Abstract
In this work, we explore how existing datasets of quantum chemical properties can be repurposed to build data-efficient downstream machine learning models, with a particular focus on predicting the activation energy of hydrogen atom transfer (HAT) reactions. Starting from a valence bond (VB) analysis of a generic HAT process, a set of informative descriptors is identified. Next, a surrogate neural network model is constructed to predict an informative representation, based on the identified VB descriptors, with the help of a publicly available dataset of (pre-computed) quantum chemical properties of organic radicals. We demonstrate that coupling the resulting on-the-fly informative representation to a secondary machine-learning model for activation energy prediction outperforms various predictive model architectures starting from conventional machine-learning inputs by a wide margin, at no additional computational cost. As a bonus, by basing their final predictions on physically meaningful descriptors, our models become inherently interpretable. Finally, because of the extreme data efficiency of our descriptor-augmented models, we are able to fine-tune and apply them to small datasets across various reaction conditions, settings and application domains, ranging from regular (liquid phase) synthesis, over metabolism and drug design, to atmospheric chemistry.