Deepcluster agent to search global minimum of nanoclusters using deep reinforcement learning: A review

Nanoclusters occupy an intermediate regime between molecules and bulk matter, where subtle changes in atomic arrangement can substantially modify optical, electronic, and catalytic behavior. Reliable identification of global minimum (GM) geometries therefore remains a central objective, yet the underlying potential energy surfaces (PES) are high-dimensional, rugged, and multi-funnel, making exhaustive searches impractical. Conventional global optimization (GO) strategies (e.g., genetic algorithms and basin hopping) can suffer from premature trapping, slow convergence, and strong sensitivity to algorithmic settings, while DFT-based local relaxations become increasingly prohibitive as system size grows. Although machine-learning surrogates can accelerate PES evaluations, their performance often depends on extensive reference data and may degrade when configurational sampling is incomplete across compositions and motifs. This review synthesizes recent progress in AI-enabled nanocluster structure search with an emphasis on deep reinforcement learning (DRL) for direct PES exploration. We comparatively assess recently reported DRL algorithms, descriptor choices for encoding local atomic environments, and PES evaluation strategies, and we highlight the on-policy TRPO–ACSFs–EMT workflow (Deepcluster agent) as a practically demonstrated route for dataset-free policy learning with DFT reserved for high-fidelity re-evaluation and refinement of low-energy candidates. Across representative monometallic, bimetallic, and trimetallic nanoclusters, this framework has been shown to recover GM structures and systematically catalogue low-lying isomers, including motifs exhibiting icosahedral (I h ) and octahedral (O h ) symmetries. Furthermore, this comprehensive review discusses current limitations—particularly energy-model bias and transferability—and outline directions for improving robustness via multi-fidelity/active-learning rewards and broader benchmarking of DRL operators for complex nanocluster PES.