Model-Driven Machine Learning for Climate and Earth Science
The “Model-Driven Machine Learning” department aims to understand and forecast the complex dynamics of the Earth’s atmosphere, the ocean, the land and the ice. Our work is characterized by the fact that we link machine learning with physical simulations, as each method on its own possesses weaknesses. We therefore combine the two to gain insights that are not available separately for each approach. Machine learning can handle large amounts of data, but ignores the physical laws of nature and is therefore difficult to transfer to new scenarios. Physical simulations, on the other hand, handle the complexity of natural processes well, but suffer from problems with data assimilation, parameter settings and uncertainty quantification. By combining physical and data-driven approaches, we compensate for these weaknesses. Through our work, we contribute to activities in the focal regions of climate and coasts.
The group is part of the Helmholtz AI initiative