Hierarchical Graph Network for Fullerene Energy Prediction

Tong Cai, Asiri Wijesinghe, Sergio José Rodríguez Méndez, Amanda J. Parker

Keywords: Fullerene, Graph Neural Networks, Molecular Property Prediction, Message Passing Neural Network, Material Discovery, Carbon, Lattice, Dual Graph

TL;DR: A novel hierarchical message passing neural network that encodes the multi-scale geometric structure of fullerene molecules.

Abstract: We propose a hierarchical message passing neural network that explicitly exploits the multi-scale geometric structure of fullerenes. State-of-the-art molecular GNNs and standard MPNNs fail to distinguish structures with nearly identical local environments, such as the repetitive lattices of carbon allotropes. Our approach successfully learns meaningful hierarchical representations for fullerene systems by encoding information from an atomic graph representation and its dual. We construct a message passing scheme that is informed by the symmetry underlying the fullerene structure (C5 and C6). When trained on a Fullerene dataset of 2487 neutral structures, we outperform MEGNet and MPNNs for electronic and structural energy predictions.

Submission Number: 5