Transferability in machine learning based interatomic potentials

Dr.  Seyed Alireza Ghasemi

Institute for Advanced Studies in Basic Sciences

Tuesday 95/11/12, 15 – 16

Ibn Al-haytham Hall, Physics Department

Abstract:
Based on an analysis of the short range chemical environment of each atom in a system, standard machine learning based approaches to the construction of interatomic potentials aim at determining directly the central quantity which is the total energy. Indeed, these methods are based on purely mathematical models and do not account for any sort of physical or chemical principles. As a consequence, such interatomic potentials perform poorly due to lack of transferability when they are applied to systems not available in the fitting database. Charge equilibration via neural network technique is a method based on electronegativity equalization method while it employs the power of machine learning techniques. Here we introduce the method and present its transferability from cluster structures to crystalline phases.