FFLUX molecular simulations driven by atomic Gaussian process regression models.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) force fields offer a computationally efficient alternative to ab initio methods for molecular dynamics (MD) simulations.
• Accurate prediction of atomic properties is crucial for reliable MD simulations.

Purpose of the Study:

• To develop and evaluate Gaussian process regression (GPR) models integrated with the FFLUX ML force field for MD simulations.
• To assess the accuracy and robustness of these GPR models for molecular simulations.

Main Methods:

• Utilized the GPyTorch library to create GPR models interfaced with the FFLUX ML force field.
• Implemented an improved kernel function to capture descriptor periodicity.
• Performed geometry optimizations and 298 K MD simulations for ammonia, methanol, and malondialdehyde.

Main Results:

• GPR models achieved highly accurate geometry optimizations with a maximum RMSD of 0.064 Å.
• Total energy predictions were highly accurate, below 1 kJ/mol.
• MD simulations showed excellent agreement with ab initio data for ammonia and methanol, with decreased accuracy for malondialdehyde.

Conclusions:

• The developed GPR models demonstrate high accuracy and robustness for MD simulations of small molecules.
• The improved kernel function enhances the prediction of molecular properties.
• Future work will focus on scaling these models for larger and more complex systems.