Data-driven sub-seasonal forecasts with a stratosphere

The stratosphere is an important source of predictive skill for sub-seasonal forecasts, yet it is poorly represented in currently published machine-learning weather prediction (MLWP) models. These models have been developed to forecast the state of the atmosphere up two weeks ahead, but they also show potential to be used for sub-seasonal forecasting. However, this requires enhancing and adapting them to be more suitable to the task. One of these enhancements is a better representation of the stratosphere. Here, we report on progress with achieving this in the AIFS-CRPS model, a probabilistic MLWP trained with a loss function based on the continuous ranked probability score which has been developed at ECMWF. We increase the model top from 50 hPa to 1 hPa, adding 6 levels in the stratosphere. This allows to model important stratospheric features such as sudden stratospheric warmings and the quasi-biennial oscillation. We discuss how improvements to the loss function scaling, a revised training data set and modified treatment of some physical variables in the stratosphere lead to improved forecast skill in the upper troposphere and lowermost stratosphere. To build up trust in the stratospheric representation in AIFS-CRPS, we investigate in more detail some recent sudden stratospheric warming events, i.e. how well they were predicted and whether AIFS-CRPS represents the expected response of surface temperatures. We conclude that a more explicit representation of the stratosphere in AIFS-CRPS is feasible and overall beneficial to the forecasts, as it provides a previously untapped source of predictive skill on sub-seasonal to seasonal time scales.