Mining of effective local order parameters for classifying crystal structures: A machine learning study.

Area of Science:

• Materials Science
• Computational Chemistry
• Statistical Mechanics

Background:

• Understanding local molecular structures is crucial for materials science and technology.
• Molecular simulations generate microscopic data, but analyzing local structures is challenging.
• Bond-orientational order parameters (BOOPs) are effective for classifying local structures like crystalline and liquid phases.

Purpose of the Study:

• To comprehensively evaluate the classification capabilities of various BOOPs, including Lechner parameters.
• To identify optimal combinations of BOOPs for distinguishing different local structures.
• To develop a systematic and automated method for discovering effective order parameters.

Main Methods:

• Evaluation of 112 species of BOOPs and 234,248 parameter combinations.
• Systematic and automated performance assessment using machine learning techniques.
• Comparison of identified parameter sets against conventional methods for classifying local structures.

Main Results:

• Identified optimal sets of two and three BOOPs for distinguishing four types of local structures with high accuracy.
• Demonstrated the effectiveness of machine learning for automated discovery of BOOPs.
• Showcased the potential of under-explored Lechner parameters for precise structural classification.

Conclusions:

• Machine learning provides a powerful framework for systematic and accurate mining of effective order parameters.
• The study offers improved sets of BOOPs for classifying complex local structures in molecular systems.
• This approach facilitates a deeper understanding of material functions through precise structural analysis.