Deep Learning of Accurate Force Field of Ferroelectric HfO2

The emergence of ferroelectricity in HfO2-based thin films opens up exciting opportunities of using ferroelectrics at the nanoscale. The performance of ferroelectric-based electronics depends on dynamical responses to external stimuli. We developed a deep neural network-based interatomic force field of HfO2, enabling molecular dynamics simulations of this silicon-compitable ferroelectrics at large time and length scales. The development of an accurate model potential using first-principles data is greatly facilitated by a concurrent learning procedure. The model potential predicts a wide range of materials properties such as elastic constants and moduli, equations of states, phonon spectra, and solid-solid phase transition barriers accurately. The temperature-driven ferroelectric-paraelectric phase transition is reproduced with isobaric-isothermal ensemble molecular dynamics simulations.