Speaker: Natalia Yurievna Lopanitsyna (École polytechnique fédérale de Lausanne).
Place and time: aud. 432 GK, 15:00
This report will highlight the main aspects of creating a model of the interatomic interaction potential for various systems using neural networks using the approach proposed by Beller and Parinello. The general concept of building such potentials will be presented: architecture and arrangement of neural networks; selection of a set of structures for training a neural network; functions used to describe structures. The main problems that may arise in the process of building such a potential will also be considered. All of the above will be illustrated by the example of creating a potential for nickel, by comparing the values of various properties calculated using this potential, with first-principle calculations, experimental data, and other forms of potential models. At the end, a small session is supposed to show directly the process of building capacity on a neural network for a small data set.