Automatic Process Exploration through Machine Learning Assisted Transition State Searches.

Area of Science:

• Computational materials science
• Surface science
• Chemical physics

Background:

• Kinetic Monte Carlo (kMC) simulations often rely on predefined elementary processes, limiting discovery.
• Human intuition can be a bottleneck in identifying all relevant atomic processes for complex systems.

Purpose of the Study:

• To introduce an efficient Automatic Process Explorer (APE) framework.
• To overcome the limitations of human intuition in defining process lists for simulations.
• To enhance the accuracy and scope of kMC simulations.

Main Methods:

• Developed an APE framework utilizing a fuzzy machine learning classification algorithm.
• Minimized redundancy in transition-state searches by targeting unexplored atomic environments.
• Applied APE to study island diffusion on a Pd(100) surface.

Main Results:

• APE identified a large number of previously disregarded low-barrier collective diffusion processes.
• These collective processes significantly increase island diffusivity compared to traditional models.
• The framework efficiently explores atomic configurations, reducing redundant searches.

Conclusions:

• The APE framework provides an efficient, automated method for discovering elementary processes.
• It enhances the predictive power of kMC simulations by including a wider range of atomic mechanisms.
• This approach is crucial for accurately modeling surface diffusion and other complex phenomena.