A machine learning based deep potential for seeking the low-lying candidates of Al clusters

Area of Science:

• Computational Materials Science
• Machine Learning Applications
• Condensed Matter Physics

Background:

• Predicting the lowest-energy structures of atomic clusters is crucial for understanding material properties.
• Traditional methods for simulating atomic interactions can be computationally expensive, limiting the scope of investigations.
• Developing accurate and efficient interatomic potentials is essential for large-scale materials simulations.

Purpose of the Study:

• To develop a novel Machine-Learning based Deep Potential (DP) model for aluminum (Al) clusters.
• To accurately predict the low-lying energy structures of Al clusters across a broad size range.
• To demonstrate the efficacy of machine learning in generating reliable interatomic potentials for complex materials.

Main Methods:

• Training a Deep Potential (DP) model using an extensive database of ab initio data for bulk and cluster Al.
• Utilizing a computational approach that required only 6 CPU hours for model development.
• Performing extensive searches for low-lying structures of 101 different sized Al clusters using the developed DP model.

Main Results:

• The developed DP model exhibits high accuracy in predicting low-lying Al cluster candidates.
• The study successfully updated the lowest-energy candidates for 69 different sized Al clusters.
• The computational efficiency of the model allowed for a broad exploration of Al cluster structures.

Conclusions:

• Machine learning provides a powerful and efficient method for generating accurate interatomic potentials.
• The developed DP model significantly advances the ability to predict and understand the properties of Al clusters.
• This work highlights the potential of AI in accelerating materials discovery and design.