Model-Driven Machine Learning for Climate and Earth Science
The “Model-Driven Machine Learning” department develops hybrid geoscientific models combining machine learning (ML) with numerical simulation. ML approaches thrive on big data, but ignore physical laws and generalizes poorly to new scenarios outside their training data. Numerical simulations incorporate scientific knowledge and generalize well, but struggle with data-oriented tasks such as parameter tuning and data assimilation.
We use hybrid models to better, forecast, simulate and understand the atmosphere and ocean. A major focus of our research is on finding new ways to represent physical, chemical and biological processes on fine spatial scales, far beneath the grid spacing of a numerical simulation. Ultimately, we aim to provide efficient, accurate and easy-to-use hybrid models as `building blocks’ for climate and weather simulations.
The group is part of the Helmholtz AI initiative