Machine-Informed Design of π-Conjugated Molecules with Tuned Optical Properties

While organic π-conjugated materials demonstrate great utility in energy generation and storage, lighting, transistors, sensors, and other optical and electronic devices, designing a new molecular material with precise properties for a specific application requires the exploration of a vast chemical space. Molecular generative models show great efficacy for in silico property prediction and design for drug-like molecules. Given this success, these models can have a natural extension to the design of π-conjugated molecules. Here, we will discuss the development of generative normalizing flow models to tune the optical properties of π-conjugated chromophores. The models are trained on the more than 25,000 chromophores and points of optoelectronic data from the Organic Crystals in Electronic and Light-Oriented Technologies (OCELOT) database. The generative models, with a focus on molecules that can be of interest for singlet fission or thermally activated delayed fluorescence, can quickly generate structures with the aim of tuning the energy gaps between low-lying singlet and triplet excited states. The model optimizes these gaps through sampling a learned chemical space and predicting the gap between a molecule’s low-lying singlet and triplet excited states using pretrained message passing neural networks that were trained on the optoelectronic data from OCELOT. Results from the generative models are verified through density functional theory (DFT) and time-dependent DFT (TDDFT) calculations to obtain their optical transitions and frontier molecular orbital energies. Though the models here are focused on robustly tuning optical transitions, the methods can be transferred to optimize additional electronic properties.