Predicting the excited-state properties of crystalline organic semiconductors using GW+BSE and machine learning

Excited-state properties of crystalline organic semiconductors are key to organic electronic device applications. Machine learning (ML) models capable of predicting these properties could significantly accelerate materials discovery. We use the sure-independence-screening-and-sparsifying-operator (SISSO) ML algorithm to generate models to predict the first singlet excitation energy, which corresponds to the optical gap, the first triplet excitation energy, the singlet-triplet gap, and the singlet exciton binding energy of organic molecular crystals. To train the models we use the “PAH101" dataset of many-body perturbation theory calculations within the GW approximation and Bethe- Salpeter equation (GW+BSE) for 101 crystals of polycyclic aromatic hydrocarbons (PAHs). The best-performing SISSO models yield predictions within about 0.2 eV of the GW +BSE reference values. SISSO models are selected based on considerations of accuracy and computational cost to construct materials screening workflows for each property. The screening targets are chosen to demonstrate typical use cases relevant for organic electronic devices. We show that the workflows based on SISSO models can effectively screen out most of the materials that are not of interest and significantly reduce the number of candidates selected for further evaluation using computationally expensive excited-state theory.