Machine learning Landau free-energy potentials for third-principles simulations

We show how to construct Landau-like free-energy potentials using a machine learning approach. For concreteness, we focus on perovskite oxide ${\mathrm{PbTiO}}_{3}$, representative of a large class of materials that undergo nonreconstructive structural phase transitions; in this case, a proper ferroelectric transformation with improper ferroelastic features. We work with a training set obtained from Monte Carlo simulations based on an atomistic ``second-principles'' potential for ${\mathrm{PbTiO}}_{3}$. We rely exclusively on data that would be experimentally accessible---i.e., temperature-dependent polarization and strain, both with and without external electric fields and stresses applied---to explore scenarios where the training set could be obtained from laboratory measurements. We introduce a scheme that allows us to identify optimal polynomial models of the temperature-dependent free-energy surface, mapped as a function of the homogeneous electric polarization and homogeneous strain. Typically, our method evaluates thousands of possible models, ranking them by accuracy and predictive power. Our results for ${\mathrm{PbTiO}}_{3}$ show that a very simple polynomial---where only two parameters depend linearly on temperature---is sufficient to yield a correct description of the material's behavior. We thus validate the usual approximations made in phenomenological studies of phase transitions of this kind. Remarkably, the obtained models also capture the subtle couplings by which elastic strain controls key features of ferroelectricity in ${\mathrm{PbTiO}}_{3}$, such as the symmetry of the polar phase and the discontinuous character of the transition, despite the fact that no effort was made to include such information in the training set. We emphasize the distinctive aspects of our methodology (which relies on an original form of validation step) by comparing it with the usual machine learning approach for model construction. Our results illustrate how physically motivated models can have remarkable predictive power, even if they are derived from a limited amount of data. We argue that such ``third-principles'' models can be the basis for predictive macroscopic or mesoscopic simulations of ferroelectrics and other materials undergoing nonreconstructive structural transitions.