Prediction Of The Curie Temperatures Of Ferroelectric Solid Solutions Using Machine Learning Methods

In this work, we studied the application of machine learning for the prediction of the Curie temperature (T-c) of ferroelectric BiMe'Me"O-3-PbTiO3 systems (where Me' and Me" are metal cations). We found that among the studied K-nearest neighbor (KNN), support vector regression (SVR) and random forest (RF) methods, RF obtains the best performance and is insensitive to the choice of hyperparameters. SVR results show a strong sensitivity to the choice of hyperparameters and obtained T-c predictions with significantly lower accuracy than RF even after hyperparameter optimization. KNN results show poor accuracy and are essentially unusable with an incomplete feature set and are only qualitatively accurate with a complete feature set. With regard to the choice of features for accurate prediction of the Ferroelectric (FE) systems, we find that Bi content and B-cation valence, ionic radius and ionic displacements form the irreducible set of features such that these features or their equivalents must be used to obtain quantitatively accurate T-c predictions. We also find that homovalent and hetemvalent BiMe'Me '' O-3-PbTiO3 solid solutions form distinct classes of compounds with different behaviors so that both types must be included in the input data set to obtain high predictive accuracy. Our work confirms that for the small data sets typically available in materials science, careful selection of the input system, features and ML methods is required to enable accurate model construction and discovery of previously unknown relationships, but can be achieved in a systematic manner.