This repository contains the code and associated datasets for the machine learning of a potential energy surface for the purpose of molecular dynamics (MD) simulations, as implemented in [1]. For more detailed information, please refer to the manuscript [1].

• Potential energy surface (PES) constructed for MD simulation, taking AlF-AlF complex as an example
• Regressor: Gaussian process regression (GPR)

scikit-learn==1.0.1

ase==3.22.1

pickle

The codes were implemented using Python 3.8.

To train a machine learning potential energy surface (PES) model and make energy predictions, you can refer to the run_AlF_dimer.py script: $ python run_AlF_dimer.py

The necessary datasets, including the training set and test set(s), are specified in run_AlF_dimer.py. For instance, data/traj_train.xyz and data/traj_test.xyz represent the training and test sets, respectively, in extxyz (extended XYZ) format. The reference energies in these datasets are expressed in electron volts (eV). The provided example training set, data/traj_train.xyz, consists of 18,732 AlF-AlF configurations combined from eight molecular dynamics (MD) trajectories. The test set, data/traj_test.xyz, is derived from a single MD trajectory with 3,633 steps. The ab initio energies are calculated at the coupled-cluster theory with single, double, and perturbation triples [CCSD(T)] level, with the aug-cc-pVQZ basis set. During the MD simulations, the forces are calculated at the second‐order Møller–Plesset perturbation theory (MP2) level.

After training, the PES model is stored by default in trained_ml_potential_model.pkl. The training and testing results are printed in the AlF_dimer.logfile .

The structural representations of the AlF-AlF complex are computed in representation.py. These representations are then used as inputs to the Gaussian process regression (GPR) model for training. If you wish to modify the Gaussian process kernels, please refer to machine_learning_potential.py. Currently, the kernel is a combination of the Matern(5/2) kernel and a dot-product kernel, with a white noise kernel indicating the noise level of the training set.

[1] X. Liu, W. Wang, J. Pérez-Ríos, Molecular dynamics-driven global potential energy surfaces: Application to the AlF dimer, J. Chem. Phys. 159, 144103 (2023)