Joint Seminar: More than better weather forecasts: The future of Earth system prediction

Forecasts and simulations of the earth system have tremendously improved over the past decades. We can predict the weather further in advance than ever before, we can forecast river levels on time scales never seen before and we can anticipate storm surges long before they batter our coasts. These advances have been achieved despite some severe deficiencies of our forecasting systems. For example, in any weather forecasting system, the water balance is severely violated due to the way we prioritise certain aspects of data assimilation or the methods we were required to choose to represent processes of the snow pack. Moving towards an earth system modelling approach, which integrates and dynamically feedbacks, between atmosphere, land, ocean and ice, allows many of these issues to be addressed head on. Earth system integration should also be user led and should allow anybody to add or modify a particular process or application of interest and run the full system to understand, analyse, improve within a context of an interactive earth. But this approach comes with several challenges. First, it requires a huge step up in computing resources. There are also other difficulties with the increasing heterogeneity of hardware infrastructure that we run our models on. There is therefore an urgent need to make forecasting model codes adaptable to future computer technologies and designed around data centric workflows.

Secondly, uncertainty remains an important focus in weather forecasting; a forecast system will always be imperfect. Despite many improvements over this past decade, the advantage of post-processed NWP outputs over raw model outputs remained constant or increased in some locations. Moving to higher resolution and implementing better representation of processes requires adequate observations to constrain the higher degrees of freedom which can stem from them.

Finally, forecast system design is constrained by resource limitations and balancing scientific and technical aspects is not always straightforward. There will be trade-offs between data assimilation, run time, size of the electrical group on HPC, number of ensemble members, competing forecast systems on the same HPC, running time step, archiving time step, speed of file system etc. Most importantly, these trade-offs need to ensure that forecasts and simulations are useful and are impact driven and led. This means users need to be enabled to manipulate, transform data by being provided with adequate cloud infrastructure, which can respond to new usage patterns of for example data hungry Machine learning. The latter will lead to new improvements which can be integratively tested in an earth system approach … closing the circle.




13:30 h


Bundesstr. 53, room 022/023
Seminar Room 022/023, Ground Floor, Bundesstrasse 53, 20146 Hamburg, Hamburg


Florian Pappenberger


Bjorn Stevens

Back to listing