Machine learning enhanced global optimization by clustering local environments to enable bundled atomic energies.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Global optimization of molecular structures is computationally intensive.
• Accurate energy calculations are crucial for determining stable molecular configurations.
• Machine learning offers potential for accelerating complex scientific computations.

Purpose of the Study:

• To develop a machine learning-based method for accelerating global optimization of molecular structures.
• To introduce an effective feature vector representation for molecular structures.
• To demonstrate enhanced efficiency in finding global minimum energy structures.

Main Methods:

• Introduction of the 'auto-bag' feature vector combining local atomic features, unsupervised clustering, and cluster counts.
• Supervised learning to assign local energies to atoms based on accumulated structure-energy data.
• Application in basin hopping searches for Lennard-Jones and density functional theory-based carbon systems.

Main Results:

• The auto-bag feature vector effectively represents molecular structures for machine learning.
• On-the-fly derivation of local energy information significantly enhances optimization speed.
• The method demonstrated improved efficiency in finding global minimum energy structures for tested systems.

Conclusions:

• The proposed machine learning approach, utilizing the auto-bag feature vector and learned local energies, effectively speeds up global structure optimization.
• This method provides a promising strategy for efficient exploration of complex energy landscapes in molecular systems.
• The approach is validated across different potential models and system sizes, indicating broad applicability.