Discovering new perovskites with artificial intelligence

An Artificial Neural Network (ANN) was developed to discover new inorganic perovskite - structures. The ANN assessed the probability to crystallize as a perovskite structure for compounds described with up to four Wyckoff sites. The ANN was also able to address the compounds independently of their crystal system. The input data needed by the ANN, also known as features, were based on the treatment of the atomic radii, electronegativity, and atom positions of the crystal compound. In this manner, the ANN was fed with information concerning the geometric and packing factors as well as the chemical environment of the atoms in the material. Quantum mechanical calculations were not required to obtain a feature for the ANN, but they were used to validate the predictions done by the ANN, such as CsBeCl3.