Designing Promising Thermally Activated Delayed Fluorescence Emitters via Machine Learning-Assisted High-Throughput Virtual Screening

Thermally activated delayed fluorescence (TADF) materials have attracted much attention due to their high performance in organic light-emitting diodes (OLEDs). However, progress in developing novel TADF materials is limited by the low-efficiency traditional trial-and-error experimental approach. In this work, by integrating machine learning (ML) and quantum mechanics (QM) calculations, we have achieved fast prediction of high-efficiency TADF emitters. 44470 molecules are first generated by virtually combining common 49 donor and 50 acceptor fragments. The ML model is trained for predicting molecular properties by using the QM results of 5136 molecules. Finally, by applying it, 384 molecules are rapidly screened out of 39334 molecules as potential TADF emitters. To validate these results, the photophysical parameters of 18 out of 384 molecules are calculated from first-principles calculations, and the obtained TADF rates are greater than 104 s–1, even up to 106 s–1, suggesting excellent TADF properties. Moreover, the TADF performance of the predicted candidates overwhelms that of existing TADF emitters with similar chemical structures. These findings amply confirm that machine learning-assisted virtual screening is an appealing means of developing excellent TADF emitters.