Accurate prediction of emission spectra is crucial for the rational design of luminescent materials in optoelectronic devices such as phosphorescent organic light-emitting diodes (OLEDs). We present a physics-informed machine learning framework that integrates the Franck–Condon progression as a differentiable module for spectrum prediction. The model estimates key physical parameters and reconstructs spectra within a continuous, differentiable, end-to-end-trainable formulation, enabling direct supervision in the spectral domain. Across multiple architectures and on multiple evaluation metrics, the approach consistently outperforms conventional baselines and maintains stable accuracy as spectral complexity increases, indicating robust gains in capturing full spectral shape and characteristic features. The learned parameters are interpretable and reveal which vibrational modes govern the spectral shape. With a physically grounded parameterization, the framework improves predictive fidelity while preserving interpretability. Consistent with the Franck–Condon principle, mode-count ablations show that a small subset of modes with large Huang–Rhys factors dominates the global spectral features. These results highlight the utility of differentiable physics-based models for enhancing accuracy, robustness, and interpretability in molecular property prediction, and suggest straightforward extension to emissive and absorptive spectroscopies.
Lee et al. (Thu,) studied this question.