arhead damage effect assessment is a critical issue in weapon system effectiveness analysis and serves as an important foundation for survivability and protection technology research. Traditional assessment methods struggle to accurately characterize the complex interdependencies among target system components and the cascade failure propagation mechanisms. This paper proposes a warhead damage assessment method based on improved Graph Attention Networks (GATv2). The method first establishes a multi-mechanism physical damage model integrating shock wave overpressure, fragmentation, and thermal radiation effects. Subsequently, the target system is modeled as a weighted directed graph, and the GATv2 network is employed to learn the dynamic dependencies among components. Furthermore, three-dimensional edge features are incorporated to enhance the model's cascade failure prediction capability. Simulation experiments conducted under three typical terrain conditions(plains, forests, and mountains)demonstrate that the proposed method can accurately predict damage levels at various distances and terrains, with damage radius prediction errors maintained within 5%, significantly outperforming traditional static assessment methods.