Machine learning-assisted accelerated research on piezoelectric response prediction of KNN-based ceramics

Potassium sodium niobate (KNN)-based lead-free piezoelectric ceramics have attracted significant attention due to the remarkable electrical properties and high Curie temperature. However, the conventional trial-and-error approach for identifying optimal doping combinations to enhance their performance is inefficient and expensive. Therefore, the rapid and precise prediction of piezoelectric coefficient (d33) exhibits great practical significance in the investigation of high-performance piezoelectric ceramics. In this work, an efficient data-driven machine learning (ML) approach is proposed to expedite the development of piezoelectric properties with the (0.92-x)K0.48Na0.52NbO3-xLiTaO3-0.08AgSbO3 (KNNAS-xLT) lead-free ceramic as the model material. Comparison of various ML models containing distributed random forest (DRF), Light Gradient Boosting Machine (LightGBM), Gradient Boosting Machine (GBM) and eXtreme Gradient Boosting (XGBoost) are conducted with a limited dataset consisting of 256 samples and 21 feature elements. Then, the DRF model is identified as the most favorable choice, demonstrating piezoelectric constants which are consistent with the experimental results along with the highest fitting accuracy of 79 %. Based on the coincident result, the mechanism is analyzed with modified phase boundary and polarization finally. This research demonstrates a powerful strategy of machine leaning for achieving high d33 in KNN ceramics, promoting designing of new materials. [Display omitted] •A novel dataset is constructed on KNN-based ceramic materials, containing composition and electrical properties.•A suitable DRF model is obtained with the largest accuracy achieved is 79% via typical models.•Predicted piezoelectric property is realized utilizing the DRF mode, along with consistent experimental data.