24AI Climate Model Tuning Without Hyperparameters

Earth system models (ESMs) feature many parameters that must be estimated, but parameter estimation is challenging because the parameters are typically poorly constrained by available observations. While parameter estimation is an important step in model development, there is no universally accepted method, and current algorithms are computationally intensive. This presentation showcases a new algorithm, KalmRidge, and demonstrates its effectiveness in tuning the Community Earth System Model (CESM2) in idealized experiments. In contrast to similar methods, KalmRidge removes the need for offline hyperparameter selection, significantly reducing computational costs. This is achieved by reformulating the update equations of the ensemble Kalman filter as an equivalent ridge regression problem, followed by adaptive selection of the regularization parameter using standard cross-validation techniques. This approach offers a promising and practical solution for parameter estimation without excessive cost.