Vibrational properties of metastable polymorph structures by machine learning

Despite vibrational properties being critical for the ab initio prediction of the finite temperature stability and transport properties of solids, their inclusion in ab initio materials repositories has been hindered by expensive computational requirements. Here we tackle the challenge, by showing that a good estimation of force constants and vibrational properties can be quickly achieved from the knowledge of atomic equilibrium positions using machine learning. A random-forest algorithm trained on only 121 metastable structures of KZnF$_3$ reaches a maximum absolute error of 0.17 eV/$\textrm\AA^2$ for the interatomic force constants, and it is much less expensive than training the complete force field for such compound. The predicted force constants are then used to estimate phonon spectral features, heat capacities, vibrational entropies, and vibrational free energies, which compare well with the ab initio ones. The approach can be used for the rapid estimation of stability at finite temperatures.