Interpretable AI Accelerates the Discovery of High‐Light‐Yield Inorganic Scintillators for X‐ray Imaging

ABSTRACT Inorganic scintillators are essential for X‐ray detection and imaging applications, but their rational design remains challenging due to limited experimental datasets and the chemical heterogeneity across oxides, fluorides, chlorides, and iodides, each with distinct luminescence mechanisms. Herein, we develop a machine learning framework based on a Mixture‐of‐Experts Kolmogorov–Arnold Network (MoE‐KAN) to accelerate the discovery of inorganic scintillators with high light yield (LY). By integrating chemistry‐aware gating with KAN‐based expert regressors, MoE‐KAN effectively captures nonlinear composition–property relationships in small and heterogeneous datasets, showing improved generalization over conventional machine learning models and MoE‐MLP on unseen scintillator systems. Through SHAP analysis and symbolic regression, we further distill an empirical composition‐property screening criterion that links LY to the interplay among effective nuclear charge, enthalpy of melting, covalent radius, and Mendeleev chemists sequence. Applying this interpretable strategy to the CsCaCl 3 system, we identify Mn 2+ as the optimal activator and demonstrate, for the first time, its X‐ray scintillation performance: CsCaCl 3 :1%Mn achieves a light yield of 21,577 photons/MeV (∼2.1× commercial BGO) with good dose‐rate linearity and X‐ray imaging capability. This work establishes an interpretable data‐driven route for discovering high‐light‐yield inorganic scintillators under data‐limited conditions.