Des réseaux de neurones pour prédire des distances interatomiques extraites d'une base de données ouverte de calculs en chimie quantique
In EGC 2019, vol. RNTI-E-35, pp.9-20
The calculation of the geometry of a molecule's fundamental state is the starting point for the vast majority of molecular quantum chemistry research. PubChemQC, an open database, provides the results of fundamental state calculations for more than three million molecules. We have extracted the converged geometries to train machine learning models. Predicting the complete geometry would be a remarkable step forward. Our initial results suggest that it is difficult to train a neural network on this complex task. On the other hand, we demonstrate that a neural network is capable of accurately predicting a distance between two atoms. The subject of this work is the most complex distance in organic chemistry, the carbon-carbon distance. The best results are obtained by limiting the amount of information through a cut-off distance around each carbon.