Simulating the ferroelectric phase transition with machine learning models

Deep neural networks trained on density functional theory data, to represent the potential energy and polarization surfaces of condensed phase systems, make possible to study the ferroelectric phase transition with ab-initio molecular dynamics accuracy. Applications to the hydrogen bonded molecular crystal potassium dihydrogen phosphate (KDP), and to the chemically disordered perovskite material potassium manganese niobate (PMN) are reported. In the first case, the role of quantum fluctuations on the transition is elucidated, after correcting empirically the quantum delocalization error of approximate functionals. In the second case, the simulations reproduce static and dynamic properties observed in experiments, and support earlier suggestions that PMN relaxor is a ferroelectric analog of magnetic spin glasses.