Reinforcement Learning for Transferable Force Fields in Binary Nano-Alloy Synthesis and Soft Landing

Traditional molecular dynamics (MD) simulations often struggle to accurately model binary nano-alloy clusters due to the limitations of force fields based on bulk crystalline data, which fail to account for the unique sizes and compositions of nanoclusters. To address this challenge, we present a novel reinforcement learning approach for developing adaptable force fields specifically tailored to the size and composition of binary nano-alloy clusters. By utilizing a comprehensive dataset derived from first-principles nanocluster data, our method optimizes a Tersoff Bond Order Potential, covering a wide range of cluster sizes and compositions. This advanced force field enhances the accuracy of dynamic and structural predictions compared to density functional theory (DFT) results while significantly improving computational efficiency.<br/>We validate the practical application of our approach through MD simulations of the gas-phase synthesis and soft landing of AuxAgy nanoclusters on graphite surfaces. These simulations explore the detailed formation mechanisms of nanoclusters from atomic vapors, demonstrating that cluster formation proceeds via sequential formations of dimers and trimers, which grow through agglomeration and coalescence. Additionally, our findings reveal that the morphology and deposition dynamics of clusters during soft landing are profoundly influenced by the strength of cluster-substrate interactions, deposition velocities, composition, and substrate temperature. These insights into cluster formation, stability, and interaction dynamics are vital for advancing technological applications in fields such as catalysis and materials science. Our tailored force fields thus offer significant potential for enhancing the predictive power and efficiency of molecular simulations.