Lightweight Extendable Stacking Framework for Structure Classification in Atomistic Simulations.

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Accurate identification of local atomic crystal structure is vital for atomistic simulations.
• Traditional methods struggle with complex materials and require manual parameter tuning.
• Machine learning offers potential but often demands extensive data and computation, especially for novel structures.

Purpose of the Study:

• To develop a highly accurate, efficient, and flexible classifier for atomic crystal structure identification.
• To overcome limitations of traditional methods and data-intensive machine learning approaches.
• To enable robust analysis of diverse and complex material systems.

Main Methods:

• Proposed a Lightweight and Extendable Stacked Structure (LESS) classifier.
• Utilized bond orientational order parameters as input features.
• Assembled multiple efficient machine learning algorithms as base models.

Main Results:

• Achieved over 98.8% accuracy in classifying amorphous, mono-, and binary crystal structures.
• Outperformed many existing methods, including some deep learning approaches.
• Demonstrated probabilistic classification for complex environments like interfaces.
• Showcased efficient retraining capabilities for novel, unknown crystal structures.

Conclusions:

• The LESS classifier provides a high-accuracy, flexible, and efficient tool for atomic structure identification.
• It surpasses current methods in accuracy and adaptability, even for unfamiliar structures.
• The model's efficiency in learning and retraining makes it suitable for complex material analyses.