Algebraic Graph-Based Machine Learning Model for Li-Cluster Prediction.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning Applications

Background:

• Determining the ground-state structure of medium-sized clusters is challenging due to numerous local minima on potential energy surfaces.
• Traditional global optimization methods using Density Functional Theory (DFT) are computationally expensive.
• Effective cluster representation for machine learning (ML) is a bottleneck for reducing computational costs.

Purpose of the Study:

• To propose an effective low-dimension representation for clusters.
• To develop an ML model for discovering cluster structure-energy relationships.
• To accelerate the search for globally stable cluster structures.

Main Methods:

• Introduced a multiscale weighted spectral subgraph (MWSS) as a novel cluster representation.
• Developed an MWSS-based machine learning model to predict structure-energy relationships.
• Combined the ML model with particle swarm optimization and DFT calculations for global structure search.

Main Results:

• Successfully implemented the MWSS representation for lithium clusters.
• The MWSS-based ML model effectively captured structure-energy relationships.
• The integrated approach successfully predicted the ground-state structure of Li20.

Conclusions:

• The multiscale weighted spectral subgraph (MWSS) is an effective low-dimension representation for clusters in ML.
• The developed MWSS-based ML model significantly aids in discovering structure-energy relationships.
• This approach accelerates the identification of globally stable cluster structures, as demonstrated for Li20.