Application of Symmetry Functions to Large Chemical Spaces Using a Convolutional Neural Network

Abstract

The use of machine learning in chemistry is on the rise for the prediction of chemical properties. The input feature representation or descriptor in these applications is an important factor that affects the accuracy as well as the extent of the explored chemical space. Here, we present the periodic table tensor descriptor that combines features from Behler-Parrinello's symmetry functions and a periodic table representation. Using our descriptor and a convolutional neural network model, we achieved 2.2 kcal/mol and 94 meV/atom mean absolute error for the prediction of the atomization energy of organic molecules in the QM9 data set and the formation energy of materials from Materials Project data set, respectively. We also show that structures optimized with a force field derived from this modelcan be used as input to predict the atomization energies of molecules at density functional theory level. Our approach extends the application of Behler-Parrinello's symmetry functions without a limitation on the number of elements, which is highly promising for universal property calculators in large chemical spaces.