Exploring Optimized Organic Fluorophore Search Through Simple Experimental Data-Driven VAE

Designing organic fluorescent molecules with tailored optical properties is challenging in decades, while the new avenue was opened by the statistical models. Inverse design has garnered considerable interest in organic materials science but concentrates on arbitrary design or theoretical properties. Here, we introduce a strategy that enables direct optimization of specific experimental properties in the inverse design process, utilizing a variational autoencoder (VAE) with a latent vector-based prediction model. Omitting the Kullback-Leibler divergence and separate training strategy successfully improved the generator's robustness and molecular diversity. We confirm the latent vectors obtained from VAE are powerful inputs for downstream prediction models of experimental properties, fluorescence energy and quantum yield. Our approach for the optimized search of organic fluorescent materials, substantiated by gradient space derived from latent vector and validated by newly synthesized and uncharacterized molecules, shows potential for broader applications in diverse organic material design.