Academy of Sciences (SICCAS); Beihang University
AI experiments, simulation, and model training were carried out on the SICCAS MatMind platform; materials synthesis and characterization on the SICCAS DreamLab automated laboratory.

For questions about the dataset, open a discussion on the Hugging Face repository.

Downloads last month: 13

Total file size:

16.7 GB