Generative Design of Singlet Fission Materials Leveraging a Fragment-oriented Database

Abstract

Recently, we leveraged the FORMED repository made up of 116,687 synthesizeable molecules to deploy fragment-based high-throughput virtual screening (HTVS) and genetic algorithm (GA) searches of singlet fission (SF) molecular candidates. With these approaches, both prototypical (e.g., acenes, boron-dipyrromethane (BODIPY)) and unreported (e.g., heteroatom-rich mesoionic) classes of chromophore candidates fulfilling specific SF energetic requirements were identified. Yet, the reliance on predefined fragments limits chemical space exploration and, thus, the discovery of truly unforeseen molecular cores. Here, we exploit FORMED to train a generative learning framework driven by reinforcement learning and property predictions. The generative model rediscovers a diverse range of previously reported SF chromophore classes, including polyenes, benzofurans, fulvenoids and quinoidal systems, but also suggests an unexpected scaffold absent from the training data, neocoumarin (2-benzopyran-3-one), characterized by two endocyclic double bonds in an extit{ortho} arrangement and capped by a lactone group. An in-depth investigation reveals a diradicaloid behavior over the conjugated core comparable to 2-benzofuran, a widely known SF compound. This work highlights the potential of using both generative and property prediction models to discover candidates beyond derivatives of known chemistry for tailored material applications.