Machine-Learning Predictions of Photoluminescence in Molecules Exhibiting Thermally Activated Delayed Fluorescence with Implicit Experimental Validation.

The application of machine learning to materials discovery is often constrained by the availability of large-scale, experimentally verified materials databases. This study presents an automatic, end-to-end framework that bridges this gap by training machine-learning predictors for materials properties on experimental data mined directly from the literature. We apply this framework to predict the photoluminescence (PL) wavelengths of thermally activated delayed fluorescence molecules. By integrating "chemistry-aware" natural language processing with automated chemical structure resolution, a dataset of 643 experimentally measured PL wavelengths was afforded. This experimentally grounded data were used to train a heterogeneous graph neural network and a ridge-regression model; both achieved mean absolute errors below 0.13 eV in less than 3 min on a personal laptop, effectively capturing complex structure-property relationships without manual feature engineering. These results demonstrate that our framework provides a fast, scalable, and generalizable pathway to generate experimentally grounded models for property predictions in organic optoelectronics.