Neural networks (NNs) are proving to be attractive alternatives to traditional interatomic potentials. Being general and flexible learning machines, they show encouragingly accurate description of interatomic interactions in different elemental and ulticomponent systems. In an effort to build a standardized library of NN models, we have developed a hierarchical training in which NNs for multicomponent systems are obtained by sequential training from the bottom up: first unaries, then binaries, and so on . The stratified procedure and a new automated data generation protocol implemented in MAISE  have been used to produce sets of accurate NNs models consistent across a large block of chemical elements. Our tests indicate that the NN interaction models are efficient enough to accelerate ab initio prediction by orders of magnitude and reliable enough to identify overlooked stable materials.
 S. Hajinazar, J. Shao, and A.N. Kolmogorov, Phys. Rev. B 95, 014114 (2017)
 A.N. Kolmogorov, http://maise-guide.org (2009)