Reinforcement learning of rare diffusive dynamics

Area of Science:

• Molecular Dynamics
• Computational Chemistry
• Statistical Mechanics

Background:

• Rare molecular dynamics events are crucial for understanding chemical reactions and material properties.
• Simulating these rare events is computationally challenging due to their low probability.
• Existing methods often struggle with efficiency and accuracy in capturing these dynamics.

Purpose of the Study:

• To develop a direct method for probing rare molecular dynamics trajectories.
• To enable efficient and accurate estimation of the likelihood of rare events.
• To apply reinforcement learning for optimizing molecular dynamics simulations.

Main Methods:

• Utilizing reinforcement learning to optimize an added force field.
• Minimizing Kullback-Leibler divergence between conditioned and driven trajectory ensembles.
• Employing low variance gradients with value functions for improved convergence.

Main Results:

• The optimized added force drives rare fluctuations to occur as typical events.
• Accurate variational estimates of rare event likelihoods were obtained.
• The method demonstrated efficiency and accuracy across various model systems.

Conclusions:

• Reinforcement learning provides a powerful tool for studying rare molecular dynamics.
• The developed method offers a significant advancement in simulating and analyzing rare events.
• This approach has broad implications for fields relying on molecular dynamics simulations.