Learning Multidimensional Electronegativity of Selected Atom Types in Organic Molecules Using Graph Neural Networks

Electronegativity can be considered a data-driven concept that has been widely used since Pauling proposed this property. However, updating the electronegativity based on the vast amount of high-quality experimental and computational data has been overlooked. Thus, advances in artificial intelligence (AI), with its ability to manage large data sets and identify underlying patterns, necessitate reconsidering data-driven concepts such as electronegativity. In this work, we present a data-driven method to generate more informative multidimensional electronegativity of atoms in organic molecules using graph neural networks. Although this electronegativity can be extended to any dimension, we focused on 2D electronegativity to do a more detailed classification of the atoms and their covalent bonds. By replacing the conventional electronegativity with the newly proposed one, we observed performance improvement in molecular machine learning tasks. We believe that our findings will be useful in understanding electronegativity and chemical bonds by applying AI-driven methods to chemical studies.