Fusing Direct and Indirect Measurements Through Multi-Fidelity Learning For Accelerated Electrocaloric Materials Discovery

The data-driven discovery of high-performance electrocaloric (EC) materials is challenged by sparse direct measurements and systematic discrepancies between direct and indirect measurements, resulting in heterogeneous datasets with varying fidelity levels. Here, a co-kriging-based multi-fidelity learning framework is developed to integrate these data sources and construct a robust predictive model for (Formula presented.) -based ferroelectric ceramics by explicitly modeling cross-fidelity correlation and discrepancy. Combined with a multi-objective active learning strategy, the framework enables efficient optimization of low-temperature EC strength and operational temperature span across the composition–processing space. Guided by this approach, a multi-element-doped (Formula presented.) -based ceramic exhibiting an EC strength of (Formula presented.) (Formula presented.) /V at (Formula presented.) together with a broad operational temperature span of 75 K is identified. Experimental characterization reveals that the enhanced performance originates from a suppressed and diffuse phase transition associated with a relaxor-like or weakly ordered state, enabling broad temperature stability together with large reversible polarization. These results demonstrate that integrating multi-fidelity learning with active learning provides an effective strategy for accelerating functional materials discovery under realistic experimental constraints.