The description of physical processes in weather and climate models involves many tunable parameters. A new strategy to optimize these parameters is presented based on surrogate models. It is applied to the one-dimensional parameterization of the boundary-layer and the benefits in terms of efficiency and outcome are demonstrated.
This new optimization strategy performs more efficiently than a commonly used approach. The resulting parameter combinations provide better results and the cost of detecting these combinations is substantially reduced. The strategy has the potential to efficiently tune tens of tunable parameters in three-dimensional atmospheric simulations using multiple physics packages.
A new optimization strategy using surrogate models based on radial basis functions is introduced to tune the free parameters of physical parameterizations in atmospheric models. This approach iterates to sample the parameter space and to detect the minimum of the cost function describing the quality of a parameter combination. Compared to a frequently used quadratic regression model this approach demonstrates a higher detection speed with more accurate optimal solutions.