Finding Predictive Descriptors for Singlet Fission by Machine Learning

Singlet fission (SF) is the conversion of one photo-generated singlet-state exciton into two triplet-state excitons. SF has the potential to significantly enhance the efficiency of solar cells by harvesting two charge carriers from one photon. Molecular crystals that undergo SF in the solid state are scarce. Computational exploration of the chemical space may accelerate the discovery of new SF materials. However, many-body perturbation theory (MBPT) methods that can reliably describe the excitonic properties of molecular crystals are computationally too expensive for large-scale materials screening. We use the sure-independence-screening-and-sparsifying-operator (SISSO) machine-learning algorithm to identify descriptors that are fast to evaluate and can accurately predict the thermodynamic driving force for SF, calculated by MBPT for a dataset of 101 polycyclic aromatic hydrocarbons (PAH101). SISSO generates interpretable models by iteratively combining physically motivated primary features. The best performing models are then selected by linear regression with cross validation. The SISSO-generated models successfully predict the SF driving force with root mean square errors as low as 0.11 eV. Based on considerations of the cost, accuracy, and classification performance of SISSO-generated models, we propose a hierarchical screening workflow for materials discovery. Three new SF candidates are discovered in the PAH101 set, which have not been previously reported.