Accelerated search for novel ferroelectric materials

We report the development of a combined machine learning and high-throughput DFT framework to accelerate the search for novel ferroelectrics. This framework is capable of predicting potential ferroelectric compounds using only compositions as input. Initially, for a given chemical composition-space, a series of machine learning algorithms predict the possible stable stoichiometries that are insulating and have a non-centrosymmetric structure, necessary for the ferroelectricity. A final classification model then predicts the point groups of these stoichiometries. Based on the point groups, a subsequent series of high-throughput DFT calculations determine the ground state crystal structure. As a final step, using group theory considerations, non-polar parent structures are identified and the polarisation values are further determined. By predicting the crystal structures as well as the polarisation values, this method provides a powerful tool to explore new ferroelectric materials beyond the existing databases.