Machine Learning for Earth system science
Institutskolloquium
- Datum: 17.04.2026
- Uhrzeit: 10:00 - 12:15
- Vortragender: Prof. Dr. Niklas Boers
- Niklas Boers works on theoretical questions of Earth system science with focus on the analysis, modelling, and prediction of extreme events and abrupt transitions (‘tipping points’). In his research, he develops methods rooted in Mathematics and Theoretical Physics, in particular Complexity Science and Machine Learning, to combine process-based and data-driven models. His work finds applications in climate dynamics, paleoclimatology, and in the context of anthropogenic climate change. Niklas Boers studied Physics and Mathematics at Ludwig Maximilian University of Munich and TUM and obtained his PhD in Theoretical Physics from the Humboldt University of Berlin. Thereafter he worked on different topics in theoretical Earth system dynamics at the Potsdam Institute for Climate Impact Research, Ecole Normale Supérieure in Paris, Imperial College London, and Freie Universität Berlin. 2021 Niklas Boers was appointed Professor of Earth system modelling at TUM.
- Ort: IPP Garching
- Raum: Arnulf-Schlüter Lecture Hall in Building D2 and Zoom
- Gastgeber: IPP
- Kontakt: stefan.possanner@ipp.mpg.de
Anthropogenic Climate Change poses three critical challenges for Earth system science: 1) Quantification of spatially resolved past trends of key variables such as temperature or precipitation from observations; 2) Projections of these variables given future anthropogenic emission and land use change scenarios, and ) Prediction of potential abrupt transitions of key Earth system components in response to anthropogenic forcing. I will present recent advances addressing these three challenges with (generative) machine learning, with focus on 1) spatiotemporal reconstructions of past climate fields from sparse and uncertain observations; 2) Bias correction and downscaling of Earth system models, including hybrid modelling approaches, and 3) Prediction of abrupt transitions with deep learning models.