SOGCN: Prediction of key properties of MR-TADF materials using graph convolutional neural networks

Abstract

The exploration of the structure and properties of the luminescent materials in OLED devices using Multiple Resonance Thermally Activated Delayed Fluorescence (MR-TADF) is constrained by challenges related to long cycles and high experimental costs, making it a key obstacle in the development of new materials. In response to this challenge, we propose an innovative approach by constructing a graph convolutional neural network model named SOGCN to quickly determine whether an unsynthesized material has the potential to become an MR material, and accurately predict its energy gap and half-peak width, thereby expediting the development process of MR-TADF materials. We constructed the MR220 dataset for training the model based on 220 MR-TADF molecules reported in experiments. To ensure the reliability of the SOGCN model in predicting new samples, we have established a rigorous set of theoretical calculation evaluation standards, providing crucial references for the model. In the prediction of the properties of 37 new samples of MR-TADF molecules, SOGCN successfully predicted the singlet–triplet energy gap (ΔE_ST) of some samples, demonstrating a good trend in FWHM prediction as well. Finally, we have synthesized our designed molecule, Design3 (DtCzB-Boz), the organic light-emitting diodes based on DtCzB-Boz exhibit an emission peak at 508 nm, with the FWHM is 27 nm. The result of photophysical characterization is highly consistent with the predicted value of SOGCN. Notably, the mean absolute errors (MAE) between our model predictions and experimental/computational values were as low as 0.037 eV and 12 nm, respectively. This indicates that SOGCN exhibits higher efficiency and accuracy in predicting the properties of MR-TADF materials.