Machine-learning-accelerated simulations to enable automatic surface reconstruction.

Area of Science:

• Computational materials science
• Surface science
• Statistical mechanics

Background:

• Accurate prediction of material surface structures is crucial for catalysis and electronics.
• Ab initio simulations offer predictive power but are computationally expensive for large phase spaces.
• Existing methods struggle with the computational cost of simulating complex multicomponent material surfaces.

Purpose of the Study:

• To develop an accelerated computational framework for predicting surface phase diagrams of multicomponent materials.
• To overcome the limitations of traditional ab initio simulations in terms of computational cost and phase space sampling.
• To enable the discovery of novel surface terminations and structures.

Main Methods:

• A bi-faceted computational loop combining accelerated energy scoring and statistical sampling.
• Training machine learning interatomic potentials using high-throughput density-functional-theory (DFT) calculations and active learning.
• Employing Markov chain Monte Carlo (MCMC) sampling in the semigrand canonical ensemble with virtual surface sites.

Main Results:

• The developed method significantly accelerates both energy evaluation and statistical sampling for surface phase diagrams.
• Predicted surface phase diagrams for GaN(0001), Si(111), and SrTiO3(001) show agreement with existing experimental and theoretical data.
• The strategy successfully models complex material surfaces and identifies previously unreported surface terminations.

Conclusions:

• The proposed computational strategy provides an efficient and accurate approach for predicting surface phase diagrams of multicomponent materials.
• This method can significantly advance research in surface science, catalysis, and materials design.
• The framework facilitates the exploration of complex material surfaces and the discovery of new structural configurations.