Machine Learning Model to Predict Free-Energy Landscape and Position-Dependent Diffusion Constant to Extend the Scale...

Area of Science:

• Computational Materials Science
• Chemical Engineering
• Physical Chemistry

Background:

• Accurate prediction of mass transport properties in heterogeneous media is crucial for optimizing devices like fuel cells and batteries.
• Traditional methods like all-atom molecular dynamics (MD) are computationally expensive for large-scale simulations.
• Coarse-grained simulations offer efficiency but require accurate local property estimations.

Purpose of the Study:

• To develop a machine learning (ML) surrogate model to predict free-energy landscapes and local diffusion constants.
• To integrate the ML model into dynamic Monte Carlo (MC) simulations for efficient prediction of mass transport properties.
• To validate the ML-assisted MC method for molecular diffusion in heterogeneous systems.

Main Methods:

• Constructed an ML model using kernel functions based on local elemental distribution functions.
• Trained the ML model using data from all-atom molecular dynamics (MD) simulations.
• Employed dynamic Monte Carlo (MC) simulations with the ML surrogate for coarse-grained dynamics.

Main Results:

• The ML-assisted dynamic MC simulation accurately predicted global diffusion constants for hydrogen in perfluorinated ionomer membranes.
• Achieved a low relative error of 3% compared to traditional MC simulations with explicit property calculations.
• Demonstrated that even small training datasets can yield high accuracy (5% relative error) for the ML model.

Conclusions:

• The developed ML-guided coarse-grained simulation approach efficiently predicts mass transport properties in heterogeneous media.
• This method significantly reduces computational cost while maintaining high accuracy.
• The approach is broadly applicable to diverse systems including fuel cells, batteries, and biological systems.