Charting Nanocluster Structures via Convolutional Neural Networks.

Area of Science:

• Computational materials science
• Nanotechnology
• Machine learning applications

Background:

• Understanding the structural landscape of nanoparticles is crucial for predicting their properties and behavior.
• Traditional methods struggle to represent the complexity of nanoparticle structures across various conditions.
• Molecular dynamics simulations generate vast datasets of atomic configurations.

Purpose of the Study:

• To develop a general method for representing nanoparticle structural landscapes using a limited set of variables.
• To create a low-dimensional, physically meaningful, and differentiable mapping of atomic positions.
• To enable clear discrimination and ordering of structural motifs and facilitate analysis.

Main Methods:

• Applied a novel method leveraging convolutional neural networks (CNNs) to analyze radial distribution functions.
• Utilized parallel tempering molecular dynamics simulations for gold, silver, and copper nanoclusters.
• Employed unsupervised clustering on the learned low-dimensional manifold for fine-grained motif analysis.

Main Results:

• Successfully generated a low-dimensional chart effectively visualizing the complex structural landscape of nanoparticles.
• Identified and clearly discriminated main structural motifs, revealing meaningful ordering.
• Unsupervised clustering resolved structural subfamilies based on subtle differences like defects and coordination.

Conclusions:

• The proposed CNN-based method provides a powerful tool for nanoparticle structure visualization and analysis.
• The generated low-dimensional variables offer potential for enhanced sampling and exploration in simulations.
• This approach enables tracking complex structural evolution, even in reactive trajectories.