RGBChem: Image-Like Representation of Chemical Compounds for Property Prediction

In this work, we introduce RGBChem, a novel approach for converting chemical compounds into image representations, which are subsequently used to train a convolutional neural network (CNN) to predict the HOMO-LUMO gap for compounds from the QM9 database. By modifying the arbitrary order of atoms present in .xyz files used to generate these images, it has been demonstrated that expanding the initial training set size can be achieved by creating multiple unique images (data points) from a single molecule. This study shows that the presented approach leads to a statistically significant improvement in model accuracy, highlighting RGBChem as a powerful approach for leveraging machine learning (ML) in scenarios where the available data set is too small to apply ML methods effectively.