Deep learning is causing a major paradigm shift in weather forecasting from physical modelling to data-driven or hybrid approaches. While the first ML-systems are now used in operational medium-range weather forecasting, the adaptation of deep learning techniques to the subseasonal-to-seasonal (S2S), decadal and multidecadal range is still lagging behind. New approaches are needed for S2S since atmospheric initial conditions are no longer the main driver of predictability. Similarly, deep-learning approaches are emerging for the decadal and multidecadal timescales, when climate forcing becomes a dominant influence on the atmosphere. The scarcity of simulation data and the lack of ground truth pose serious challenges at these time ranges. For this session, we solicit abstracts that present current ideas and results at the intersection of deep learning, S2S-forecasting and climate emulation and simulation.

In particular, we welcome contributions from the following areas.
- Generative modelling for time-scales beyond medium-range
- Calibration and uncertainty estimation for data-driven long-term forecasting
- Coupled generative models of earth system components
- Predictability limits and failure modes in data-driven long-term forecasting
- Out-of distribution emulation and forecasting
- Benchmarking and validation of long-term forecasts with data-driven models
- Identification of drivers of increased predictability and forecast skill for data-driven models (XAI)