Correlation between the Stability of Substituted Cobaltocenium and Molecular Descriptors.

Metallocenium cations, used as a component in an anion exchange membrane of a fuel cell, demonstrate excellent thermal and alkaline stability, which can be improved by the chemical modification of the cyclopentadienyl rings with substituent groups. In this work, the relation between the bond dissociation energy (BDE) of the cobaltocenium (CoCp) derivatives, used as a measure of the cation stability, and chemistry-informed descriptors obtained from the electronic structural calculations is established. The analysis of 12 molecular descriptors for 118 derivatives reveals a nonlinear dependence of the BDE on the electron donating-withdrawing character of the substituent groups coupled to the energy of the frontier molecular orbitals. A chemistry-informed feed-forward neural network trained using k-fold cross-validation over the modest data set is able to predict the BDE from the molecular descriptors with the mean absolute error of about 1 kcal/mol. The theoretical analysis suggests some promising modifications of cobaltocenium for experimental research. The results demonstrate that even for modest data sets the incorporation of the chemistry knowledge into the neural network architecture, e.g., through mindful selection and screening of the descriptors and their interactions, paves the way to gain new insight into molecular properties.