Predicting room temperature phosphorescence of new organic molecules by combining physical models and data-driven methods.

The rational discovery of new RTP emitters remains challenging because phosphorescence depends on multiple competing processes, including intersystem crossing, nonradiative decay, and competition with fluorescence. While several mechanistic hypotheses have been proposed, like enhanced spin-orbit coupling, reduced energy gap between singlet and triplet states, and favourable electronic transition properties, the extent to which these factors can quantitatively predict RTP behaviour remains unclear. In this work, we have collected a dataset of 218 metal-free organic RTP emitters from the literature. We first explored physical approaches to predict phosphorescence yield using descriptors derived from established mechanisms and hypotheses, such as favourable spin-orbit coupling, reorganization energy, and oscillator strength. Although these quantities show statistically meaningful correlations with RTP behaviour, the physical models alone are insufficient to construct a robust quantitative predictive tool. We therefore developed a data-driven machine learning (ML) model based on the Random Forest algorithm, combining these physically motivated descriptors with additional topological descriptors derived from the molecular structure. After feature selection, the optimized models achieved prediction accuracies of 0.774 and 0.820 identifying long-lifetime (lifetime > 90 ms) and long-wavelength (RTP emission > 525 nm) RTP materials, respectively. With the boundary from the physical process and ML analysis, 65 candidates are identified from 48 168 molecules with predicted probabilities greater than 0.75 for both long-lifetime and long-wavelength properties, providing a focused set of promising materials for further experimental validation.