Machine learning-enabled design of ferroelectrics with multiple properties via a Landau model

Abstract

A physics based model often allows us to calculate several properties if the parameters for a given material are known. Here we address the question of making predictions of several properties from a Landau model for unexplored materials for which we do not know the material parameters. This is necessary if we are to predict new materials with targeted response with a physics based model than merely from data. We show how machine learning can be employed to learn parameters with an initial data set that need not be directly connected to the target properties. We demonstrate the approach by searching for BaTiO${}_3$-based ceramics to predict properties relevant for the electrocaloric effect, dielectric tunability and pyroelectricity, starting from polarization and permittivity data only. The predictions are experimentally validated by synthesizing eight ceramics with a combination of competing properties, such as large adiabatic temperature change and wide temperature window at given temperatures. Five of the compounds show enhanced refrigeration capacity, outperforming reported counterparts. The approach shows promise for problems where adequate physics based models are available and there is limited data for properties.