wB97X-ML

Robust machine learning interatomic potentials (MLIPs) that achieve accuracy comparable to the ωB97X-D3BJ/def2-TZVPP quantum mechanical method on non-covalent interactions.

---

Requirements

wB97X-ML is built on NequIP, please install NequIP first.

• Python >= 3.9
• NequIP == 0.5.6

Quick Setup

# Create and activate a conda environment (recommended)  
conda create -n nequip-env python=3.10
conda activate nequip-env

# Install PyTorch with CUDA 11.3 (adjust based on your driver)  
conda install pytorch==1.11.0 cudatoolkit=11.3 -c pytorch

# Install Nequip 0.5.6 and dependencies  
wget https://github.com/mir-group/nequip/archive/refs/tags/v0.5.6.tar.gz
tar -xvzf v0.5.6.tar.gz
cd nequip
pip install . 

Installation and usage

• Download pretrained models.

git clone git@github.com:hnlab/wB97X-ML.git
cd models
# download all models from https://zenodo.org/records/15514804
pip install zenodo_get
zenodo_get 10.5281/zenodo.15514804 -g "[A-Z]*.tar.gz"
tar -xzvf ./*.tar.gz

• Run Energy Prediction Example:
  Note: Please refer to the corresponding model's README.md for applicable dimer.

  • Basic (No Multiprocessing)
    in Windows/Jupyter environments where multiprocessing.Pool is unstable.

  cd scripts
  python predict_energy.py -xyz dataset/ACET_ETOH.xyz -md ./models -m ACET -od ./examples

  • Parallel Accelerated
    Leverages multiprocessing.Pool for speedup on multicore systems.

  cd scripts
  # 2 cores
  python predict_energy.py -xyz dataset/ACET_ETOH.xyz -md ./models -m ACET -od ./examples --mlp -w 2

Training set: PDB-FRAGID

Citation