Data-Driven Cloud Cover Parameterizations for the ICON Model
DescriptionA promising approach to improving cloud parameterizations in climate models, and thus climate projections, is to train machine learning (ML) algorithms on coarse-grained output of global storm-resolving model simulations. The ICOsahedral Non-hydrostatic (ICON) modeling framework enables simulations ranging from weather prediction to climate projections, making it an ideal target for developing ML-based parameterizations. We develop four different neural network (NN) types, distinguished by the degree of vertical locality they assume for diagnosing cloud cover. We enforce sparsity by sequentially selecting features based on the models' performance gains for the NN type that offers the best tradeoff between complexity and accuracy. In their training domain, the NNs already achieve excellent predictive performance (R^2 >0.9) with as few as three features. Using symbolic regression, we discover an interpretable equation with superior performance (R^2 =0.94) that satisfies physical constraints. These ML-based cloud cover parameterizations are coupled to the atmospheric component of the ICON Earth system model via the Fortran-Keras Bridge. To provide the most appropriate measure of cloudiness for ICON's microphysics scheme, the NNs provide cloud volume fraction in addition to cloud area fraction in each grid cell. We evaluate the resulting coupled ICON-ML model using the Earth System Model Evaluation Tool (ESMValTool).
TimeWednesday, June 2815:00 - 15:30 CEST
Climate, Weather and Earth Sciences