Optimization of Silver Nanocluster Geometries: A Deep Reinforcement Learning Approach to Identifying the Most Stable Configurations in Ag15 Cluster

Abstract

Silver nanoclusters (Ag₁₅) are of huge interest due to their unique electronic, optical, and catalytic properties, which all strongly depend on their atomic geometries. Finding their most stable configurations is an important step toward understanding and exploiting such properties. Traditional optimization methods, including genetic algorithms and basin hopping, usually perform poorly for systems with complex potential energy surfaces. This paper reports on a Deep Reinforcement Learning based strategy to explore and optimize Ag₁₅ clusters. We show that the navigation of the potential energy surface of such silver nanoclusters through deep reinforcement learning allows us to identify the most stable configurations more efficiently than through conventional methods. This method has been demonstrated to perform much better than previous approaches regarding computational efficiency and quality of the identified configurations, hence providing new insights into nanocluster stability.