Teaching a Neural Network to Attach and Detach Electrons from Molecules

Physics-inspired Artificial Intelligence (AI) is at the forefront of methods development in molecular modeling and computational chemistry. In particular, interatomic potentials derived with Machine Learning algorithms such as Deep Neural Networks (DNNs), achieve the accuracy of high-fidelity quantum mechanical (QM) methods in areas traditionally dominated by empirical force fields and allow performing massive simulations. The applicability domain of DNN potentials is usually limited by the type of training data. As such, transferable models are aimed to be extensible in the description of chemical and conformational diversity of organic molecules. However, most DNN potentials, such as the AIMNet model we proposed previously, were parametrized for neutral molecules or closed-shell ions due to architectural limitations. In this work, we extend our AIMNet framework toward open-shell anions and cations. This model explores a new dimension of transferability by adding the charge-spin space. The resulting AIMNet model is capable of reproducing reference QM energies for cations, neutrals and anions with errors of 4.1, 2.1, 2.8 kcal/mol, respectively, compared to the reference QM simulations. The spin-charges have errors 0.01-0.06 electrons for small organic molecules containing nine chemical elements {H, C, N, O, F, Si, P, S and Cl}. Thus the proposed AIMNet model allows researchers to fully bypass QM calculations and derive the ionization potential, electron affinity, and conceptual Density Functional Theory quantities like electronegativity, hardness, and condensed Fukui functions. We show that these descriptors, along with learned atomic representations, could be used to model chemical reactivity through an example of regionselectivity in electrophilic aromatic substitution reactions.