Downscaling near-surface atmospheric fields with multi-objective Genetic Programming
The coupling of models for the different components of the Soil-Vegetation-Atmosphere-System is required to investigate component interactions and feedback processes. However, the component models for atmosphere, land-surface and subsurface are usually operated at different resolutions in space and time owing to the dominant processes. The computationally often more expensive atmospheric models, for instance, are typically employed at a coarser resolution than land-surface and subsurface models. Thus up- and downscaling procedures are required at the interface between the atmospheric model and the land-surface/subsurface models. We apply multi-objective Genetic Programming (GP) to a training data set of high-resolution atmospheric model runs to learn equations or short programs that reconstruct the fine-scale fields (e.g., 400 m resolution) of the near-surface atmospheric state variables from the coarse atmospheric model output (e.g., 2.8 km resolution). Like artificial neural networks, GP can flexibly incorporate multivariate and nonlinear relations, but offers the advantage that the solutions are human readable and thus can be checked for physical consistency. Using the Strength Pareto Approach for multi-objective fitness assignment allows us to consider multiple characteristics of the fine-scale fields during the learning procedure.