Utilizing Data-Driven Optimization to Automate the Parametrization of Kinetic Monte Carlo Models

Area of Science:

• Computational Physics
• Materials Science
• Data Science

Background:

• Kinetic Monte Carlo (kMC) simulations are crucial for studying dynamic stochastic systems.
• High computational costs and challenges in parameterization limit kMC applicability, especially for complex systems.
• Automating parameterization is essential for efficient kMC model utilization.

Purpose of the Study:

• To develop a data-driven methodology for efficient and systematic input parameterization of kMC simulations.
• To reduce the computational burden associated with finding optimal parameters for kMC models.
• To enhance the usability of kMC simulations in complex scientific and industrial applications.

Main Methods:

• Coupling kMC simulations with Gaussian Processes (GPs) and Bayesian Optimization (BO) in a feedback loop.
• Utilizing fast kMC simulation results to train a cheap-to-evaluate GP surrogate model.
• Employing a system-specific acquisition function for BO to guide parameter prediction.

Main Results:

• The developed methodology enables systematic and data-efficient input parametrization for kMC models.
• Demonstrated effectiveness in parameterizing space-charge layer formation in solid-state electrolytes.
• Achieved accurate parameter reconstruction within 1-2 iterations and successful extrapolation beyond training data.
• The surrogate model's accuracy was validated, potentially making original kMC simulations obsolete.

Conclusions:

• The integration of GPs and BO offers a powerful solution for overcoming kMC computational limitations.
• This data-driven approach significantly accelerates the parameter discovery process for complex simulations.
• The methodology shows high potential for applications in materials science, particularly in areas like solid-state batteries.