SevenNet_dFS

SevenNet_dFS is an extended version of SevenNet, a machine learning force field model tailored for molecular dynamics (MD) simulations, incorporating a direct output head for force and stress.

Key Differences

+ Direct force and stress prediction output
+ Integrated derivative values during training for higher accuracy
+ Up to 5x faster inference speed
+ Up to 250x faster MD simulation speed
- Non-conservative MD simulation
- Slower training speed

Performance

Metric	Improvement
Training Speed	⚠️ 1.59x slower
Inference Speed	✅ 4.91x faster
MD Simulation Speed	✅ 255.28x faster
Prediction Accuracy	✅ Excellent
MD/Inference Computational Cost	✅ Significantly reduced

Our example dataset for Li-argyrodite (Li₆PS₅Cl) consists of 2,000 configurations, each containing 416 nodes (Li₁₉₂P₃₂S₁₆₀Cl₃₂) and 8,332 edges (with a cutoff radius of 4.5 Å).

The model achieved excellent performance on test MD simulation:

• Energy MAE: 0.5131 meV/atom
• Force MAE: 0.0396 eV/Å
• Stress MAE: 0.2903 kbar

Quick Start

# Clone Repository
git clone https://github.com/hyukjunlim/SevenNet-dFS.git
cd SevenNet_dFS

# Install dependencies
pip install sevenn

# Run tutorial example
cd sevennet_tutorial
python tuto.py

Citation

If you use this code, please cite the SevenNet paper:

@article{park_scalable_2024,
	title = {Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations},
	volume = {20},
	doi = {10.1021/acs.jctc.4c00190},
	number = {11},
	journal = {J. Chem. Theory Comput.},
	author = {Park, Yutack and Kim, Jaesun and Hwang, Seungwoo and Han, Seungwu},
	year = {2024},
	pages = {4857--4868},
}