Predicting Phosphorescence Energies and Inferring Wavefunction Localization with Machine Learning

Long-time excited state dynamics of triplet states and subsequent emission via phosphorescence are commonly utilized for applications including light-emitting diodes and photovoltaics. Machine learning (ML) approaches trained using ab initio datasets may expedite the discovery of phosphorescent compounds. However, we show that standard ML approaches for modeling potential energy surfaces that succeed on small molecules do not generalize to molecules of larger sizes, due to the failure to account for spatial localities in spin transitions. To solve this, we introduce localization layers in a neural network model that weight atomic contributions to the transition energy. Trained on phosphorescent transition energies of organic molecules, the model achieves prediction accuracies of ~4 kcal/mol on the held-out test set and ~13 kcal/mol on an out-of-sample test set of large phosphorescent molecules. These localization weights have a strong relationship with the ab initio spin density of the triplet to singlet state transition, and thus infer localities of the molecule that determine the spin transition, despite that no direct electronic information was provided during training.