Causal discovery in the presence of latent variables


This is a theoretically and methodologically oriented project. You will develop and implement cutting-edge data-driven methods that will enhance our understanding of the complex Earth system. You will be part of a research group ( of physicists, environmental scientists, mathematicians, and computer scientists that work together on exciting topics in causal inference, machine learning, and dynamical systems for climate research.

Causal inference and discovery address the problem of learning cause and effect relationships from observational data. These methods help to deepen our understanding of complex dynamical processes in the climate system such as the interactions of aerosols and clouds. A fundamental problem in attributing causal information from large-scale remote sensing time series is that some important driving variables, so called confounders, may not be observed. This can lead to spurious correlations and apparent non-stationarity. These problems can be addressed by certain causal inference algorithms, which in many cases are able to decide whether an estimated dependency is in fact causal or merely due to an unobserved confounder. However, such algorithms are currently not well adapted to the particular challenges occurring in large-scale climate time series.

This project will investigate these problems and develop new causal inference methods suitable for large-scale time series with unobserved confounders. These methods will be systematically evaluated and improved on artificial data from the causality benchmark database. In collaboration with our partners at the University of Oxford, the new techniques will further be applied to aerosol-cloud interaction data from large-eddy simulations. By systematically including and excluding selected variables, we will examine the effect of unobserved variables on the estimation of aerosol-cloud interactions. Finally, the methods will be applied to real-world climate observations such as satellite imagery to investigate the impact of aerosols on cloud albedo and the associated radiative effects. The newly developed approaches will also be put to use in a secondment at Amazon Cambridge.

This project plan is flexible and very much open to your own ideas! To grow as a scientist, we will provide you with the support you need but also grant the freedom you want. The project is hosted at the German Aerospace Center (DLR), which is Germany’s national research institution for aerospace, energy and transportation. Our group is based at the DLR Institute of Data Science in Jena, a lovely mid-sized university town that is home to various research facilities and technology-focused companies. DLR supports a healthy work-life balance, including remote working options.


Jakob Runge (supervisor), Joachim Denzler (co-supervisor), Philip Stier (co-supervisor), Javier Gonzalez Hernandez (Amazon, non-academic advisor)


3 months at University of Oxford & 6 months at Amazon Cambridge

Enrolment in Doctoral degree

PhD in Data Science at University of Jena

tile dlrwhite bg