Simulation-Assisted Deep Learning Techniques for Commercially\nApplicable OLED Phosphorescent Materials

Phosphorescent light-emitting\nmaterials play a central role in\norganic light-emitting diode (OLED) devices. Due to their synthesis\ndifficulties, unsystematic trial-and-error synthesis is prohibitively\nchallenging. For this reason, deep learning (DL), which has shown\nsuccess in various fields, is being actively applied to materials\ndiscovery. However, one challenge in applying DL to phosphorescent\nmaterials is the limited amount of experimental data set. One way\nto circumvent this issue is to apply powerful DL techniques that have\nbeen successfully implemented in several domains. Another solution\nwould be to use a large amount of data set for pretraining DL models\nwith simulated properties highly relevant to target properties. In\nthis work, phosphorescent materials are represented as strings, molecular\ngraphs, and point clouds, which are employed by language models, two-dimensional\ngraph, and three-dimensional graph neural networks. In addition, more\nthan 200 000 molecules with simulated properties highly relevant\nto experimental properties are used for pretraining the DL models.\nOur work shows high performance in the prediction of five experimental\nproperties that are importantly considered when commercializing OLED\ndevices. This means that faster material discovery for OLEDs can be\nachieved through DL models that are trained with simulation information\nthat is highly correlated with experimental properties.