PRDNet: Pseudo-particle Ray Diffraction Network

English · 简体中文 · 日本語 · 한국어

PRDNet is a physics-informed graph neural network for crystal property prediction. It combines graph representations with a learnable pseudo-particle ray-diffraction mechanism in reciprocal space, capturing long-range structural interactions while preserving crystallographic symmetry invariance.

Highlights

• Graph neural networks for crystal structures
• Pseudo-particle ray-diffraction physics in reciprocal space
• Multi-head attention and symmetry-invariant representations

Quick Start

git clone https://github.com/Bin-Cao/PRDNet.git
cd PRDNet
pip install -r requirements.txt
python -c "import prdnet; print('PRDNet installed successfully!')"

Data

Use ASE database files. Each structure can include numeric target properties such as formation_energy, band_gap, bulk_modulus, or shear_modulus.

from ase.db import connect
from ase.build import bulk

db = connect("my_data.db")
db.write(bulk("Si", "diamond", a=5.43), formation_energy=-5.42, band_gap=1.12)

Datasets: CPPbenchmark / Materials Project, JARVIS-DFT, or any ASE-compatible database.

Training

Set your database paths and target in trainer.py, then run:

python trainer.py

For multi-GPU training:

torchrun --nproc_per_node=4 trainer.py

Important configuration options include epochs, batch_size, learning_rate, conv_layers, node_features, use_diffraction, and diffraction_max_hkl. See the main README for the complete configuration and troubleshooting guide.

Citation

@article{cao2025beyond,
  title={Beyond Structure: Invariant Crystal Property Prediction with Pseudo-Particle Ray Diffraction},
  author={Cao, Bin and Liu, Yang and Zhang, Longhan and Wu, Yifan and Li, Zhixun and Luo, Yuyu and Cheng, Hong and Ren, Yang and Zhang, Tong-Yi},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

License

MIT. See LICENSE.