Superior Energy-Storage Performance Enabled by Machine Learning Accelerated Composition Design for Lead-Free...

Development of high-performance lead-free AgNbO₃ (AN)-based antiferroelectrics (AFEs) have emerged as promising candidate for high-power energy-storage capacitors. Routine trial-and-error method in enhancing energy-storage density (W_rec) and efficiency (η) encounters great challenges since extensive latent space are explored for composition screening. Using machine learning (ML) algorithms, a two-layer stacking framework termed as SS-PAN (stacking strategy for predicting AN-based ceramics) is proposed here for designing high-performance AN-based AFEs. This framework achieves a high R² score of 0.82 through cross validation and outperforms individual ML model. The predicted composition represented by Li_0.01Ag_0.99Nb_0.5Ta_0.5O₃, possesses a quasi-linear P-E loop and an ultrahigh W_rec of 16.6 J cm^-3 and η of 92.6% with an excellent figure of merit of 224.3 J cm^-3 is achieved at electric field of 108 kV mm^-1 in MLCC. Based on SHapley Additive exPlanations analysis, high prediction accuracy is enabled by precisely selecting features of tolerance factor and electron affinity of B-site element. Notably, local structure for Li_0.01Ag_0.99Nb_0.5Ta_0.5O₃ composition is thoroughly decoded by STEM and DFT calculations, where highly polar short-range antiferroelectric nanodomains with strong localized dipole moments are induced by Li/Ta co-doping. This work imparts a potent potential of data-driven methodology for seeking emergent relaxor AFEs for advanced dielectric capacitor applications.