Calibrating the GAMIL3-1° climate model using a derivative-free optimization method

Abstract. Parameterization in climate models often involves parameters that are poorly constrained by observations or theoretical understanding alone. Manual tuning by experts can be time-consuming, subjective, and prone to underestimating uncertainties. Automated tuning methods offer a promising alternative, enabling faster, objective improvements in model performance and better uncertainty quantification. This study presents an automated parameter-tuning framework that employs a derivative-free optimization solver (DFO-LS) to simultaneously perturb and tune multiple convection-related and microphysics parameters. The framework explicitly accounts for observational and initial condition uncertainties (internal variability) to calibrate a 1-degree resolution atmospheric model (GAMIL3). Two experiments, adjusting 10 and 20 parameters, were conducted alongside three sensitivity experiments that varied initial parameter values for a 10-parameter case. Both of the first two experiments showed a rapid decrease in the cost function, with the 10-parameter optimization significantly improving model accuracy in 24 out of 34 variables. Expanding to 20 parameters further enhanced accuracy, with improvement in 25 of 34 variables, though some structural model errors emerged. Ten-year AMIP simulations validated the robustness and stability of the tuning results, showing that the improvements persisted over extended simulations. Additionally, evaluations of the coupled model with optimized parameters showed–compare to the default parameter setting–reduced climate drift, a more stable climate system, and more realistic sea surface temperatures, despite a slight energy imbalance and some regional biases. The sensitivity experiments underscored the efficiency of the tuning algorithm and highlight the importance of expert judgment in selecting initial parameter values. This tuning framework is broadly applicable to other general circulation models (GCMs), supporting comprehensive parameter tuning and advancing model development.