Verändere deine Karriere: Werde Scientist for Machine Learning - two positions bei ECMWF in Bonn

The roles

ECMWF are looking to fill two positions that will develop cutting-edge machine learning applications to support future climate projections. This will allow you to apply and extend the latest generative machine learning approaches and train these at scale on very large datasets. Simultaneously, you will help to define the state-of-the-art in Earth system modelling and support Europe’s efforts to better understand and adapt to climate change.

Machine learning techniques for weather forecasting have made tremendous progress in the last two years, with state-of-the-art models now providing skill comparable to the best equation-based models at a fraction of the compute costs. For climate applications, large-scale machine learning techniques are still in their infancy and have so far mainly played a supporting role. The open positions give you the possibility to contribute to the development of future machine learning models for weather and climate. One position is climate-centred and funded by two related projects, WarmWorld ICON-Rep and EXPECT. The second position is focused on building a large machine learning model using a variety of data sources and applying it to short-term weather forecasting.

The WarmWorld ICON-Rep project will extend the AtmoRep model (https://github.com/clessig/atmorep), a large-scale, self-supervised representation learning model for atmospheric dynamics. In particular, you will enable native support for ICON mesh data in AtmoRep and than train the adapted model on ICON climate simulations using large scale compute infrastructure, including JUPITER - the first European exa-scale supercomputer at the Jülich Supercomputing Center. With the trained model, it will be explored to what extent an interpolation or extrapolation of climate scenarios is possible, e.g. if training on a small number of CMIP scenarios will allow to generate samples also from other ones not seen during training. Such a capability would hold enormous potential to speed-up and improve climate projections.

The EXPECT project develops European climate change projection and attribution capabilities. The position will support this by implementing a novel downscaling model that can take low-resolution climate model simulation data as input and produce high-resolution data that provides climate information at the local scale where most adaptation and mitigation efforts take place. The localized data should be statistically consistent with high-resolution climate simulation, e.g. from the EERIE project (https://eerie-project.eu), the Destination Earth Climate Digital Two and NextGEMS (https://nextgems-h2020.eu) but be produced at a small fraction of the computational costs.

The RAINA project aims to develop a statically robust, large-scale machine learning model for weather applications that supports a very fine, km-scale resolution. Through your work on the project, you will help to provide more accurate and reliable short-term forecasts of extreme events, such as strong wind and precipitation events with their impact on individuals and societies, than existing models. You will develop the RAINA core model that builds on the AtmoRep project and use self-supervised learning for model training. This will ensure that the required robust representations of extreme events are learned, and it will also be a key to obtain a model that can handle both coarser, global data as well as high-resolution, local data. You will also train the model at scale on large supercomputing infrastructure and together with the project partners implement the short-term forecasting downstream application.

The positions give you the possibility to be part of and shape the exciting developments on machine learning for Earth system modelling that are currently taking place at ECMWF, e.g. in the WeatherGenerator project and with AIFS (https://www.ecmwf.int/en/about/media-centre/aifs-blog), as well as in the WarmWorld and EXPECT projects, and in the wider community.

The team

The position will be part of the Earth System Modelling section at ECMWF and tightly linked to the different machine learning efforts at the Centre, e.g. the AIFS, ECMWF's data-driven forecasting model and the WeatherGenerator project.

About ECMWF

The European Centre for Medium-Range Weather Forecasts (ECMWF) is a world-leader in weather and environmental forecasting. As an international organisation we serve our members and the wider community with global weather predictions and data that is critical for understanding and solving the climate crisis. We function as a 24/7 research and operational centre with a focus on medium and long-range predictions, holding one of the largest meteorological data archives in the world. The success of our activities builds on the talent