Learning Intermolecular Electronic Coupling with Molecular-Orbital-Based Descriptors.

Electronic coupling is the key parameter to determine the rate of intermolecular electron transfer and energy transfer at excited states. When excited states are involved, the couplings are state-specific, as they originate from the interactions between different molecular orbitals (MOs). Based on the MO overlap description of the electron transfer (ET) and excitation energy transfer (EET) couplings taken as the domain knowledge, here we propose a graphic molecular orbital (MO) based descriptor to predict intermolecular electronic couplings. As the MOs are characterized by the spatial distribution of their wave functions, namely, the size and sign of the lobes, we transform the grid points of a MO into two feature vectors containing the quantum and spatial information. Then, inspired by the MO overlap description of the electronic couplings, we build the descriptors by multiplying the vectors for paired MOs. Together with a deep neural network (DNN) model, we learn the couplings of hole transfer (HT), electron transfer (ET), and Dexter energy transfer (DET). For the couplings of naphthalene dimers, high accuracy of learning is achieved by our approach compared with the results from quantum chemical (QC) calculations with a small size of training data. Therefore, the MO-pair-based descriptor shows the ability to characterize MO interaction for the high performance learning of electronic couplings and implies the potential of this strategy for other state-specific properties of excited molecular systems.