Materials datasets are often assembled from multiple experimental and literature sources that differ substantially in quality and reliability. We present an uncertainty-aware framework for materials property prediction that accounts for such data heterogeneity. A resampling-based approach is used to quantify per-source uncertainty, and the resulting data-driven uncertainty proxies are embedded into a Kriging surrogate model, enabling adaptive weighting of data according to reliability. Applied to four ferroelectric datasets, including electrostrain, piezoelectric coefficient, recoverable energy storage density, and electrocaloric strength, the framework consistently improves predictive accuracy relative to uncertainty-agnostic baselines. Experimental validation on five newly synthesized compositions further confirms its robustness and predictive reliability. This data-efficient framework provides a practical route for integrating heterogeneous and variable-fidelity datasets in materials informatics.

Close

PrevNext

View figure in article

• Ferroelectricity
• Ceramics
• Ferroelectrics
• Machine learning