Prediction of UV-vis absorption spectra of thiolate-protected gold nanoclusters based on graph convolutional neural networks

Accurate and efficient prediction of the absorption spectra of gold nanoclusters (Au_<i>n</i>(SR)_<i>m</i>) is essential for elucidating the structure-activity relationship between the cluster structure and optical properties, as well as guiding the synthesis of new clusters. In this study, we develop a graph convolutional neural network (GCNN) method that directly predicts the absorption spectrum and properties of Au_<i>n</i>(SR)_<i>m</i> clusters based on their atomic-level structural characteristics. In this method, Au_<i>n</i>(SR)_<i>m</i> clusters are represented as topological graphs and various features of the atomic surrounding environment are generated, including tensors such as the bond length and bond angle. The model framework can accurately predict the properties of nanoclusters with limited training data. Based on this neural network framework, we propose four extension methodologies for the precise prediction of the UV-vis absorption spectral curves of gold nanoclusters: prediction through key points, oscillator intensity, the discretized spectral curve, and encoder dimensionality reduction representation. We demonstrate the graph-based characterization method for gold nanoclusters and the ability of GCNN to predict complex targets, such as absorption spectra.