An Interpretable Multimodal Architecture for the Discovery of High-Efficiency Thermally Activated Delayed Fluorescence Emitters

The scarcity of high-quality experimental data often hampers data-driven materials discovery. This work introduces HoloMat-TADF, a holographic machine learning framework that provides predictive and interpretable insights from limited data sets. Its core feature is the synergistic integration of physicochemical data with robust features distilled via self-supervised pretraining on approximately 650,000 sequences and 130,000 molecular graphs and images. Combined with film-based host properties, this approach bridges the gap from theoretical molecular models to nanoscale organic light-emitting diode (OLED) devices, achieving a predictive R 2 of 0.847. Furthermore, the deeply interpretable architecture moves beyond black-box predictions to elucidate learned photophysical principles governing nanoscale exciton behaviors. This dual power was prospectively validated through in silico discovery and synthesis of two thermally activated delayed fluorescence (TADF) emitters, Mol-5 and Mol-9. Experimental validation confirmed their favorable properties, including high photoluminescence quantum yield (PLQY >98%), small singlet–triplet energy splitting (Δ E ST ≈ 0.1 eV), and short delayed fluorescence lifetime (τ d < 6 μs), congruent with predictions. Consequently, the resulting OLEDs achieved a maximum external quantum efficiency (EQE) of 31.3%. This work establishes an effective framework to accelerate discovery of functional materials in OLEDs and offers a generalizable approach for other data-scarce material systems.