Understanding the fracture behavior of an interface between two different materials is crucial for many advanced devices and composite materials. To analyze crack propagation at atomic level, molecular dynamics simulation is a powerful tool. However, few empirical interatomic potentials can accurately capture the interatomic interactions near interfaces, which differ significantly from those in the bulk. Recently, the application of machine learning techniques to develop highly accurate interatomic potentials from density functional theory (DFT) calculation data has gained much attention. These machine-learning potentials (MLPs) have been demonstrated to achieve sufficient accuracy even for lattice defects such as dislocations and grain boundaries[1,2]. In this study, we develop MLP of Si/Ge interfaces based on artificial neural network, enabling highly accurate descriptions of interatomic interactions in the vicinity of Si/Ge interfaces and the associated stress fields around an interfacial crack. The dataset used for training MLP were systematically constructed using DFT calculations. The developed MLP can describe the interface energy curve of coherent Si/Ge interface structure. Utilizing the MLP, we conducted tensile simulations on pre-cracked atomic models of Si/Ge interfaces. At the critical load where an interfacial crack begins to propagate along the interface, the atomic stress at the crack tip remains almost constant regardless of the crack length. The critical tensile stress has showed the crack length dependence of 1/√a consistent with the Griffith’s theory.