Machine Learning-Assisted Discovery of Thermally Activated Delayed Fluorescence Emitters

This study presents a machine learning (ML)-assisted framework for the discovery and screening of novel TADF emitters. A dataset of 366 known compounds was used to train regression models based on molecular descriptors calculated via RDKit. Among several algorithms tested, the CatBoost model demonstrated superior performance with an R² of 0.845 on the test set. The trained model was subsequently employed to predict TADF-likeness scores for over 50,000 compounds from the Harvard Organic Photovoltaic Database (HOPV15). Using descriptor-based filtering and synthetic accessibility analysis, 50 high-potential TADF candidates were identified. Structural clustering using t-SNE analysis revealed diverse donor-acceptor frameworks favorable for TADF behavior. The integration of cheminformatics and ML enables rapid screening of chemical libraries and accelerates the discovery of TADF materials with high efficiency and practical synthetic feasibility.