Accelerating atomic structure search with cluster regularization.

Area of Science:

• Materials Science
• Computational Chemistry
• Surface Science

Background:

• Global structure optimization is crucial for predicting material properties.
• Current methods can be computationally expensive and may get trapped in metastable states.
• Efficiently exploring the energy landscape of atomic compounds is a significant challenge.

Purpose of the Study:

• To develop and demonstrate a novel method for accelerating the global structure optimization of atomic compounds.
• To improve the efficiency of finding stable surface reconstructions.
• To leverage machine learning for enhanced materials discovery.

Main Methods:

• Utilized unsupervised machine learning for categorizing atoms based on local atomic environments in disordered structures.
• Developed a hybrid optimization approach combining gradient-based energy minimization with a novel gradient-based minimization of summed cluster distances.
• Applied the method within a density functional tight-binding theory framework using an evolutionary algorithm.

Main Results:

• Successfully accelerated the identification of the anatase TiO2(001)-(1 × 4) surface reconstruction.
• Demonstrated a correlation between total energy and the summed distances of atomic environments to cluster centers.
• The proposed method effectively escapes metastable energy basins, enhancing global optimization performance.

Conclusions:

• The presented machine learning-guided method significantly speeds up global structure optimization for atomic compounds.
• This approach offers a promising strategy for efficient materials design and discovery.
• The technique effectively navigates complex energy landscapes, overcoming limitations of traditional optimization methods.