3D microstructure reconstruction of heterogeneous material from slice descriptors using explicit neural network

Heterogeneous materials such as rocks, concrete, and composites exhibit random microstructures that strongly influence their physical properties. These structures can be captured using imaging techniques such as micro-Computed Tomography (micro-CT) and Scanning Electron Microscopy (SEM), but such methods are costly. As an alternative, digital reconstruction offers a way to generate synthetic microstructural images. Traditional reconstruction methods provide geometric control but lack flexibility and generality. In contrast, Computer Vision (CV) approaches offer strong generative capabilities, but often lack interpretability and material-specific constraints. Conditional generation introduces descriptors as labels, yet struggles with control during training. In this work, we enhance conditional generation by decomposing the loss into two parts: an implicit CV-based component and an explicit, descriptor-driven component inspired by traditional methods. The explicit loss includes a targeted loss for individual inputs and a distribution loss for overall output quality, ensuring both accuracy and diversity. To support fast training, a rapid descriptor regression model is developed and integrated into our digital reconstruction workflow. We validate our method on Fontainebleau sandstone, Polytetrafluoroethylene (PTFE), and SOFC electrodes, demonstrating improved reconstruction performance compared to the current state-of-the-art 2D-to-3D method. Our approach accurately captures statistical descriptors, even for complex geometries, and maintains strong consistency across multiple 2D slices representing 3D structures.