Clustering methods for the optimization of atomic cluster structure.

Area of Science:

• Computational Chemistry
• Machine Learning
• Optimization

Background:

• Clustering methods were once efficient global optimization techniques but declined due to challenges in high-dimensional spaces.
• Existing methods struggle with large-scale cluster conformation problems.

Purpose of the Study:

• To redesign clustering methods for molecular structures in a reduced feature space.
• To develop an effective optimization tool for finding cluster conformations with high success rates.

Main Methods:

• Inspired by machine learning, clustering methods were adapted for reduced feature spaces.
• Utilized geometrical features and an efficient descent method for optimization.
• Applied the method to Lennard-Jones and Morse potentials for medium-sized clusters.

Main Results:

• The revised method successfully finds known putative optima for medium-sized clusters without prior information.
• A significant reduction in computational searches was achieved by performing expensive local searches only when necessary.
• The method demonstrates reliability and efficiency in global optimization.

Conclusions:

• The proposed clustering scheme in a reduced variable space offers a straightforward way to save local searches.
• This approach can be integrated into various modern optimization methods.
• The method is effective for large-scale cluster conformation problems.