Machine learning assisted composition design of high-entropy Pb-free relaxors with giant energy-storage.

The high-entropy strategy has emerged as a prevalent approach to boost capacitive energy-storage performance of relaxors for advanced electrical and electronic systems. However, exploring high-performance high-entropy systems poses challenges due to the extensive compositional space. Herein, with the assistance of machine learning screening, we demonstrated a high energy-storage density of 20.7 J cm with a high efficiency of 86% in a high-entropy Pb-free relaxor ceramic. A random forest regression model with key descriptors based on limited reported experimental data were developed to predict and screen the elements and chemical compositions of high-entropy systems. Following basic experiments, a (BiNa)TiO-based high-entropy relaxor characterized by fine grains, weakly-coupled and small-sized polar clusters was identified. This resulted in a near-linear polarization behavior and an ultrahigh breakdown strength of 95 kV mm. Further, this high-entropy realxor presented a high discharge energy density of 7.7 J cm under discharge rate of about 27 ns, along with superior temperature and fatigue stability. Our results present the data-driven model for efficiently exploring high-performance high-entropy relaxors, demonstrating the potential of machine learning in developing relaxors.