Predicting Interatomic Distances of Molecular Quantum Chemistry Calculations

Jules Leguy, Thomas Cauchy, Béatrice Duval, Benoit Da Mota

Abstract: Geometry calculation of a molecule’s fundamental state is the starting point for the vast majority of molecular quantum chemistry researches. Few databases provide the results of these fundamental state calculations for large numbers of molecules. Our long term objective is to train machine learning models on such data to predict different kinds of molecules properties. Predicting the complete geometry would be a remarkable step forward. We first present results suggesting that it is difficult to train a neural network on this complex task. Then we demonstrate that a neural network can accurately predict the distance between atom pairs. The best results have been obtained by considering a neighborhood around each atom. This neighborhood depends on a cut-off distance and contains a limited number of atoms.