General reactive machine learning potentials for CHON elements

This work presents a scalable workflow for developing general and reliable reactive machine learning potentials for systems containing C, H, O, and N elements, achieving high accuracy and transferability for large-scale chemical simulations.

---

➡️ Read the full paper on ChemRxiv

---

How to Cite

If you use this work, please cite the following publication:

@article{li2025general,
  title={General reactive machine learning potentials for CHON elements},
  author={Li, Bowen and Mi, Sixuan and Xiao, Jin and Duo, Zhang and Shuwen, Zhang and Zhang, John and Wang, Han and Zhu, Tong},
  year={2025}
}

---

DPA-3 Environment Setup and Model Usage

This document outlines the steps to set up the deepmd-kit environment and use the DPA-3 models. This guide is based on deepmd-kit-3.1.0.

1. Environment Setup

For detailed installation instructions, please refer to the official deepmd-kit v3.1.0 release page: https://github.com/deepmodeling/deepmd-kit/releases/tag/v3.1.0.

Environment Activation and Configuration

1. After installation, activate the environment. By default, it is located in the /home/xxx/deepmd-kit directory.

   # Note: 'xxx' should be replaced with your actual user directory.
   source activate /home/xxx/deepmd-kit

2. The default environment does not include the ASE and xtb-python packages, which must be installed manually.

   # Install ase
   pip install ase
   
   # Install xtb-python from the conda-forge channel
   conda install xtb-python -c conda-forge

2. Model Usage

The released model files use the following names:

• DPA-3-F.pt: directly trained DFT-level model
• DPA-3-DF.pt: DFT-level $\Delta$-learning model
• DPA-3-F@DeePHF.pt: directly trained DeePHF-level model
• DPA-3-DF@DeePHF.pt: DeePHF-level $\Delta$-learning model
• DPA-3-pretrain.pt: pretrained model

2.1. DPA-3-F.pt (Directly Trained Model)

This model is trained directly and can be used in Python through the ASE interface.

from deepmd.pt.utils.ase_calc import DPCalculator
from ase.io import read

# Initialize the calculator
calc = DPCalculator(model="DPA-3-F.pt", device='cuda')

# The remaining steps are consistent with other ASE calculators
data = read('test.xyz')
data.calc = calc

# Example calculations
energy = data.get_potential_energy()
forces = data.get_forces()

2.2. DPA-3-DF.pt ($\Delta$-Learning Model)

This is a $\Delta$-Learning model, and its use requires combination with GFN2-xTB calculations. To use it with the ASE interface in Python, a helper script (deepmd_xtb.py) is needed.

import sys

# Add the directory containing 'deepmd_xtb.py' to the Python path.
# This example assumes the script is in '/home/xxx/ase_interface'.
sys.path.append('/home/xxx/ase_interface')

from deepmd_xtb import DP_xTB

# Initialize the calculator
calc = DP_xTB(model="DPA-3-DF.pt", device='cuda')

For the DeePHF-level models, use DPA-3-F@DeePHF.pt or DPA-3-DF@DeePHF.pt. If selecting the DeePHF head explicitly, use head="CC".