Traditional trial-and-error tuning of uncertain parameters in climate models is time-consuming and subjective. Although auto-tuning has been employed in atmosphere models, coupled models pose significant challenges due to their higher computational cost. This study explores the feasibility of using machine learning for automatic tuning of parameters in the E3SMv3 coupled model. In the current phase of coupled tuning, we still focus on the atmospheric parameters. However, to measure the coupling effect, we take into account not only the responses in the atmospheric properties such as precipitation, radiative fluxes, temperature and large-scale circulation, but also sea surface temperature, sea-ice area and volume to reflect the impacts on the upper oceans. These properties are carefully selected using the hierarchical clustering machine learning method to balance the use of a more comprehensive set of variables and the efficiency of the optimization algorithm. Conventional machine learning-based tuning methods are offline, involving collection of samples to construct an offline surrogate model, which is then used by an optimization algorithm to tune the parameters. However, the surrogate model could not achieve high accuracy in the small domain around a local optimal solution due to insufficient sampling and the presence of nonlinearity and non-smoothness in this small domain. To enhance the fidelity of the surrogate model, we adopt a trust region Bayesian optimization approach (Turbo). This algorithm is an online tuning process, which utilizes a probability-based acquisition function to update the surrogate model during the optimal iterations. In Turbo, the surrogate model is updated in the trust region, rather than the entire search space. By focusing on a smaller region, the algorithm can achieve better accuracy and efficiency in finding the optimal solution. Overall, the method has the potential to reduce the auto-tuning computational cost and improve the fidelity of E3SMv3 coupled simulation.
