Integrating Artificial Intelligence via Deep Reinforcement Learning for the Autonomous Discovery of High-Entropy Nanoclusters

The rational design and prediction of stable configurations for high-entropy alloy (HEA) nanoclusters present a formidable challenge due to the vast combinatorial space defined by size, shape, composition, and chemical ordering. While traditional global optimization algorithms are widely used, they face critical limitations in evaluator accuracy and computational efficiency for realistic systems. To overcome these constraints, we employed a deep reinforcement learning (DRL) framework designed for the rapid and accurate prediction of stable nanostructures. We demonstrate the efficacy of this approach on quaternary and quinary cluster systems. The agent explores the potential energy surface (PES) through stochastic configuration generation and atomic displacements. The local environment is encoded using atom-centered symmetry functions and integrates energy and force data for an accurate state evaluation, guided by a tailored reward function that balances exploration and exploitation. The agent is trained with the Trust Region Policy Optimization (TRPO) algorithm to ensure stable and efficient training, leading to rapid convergence, a strong bias toward low-energy states, and the effective sampling of diverse minima─outperforming conventional search methodologies. Subsequent first-principles validation confirms the stability of the identified nanoclusters, revealing negative formation energies. This thermodynamic stability is further analyzed by molecular dynamics simulations at elevated temperatures (500 and 700 K), which show no structural degradation. Sharp peaks in the radial distribution function (RDF) confirm the preserved crystallinity with minimal atomic displacement. The electronic structure analysis further reveals the emergence of pronounced pseudogaps near the Fermi level, a feature strongly associated with enhanced catalytic potential. Collectively, these results demonstrate that the DRL agent establishes a powerful and generalizable computational framework for the accelerated design of thermally robust functional materials for advanced applications.