Molecules as Neural Networks

Given a quantum mechanical electron density for a molecule, an atom-in-molecule (AIM) decomposition partitions the molecular system into constituent atomic-like entities. This decomposition is not unique, and various theoretical methods have been proposed for defining chemically-reasonable AIMs. We explore the use of machine learning techniques, specifically radial basis function (RBF) neural networks, to analyze molecular electron density distributions as AIM ensemble decompositions. This AIM framework is expressed in terms of weighted superpositions of electron densities for individual ground and charged states of the constituent atoms. The neural network optimizes these weights to define the contributions of individual atomic states to the overall molecular density. Development of this framework leads to a deeper understanding of the interplay of ionic and covalent contributions of individual atoms to various molecular properties and interactions, as a function of molecular structure. We will present results on LiF as an exemplar of ionic bonding, and CO for covalent bonding, and discuss insights into different bonding regimes derived from the neural network representation.