Reinforcement Learning for Chemical Ordering in Alloy Nanoparticles

Area of Science:

• Materials Science
• Computational Chemistry
• Nanotechnology

Background:

• Determining optimal element ordering in bimetallic alloy nanoparticles (NPs) is crucial for their properties.
• Combinatorial complexity makes traditional search methods inefficient for large nanoparticle systems.

Purpose of the Study:

• To frame the search for optimal element ordering in bimetallic alloy NPs as a reinforcement learning (RL) problem.
• To develop and train an RL agent capable of performing global optimization of NP structures.

Main Methods:

• Utilized a reinforcement learning (RL) approach with a geometric graph representation for NPs.
• Trained an RL agent using composition-conserving atomic swap actions on icosahedral nanoparticle structures.
• Employed randomized Ag_XAu_309‑X compositions and orderings for agent training.

Main Results:

• The trained RL agent successfully discovered known ground-state structures for bimetallic NPs.
• Optimization demonstrated robustness across varied initial NP orderings.
• The RL policy showed effective extrapolation to NPs of unobserved sizes.
• Limited efficacy was observed when dealing with multiple alloying elements.

Conclusions:

• Reinforcement learning, combined with pretrained equivariant graph encodings, can effectively navigate complex ordering spaces in nanoparticles.
• This approach offers a transferable optimization strategy with potential for generalization across compositions.
• The RL method can significantly reduce the cost associated with repeated individual searches for optimal NP structures.