Machine Learning-Aided Prediction of Molecular Self-Assembly on Metal Surfaces

The self-assembly of small organic molecules on metal surfaces presents a promising approach for fabricating numerous functional nanostructures. However, the diversity of precursor molecules and the large-scale self-assembly process pose a challenge for investigating molecular self-assembly via scanning tunneling microscopy (STM) techniques and density functional theory (DFT) calculations. We propose a data-driven random forest classification (RFC) algorithm to predict the self-assembly behavior of various precursor molecules on metal surfaces. Taking nucleobases and their derivatives as representatives, we have constructed a data set consisting of both experimental STM characterizations and DFT calculations. The RFC model is well-trained with 13 features and shows desirable prediction on determining the molecules arrangement and identifying the diversity of self-assembled structures. The importance of these features in predicting the targets of molecular self-assembly are analyzed based on the RFC model. To validate our RFC model, the self-assembly behavior of three new molecules that are not involved in the training data set are predicted on Au and Cu surfaces, which agrees well with the experimental observations. Our strategy provides essential insights into understanding the origin of molecular self-assembly and aids the rational design of the targeted nanostructure.