Prediction of the overlap population diagram of organic molecules based on a graph neural network

Abstract Chemical bonds and electronic structures govern properties of materials and molecules. Overlap population (OP) diagrams, a key tool for understanding the chemical bonding, typically relies on computationally expensive quantum simulations. We developed a graph neural network model to directly predict OP diagrams from molecular structures, using a chemical bonding database of 2,370,000 bonds in 130,000 molecules. The model achieved high accuracy across diverse bond types and environments, reproducing qualitative trends even for unknown structures.