Abstract
The global minimum search on nanocluster potential energy surfaces remains a formidable challenge due to exponential scaling of local minima with cluster size and composition. Deep reinforcement learning (DRL) offers a promising alternative to traditional genetic algorithms and basin hopping, but prior DRL frameworks are restricted to seven noble metals [via effective medium theory (EMT) potentials] and lack systematic validation. Here, we integrate proximal policy optimization (PPO) with an embedded atom method (EAM) potential to create PPO–EAM–DRL for autonomous nanocluster optimization. The framework is applied to Ag1Ni12, Ag6Ni7, Ag7Ni6, Ag12Ni1, and Pt6Co7 nanoclusters. The agent learns stable exploration policies within 300–500 episodes and consistently identifies lower-energy global minima than those obtained using a Birmingham cluster genetic algorithm, with particularly significant improvements for Ag6Ni7 and Pt6Co7 systems. Machine learning atomic cluster expansion re-optimization further confirms that EAM-derived minima are energetically more stable or equivalent to EMT-derived minima across all investigated Ag–Ni compositions. By replacing EMT with EAM, the proposed framework extends DRL-driven nanocluster optimization from a limited set of metallic systems to a broader transition-metal chemical space. This work establishes a transferable and extensible platform for autonomous materials discovery and global minimum search in multimetallic nanoclusters.