Accelerating global search of gold-silver clusters using equivariant graph neural network.

Area of Science:

• Computational chemistry
• Materials science
• Nanotechnology

Background:

• Medium-sized gold-silver clusters are computationally challenging to study using traditional methods like density functional theory (DFT).
• Graph neural networks (GNNs) offer efficient and accurate fitting of potential energy surfaces for complex molecular systems.
• Equivariant GNNs enhance information extraction without substantial computational overhead.

Purpose of the Study:

• To develop an efficient and accurate equivariant graph neural network (GNN) model for predicting properties of gold-silver clusters.
• To accelerate the discovery of global minimum structures for medium-sized gold-silver clusters.
• To uncover novel structural motifs and understand stability principles in Au-Ag nanoclusters.

Main Methods:

• Development of CCCNet, an equivariant GNN requiring only atomic coordinates and elemental information.
• Training CCCNet on a large dataset of over 1.4 × 10^6 gold-silver cluster structures.
• Integration of CCCNet with a comprehensive genetic algorithm (CGA) for global structure searches.

Main Results:

• CCCNet achieved high prediction accuracy for binding energies (MAE = 6.5 meV/atom) and atomic forces (MAE = 25.4 meV/Å).
• Global minimum structure searches for Au_mAg_n clusters (m + n = 20, 24, 30) were performed with computational costs reduced by three orders of magnitude compared to DFT.
• Several previously unknown low-energy configurations and novel structural motifs were discovered.

Conclusions:

• Equivariant GNNs, like CCCNet, are powerful tools for accelerating structural discovery in medium-sized clusters.
• The developed model provides new insights into the stability and design principles of gold-silver nanoclusters.
• This approach significantly reduces the computational burden, enabling broader exploration of complex nanomaterials.