Projects

Following up on the findings from the UNITE project, we aim to diagnose the hydrological consistency of hybrid models through mapping between model components and physical processes. We envision a model evaluation space “UNI-BENCH” that features three axes pivotal for model evaluation: performance, interpretability, and process consistency. Equipped with universal anchors for benchmarking, the evaluation framework shall help ensure that hybrid models perform well for the right reasons, such that we can confidently promote their use in science and practice.

Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), project 507884992

Geomodeling is a valuable tool for answering urgent questions, e.g. with respect to climate change or flood protection. Models are typically constructed with a specific purpose in mind, which limits their range of applicability. The aim of this project is to develop a fundamentally different, open-purpose type of model through an innovative combination of ML methods: neural stars. They can also be applied to unforeseen questions, thus saving resources in model (re-)building and modification, and enable efficient "storage" and extension of knowledge.

Funded by the Vector Stiftung in the program MINT-Innovationen 2024

Merging data-driven with physics-based modeling requires also merging on the level of statistical evaluation: we propose a novel model evaluation framework that is able to fuse uncertainty assessment with information-theoretic measures to bring (lack of) information in physics-based models and information in data to the same scale. The developed framework enables an objective comparison of the various emerging approaches to hybrid modelling and provides guidance in model development and knowledge discovery.

Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), project 507884992

Proper sensitivity and uncertainty analysis for complex environmental systems models may become computationally prohibitive. Cheap surrogate models can be an alternative to enable such analyses; however, it is an open research question how to correctly propagate the uncertainties related to surrogate modelling. We propose a fully Bayesian approach to surrogate-based uncertainty propagation, parameter inference, and uncertainty validation. Further, we introduce an innovative data-integrated training scheme for uncertainty-aware surrogates.

Funded by SimTech