<p>Three-dimensional microstructural representations are essential for accurately predicting the mechanical response of heterogeneous materials, yet experimental 3D characterization techniques are often costly and time-consuming. This work presents a framework that integrates deep learning–based microstructure segmentation with three-dimensional reconstruction and mechanical modeling of gray and nodular cast irons from 2D metallographic images acquired at different depths. The methodology combines advanced AI-based image analysis with conventional reconstruction techniques to overcome the limitations of conventional 2D characterization. Several segmentation approaches were evaluated to distinguish the metallic matrix and graphite phases, with a modified architecture combining the Segment Anything Model and a convolutional encoder–decoder network showing the best performance, particularly for complex lamellar and spheroidal graphite morphologies. The segmented image stacks were aligned and compiled to generate consistent 3D microstructures that preserve phase distribution, size, and connectivity. Mechanical response predictions were obtained through finite element simulations of indentation loading. The numerical results, validated against experimental microhardness measurements and nanoindentations, demonstrate that the reconstructed microstructures capture the key features governing local mechanical behavior, supporting the proposed framework as an efficient alternative for 3D microstructure reconstruction and mechanical property prediction.</p>