Incorporating Noncovalent Interactions in Transfer Learning Gaussian Process Regression Models for Molecular...

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• FFLUX uses machine learning for fast molecular dynamics.
• Current models trained on monomers miss many-body effects like charge transfer.
• Lennard-Jones potentials are used for dispersion and repulsion, requiring extensive parameterization.

Purpose of the Study:

• To improve FFLUX simulations by incorporating two-body effects.
• To develop and benchmark a formamide dimer model for FFLUX.
• To enable FFLUX simulations that better approximate quantum mechanics.

Main Methods:

• Trained FFLUX models on a formamide dimer instead of monomers.
• Utilized hyperparameter transfer for efficient training of higher-dimensional models.
• Benchmarked the dimer model against quantum mechanics calculations.

Main Results:

• The dimer model enables FFLUX simulations to include two-body effects like intermolecular polarization and charge penetration.
• The new approach eliminates the need for Lennard-Jones potentials.
• Hyperparameter transfer reduced training time by an order of magnitude.

Conclusions:

• Training FFLUX models on clusters (dimers) significantly enhances simulation accuracy.
• This work represents a key step towards quantum mechanical accuracy in molecular dynamics.
• The developed methods allow for more accurate simulations of intermolecular interactions.