In the past year, it has been demonstrated that machine learning models trained on the ERA5 reanalysis can
outperform even the best conventional medium-range forecasting systems. One of the next frontiers is the
development of machine learning-based Earth system models that cover not only the atmosphere but also
the other subsystems of the Earth and that are able to represent processes on much wider range of scales in
space and time. This poses many challenges in terms of Earth system modeling, machine learning, and highperformance
computing, e.g. how training on many different datasets with different resolutions and quality
characteristics is possible. In my talk, I will discuss these challenges and first steps towards addressing
these. I will also present results on training directly from observations, which is a tantalizing information
source for building models that can potentially outperform conventional ones.