Hydrogenoxalate Potential Energy Surfaces
Meuwly Group, University of Basel
The present repository provides access to the raw data and potential energy surfaces for hydrogenoxalate, which are described in detail in Reference [1]. The PESs are obtained following a rational procedure based on transfer learning to reach CCSD(T) quality and are based on PhysNet [2]. This repository contains instructions for the installation and dependencies, a description of the different PESs and corresponding raw data. This is followed by examples on using the neural network-based PESs. The ab initio raw data is available in data/mp2 and data/ccsd. MP2/aug-cc-pVTZ level of theory PES could be found in evaluation/model_mp2_f32 directory and CCSD(T)/aug-cc-pVTZ models (f32 and f64) could be found in evaluation/model_ccsd_f32 and evaluation/model_ccsd_f64 respectively.
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). The installation of the dependencies (excluding the installation of Miniconda) takes less than 5 min.
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
pip install tensorflow_estimator
The PESs for hydrogenoxalate is obtained using transfer learning (TL): TL builds on the knowledge acquired by solving one task (representing a lower level PES) to solve a new, related task (representing a higher level PES). This encompasses training a lower level model on quantum chemical reference data that is rapid to evaluate (here we used MP2/aug-cc-pVTZ). This allows us to generate large datasets with structures that span the entire (relevant) configuration space. Then, the optimal set of weights and biases for the low-level data is retrained based on small amouts of high level ab initio data (here CCSD(T)/aug-cc-pVTZ).
The raw ab initio data (Molpro outputs) are given for both the MP2 and the CCSD(T) level of theory in data/mp2 and data/ccsd. For oxalate this includes 22100 MP2 and 2688 CCSD(T) data points.
Most Python scripts that are used to evaluate the PhysNet PESs make use of the atomic simulation environment (ASE) [3] and are written in Python. It is important to get used to ASE, which has very good tutorials online (https://wiki.fysik.dtu.dk/ase/tutorials/tutorials.html#ase). Scripts on how to use the PESs that have been used throughout the evalulation of Reference [1] are given in the evaluation folder. These can for example be used to
-
predict the energy of a given structure in .xyz format (predict_mol.py)
./predict_mol.py -i hoxa_min.xyz -
to optimize a given structure in .xyz format (optimize.py)
./optimize.py -i hoxa_min.xyz -o opt_hoxa_min.xyz --fmax 0.000005 -
to calculate the harmonic frequencies of an optimized molecule (NN_NM.py).
./NN_NM.py -i opt_hoxa_min.xyz -
to calculate the anharmonic frequencies of an optimized molecule using VPT2 method via Gaussian Interface following two steps:
i) Geometry optimization using Gaussian with external potential:
gview opt_hoxa_min.comctrl + g
submit
submit
YES
ii) After optimization, replace the coordinates in nnvpt2_hoxa.com with the Cartesian coordinates of your optimized structure. Then submit the VPT2 calculation using gview:
gview nnvpt2_hoxa.comctrl + g
submit
submit
YES
When using the PhysNet or the Hydrogenoxalate PESs, please cite the following papers:
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
Andreichev, Valerii, et al. "Dynamics of Protonated Oxalate from Machine-Learned Simulations and Experiment: Infrared Signatures, Proton Transfer Dynamics and Tunneling Splittings." arXiv preprint arXiv:2508.06419 (2025). https://arxiv.org/abs/2508.06419
[1] Andreichev, Valerii, et al. "Dynamics of Protonated Oxalate from Machine-Learned Simulations and Experiment: Infrared Signatures, Proton Transfer Dynamics and Tunneling Splittings." arXiv preprint arXiv:2508.06419 (2025). https://arxiv.org/abs/2508.06419
[2] 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
[3] Ask Hjorth Larsen et al, "The atomic simulation environment—a Python library for working with atoms", 2017, J. Phys.: Condens. Matter, 29, 273002, DOI 10.1088/1361-648X/aa680e
If you have any questions about the codes feel free to contact Valerii Andreichev (valerii.andreichev@unibas.ch) or Prof. Markus Meuwly (m.meuwly@unibas.ch)