Scale-aware parameterization of cloud fraction and condensate for a global atmospheric model machine-learnt from coarse-grained kilometer-scale simulations

Kilometer grid-length simulations over a variety of different locations worldwide are used as training data for a deep-learning model designed to predict clouds in a global climate model. The inputs to the neural network are profiles of temperature, humidity and pressure from the high-resolution model, averaged to the scale of the climate model. The outputs are profiles of cloud fraction, liquid water content and ice water content. The high-resolution data is coarse-grained to a range of sizes, allowing the model to learn how the cloud formation depends on the size of the area being considered. The machine-learnt cloud cover and condensate scheme is coupled to a global climate model and used to run multi-year simulations where the clouds predicted by the neural-network are fully interacting with the rest of the model.