Data-driven and interpretable machine learning accelerates discovery of Li₄SrCaSi₂O₈-based microwave ceramics for high-performance dielectric antennas

High-performance dielectric antennas have pursued a high quality factor (Q×f) for microwave ceramics. Nevertheless, the cross-laboratory inconsistency in reported Q×f would confuse the invention of materials systems, owing to divergent preparation and measurement protocols. Herein, based on a self-consistent dataset, an interpretable machine learning framework is proposed to unveil the structure–property relationship and consequently guide the compositional design of candidate microwave ceramic Li₄SrCaSi₂O₈. Through feature engineering, nine critical features are identified, in which the Si/Li atomic mass ratio (Si/Li-AW), Si/Sr ionic radius ratio (Si/Sr-IR), and total electronegativity of cations (TEC) are found to be predominant. Interpretability technologies further reveal that a higher Si/Li-AW coupled with lower Si/Sr-IR and TEC is conducive to the increase in the Q×f value for the chosen decision tree (DT) model. Guided by these insights, Sn⁴⁺-doped microwave ceramic Li₄SrCaSi_1.98Sn_0.02O₈ is created with a Q×f value up to 83,526 GHz, the origin of which is elucidated by P–V–L theory combined with first-principles calculations and infrared spectroscopy. Such an optimized material is ultimately verified by a microstrip patch antenna with a high radiation efficiency of 81.12% and a gain of 5.94 dB in the C-band.