On Sampling Minimum Energy Path.

Area of Science:

• Computational Chemistry
• Materials Science
• Statistical Mechanics

Background:

• Sampling the minimum energy path (MEP) between metastable states is challenging due to energy barriers, making direct molecular dynamics or Markov Chain Monte Carlo inefficient for rare events.
• Augmented sampling methods exist but often rely on arbitrary choices for dimensionality reduction algorithms like collective variables or reaction coordinates.

Purpose of the Study:

• To develop a method that simultaneously learns the minimum energy path (MEP) and the collective variable.
• To overcome the limitations of arbitrary dimensionality reduction in augmented sampling techniques.

Main Methods:

• Coupling statistical sampling techniques (metadynamics) with invertible neural networks and autoencoders.
• A two-step learning process involving statistical sampling of the most probable path and redefinition of the collective variable based on updated data points.

Main Results:

• Successfully demonstrated the ability to learn complex minimum energy paths (MEPs).
• The coupled approach effectively unravels complex MEPs, as shown using the Mueller potential.

Conclusions:

• The proposed method efficiently learns the MEP and collective variable simultaneously, overcoming limitations of existing techniques.
• This approach offers a powerful tool for studying rare events and complex energy landscapes in various systems.