Wood fiber mats are currently of high interest for building insulation as they are recyclable and made from a renewable resource. However, they perform worse than conventional materials. The microstructure of the mats is a key influencing factor for their thermal insulation [2]. We therefore aim to optimize the microstructure by creating digital twins for thermal conductivity simulations. Based on image data, simplified geometric structural models are developed. They reflect essential geometric properties like orientation and size distributions, and the distribution of solid mass in individual fibers and bundles while allowing for quick generation of large representative volume elements. Being a natural material, the microstructure of the wood fibers varies strongly, so that large volumes have to be analyzed and modeled to ensure representativeness. Moreover, the extremely high porosity makes it almost impossible to prepare sufficiently small samples for 3D imaging using computed tomography with sufficient resolution to preserve the thin walls of the hollow cellulose fibers. Synchrotron radiation-based computed tomography at ID19 of the ESRF allowed for completely preserved fiber walls thanks to sub-μ voxel sizes while keeping a rather large field of view. Based on these images, we separate individual fibers and bundles, identify their contact regions similarly to Viguié et al. [3], and finally fit a modified Altendorf-Jeulin fiber model [1]. Numerical simulations of both conduction and radiation run on realizations of this fine-scale model will be incorporated into a simplified geometry model on the next coarser scale. The cuboids forming the coarser model are calibrated to have a quasi-effective heat conduction. Virtual samples representative of the microstructure will be generated from these cuboids, on which heat transfer can be numerically simulated to ultimately optimize the structure on this scale. Here, machine learning methods are planned to guide the variation of parameters in the stochastic models.