A machine learning approach for evaluating Southern Ocean cloud radiative biases in a global atmosphere model

• Published in: Geoscientific Model Development
• Main Authors: S. L. Fiddes, M. D. Mallet, A. Protat, M. T. Woodhouse, S. P. Alexander, K. Furtado
• Format: Article in Journal/Newspaper
• Language: English
• Published: Copernicus Publications 2024
• Subjects: Geology; QE1-996.5; Southern Ocean
• Online Access: https://doi.org/10.5194/gmd-17-2641-2024; https://doaj.org/article/d115712bb54e4672a427a0305b92f8dd

The evaluation and quantification of Southern Ocean cloud–radiation interactions simulated by climate models are essential in understanding the sources and magnitude of the radiative bias that persists in climate models for this region. To date, most evaluation methods focus on specific synoptic or cloud-type conditions that do not consider the entirety of the Southern Ocean's cloud regimes at once. Furthermore, it is difficult to directly quantify the complex and non-linear role that different cloud properties have on modulating cloud radiative effect. In this study, we present a new method of model evaluation, using machine learning that can at once identify complexities within a system and individual contributions. To do this, we use an XGBoost (eXtreme Gradient Boosting) model to predict the radiative bias within a nudged version of the Australian Community Climate and Earth System Simulator – Atmosphere-only model, using cloud property biases as predictive features. We find that the XGBoost model can explain up to 55 % of the radiative bias from these cloud properties alone. We then apply SHAP (SHapley Additive exPlanations) feature importance analysis to quantify the role each cloud property bias plays in predicting the radiative bias. We find that biases in the liquid water path are the largest contributor to the cloud radiative bias over the Southern Ocean, though important regional and cloud-type dependencies exist. We then test the usefulness of this method in evaluating model perturbations and find that it can clearly identify complex responses, including cloud property and cloud-type compensating errors.