Parameter Estimation in CLM4.0 Using Surrogate Model Based Global Optimization

Thursday, 18 December 2014
Juliane Mueller1, Christine A Shoemaker1, Natalie M Mahowald2 and Rajendra Paudel1, (1)Cornell University, Ithaca, NY, United States, (2)Earth and Atmospheric Sciences, Cornell University, Ithaca, NY, United States
Methane is a very important greenhouse gas that traps more radiation per molecule than CO2 and is a driver of tropospheric chemistry. Over the anthropocene methane has increased dramatically, but the role of changes in wetland emissions is not well constrained. The Community Land Model (CLM) of the Community Earth System Models is able to estimate methane emissions from wetlands and rice paddies. Our comparison of CH4 emission observations at 16 sites around the planet reveals, however, that there are large discrepancies between the CLM predictions and the observations. The goal of our study is to use optimization methods to adjust sensitive model parameters in order to minimize the root mean squared error (RMSE) between model predictions and observations. These parameters have been selected based on a sensitivity analysis. We run CLM4.0 on the Yellowstone supercomputer and a single CLM simulation for obtaining the CH4 emission predictions at the 16 locations takes on average between 15 and 30 minutes. Thus, only few simulations can be allowed in order to find a near optimal solution within an acceptable time. To achieve this goal, we use a recent computationally efficient global optimization algorithm that uses a radial basis function (RBF) surrogate model to approximate the objective function in a computationally cheap way. We use the information from the RBF to select new sample points in the variable domain that are most promising with respect to improving the objective function value. In every iteration of the algorithm, we update the RBF to improve its fit and predictions. We show with pseudo data (generated by using the default model parameters) that our optimization algorithm is able to make excellent progress with respect to decreasing the RMSE. Using the true CH4 emission observations, we show that by optimizing the parameters we are able to significantly reduce the RMSE between observations and model predictions.