Deep reinforcement learning unveils ternary nanocluster configurations: A case study on Ag6Pd5Cu4

Abstract

The identification of global minimum configurations in ternary nanoclusters faces challenges due to the inherent complexity of multimetallic potential energy surfaces. Traditional methods, such as genetic algorithms, suffer from computational inefficiency and premature convergence. In this study, we utilized a deep reinforcement learning framework integrating trust region policy optimization to balance exploration and exploitation in potential energy surface navigation. Applied to Ag6Pd5Cu4, the deep reinforcement learning framework predicted a global minimum configuration with the energy of 0.5324 eV/atom, validated by density functional theory. Effective medium theory potentials reduced computational costs as compared to density functional theory, enabling the rapid discovery of 15 distinct low-energy configurations. Ab initio molecular dynamics simulations confirmed thermal stability at 300 K. This work demonstrates the capability of deep reinforcement learning to autonomously resolve multimetallic complexity, offering a scalable pathway for accelerating nanocluster design in materials science.