PhysNet-TF2

TensorFlow 2 PhysNet implementation

Patching of PhysNet for Tensorflow 2.

Git repository for original (TensorFlow 1) version of PhysNet: https://github.com/MMunibas/PhysNet.

Requirements:

• Tensorflow 2.20
• Python 3.13
• CUDA 12 (for running on GPUs only)

Pre-requisite setup instructions:

1. Install Anaconda or Miniconda on the system. Link: www.anaconda.com.
2. In terminal, initialize the conda (base) environment.
3. Create a environment of PhysNet with Tensorflow2.

conda create -n physnet-tf2 python=3.13
conda activate physnet-tf2

For system with GPU:

pip install tensorflow[and-cuda] ase tensorboard

For system without GPU:

pip install tensorflow ase tensorboard

Training PhysNet:

1. Create a file all.npz, which contains the following information about the set of moleculs (for a set containing n configurations each with m atoms):

Property	Symbol	Dimension of array	Data type	Units	Remarks
No. of atoms	N	n	int
Energy	E	n	float	eV	Individual atomic contributions removed
Charge	Q	n	float	e	Total charge on molecule
Dipole moment	D	n x 3	float	a.u.	Total dipole moment of molecule
Atomic number	Z	n x m	int
Positions	R	n x m x 3	float	Angstrom
Forces	F	n x m x 3	float	eV/Angstrom

2. Modify the config.txt file according to specification. Ponder on the following:
   • num_train and num_valid: Number of points in training and validation sets respectively. The rest are used for test set. A general rule of thumb for split can be 80%, 10% and 10% for training, validation and testing.
   • seed: The initialization seed for random number. It is essential to change this if multiple models are trained on the same dataset.

For local computer:

3. Load the physnet-tf2 conda environment, and then run using:

python3 train.py @config.txt

For running on cluster (slurm queuing system):

3. Modify the submit.run file according to cluster specifications.
4. Submit using the following command:

sbatch submit.run

GPU Compatibility:

Warning

Although the latest CUDA version released is CUDA 13, Tensorflow 2 currently supports only until CUDA 12. Therefore, please ensure that CUDA 12 is enabled in your system during runtime.

Tip

To check the runtime CUDA version of the local computer (or node), one can use nvidia-smi command.

Note

This version has been tested on NVIDIA RTX 2080Ti, RTX 3080, RTX 3090, RTX 3090 Ti and RTX 4090 GPU cards, along with their submodels.

Citation:

@article{unke2019physnet,
  title={PhysNet: A neural network for predicting energies, forces, dipole moments, and partial charges},
  author={Unke, Oliver T and Meuwly, Markus},
  journal={Journal of chemical theory and computation},
  volume={15},
  number={6},
  pages={3678--3693},
  year={2019},
  publisher={ACS Publications}
}