Generation and optimization of gold nanoclusters via reinforcement learning

The identification and prediction of atomic cluster structures are crucial in nanocluster and materials research, as the molecular structure significantly influences the properties of nanoclusters. This study presents an innovative approach for generating and optimizing gold Au₁₃, Au₇, Au_6, and Au₅ nanoclusters using reinforcement learning (RL). Conventional techniques for optimizing nanoparticle structures are significantly expensive in computation and have some restrictions when exploring a broad range of design possibilities. To overcome these challenges, we used a policy-based RL model that learns how to arrange atoms on a canvas to minimize the potential energy of the nanocluster, like an actor–critic model. The agent works under a reward function based on the molecule’s energy, systematically positioning atoms on a canvas until it reaches convergence. The performance and evaluation of our RL model are assessed by local optimization techniques, specifically the BFGS optimization algorithm and simulated annealing. We conclude that the RL method is effective for identifying the configuration of Au₁₃ nanoparticles and achieving a stable and low-energy icosahedral structure. The complexity of the energy landscape of nanoalloys renders the determination of their structure a complicated task. This study points out the potential of reinforcement learning in materials science for designing and optimizing nanoparticles with stability characteristics.

Graphic abstract

A schematic representation of the actor-critic reinforcement learning model. The input data is processed into a state, which the critic evaluates to estimate the value function. The actor uses the state to determine the action parameters, influencing the next state. The process continues as the agent learns to maximize the reward