Application Of Adjoint Model In CICE6 Sea Ice Mode

Modern earth system models usually include multiple components,such as atmosphere,ocean,land and sea ice.The sea ice model is an important component in an earth system model.But its development is relatively slow compared to other models.In the past few decades,the model description of thermodynamic and dynamic processes of sea ice has seen great improvements.For some sub-grid physical processes,however,parametric methods are still needed,which are limited by the inaccuracy of the description of the physical process and the scarcity of in-site observation data.There are still great uncertainties in the parameters of the physical process of sea ice.Usually,these crucial parameters in sea ice require a lot of sensitivity experiments to determine their value.However,sensitivity experiments require a lot of computing resources.Another way to determine parameter value is adjoint method.In the study,adjoint model method has been used for parameter adjustment and tuning of the TOPO melt pond scheme in the CICE6 sea ice model.Since it takes considerable human resources to attain the adjoint code from the original model,and once the model code is changed,the adjoint code is no longer valid,our study attempts to combine the automatic differentiation tool with the original model.First,we verify the feasibility of bringing the automatic differentiation tool into the modern sea ice model,and adjust the parameters in the melt pond scheme of sea ice model.The adjusted parameters were then used in the model to evaluate its impact in the simulation of melt pond and other sea ice features.In order to adapt to the automatic differentiation tool,parts of the model codes were rewritten.Using the Shell script,the generation of the adjoint code,the handling of errors and abnormal situation,and the integration of adjoint code with the original code were automatically realized.Taking the thermodynamic process and radiative transmission process of the sea ice model as examples,the tangent linear model and the adjoint model were tested separately,verifying the usability of the automatic differentiation tool in the CICE6 sea ice model.In next step,by combining the original TOPO melt pond parameterization scheme code,the adjoint code and the L-BFGS minimization algorithm,the whole parameter optimization scheme was constructed through the cost function.The model is integrated with the original parameter settings in the first time,and the model results is assumed to be the “true value”.The correctness of the parameter optimization scheme is evaluated,the results show that that the parameter optimization scheme is effective.On the basis of the above optimization scheme,the melt pond fraction data retrieved from MODIS is used in cost function.Two typical Arctic regions of firstyear ice(FYI)and multi-year ice(MYI)are selected.Parameter estimates at different times from different location are attained and then used in the sea ice model.The results shows that the simulation results using the optimized parameters are closer to the observation than the results under the default settings.We demonstrated that this parameter optimization scheme can be used to adjust the parameters in the sea ice model,and further,to optimize the performance of the model.