Structure prediction of ionic materials using the Minima Hopping method and the CENT machine learning potential

Ionic materials and in particular oxides are the dominating class of materials on earth. Because of the huge number of structures in this class, a theoretical exploration of novel low-energy phases is desirable. Up to now, only density functional theory was sufficiently accurate for general purpose structure predictions. Hence they were numerically quite expensive and could only be applied to relatively small systems. The charge equilibration via neural network technique (CENT) potential allows to calculate energies and forces faster by several orders of magnitude and enables large scale structure predictions. Being based on a charge equilibration scheme, the CENT potential allows us to describe accurately the energy associated to the charge transfer that is the dominant bonding mechanism in ionic materials. Even though it is trained only on previously known structures it can reliably predict the energy of entirely new structures. The new structures are found by the Minima Hopping structure prediction method which can escape in an efficient way from the funnels of the known input structures. We will present novel structures for TiO2 sheets, bulk and surfaces of CaF2, stoichiometric and non-stoichiometric phases of MgO, as well as crystalline structures of SrTiO3 and LiCl.