Towards a fully machine learned Earth system model at ECMWF

Today, pure machine learned models have a significant impact on numerical weather predictions and the European Centre for Medium-Range Weather Forecasts (ECMWF) has started to produce daily predictions with the Artificial Intelligence Forecasting System (AIFS) – a purely data-driven model of the atmosphere. While results are very good for weather predictions into the medium-range, it is still unclear how much impact pure machine learning models will have on ocean modelling and climate predictions, which will have less training data available and require extrapolation into unseen weather regimes and climate states. To explore these aspects, ECMWF has started developing pure machine learning models for the ocean, sea ice, surface ocean waves, and land as part of the EU-funded Destination Earth programme, and will eventually integrate these components into the AIFS to design a full machine learned model of the Earth system. This talk will summarise the ongoing work and experience with the AIFS and present how the expertise and infrastructure that ECMWF is developing can aid the system's expansion towards more machine learned Earth system model components.