Abstract: The optimization of ultra-low-loss microwave dielectric ceramics is hindered by the complex interplay of factors governing the dielectric loss and inefficient trial-and-error screening. This study introduces an interpretable machine learning framework that combines accelerated materials discovery with quantitative structure-property analysis for the Li₂TiO₃-Li₃NbO₄-MgO system. A Gaussian process regression-based active learning strategy was employed to explore the compositional space, identifying the near-optimal composition of (1-x)Li₃MgNbO₅-xLi₂TiO₃ at x = 0.80 using only six compositions (from a total of 21) over three iterations, closely aligning with the experimentally determined optimum at x = 0.85. Beyond optimization, a high-fidelity ensemble machine learning model (R² = 0.8285) was developed to quantitatively evaluate the influence of twelve physically meaningful structural descriptors on dielectric loss, utilizing permutation importance and SHapley Additive exPlanations. Both methods consistently identified internal strain (IS) and configurational entropy (CE) as the most dominant factors influencing Q × f. Maximum Q × f value arises from a synergistic combination of moderate CE associated with order-disorder transitions, minimized IS, and a densely packed microstructure with elevated relative density. Experimental and model validation confirms that the composition at x = 0.85 achieves outstanding dielectric performance, with ε_r = 16.17, Q × f = 104,300 GHz at 9.2 GHz, and τ_f = -3.47 ppm/°C. This closed-loop, interpretable, and experimentally supported methodology offers a generalizable route for the rapid design of high-performance dielectric ceramics.

Key words: Microwave dielectric ceramics, Ultra-low loss, Interpretable machine learning, SHAP analysis, Quantitative microstructure-property relationships