This study proposes an artificial intelligence based framework for reconstructing the 3D multi-phase cement paste microstructure to evaluate its mechanical properties using simulation. The reconstruction of cement paste microstructures is performed using modified generative adversarial networks (GANs) based on microstructural images from micro-CT. For computational efficiency, 2D microstructures are first reconstructed and then extended to 3D microstructures. The reconstructed microstructures exhibit the same microstructural features as the original microstructures when characterized by probability functions. Mechanical properties such as stiffness and tensile strength are evaluated for the original and reconstructed specimens using a phase-field fracture model, and similar behaviors are observed. The results confirm that the reconstructed virtual microstructures can be used to supplement the real microstructures in evaluating the mechanical properties of 3D multi-phase cement paste. This approach thus provides a critical element of a data-driven approach to correlating its microstructure and properties.