Machine learning combined with feature engineering to search for BaTiO3 based ceramics with large piezoelectric constant

Machine learning based strategies have been increasingly applied in materials science to accelerate the discovery process. Regression algorithm learns the mapping from compositions/features to targeted property and makes prediction for unknown compositions. The quality of features, in some degree, determines the upper limit of the surrogate model performance and the associated search efficiency for desired candidates. We herein propose a data-driven framework combining feature engineering, machine learning, experimental design and synthesis, to optimize the piezoelectric constant of BaTiO3 based ceramics, with the emphasis on feature engineering realized by four strategies. The search for improved piezoelectric constant in the initial data set behaves differently compared to that in the whole unknown space, indicating that the initial data set might be biased to a local scheme. The best composition with a piezoelectric constant of ~ 430 pC/N is synthesized in the second iteration, better than the majority in the initial data set. Insight for the change of piezoelectric constant for the newly synthesized 12 compositions is provided by examining the corresponding evolution of dielectric permittivity within the thermodynamic theory.