On-the-Fly Machine Learning of Atomic Potential in Density Functional Theory Structure Optimization.

Area of Science:

• Materials Science
• Computational Chemistry
• Surface Science

Background:

• Density Functional Theory (DFT) is crucial for understanding material properties.
• Accurately predicting stable surface reconstructions, like the SnO2(110)-(4x1) case, is computationally intensive.
• Evolutionary Algorithms (EA) are effective for global minimum energy searches but can be slow.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting local stability in DFT calculations.
• To integrate ML with an EA for accelerated structure prediction of the SnO2(110)-(4x1) reconstruction.
• To analyze the chemical insights provided by the ML model regarding atomic potentials.

Main Methods:

• Training a machine learning model on (structure, total energy) data from EA global minimum energy searches.
• Employing the ML model to guide the evolutionary algorithm during the search process.
• Analyzing the local atomic potentials derived from the trained ML model.

Main Results:

• The ML-guided EA significantly increased the speed of finding stable structures.
• The ML model successfully predicted local stability information for the SnO2 system.
• Chemically intuitive patterns were observed in the local atomic potentials generated by the ML model.

Conclusions:

• Machine learning provides an effective strategy to accelerate DFT calculations and structure prediction.
• The synergy between ML and EA offers a powerful computational tool for materials discovery.
• The derived atomic potentials offer valuable chemical insights into surface reconstructions.