A Divide-and-Conquer Approach to Nanoparticle Global Optimisation Using Machine Learning.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Global optimization of atomic nanoparticles is challenging due to complex potential energy surfaces with multiple funnels.
• Narrow funnels on these surfaces are difficult to explore, hindering the identification of the global minimum structure.

Purpose of the Study:

• To develop a machine learning-based divide-and-conquer approach to overcome the multifunnel effect in nanoparticle structure optimization.
• To improve the efficiency of locating global minima without prior knowledge of the potential energy surface.

Main Methods:

• A coarse-grained exploration of the potential energy surface was performed.
• A machine learning Gaussian mixture model was trained to partition the surface into distinct regions.
• Each region was then explored independently using a divide-and-conquer strategy.

Main Results:

• Significant improvements in the time required to find global minima were observed for Lennard-Jones (LJ) nanoparticles (LJ75, LJ104) and metallic clusters (Au55, Pd88).
• The approach successfully navigated multifunnel systems, demonstrating its effectiveness.

Conclusions:

• The machine learning-driven divide-and-conquer method effectively addresses the multifunnel problem in atomic nanoparticle optimization.
• Further refinements are suggested based on observed difficulties with specific systems like LJ98.