Tropolone and PFD Potential Energy Surfaces
Meuwly Group, University of Basel
The present repository provides access to the raw data and potential energy surfaces for tropolone and (propiolic acid)–(formic acid) dimer (PFD), 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 for large molecules 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 explicity examples on how to use the PESs are based on tropolone; however, the same evaluations can be done for PFD, too. The ab initio raw data is available in tropolone/raw_data and pfd/raw_data.
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 both systems are 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/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 very 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 tropolone/raw_data and pfd/raw_data. For tropolone this includes 19997 MP2 and 100 CCSD(T) data points while for PFD this includes 40290 (this also includes monomer structures) MP2 and 406 (106 PFD + 300 monomer structures) CCSD(T) data points.
Due to the rather small high-level data set sizes, TL was repeated 10 times on different (random) splits of the data. For each of the systems, the 10 resulting TL models are provided in Tropolone/evaluation/models_tropo_tlcc/ and PFD/evaluation/models_pfd_tlcc/, respectively. The corresponding MP2 level models are also provided in Tropolone/evaluation/models_tropo_mp2/ and PFD/evaluation/models_pfd_tlcc/
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 min-tropo.xyz
-
or to optimize a given structure in .xyz format (optimize.py)
./optimize.py -i min-tropo.xyz -o opt_min-tropo.xyz --fmax 0.000005
-
or to calculate the harmonic frequencies of an optimized molecule (NN_NM.py).
./NN_NM.py -i opt_min-tropo.xyz
To allow reproduction of the results, a text file (Tropolone/evaluation/outputs_for_reproduction and PFD/evaluation/outputs_for_reproduction) is available for exemplary observables (energy barrier for hydrogen transfer, harmonic frequencies, ...).
The models and evaluation scripts are also available for PFD, see PFD/evaluation.
When using the PhysNet or the Tropolone/PFD 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
Silvan Käser, Jeremy Richardson, and Markus Meuwly "Accurate Tunneling Splittings for Ever-Larger Molecules from Transfer-Learned, CCSD(T) Quality Energy Functions" https://arxiv.org/abs/2407.21366
[1] Silvan Käser, Jeremy Richardson, and Markus Meuwly "Accurate Tunneling Splittings for Ever-Larger Molecules from Transfer-Learned, CCSD(T) Quality Energy Functions" https://arxiv.org/abs/2407.21366
[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 free to contact Prof. Markus Meuwly ([email protected])