Stable and accurate atomistic simulations of flexible molecules using conformationally generalisable machine learned...

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Machine learned potentials (MLPs) offer revolutionary potential for molecular shape prediction.
• Current adoption is hindered by limitations in generating comprehensive training data.
• Accurate representation of conformational degrees of freedom is crucial.

Purpose of the Study:

• To present a novel approach for generating molecular datasets that properly represent key conformational degrees of freedom.
• To develop generalizable MLPs that achieve quantum chemical accuracy for all conformers.
• To enable stable, long molecular dynamics (MD) simulations and accurate free energy calculations.

Main Methods:

• Generating reference molecular datasets with complete conformational coverage, including barrier regions.
• Training MLPs using a global descriptor scheme on these comprehensive datasets.
• Deploying MLPs in well-tempered metadynamics simulations to compute conformational free energy surfaces.

Main Results:

• MLPs trained on complete datasets demonstrate generalizability across conformational space.
• Achieved quantum chemical accuracy for all molecular conformers.
• Successfully propagated long, stable molecular dynamics trajectories, a significant advancement.
• Obtained converged conformational free energy surfaces for flexible molecules.

Conclusions:

• MLPs trained on comprehensive conformational datasets are essential for stable MD simulations.
• This approach provides accurate computation of structural, dynamical, and thermodynamical properties for flexible molecules.
• The method offers a new route to understanding complex molecular systems.