A major bottleneck in the application of machine learning (ML) to organic electronic materials has been the lack of relevant data. In fact, most developments in ML for chemistry have focused on few selected datasets – e.g., QM9 and MD17 – and on ground state properties. While large in size, these datasets lack the compositional complexity and diversity that are representative of the chemical space required for the design of new materials with tailored optoelectronic properties. The other side of the problem is that existing datasets may even contain chemically unrealistic or impossible compounds. To address these challenges, we developed a dataset of over 110K experimentally- reported organic molecules, curated from a crystal structure repository, as well as protocols to automate the coupling of these molecules together in chemically meaningful ways.[1] The size and chemical diversity of the dataset enable the training of extreme gradient boosting regressors for the prediction of molecular excited state energies with excellent accuracy across the broadest range of known experimental compound space. These models enable on-the-fly evaluation of excited states in the screening of over a million potential candidates for singlet fission (SF) - a multiexciton- generating process that provides a promising route to increase the efficiency of solar cells through the conversion of one excited singlet state into two triplet states. The models and dataset are then redeployed in the multiobjective genetic optimization of SF candidates.[2] We show that calibrated uncertainty in model predictions can be used to guide the genetic algorithm towards low- or high-uncertainty regions of chemical space. In the former (exploitative) setup, variations of known SF compounds are uncovered. In the latter (explorative) mode, altogether new candidates are generated which display excellent excited state properties for SF but remain synthetically feasible.
J. Terence Blaskovits