Generative graph-diffusion optimization of hierarchical stitch-needle hybrid textile composites for manufacturable interlaminar toughness and delamination-resistant lightweight structures

High interlaminar toughness is a critical requirement for 3D textile polymer composites, but its improvement remains challenging because of limited through-thickness reinforcement and the complexity of stitch-needle interactions. Existing computational approaches, including finite-element-based homogenization, surrogate models, and deep learning methods, are often limited by dependence on predefined datasets, insufficient incorporation of fracture physics, and the absence of integrated multi-objective optimization. To address these limitations, this study proposes the Hierarchical Architecture Synthesis for Textile Enhancement using Neural Intelligence (HASTEN) framework, a unified generative–predictive–optimization approach for designing hierarchical stitch-needle architectures. The framework was implemented using Python 3.11, PyTorch Geometric, and CUDA-enabled GPU acceleration, and trained on a high-fidelity dataset comprising Mode I/II interlaminar toughness measurements, resin-fiber characteristics, and microstructural information. HASTEN integrates a Generative Graph Neural Network (G-GNN), a graph diffusion refinement module, a Physics-Guided Transformer (PGT), and a hybrid NSGA-III with Differential Evolution optimizer. The optimized candidate architectures achieved predicted improvements of 33% in Mode I toughness, 34% in Mode II toughness, 26% in energy absorption, 48% reduction in defect probability, and 10% reduction in structural weight, while reducing the prediction error to 4.5%, outperforming representative benchmark models. These improvements represent computational predictions generated by the trained physics-guided model and multi-objective optimization framework rather than experimental measurements of newly fabricated specimens. Overall, the proposed framework demonstrates the potential of integrating graph-based learning, physics-guided prediction, and evolutionary optimization to generate manufacturability-compliant candidate architectures for next-generation textile composite design, providing a foundation for future experimental validation.