Aerosol-cloud interactions occur on a variety of time scales that are all mixed in the observed large-scale time series signals. If not taken into account in statistical modelling, this leads to temporal confounding. Commonly, variability between different time scales can be analysed using spectral approaches, e.g., cross-coherence, but such approaches fail and yield spurious dependencies if common drivers or indirect dependencies are involved. There are some approaches for spectral versions of Granger causality, but these are not suited for large-scale and nonlinear time series datasets. We will investigate and develop new methodologies for causal discovery that take into account different time scales involving filtering techniques for such interactions. To this end, causal discovery methods suitable for large-scale linear and nonlinear time series datasets will be extended using Wavelets and Gaussian Processes (together with partner UVEG). Methods will be tested and improved on artificial data coming from the http://www.causeme.net causality benchmark database (initiated by DLR and UVEG). Then the new techniques will be applied and evaluated on simulated aerosol-precipitation interactions using CMIP5 climate models (in particular existing experiments with the GFDL model) together with partner UEDIN. Finally, also in cooperation with the partner EDIN, the method will be used to detect and attribute the aerosol influence on global precipitation patterns in the GPCP and MODIS datasets. In a secondment at GAF AG the student will utilize the method for parameter estimation in operational remote-sensing products.
Jakob Runge (supervisor), Joachim Denzler (co-supervisor), Gabi Hegerl (co-supervisor), Gustau Camps-Valls (co-supervisor), Sebastian Carl (GAF AG, non-academic advisor)
2 months at University of Valencia, 4 months at University of Edinburgh, 4 months at GAF AG
Enrolment in Doctoral degree
PhD in Data Science at University of Jena