Abstract: To tackle the key challenges in product quality modeling for complex manufacturing systems—such as nonlinear dependencies across multiple processes, intricate structural relationships, long-range error propagation, and the absence of engineering knowledge on the production process—this paper proposes an improved graph attention network modeling method called ICLF-GAT (Inter-layer Contrastive Loss Filtering Graph Attention Network). The method systematically addresses these issues: it adopts a data-driven deep learning framework that eliminates dependence on prior physical knowledge; constructs a directed graph model of the manufacturing system based on Graph Attention Networks (GAT) to effectively capture complex structural features and nonlinear dependencies between processes; introduces an inter-layer contrastive loss filtering mechanism that dynamically evaluates and filters the quality of node features to significantly mitigate the over-smoothing problem in deep GATs and enhance long-range error propagation modeling; and designs a target attention decoder to further improve the accuracy in modeling complex system-wide dependencies. The results in the simulation experiments and a real-world industrial case show that ICLF-GAT significantly reduces Root Mean Squared Error (RMSE) compared to existing benchmark methods, particularly excelling in long-range error propagation tasks.