69901: PhD candidate Computer Scientist, Physicist, Mathematician or similar (f/m/x) - Causally-constrained Machine Learning Based Parametrizations for a Climate Model

German Aerospace Center (DLR)

Oberpfaffenhofen, Germany

Area of research:

Other,PHD Thesis


Part-Time Suitability:

The position is suitable for part-time employment.


Job description:

The Department of Earth System Model Evaluation and Analysis at the German Aerospace Center (DLR) invites applications for a PhD Position in the field of “Causally-constrained machine learning (ML) – based parametrizations for climate models” under the supervision of Prof. Veronika Eyring.

Despite significant progress in climate modelling over the last few decades, systematic biases and substantial uncertainty in the model responses remain. For example, the range of simulated effective climate sensitivity – the change in global mean surface temperature for a doubling of atmospheric CO2 – has not decreased since the 1970s. A major cause of this is differences in the representation of clouds and other processes occurring at spatial scales smaller than the model grid resolution. These need to be approximated through parametrisations that represent the statistical effect of that process at the grid scale of the model. This impacts the model’s ability to accurately project global and regional climate change, climate variability, extremes and impacts on ecosystems and biogeochemical cycles. High-resolution, cloud resolving models alleviate many biases of coarse-resolution models for deep clouds and convection, wave propagation and precipitation, but they cannot be run at climate timescales for multiple decades or longer due to high computational costs. New approaches are required that exploit opportunities from increasing computational power while building on and expanding the knowledge gained from theory and observations and continuing the inclusion of missing processes in the models.

While efforts are underway to develop convection resolving high resolution global climate models where some of the physical processes can be explicitly modelled, smaller scale effects will continue to be parameterized in models covering long time periods or in those that require additional complexity beyond the traditional physical model setup. A promising way forward that is addressed in this thesis is to develop innovative methods combining causal discovery and deep learning via neural nets in a consistent framework (ensuring the neural nets respect the causal structure of the system).

This work will support and enable the development of causally-constrained machine learning and especially deep learning technique-based atmospheric parametrizations for the Icosahedral nonhydrostatic (ICON) atmospheric general circulation model, which have the potential to solve problems arising from using conventional neural nets that do not take causal relationships into account.

The candidate will be part of an international team of the European Research Council (ERC) Synergy Grant on „Understanding and Modelling the Earth System with Machine Learning (USMILE, www.usmile-erc.eu/)“ and will be co-supervised by Prof. Pierre Gentine (Columbia University, New York, USA, gentinelab.eee.columbia.edu) and Prof. Jakob Runge (DLR Institute of Data Science, Jena, climateinformaticslab.com/). The project will entail visits at Prof. Runge’s Causal Inference and Climate Informatics Group.

During your PhD you will utilize high-resolution ICON simulations (large eddy simulations – LES and cloud resolving simulations) to develop causal neural net approaches that can be applied in the development of machine learning parameterizations, and will implement the new causally-constrained parameterization into ICON to perform climate projections.

  • Development of a new class of deep learning techniques, namely causally-constrained neural nets, that can emulate subgrid processes in coarse resolution state information (e.g. mean temperature, humidity, tracer transport), ensuring the predictions of the algorithm do not violate the underlying causal structure of the system.
  • Implementation into the general circulation model ICON
  • Performing corresponding ICON-ML simulations
  • Evaluation of the resulting ICON-ML with the Earth System Model Evaluation Tool (ESMValTool)
  • Documentation and software as open source

At the DLR Institute of Atmospheric Physics we provide excellent facilities with opportunities to work with world-renowned experts in the field of Earth system modelling and observations. You will be part of the Earth System Model Evaluation and Analysis Department which develops and applies innovative methods, including ML techniques, for the analysis of Earth system models in comparison to observations with the aim to better understand and project the Earth system. The department is strongly linked to international research activities within the World Climate Research Programme (WCRP), with substantial contributions to the Coupled Model Intercomparison Project (CMIP). We are striving to increase the proportion of female employees and therefore particularly welcome applications from women.

This research center is part of the Helmholtz Association of German Research Centers. With more than 42,000 employees and an annual budget of over € 5 billion, the Helmholtz Association is Germany’s largest scientific organisation.

Please apply via recruiter’s website.

Quote Reference: 15165991