Single- and double-precision Codes for PhysNet
Meuwly Group, University of Basel

General

The present repository provides access to two implementations of the PhysNet [1] codes (single (F32)- and double (F64)-precision), which can be used to learn molecular potential energy surfaces (PESs). The resulting single- and double-precision PESs were assessed in Reference [2]. The required installation and dependencies are outlined below and are followed by examples for training and using the neural network-based PES for H₂CO. Therefore, the repository also contains the ab initio reference MP2/aug-cc-pVTZ level data for H₂CO [3], as well as ready-to-use models [2].

Installations & dependencies

The following installation steps were tested on a Ubuntu 20.04 workstation and using Conda 23.7.2 (see https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html)

a) If not installed already, install Miniconda on your machine (see https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html)

b) Create an environment named (e.g.) physnet_env, install Python 3.6:

conda create --name physnet_env python=3.6

Activate it:

conda activate physnet_env

(deactivating it by typing: conda deactivate)

c) With activated environment, all dependencies can be installed.

pip install ase==3.19.1
pip install tensorflow==1.12

Examples

Training and using the single- and double-precision PESs follows the same procedure with the required adaptations made in the f32 and f64 folders and corresponding source codes.

PhysNet training

In either of the f32 or f64 folders, training can be started with activated conda environment by running

./train.py @run_ch2o_mp2.inp

Once the training is converged, the models can be extracted from the "best/" folder, which can be found in a newly creaded folder with a timestamp.

Evaluation

In either of the f32/eval or f64/eval folders, scripts for exemplary evaluations are given. These can for example be used to predict the energy of a given structure in .xyz format (predict_mol.py)

./predict_mol.py -i h2co.xyz

or to optimize a given structure in .xyz format (optimize.py)

./optimize.py -i h2co.xyz

or to calculate the harmonic frequencies of an optimized molecule (ase_vibrations.py)

./ase_vibrations.py -i opt_h2co.xyz

or to calculate the energy along a stretch of the C-H bond (predict_stretch.py) - see Fig. 1 of Ref [2].

./predict_stretch.py -i opt_h2co.xyz

How to cite

When using the PhysNet or the H₂CO PES, please cite the following papers:

For PhysNet:

Oliver T. Unke and Markus Meuwly "PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges", J. Chem. Theory Comput., 2019, 15, 6, 3678–3693

For the H₂CO dataset/PES

Silvan Käser, Debasish Koner, Anders S. Christensen, O. Anatole von Lilienfeld, and Markus Meuwly "Machine Learning Models of Vibrating H₂CO: Comparing Reproducing Kernels, FCHL, and PhysNet" J. Phys. Chem. A 2020, 124(42), 8853-8865, DOI: 10.1021/acs.jpca.0c05979

For the double-precision PES

Silvan Käser and Markus Meuwly "Numerical Accuracy Matters: Applications of Machine Learned Potential Energy Surfaces", 2023, arXiv e-prints, DOI: 10.48550/arXiv.2311.17398

References

[1] Oliver T. Unke and Markus Meuwly "PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges", J. Chem. Theory Comput., 2019, 15, 6, 3678–3693

[2] Silvan Käser and Markus Meuwly "Numerical Accuracy Matters: Applications of Machine Learned Potential Energy Surfaces", 2023, arXiv e-prints, DOI: 10.48550/arXiv.2311.17398

[3] Silvan Käser, Debasish Koner, Anders S. Christensen, O. Anatole von Lilienfeld, and Markus Meuwly "Machine Learning Models of Vibrating H₂CO: Comparing Reproducing Kernels, FCHL, and PhysNet" J. Phys. Chem. A 2020, 124(42), 8853-8865, DOI: 10.1021/acs.jpca.0c05979

Contact

If you have any questions about the PESs free to contact Silvan Käser (silvan.kaeser@unibas.ch) or Markus Meuwly (m.meuwly@unibas.ch)