Reinforcement learning based hybrid bond-order coarse-grained interatomic potentials for exploring mesoscale...

Area of Science:

• Computational chemistry
• Materials science
• Statistical mechanics

Background:

• Mesoscopic physical phenomena are challenging for all-atom molecular dynamics simulations due to timescale limitations.
• Coarse-graining offers a solution by reducing resolution while preserving molecular structure, unlike continuum methods.

Purpose of the Study:

• To develop a novel coarse-grained forcefield for modeling mesoscale aggregation phenomena.
• To create an interpretable and accurate model for liquid-liquid mixtures.

Main Methods:

• Introduced a hybrid bond-order coarse-grained forcefield (HyCG).
• Parameterized the HyCG potential using continuous action Monte Carlo Tree Search (cMCTS), a reinforcement learning (RL) algorithm.
• Trained the model using data from all-atom simulations.

Main Results:

• The developed RL-HyCG model accurately describes mesoscale critical fluctuations in binary liquid-liquid extraction systems.
• The cMCTS algorithm effectively captured molecular geometrical properties not explicitly included in the training data.

Conclusions:

• The RL-HyCG model provides an interpretable and accurate approach for mesoscale simulations.
• This method and training workflow can be extended to study other mesoscale phenomena inaccessible to all-atom simulations.