Phase-change memory materials are semiconductors with a fast and re- versible transition between a resistive amorphous phase and a conductive crystalline phase. A sufficiently large contrast in electrical resistivity can be used to read and write bits of information, as encoded in the struc- ture of the material. Modeling the structure of these materials is therefore crucial for the design and optimization of devices that exploit these proper- ties. The relevant atomic interactions required to run computer simulations and to obtain structural models are governed by the laws of quantum me- chanics. These interactions can be calculated from first-principles using approaches such as density-functional theory. However, the estimation of macroscopic properties of materials requires extensive simulations that are not yet tractable for these methods. This is all the more important for disordered phases, such as liquids or glasses, where these properties have broad statistical distributions that require significant sampling of configu- ration space. Second-principles methods such as force-fields derived from quantum mechanical data, can offer a powerful method for extending the scope of atomistic simulations and can improve, through better scaling, the estimates of macroscopic properties, without losing the accuracy of the quantum-mechanical potential-energy surface. This thesis presents the de- velopment, validation and applications of a machine-learned interatomic potential, trained on data obtained from density-functional theory calcu- lations of Ge−Sb−Te materials, which are widely used in phase-change memory and data-storage electronic devices.