Deep Reinforcement Learning for Structural Optimization and Potential Energy Landscape of 13-Atom Gold Nanoclusters for Application in Nanomaterial Discovery

The properties of atomic clusters strongly depend on their structure; therefore, identifying stable structures is critical for materials discovery. In this study, we implemented a deep reinforcement learning (DRL) framework to generate and optimize the gold nanoclusters (Au 13 ). The DRL framework is a policy-based actor-critic model with trust region policy optimization (TRPO). The DRL agent effectively navigates the potential energy surface (PES), starting from randomized configurations and progressively learning to identify low-energy and thermodynamically favorable states. The deep reinforcement learning model successfully predicts the global minimum cluster of Au 13 with icosahedral (Ih) symmetry and discovers 25 unique low-energy minima configurations, demonstrating its ability to balance exploration and exploitation effectively. Furthermore, the density functional theory (DFT) confirms the stability of the Au 13 nanocluster obtained from DRL. This work presents a significant advancement in the autonomous structural optimization of nanoclusters, offering insights into their stability and energy landscapes. Moreover, the proposed DRL framework demonstrates strong potential for accelerating nanomaterial discovery by identifying stable nanocluster structures through the efficient exploration of potential energy surfaces (PES).