Deep Learning Framework for Atomic-Level Design and Presynthesis Prediction of Coinage-Metal Nanoclusters

The atomically precise nature of coinage-metal nanoclusters (CMNs) enables systematic exploration of structure-property relationships and motivates application oriented inverse design. However, the synthesis of CMNs typically relies on trial-and-error methods, with atomic-level structures only revealed through crystallography (postsynthesis), posing a major challenge to the deterministic synthesis of predesigned cluster structures, which is known as <b>inverse synthesis</b>. Here, we introduce CoLiM, a deep neural network framework that predicts the chemical compatibility between the unexplored inorganic core and ligands <i>before synthesis</i>. CoLiM employs a dual-encoder architecture and is trained on a newly constructed dataset comprising 1,989 reported CMN structures, supplemented by an additional gas-phase cluster dataset. The optimal CoLiM model achieves an area under the curve (AUC) exceeding 0.83 on a held-out test set, outperforming all of the baseline methods. To demonstrate its practical utility, CoLiM is applied to address the long-standing challenge of achieving atomically precise structural tailoring. Starting from [Cu₂₀Cl-(PET)₁₂(PPh₃)₄-(MeCOO)₆]⁺, we successfully performed single-atom editing on its inorganic core to synthesize [Cu₁₉Cl-(PET)₁₂(PPh₃)₃-(HCOO)₆] guided by the prediction of CoLiM, validating the model's generalizability under real experimental conditions. Our framework facilitates the inverse synthesis and precise atomic-level modification of nanoclusters, underscoring its substantial potential to accelerate rational nanocluster discovery.