Efficient Global Structure Optimization with a Machine-Learned Surrogate Model.

Area of Science:

• Computational materials science
• Chemical physics
• Surface science

Background:

• Accurate prediction of atomistic structures requires computationally intensive first-principles energy calculations.
• Global optimization of complex energy landscapes is a significant challenge in materials discovery.
• Existing methods like evolutionary algorithms can be computationally expensive for large search spaces.

Purpose of the Study:

• To develop an efficient global optimization scheme for atomistic structures using first-principles energy expressions.
• To accelerate the discovery of stable and metastable material configurations.
• To investigate the initial stages of graphene oxidation and intercalation on Ir(111).

Main Methods:

• A hybrid approach combining surrogate modeling (Gaussian processes) with first-principles calculations.
• Active learning of the potential energy landscape using an acquisition function for guided exploration.
• Dual kernel widths in Gaussian processes to capture both broad features and local minima.
• Application to surface reconstruction problems and graphene/Ir(111) surface phenomena.

Main Results:

• The proposed method achieves global optimization with first-principles accuracy.
• Demonstrated a two-orders-of-magnitude improvement over a first-principles evolutionary algorithm for surface reconstructions.
• Successfully identified initial stages of edge oxidation and oxygen intercalation in graphene on Ir(111).

Conclusions:

• The developed global optimization scheme offers a significant advancement in computational materials science.
• This approach accelerates the identification of complex surface phenomena and material structures.
• The method provides a powerful tool for exploring energy landscapes in atomistic systems.