Machine Learning‑Assisted Development of High-Performance Na0.5Bi2.5Nb2O9 Piezoceramics via Synergistic Multisite Codoping

Abstract Bismuth layer-structured ferroelectrics (BLSFs) are critical candidates for high-temperature piezoelectric applications, yet their inherently low piezoelectric activity (d33) restricts their practical implementation. In this study, a machine-learning-guided strategy was employed to rationally design multisite Li+/Ce3+ and W6+ codoped (Na0.5Bi2.5)Nb2O9 piezoceramics. Using Random Forest algorithm, we efficiently screened a large virtual composition space and identified the optimal doping window. The optimized composition exhibits a high d33 of 30 pC/N, markedly enhanced compared with the 16 pC/N of the pure composition, while retaining an ultrahigh Curie temperature (TC) of 776 °C and excellent thermal stability. Structural analyses reveal that multisite doping increases local polar disorder, which drives a significant reduction in domain size, drastically lowering the energy barrier for domain switching. Ultimately, this work successfully delivers a high-performance material for high-temperature piezoelectric devices while providing an AI-assisted paradigm for the targeted design of bismuth layer-structured ceramics.