Integrating Theory and Experiment with Graph Neural Networks to Classify Molecular Self-Assembly on Metal Surfaces.

Abstract

We introduce a graph neural network framework that integrates density functional theory (DFT) calculations and scanning tunneling microscopy experimental data to predict molecular self-assembly behavior on metal surfaces. We constructed a data set comprising 20 diverse aromatic precursor molecules and their corresponding assembled structures on Au, Ag, and Cu substrates. Leveraging DFT-derived descriptors, feature importance analysis identified the molecule-substrate interactions and interfacial charge transfer as the dominant factors governing assembly behavior. A modified graph attention network model, trained on this multisource data set, achieved exceptional predictive accuracy, which exceeded 95% in classifying molecular arrangements and attained a coefficient of determination (<i>R</i>²) of 0.985 for adsorption energy regression. The model's generalizability was further validated by accurately classifying the self-assembled layers of three previously unseen molecules. This study establishes a machine learning framework that bridges computational and experimental insights, paving the way for the rational design of surface-supported functional nanostructures.