Composition-based machine learning for predicting and designing Mn⁴⁺-doped phosphors

We present a data-driven approach to predict the excitation wavelength, emission wavelength, and crystal field energy levels (⁴T₁, ⁴T₂) in Mn⁴⁺-doped phosphors based solely on elemental composition. For the first time, we construct the largest and most compherensive experimental dataset of Mn⁴⁺-activated phosphors to train and accurately predict the properties without relying on complex structural descriptors. Among several evaluated models, the K-Nearest Neighbors and Extra Trees Regressors achieved the highest accuracy for predicting excitation and emission wavelengths, respectively. Importantly, to evaluate generalization, we test these models on Eu³⁺-doped systems and achieve high predictive accuracy. An inverse design model is further developed to suggest candidate phosphor compositions for target optical outputs. By avoiding complex descriptors while preserving accuracy and interpretability, this work provides a foundation for theory-informed discovery of luminescent materials.