Physics Parameter Optimization in a Setting with Multiple Response Functions

The parameter optimization problem for numerical weather prediction (NWP) models can be stated in simple terms as that of finding the parameter values that minimize or maximize some function of the forecast (the so-called response function), e.g. “What is the choice of parameter values that minimize the bias of the forecast air temperature in the Tropics?” However, the problem in practice is more complicated, as the model developer typically must consider multiple response functions simultaneously. E.g., the problem might become, “What is the choice of parameter values that minimize the bias of the forecast air temperature in the Tropics while maximizing the anomaly correlation of 500 hPa geopotential height in the Midlatitudes?” In this presentation the problem of physics parameter optimization in a setting with multiple response functions is examined using a realistic global NWP model and 14 physics parameters from six different parameterizations including deep and shallow convection, radiation, boundary layer, and clouds. Emphasis is placed on understanding the joint relationship between the different response functions and the implied tradeoffs for the parameter optimization. Additionally, the sensitivity of the optimization to the choice of the response functions’ metric and region is investigated. Finally, the structure of the optimization space is elucidated by examining the optimization results as a function of different choices of starting point and optimization algorithm.