Graph Neural Networks on One- and Two-Body Integrals for Molecular Energy Prediction

We present a graph neural network (GNN) framework for predicting molecular energies from molecular orbital graphs. Our approach leverages information from one- and two-electron integrals encoded as graph features, together with pooling strategies that map orbital-level predictions to molecular energies. The proposed approach was tested on 6 diatomic molecules, a total of 132 geometries, where the target molecular energy was computed using Full Configuration Interactions.Our results illustrate that a GNN architecture, even for this small dataset, is capable of learning the energy value given the molecular representation through the molecular orbitals.