Generative Modeling of Dynamic Systems
The fast, time-resolved prediction of turbulent flows, particularly in complex natural environments such as urban canopies or wind farms, remains a major challenge. Traditional high-fidelity approaches, including DNS and LES, are often prohibitively expensive and rely on uncertain model inputs. Recent advances in generative AI provide a promising new avenue for capturing flow dynamics efficiently while quantifying predictive uncertainties. We investigate how state-of-the-art generative models can be adapted for scientific modeling to enable reliable and scalable turbulence prediction, as explored in the projects below.
Turbulence Generation and Data Assimilation with Reduced-Dimensionality Models
We propose replacing conventional sequential data assimilation pipelines—typically consisting of numerical solvers followed by a correction stage—with a scalable, data-driven alternative. Our framework integrates dimensionality-reduction techniques with a transformer-based diffusion model operating in a compact latent space, enabling efficient prediction and data assimilation of four-dimensional turbulent flow fields.
Physics-Guided Generative Modeling for Wind Turbine Wake Prediction
Current simulation methods of wind farms wakes lack physical realism and are limited in the integration of multi-source data. Here, we explore and develop a generative modeling framework for wind turbine wake prediction that is able to integrate physical constraints and multi-source measurements e.g., UAV and LiDAR.
Uncertainty Quantification in Urban Planning
Urban planning requires reliable assessments of how land-use changes affect local climate, yet high-resolution simulations are too costly for large ensembles and comprehensive uncertainty quantification. We propose a conditional diffusion model with a given land-use configuration as the conditioning input.
Funding Information and Acknowledgement
Funded by Carl-Zeiss-Stiftung (CZS, Carl Zeiss Foundation), project number P2021_0401. Funding and support is highly acknowledged.
Funded by Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2075 - 390740016. We acknowledge the support by the Stuttgart Center for Simulation Science (SC SimTech).