Reinforced dynamics for enhanced sampling in large atomic and molecular systems.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Efficiently exploring configuration space and computing free energy are crucial for understanding large atomic and molecular systems.
• Traditional methods face challenges with system size and the selection of collective variables.

Purpose of the Study:

• To propose a new approach for efficient exploration of configuration space and free energy computation in large systems.
• To leverage deep reinforcement learning for adaptive biasing potentials in molecular dynamics.

Main Methods:

• The method combines principles from metadynamics and deep reinforcement learning.
• An adaptively computed biasing potential is added to the original dynamics, trained on-the-fly using deep neural networks.
• Uncertainty from the neural network model serves as the reward function.

Main Results:

• The approach enables efficient exploration of configuration space for large atomic and molecular systems.
• Neural network parameterization allows handling a large set of collective variables, reducing the criticality of their precise selection.
• Demonstrated effectiveness on full-atom explicit solvent models of alanine dipeptide, tripeptide, and a polyalanine-10 molecule.

Conclusions:

• This novel method offers a powerful tool for advancing molecular simulations.
• The integration of deep reinforcement learning provides a more robust and flexible approach to free energy calculations.
• The technique shows promise for studying complex structural transformations in large molecular systems.