Searching for Singlet Fission Candidates with Many-Body Perturbation Theory and Machine Learning

Singlet fission (SF) is a photophysical process where one singlet-state exciton converts into two triplet-state excitons. SF is considered as a possible approach to surpass the Shockley-Queisser limit and has started wide discussions in the past decade. However, commercialization of SF-based photovoltaics remains in incubation due to the lack of practical SF materials. To tackle this bottleneck, performing large-scale simulation and screening molecular materials database to search for SF candidates with promising properties is suggested. One of the decisive excitonic properties directing the fission process, the SF thermodynamic driving force can be calculated with the state-of-the-art, many-body perturbation theory (MBPT) under the GW approximation paired with Bethe-Salpeter equation (BSE). However, GW+BSE calculation is too cumbersome to be selected as the screening scheme for a database with tens of thousands of molecular crystals. Statistical inference is hence introduced to maximize the probability of discovering SF candidates with minimized computational cost. To realize this process, a hierarchical screening workflow incorporating materials science and machine learning (MatML Workflow) was designed and implemented.