Analysis of polynuclear aromatic hydrocarbons via the Voronoi tessellation approach: classification of atom types using artificial neural networks

When identifying the correct atom types to occupy the specific atomic locations within newly observed structures or when assessing the plausibility of new suggested structures with specific locations for specific types of atoms, any information quantifying geometrically the local environments around those locations is valuable, provided known characteristic differences exist, with respect to this geometric information, between the different atom types. A powerful tool for quantifying such geometries is the Voronoi tessellation; this has been used in a pilot study of polynuclear aromatic hydrocarbons. It has been found that perfect identification of all C and H atoms may be achieved through the examination of polyhedral volumes and surface areas. The use of a weighted face-area average is also found to be a useful measure of local structure. Simple neural network models that may be used for atom-type prediction are given in the paper. It is expected that the present approach will be useful in distinguishing between atoms that have close scattering curves whilst displaying similar crystallographic behaviour.