Beyond group additivity: Transfer learning for molecular thermochemistry prediction

Abstract

The accuracy of thermochemical prediction methods is strongly dependent on the size of the set of training data. Group additivity is an interpretable modeling strategy that can be developed from a limited dataset, but fails to consider delocalized molecular effects such as inductive stabilization, delocalized resonance stabilization, and steric effects. In contrast, machine learning allows the incorporation of these effects but requires an extensive amount of high-quality data. Therefore, a new transfer learning approach is proposed, uniting group additivity with machine learning. First, a machine learning model is pretrained on a large set of group additive predictions, after which it is refined on a limited high-quality dataset with transfer learning. The proposed approach was tested to predict the standard enthalpy of formation, standard molar entropy, and heat capacity of a wide range of hydrocarbons, hydrocarbon radicals, and carbenium ions. By using transfer learning, chemically accurate predictions for hydrocarbons, radicals, and carbenium ions could be obtained, drastically reducing the group additive error using less than 450 molecular datapoints per model. A SHapley Additive exPlanations analysis reveals that a data-efficient but interpretable transfer learning methodology is obtained, achieving chemically accurate predictions for a wide range of hydrocarbons.