(502e) Neural Network Potential for Predicting Bond Dissociation Energies

Machine learning has become a ubiquitous tool in data science and is becoming increasingly common in fundamental computational chemistry. Of particular interest to this study are neural network potentials (NNPs). NNPs are trained to predict ab initio molecular energies and forces, similar to classical force fields, at a fraction of the computational cost of traditional quantum mechanical (QM) calculations. The recently developed family of NNPs known as the ``Accurate NeurAl networK engINe for Molecular Energies'' (ANAKIN-ME or ANI) was trained to a data set consisting of equilibrium and near-equilibrium structures of organic compounds. This study applies active learning (query by committee) and transfer learning to retrain ANI with an augmented data set consisting of dissociated radical species for the purpose of predicting bond dissociation energies (BDE). An extensive comparison is presented between the resulting NNP, traditional BDE prediction methods, and a novel quantitative structure-property relationship neural network model.