<p>This study presents a multi-scale deep-learning framework that integrates several advanced neural models to optimize hybrid three dimensional (3D) printed self-sensing nano-carbon cementitious composites. The first step was undertaken by Multi-Scale Graph Neural Network, where special conductive pathways were taught ensuring the uniform work on nano-carbon learning patterns, improving electrical conductivity by 25%–35%. four-dimensional Spatiotemporal Transformer Network decoded printing parameters achievements with an interlayer conductivity improvement of 40%–50%, avoiding anisotropic print by aiming for defects prediction on print Induced anisotropic behavior. High-fidelity artificial microstructures have been generated with Physics Informed Generative Adversarial Networks; these synthetic methods realize an experimental cost-cutting of about 50% while conserving about 98% fidelity to the characteristics of real microstructures. Fifth, Self-Supervised Contrastive Learning automatically classifies small macro and microdefects with over 95% detection reliability. There has been reduction of as much as 35% in the number of false positives. Predicted kinetics of hydration and long-term electrical stability can now be predicted with speed improvements of 15% and resistance drift reduction by 20% over six months. This approach for the first time combines different hybrid models of deep learning for the self-sensing cementitious composites, thus significantly increasing percolation of electrical networks, accuracy in crack detection, and predictions on long-term durability. The developed framework creates a new paradigm in the real-time structural health monitoring world, providing enhanced reliability in structures while also reducing costs at a level for the next generation of smart infrastructure sets.</p>