• 📢 News
• 🧪 1. Summary
• 🧬 2. EDBench Database
• 🧪 3. Benchmark Tasks
  • 🔮 3.1 Prediction Tasks
  • 🔁 3.2 Retrieval Task
  • 🧬 3.3 Generation Task
  • 📂 3.4 Dataset File Format
• 🚀 4. Running Benchmarks
  • ⚛️ 4.1 Quantum Prediction Tasks
  • ⚡ 4.2 Retrieval and Generation Tasks
• 📬 Contact
• 📘 License

• Release the full ED dataset (>3M)
• Release model checkpoints
• Release model code for quantum property prediction task
• Release model code for retrieval between molecule and ED task
• Release model code for ED generation task
• Release benchmark dataset

Note: The full dataset has not been released yet. Due to the increased data volume, we are currently coordinating with data storage providers to handle and host the large-scale data efficiently. Once this is completed, the full dataset will be made publicly available.

• [2025/09/19] 🎉 Paper was accepted by NeurIPS 2025 !

• [2025/05/13] Uploaded code of prediction tasks with X-3D and PointVector.

• [2025/05/10] Repository initialized!

Most existing molecular machine learning force fields (MLFFs) focus on atom- or molecule-level properties like energy and forces, while overlooking the foundational role of electron density (ED), denoted as $\rho(r)$. According to the Hohenberg–Kohn theorem, ED uniquely determines all ground-state properties of many-body quantum systems. However, ED is expensive to compute via first-principles methods such as Density Functional Theory (DFT), limiting its large-scale use in MLFFs.

EDBench 🐋 addresses this gap by providing a large-scale, high-quality dataset of electron densities for over 3.3 million molecules, based on the PCQM4Mv2 standard. To benchmark electronic-scale learning, we introduce a suite of ED-centric tasks covering:

• Prediction of quantum chemical properties
• Retrieval across structure and ED modalities
• Generation of ED from molecular structures

We demonstrate that ML models can learn from ED with high accuracy and also generate high-quality ED, dramatically reducing DFT costs. All data and benchmarks will be made publicly available to support ED-driven research in drug discovery and materials science.

📄 Citation

@inproceedings{xiang2025edbench,
  title         = {EDBench: Large-Scale Electron Density Data for Molecular Modeling},
  author        = {Hongxin xiang and Ke Li and Mingquan Liu and Zhixiang Cheng and Bin Yao and Wenjie Du and Jun Xia and Li Zeng and Xin Jin and Xiangxiang Zeng},
  booktitle     = {The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year          = {2025},
  url           = {https://openreview.net/forum?id=pAd7qVrYPG}
}

@misc{xiang2025edbenchlargescaleelectrondensity,
  title         = {EDBench: Large-Scale Electron Density Data for Molecular Modeling},
  author        = {Hongxin Xiang and Ke Li and Mingquan Liu and Zhixiang Cheng and Bin Yao and Wenjie Du and Jun Xia and Li Zeng and Xin Jin and Xiangxiang Zeng},
  year          = {2025},
  eprint        = {2505.09262},
  archivePrefix = {arXiv},
  primaryClass  = {physics.chem-ph},
  url           = {https://arxiv.org/abs/2505.09262}
}

Built on PCQM4Mv2, the EDBench dataset contains accurate DFT-computed EDs for 3.3M+ molecules, enabling deep learning at the electronic scale.

We design a suite of benchmark tasks centered on electron density (ED):

Predict quantum chemical properties from ED representations.

{benchmark_root}
{benchmark root}
├── ed_energy_5w
│   ├── raw
│   │   ├── ed_energy_5w.csv
│   │   ├── readme.md
│   │   └── psi4_grid0.4_cubes
│   │       └── {mol_index}
│   │           ├── Mol1_Dt.cube
│   │           ├── timer.dat
│   │           ├── Mol1.sdf
│   │           ├── Mol1_ESP.cube
│   │           └── {mol_index}_Psi4.out
│   └── processed
│       └── mol_EDthresh{thresh}_data.pkl
├── ed_homo_lumo_5w
│   ├── raw
│   │   ├── ed_homo_lumo_5w.csv
│   │   ├── readme.md
│   │   └── psi4_grid0.4_cubes
│   └── processed
│       └── mol_EDthresh{thresh}_data.pkl
├── ed_multipole_moments_5w
│   ├── raw
│   │   ├── ed_multipole_moments_5w.csv
│   │   ├── readme.md
│   │   └── psi4_grid0.4_cubes
│   └── processed
│       └── mol_EDthresh{thresh}_data.pkl
└── ed_open_shell_5w
    ├── raw
    │   ├── ed_open_shell_5w.csv
    │   ├── readme.md
    │   └── psi4_grid0.4_cubes
    └── processed
        └── mol_EDthresh{thresh}_data.pkl

Cross-modal retrieval between molecular structures (MS) and electron densities (ED).

{benchmark_root}
├── ed_retrieval_5w/
│   ├── raw/
│   │   ├── ed_retrieval_5w.csv
│   │   ├── readme.md
│   │   └── psi4_grid0.4_cubes/
│   └── processed/
│       └── mol_EDthresh{thresh}_data.pkl

Generate ED representations from molecular structures.

{benchmark_root}
├── ed_prediction_5w/
│   ├── raw/
│   │   ├── ed_prediction_5w.csv
│   │   ├── readme.md
│   │   └── psi4_grid0.4_cubes/
│   └── processed/
│       └── mol_EDthresh{thresh}_data.pkl

Each raw/ directory includes a .csv summary file describing each molecule.

• index: Molecule index
• smiles: Original SMILES
• canonical_smiles: Canonicalized SMILES
• scaffold_split: Scaffold-based split (80% train / 10% valid / 10% test)
• random_split: Random split (80% train / 10% valid / 10% test)

• Prediction:
  • label: Ground-truth values (space-separated if multi-task)
• Retrieval:
  • negative_index: Space-separated indices of 10 negative samples

Description:

This task focuses on predicting quantum properties of molecules using point-cloud representations of electronic density. It is designed to evaluate how accurately models can capture molecular electronic behavior and properties.

Description:

This task involves two objectives:

1. Retrieval between molecular structures and their electronic densities.
2. Predicting electronic densities from molecular conformations.

These tasks assess the model’s ability to connect molecular structures with their electron distributions and to generate accurate electronic density predictions.

Feel free to open an issue or pull request for questions or contributions. For academic inquiries, contact the authors upon paper publication.

Released for research use under an open-source MIT license.