Bridging Machine Learning and Zintl Phase Thermoelectric Materials: The Ca9– x Eu x Zn4.5– y Cu y Sb9 System

Abstract Navigating the multifaceted compositional landscape of quinary thermoelectric (TE) materials via classical trial-and-error is impractical due to high labor and material costs. To address this, we present an AI-guided methodology to accelerate the discovery of high-performance Zintl phases in the Ca9–xEuxZn4.5–yCuySb9 system, leveraging a synergistic Eu–Cu cosubstitution strategy. Three machine learning (ML) models were developed to predict electrical conductivity, Seebeck coefficient, and thermal conductivity, enabling systematic screening for a high figure-of-merit, ZT value. Experimental validation confirmed the orthorhombic Ca9Mn4Bi9-type crystal structure for the title compounds, with Eu showing a strong site preference for the interlayer A4 site. This substitution, combined with charge compensation of Cu+ for Zn2+, increased interstitial site occupancy, enhancing anionic framework connectivity and charge transport. Density functional theory and crystal orbital Hamilton population analyses also indicated that cosubstitution optimized interatomic interactions, yielding simultaneous enhancements in σ and S via boosted carrier mobility. A maximum ZT of 0.60 at 810 K was achieved for Ca8.5Eu0.5Zn4.4Cu0.1Sb9, a substantial improvement over the ternary parent Ca9Zn4.5Sb9, that aligned with ML predictions within a 7.7% margin. This work demonstrated that an integrated AI-experimental workflow effectively bridged the gap between ML prediction and the realization of high-performance TE materials.