Physics-guided machine learning of excited-state properties for the design of high-performance TADF emitters

The rational design of thermally activated delayed fluorescence (TADF) and inverted singlet-triplet (INVEST) emitters demands accurate prediction of critical photophysical properties, particularly singlet-triplet energy gaps (ΔE<inf>ST</inf>) and oscillator strengths (f). Conventional machine learning (ML) models often neglect the underlying physics, limiting their transferability and interpretability across chemical space. In this work, we develop a physics-informed machine learning (PIML) framework that leverages physically meaningful molecular descriptors to predict ΔE<inf>ST</inf> and f with high accuracy and robust generalization. Training on a chemically diverse dataset of over 39 000 compounds, our models achieve correlation coefficients (r) between 0.77 and 0.88 and mean absolute errors (MAE) below 0.1 eV for ΔE<inf>ST</inf> and 0.02 for f on unseen test data. The reliability of the PIML models is further validated via leave-one-out cross-validation and external datasets, including 28 experimentally reported emitters, for which our model outperforms state-of-the-art quantum chemical and ML approaches. Beyond predictive accuracy, integrating interpretability tools reveals the exchange integral, dynamic spin polarization, and excited-state energies as dominant factors controlling the target properties—offering mechanistic insights often inaccessible in standard black-box models. Finally, leveraging the predictive power of the trained models, we performed high-throughput screening of 400 newly designed TADF emitters, successfully identifying promising candidates with optimal ΔE<inf>ST</inf> and f combinations for OLED applications. This study highlights the strength of combining physical intuition with data-driven modeling, offering an efficient, scalable, and interpretable route for accelerating the discovery of next-generation optoelectronic materials.