Abstract: Multiple resonance thermally activated delayed fluorescence (MR-TADF) materials have attracted significant attention in organic electroluminescent devices due to their high exciton utilization efficiency and narrow emission spectra. However, their key performance parameters, singlet-triplet energy gap (ΔE_ST) and emission spectral full width at half maximum (FWHM), exhibit complex nonlinear relationships with molecular structures. To overcome the challenges of time-consuming, costly experiments and the limitations and insufficient accuracy of theoretical calculations, this study proposes a lightweight prediction framework based on SMILES encoding. By integrating Morgan fingerprints and physicochemical descriptors, end-to-end predictive models for ΔE_ST and FWHM were established. Skeleton similarity constraints were introduced to prevent data leakage, while feature selection and Bayesian optimization were applied to further enhancing model performance. Several machine learning algorithms were explored, among which the XGBoost model demonstrated the best predictive ability. Shapley additive explanations (SHAP) analysis revealed that ΔE_ST is mainly associated with electronic distribution, whereas FWHM is influenced by local skeleton structures and polar surface area. This approach achieves high-accuracy predictions without relying on three-dimensional structural information, providing an efficient solution for the rational design of MR-TADF materials.

Key words: Multiple resonance thermally activated delayed fluorescence (MR-TADF), Photophysical property prediction, SMILES, Shapley additive explanations (SHAP)