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

Area of Science:

• Nanomaterials Science
• Computational Chemistry
• Artificial Intelligence in Chemistry

Background:

• Coinage-metal nanoclusters (CMNs) offer precise structure-property relationships but their synthesis relies on trial-and-error.
• Deterministic synthesis of predesigned CMNs (inverse synthesis) is challenging due to post-synthesis structural determination.
• Current methods lack predictive power for guiding synthesis before experimentation.

Purpose of the Study:

• To introduce CoLiM, a deep neural network framework for predicting chemical compatibility before CMN synthesis.
• To enable inverse synthesis and precise atomic-level modification of nanoclusters.
• To accelerate the rational discovery and design of novel CMNs.

Main Methods:

• Developed a deep neural network (CoLiM) with a dual-encoder architecture.
• Trained CoLiM on a dataset of 1,989 reported CMN structures and gas-phase cluster data.
• Validated CoLiM's predictive performance using an area under the curve (AUC) metric on a held-out test set.

Main Results:

• The optimal CoLiM model achieved an AUC exceeding 0.83, outperforming baseline methods.
• CoLiM successfully guided the single-atom editing of a copper nanocluster, synthesizing a predesigned structure.
• Demonstrated the framework's generalizability and practical utility in real experimental conditions.

Conclusions:

• CoLiM facilitates the inverse synthesis of CMNs by predicting pre-synthesis chemical compatibility.
• The framework enables precise atomic-level modification of nanocluster structures.
• CoLiM significantly accelerates rational nanocluster discovery and design.