In the proposed project, I want to provide a faster and more flexible method to downscale global climate models with respect to regional data. The emphasis is on incorporating the uncertainty of the numerical model approaches, i.e., of the approximation schemes such as local discretizations. Additionally, I want to democratize access to such forecasts by providing an implementation that respects in its choices the computational power of the end-user. At the same time, the project serves as the motivation and benchmark for fundamental methods development: I will investigate mesh-refinement for PDE solvers from a probabilistic perspective, aiming in particular to quantify uncertainty arising through local refinement.