Generative Machine Learning for Accurate and High-Resolution Climate Projections

• Applicant: Dr. Philipp Hess
• Subject Area: Atmospheric Science
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 556377191

Anthropogenic climate change poses severe risks to natural and human systems, making the assessment of its impacts, in particular of extreme weather events, an urgent task of fundamental importance for our society. Dynamical climate impact models have been developed to represent our process understanding of the biophysical response to weather fluctuations in future climate projections, e.g., for crop yields, runoff, river discharge, or fire. Numerical Earth system models (ESMs), on the other hand, are our primary tool for projecting possible future climates by simulating relevant physical processes of the different Earth system components and their interactions. However, to make global ESM simulations computationally feasible, the horizontal spatial resolution of ESMs is currently limited to around 100 kilometers. Key processes, such as precipitation, that involve smaller spatial scales or are otherwise computationally too expensive to be modelled explicitly in the ESM need to be approximated. This can lead to systematic errors (biases) in the simulations. Therefore, estimating the historical impacts of climate change and projecting them into the future with impact models requires bias correction and downscaling of the ESM simulations. Statistical methods have been developed that use observational data to correct for biases in ESM simulations and to resolve local scales. However, current statistical approaches are restricted to univariate distributions for each grid cell and variable. Spatial and inter-variable correlations in high-dimensional simulations and observations are not taken into account, which can lead to unphysical inconsistencies, causing distorted impact model simulations. In this project, a fully generative machine learning-based approach will be developed for a multivariate treatment of the most important climate impact variables to yield global, physically consistent, and accurate high-resolution climate projections. Building on the preliminary work that shows the effectiveness of generative machine learning for bias correcting and downscaling of global precipitation fields from ESM simulations, the proposed project will significantly advance these proof-of-concept studies. In particular, a focus will be on multivariate physical consistency, the preservation of trends, temporal dynamics, and scalability to high resolutions. A comprehensive evaluation of the post-processed climate projections, including statistical metrics and dynamical impact models, will ensure that the high quality standards of the Inter-Sectoral Impact Models Intercomparison Project (ISIMIP) are fulfilled. ISIMIP, with its global network of climate impact modeling teams, will provide an important platform for making the novel climate dataset available.

• DFG Programme: WBP Position