J1.1 Testing the Potential of Machine Learning for Weather Prediction

Wednesday, 9 January 2019: 8:30 AM
North 124B (Phoenix Convention Center - West and North Buildings)
Peter Dueben, ECMWF, Reading, United Kingdom; and P. Bauer

From the perspective of Earth System modelling, the use of machine learning, and in particular deep learning, is still in its infancy. There are many possible ways how deep learning could improve model quality or generate significant speed-ups for simulations. However, it has yet to be shown that deep learning can hold what it is promising for this application and its specific needs. We will present an example of deep learning methods applied to a toy model of atmospheric dynamics (the Lorenz'95 model) aiming to characterize skill for a highly non-linear, chaotic system but with a limited number of degrees of freedom. The assessment is then extended to full ECMWF forecast model output for selected parameters at reduced spatial resolution. Lastly, we show examples of deep learning methods applied to parts of the forecast model at full complexity targeting computational efficiency gains.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner