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 insufficient process engineering knowledge, this paper proposes a modelling method based on an improved graph attention network called Inter-layer Contrastive Loss Filtering Graph Attention Network (ICLF-GAT). Initially, a data-driven deep learning framework is adopted to avoid reliance on prior physical knowledge. Subsequently, a directed graph model of the manufacturing system is constructed based on Graph Attention Networks (GAT) to effectively capture complex structural features and nonlinear dependencies between processes. On this basis, a novel inter-layer contrastive loss filtering mechanism is introduced, which dynamically evaluates and filters the quality of node features to significantly mitigate the over-smoothing problem in deep GATs and enhance the modeling capability for long-range error propagation. Finally, a target attention decoder is designed to further improve the modeling accuracy of the systematic complex dependencies. Simulation experiments and a practical industrial case demonstrate that ICLF-GAT significantly reduces the Root Mean Squared Error (RMSE) compared to existing benchmark methods, with particularly advantages in long-range error propagation.