Machine-learning based interatomic potential for studying of crystal structures properties

Abstract

In the process of modeling multilayer semiconductor nanostructures, it is important to quickly obtain accurate values the characteristics of the structure under consideration. One of these characteristics is the value of the interaction energy of atoms within the structure. The energy value is also important for obtaining other quantities, such as bulk modulus of the structure, shear modulus etc. The paper considers a machine learning based method for obtaining the interaction energy of two atoms. A model built on the basis of the Gaussian Approximation Potential (GAP) is trained on a previously prepared sample and allows predicting the energy values of atom pairs for test data. The values of the coordinates of the interacting atoms, the distance between the atoms, the value of the lattice constant of the structure, an indication of the type of interacting atoms, and also the value describing the environment of the atoms were used as features. The coordinates of the atoms, the distance between the atoms, the lattice constant of the structure, an indication of the type of interacting atoms, the value describing the environment of the atoms were used as features. The computational experiment was carried out with structures of Si, Ge and C. There were estimated the rate of obtaining the energy of interacting atoms and the accuracy of the obtained value. The characteristics of speed and accuracy were compared with the characteristics that were achieved using the many-particle interatomic potential — the Tersoff potential.