Earth System Model Tuning Without Hyperparameters

Abstract

This article introduces a new algorithm, KalmRidge, and demonstrates its ability to tune an Earth system model using idealized experiments. Unlike similar algorithms, KalmRidge eliminates the need for offline hyperparameter selection, thereby substantially reducing computational expense. This is done by rewriting the update equations for the ensemble Kalman filter as an equivalent ridge regression problem, then applying standard cross-validation techniques to adaptively choose the regularization parameter. We propose that this algorithm, with time-mean spherical harmonic projections as tuning targets, provides a promising, tractable approach for parameter estimation.