Investigating the Structure-Property Relationship of Relaxor Ferroelectrics via Machine Learning

Relaxors ferroelectrics are an intensively studied field of research that are of great interest owing to their large dielectric permittivity and electromechanical coupling. The polarization response of relaxors is believed to be correlated with the presence of polar nanoregions (PNRs) in the material, which give origin of their unique behavior. After decades of research, however, PNRs and their relationship to relaxor dynamics is a discussion that is still actively disputed. Given both the computational and experimental challenges that impede progress on the atomistic insight into PNRs dynamics, it is hypothesized that machine learning (ML), a nontraditional computational approach, is the way to tackle the problem. We expect that ML can be used to analyze the thousands of dipole patterns within PNRs produced by Molecular Dynamics (MD) simulations of relaxors and provide insight into their intrinsic dynamics. We begin by testing various ML toy models to classify or group the electric response of relaxors, which will allow for assessment of the ML algorithm performance for the given problem. The ML algorithms with the most promising performance will be applied to study the structure-property relationship in relaxors. The aims of this work are therefore to (i) gain insight into the presence and properties of polar nanoregions in relaxor ferroelectrics via ML and atomistic MD, (ii) demonstrate the potential of ML as a predictive tool in relaxor ferroelectrics research, and (iii) develop a multifunctional ML model that can be applied to a wide range of material properties originating from dipolar interactions