PiezoTensorNet: Crystallography informed multi-scale hierarchical machine learning model for rapid piezoelectric performance finetuning

Piezoelectric devices offer numerous opportunities for sustainable harvesting of wasted mechanical energy, leading to a significant interest in data-driven research on these materials. This study presents the design of PiezoTensorNet, a comprehensive framework that encompasses a hierarchical classification neural network for crystal point group determination and modular ensembles of regression-based multi-dimensional models for predicting piezoelectric tensors. The machine learning models capable of forecasting tensors for dopant element alloying and various crystallographic transformations is integrated along with finite element analysis for electromechanical performance evaluation. The efficacy of integrated toolkit is demonstrated through the computational design and discovery of a lead-free microelectromechanical system based on AlN. The introduction of Boron and Erbium dopants in AlN enhances its piezoelectric performance, particularly when the crystal undergoes rotations along a preferred axis. Specifically, under a vertical loading of 5 × 10−5 N/m2 applied to a cantilever beam, the preferentially oriented B0.3Er0.5Al0.2N material generates a power 9.96 times larger than that of AlN ceramics.