Machine Learning–Guided Rational Design Platform of Cr 3 + ‐Doped AB 2 O 4 Broadband Emitters via SHAP‐Interpreted Transfer Learning

ABSTRACT The design of near‐infrared materials, especially Cr 3 + ‐doped phosphors with specific emission linewidths (full width at half maximum [FWHM]) is hampered by two fundamental challenges: the intricate, nonlinear dependence of FWHM on host lattice chemistry and microstructure, and the scarcity of high quality. Addressing this gap, we present a novel interpretability‐driven machine learning framework that synergistically combines transfer learning with targeted experimental synthesis. A predictive model, initially trained on a literature‐sourced domain of 73 Cr 3 + ‐doped luminescent materials, was successfully adapted via transfer learning to our custom domain of 12 synthesized AB 2 O 4 : Cr 3+ spinel ceramics, achieving a high cross‐validated R 2 of 0.862 for FWHM prediction. Crucially, Shapley Additive exPlanations (SHAP) analysis transforms the model into a discovery tool, quantitatively identifying lattice parameter and effective nuclear charge as the primary structural features governing linewidth broadening. These computational insights are experimentally validated, demonstrating how cation substitution and annealing tune local crystal field heterogeneity. This work establishes a closed‐loop, hypothesis‐generating methodology that moves beyond black‐box prediction to deliver fundamental design principles for tailored luminescent materials.