Deep reinforcement learning for AgPd-based multimetallic nanoclusters: Accelerating global minimum discovery in high-entropy alloy systems

Abstract

The combinatorial complexity and rugged potential energy surfaces of multimetallic nanoclusters, particularly high-entropy alloys, challenge traditional optimization methods, such as genetic algorithms and basin hopping. This work utilized a deep reinforcement learning framework to predict stable atomic arrangements for AgPd-based multimetallic systems, including high-entropy alloys, such as Ag3Pd3Au3Pt3Cu3 and Ag5Pd5Au5Pt5Cu5 nanoclusters. The deep reinforcement learning agent employs atom-centered symmetry functions with element-specific descriptors to encode nanocluster geometries and strategically navigate the configuration space via a hybrid action space, balancing atom selection and displacement. Predicted configurations are validated through density functional theory calculations of formation energies, confirming the thermodynamic stability of DRL-generated structures. Ab initio molecular dynamics simulations confirmed the thermal stability of high entropy alloy nanoclusters at 300 K. This approach demonstrates the potential of deep reinforcement learning to accelerate the discovery of stable multimetallic nanoclusters, offering a pathway for rational design of advanced nanomaterials.