The mechanical switching of ferroelectric domains is achieved in PbZr0.2Ti0.8O3 thin films obtained by the sol-gel method for thicknesses up to 200 nm. The dielectric polarization can be switched when a force higher than a given threshold value in the order of some µNewtons is applied with the tip of an atomic force microscope. This threshold is determined as a function of the thickness of the films, and local hysteresis loops are recorded under mechanical stress. The possibility of switching the polarisation in such unusually thick films is related to the existence in their volume of physical nanoscale defects, which might play the role of pinning centers for the domains.

Phase-field modeling is a powerful technique for predicting domain structure evolution and electromechanical properties of ferroelectric materials. However, it remains computationally very expensive, thus demanding high computing resources and restraining its use for exploring large systems. Some machine learning approaches have already been proposed to accelerate general phase-field simulations. Here, we present a specifically neural-network-trained model for ferroelectric phase-field modeling, including supervised and nonsupervised learning from Landau energy. A surrogate model predicts the microstructural polarization field evolution determining the electrostatic and mechanical equilibrium at each time step. The model produces accurate and stable rollout predictions, up to hundreds of frames even when starting from the beginning of the simulation. With a relative error contained below 5% compared to a high-fidelity phase field, we show that our model can be used instead of real solvers to conduct full simulations. While being at least 685 times faster than the classical phase-field computation, our approach opens a path to explore ferroelectric materials at a larger scale with fewer computing resources.