Machine learning-assisted design of AlN-based high-performance piezoelectric materials

Abstract

Dopants play an important role in improving the piezoelectric stress coefficient (e₃₃) of aluminum nitride (AlN)-based piezoelectric materials. However, the existing experimental or computational approaches cannot provide generalized design criteria or fast predictive capabilities for screening high-performance piezoelectric materials over a wide range of composition space. To address this demand, we have designed a general machine learning (ML) strategy to make a comprehensive prediction and exploration of AlN-based piezoelectric materials of various concentrations and compositions. The predicted piezoelectric strain coefficient (d₃₃) was verified to be remarkably consistent with the experimentally available values of Sc-, MgTi-, and MgZr-doped AlN compounds. It is worth noting that an extremely large d₃₃ of 202 pC N⁻¹ was discovered in Sc_0.5Al_0.5N. Besides, the first ionization energy, the formation energy of decomposition products, and the number of out-of-plane first-nearest-neighbor cation bonds were revealed to be critical physical quantities to facilitate the prediction of the piezoelectric coefficient based on a detailed investigation of the physical mechanism. This study demonstrates the feasibility of the fast prediction and design of high-performance piezoelectric materials with easily accessible features.