The goal of this thesis is to study the phase change of Ge-rich GST with molecular dynamics (MD) simulations using equivariant graph neural networks interatomic potentials (GNN-IP). The candidate will train a GNN-IP model on ab initio reference calculations of Ge-rich GST in order to describe amorphous and crystalline phases. The GNN-IP will be used to compute thermodynamic and kinetic properties of the phase transitions. In a second step, further developments including the effect of impurities and the impact of an electrical field on the phase change will be addressed. Finally, physical parameters computed from MD simulations will be employed to improve our in-house mesoscopic model based on the phase-field method.