Machine-Learning-Assisted Composition Design of High-Entropy CaBi 4 Ti 4 O 15 -Based Piezoceramics with Ultrahigh Piezoelectricity

High-entropy strategy presents a promising approach for enhancing the piezoelectric properties of bismuth layer-structured ferroelectrics, yet the vast compositional space poses significant challenges for traditional trial-and-error methods. In this work, a machine learning (ML)-assisted design strategy is proposed to accelerate the discovery of high-performance, high-entropy CaBi₄Ti₄O₁₅ ceramics. Based on a small data set containing 114 compositions, seven key elements closely related to the piezoelectric coefficient (d₃₃) were identified through feature engineering. Using the eXtreme Gradient Boosting regression model, the d₃₃ values of 22500 candidate compositions were predicted, thereby guiding the synthesis of Ca_1-x(Na_1/6Bi_1/6Li_1/6Ce_1/6K_1/6Bi_1/6)_xBi₄Ti_3.9(Nb_1/2Mn_1/2)_0.1O₁₅ ceramics. The optimal composition (x = 0.3) achieved a record-high d₃₃ of 28.1 pC/N, along with outstanding temperature stability (93% retention after 500 °C annealing). Structural characterization indicates that the high-entropy-induced chemical disorder promotes the formation and switching of nanodomains. On the other hand, the significant lattice distortion it induces enhances the spontaneous polarization, with both aspects synergistically improving the piezoelectric performance. This work demonstrates the effectiveness of ML in navigating complex composition-performance relationships in high-entropy piezoceramics and provides a viable route for designing piezoelectric materials with superior integrated performance for high-temperature applications.