Active learning-driven global search for neutral gold clusters via neural network potential

Area of Science:

• Computational chemistry
• Materials science
• Nanotechnology

Background:

• Predicting metal nanocluster structures is computationally challenging due to complex energy landscapes.
• First-principles calculations are often too costly for extensive structure prediction.

Purpose of the Study:

• To develop an efficient framework for metal nanocluster structure prediction.
• To accurately identify low-energy structures of gold nanoclusters (Au_n, n=30-45).

Main Methods:

• Integrating machine learning interatomic potentials (MLIPs) with global optimization algorithms.
• Iteratively training neural network atomic potentials to density-functional-theory (DFT) accuracy.
• Employing a genetic algorithm for exploring complex energy landscapes.

Main Results:

• The framework successfully identified low-energy structures for Au_n clusters at reduced computational cost.
• A non-monotonic structural evolution was observed, transitioning from hollow cages to multi-core cages.
• Significant differences in structural evolution were noted between neutral and anionic gold clusters.

Conclusions:

• The proposed active-learning workflow offers an efficient and extensible strategy for metal cluster structure prediction.
• The findings highlight the impact of electronic structure on nanocluster morphology.
• This approach can be applied to investigate metal clusters with complex electronic properties.