JuLES: A Data-Driven Framework for Subfilter Model Development on Supercomputers at Scale

The transformation of all industrial processes resulting from the fight against climate change can initially be described by the fact that already existing climate-neutral technologies are developed to market maturity and made more efficient as quickly as possible. Examples include renewable energy generation from wind power and the hydrogen industry. This is inconceivable without predictive simulations and supercomputers to drastically reduce the time-to-innovation. JuLES (JUelich Large-Eddy Simulation) is a disruptive Software-as-a-Service (SaaS) platform for developing reduced models of energy flows (e.g. flows in turbines) in the shortest possible time and with unprecedented predictive accuracy. It is integrated into the supercomputing environment at the Jülich Supercomputing Centre, including the JupyterLab cloud, HPC resources, and quantum computing, to provide maximum opportunities with minimum barriers to entry. This is seen as an absolute key technology. Technically, JuLES uses Physics-Informed Enhanced Super-Resolution Generative Adversarial Networks (PIESRGAN) for Large-Eddy Simulation (LES). PIESRGAN for LES enables the data-driven development of reduced models for energy flows that have been shown to be significantly superior to existing models in terms of prediction accuracy, robustness, and universality, while reducing simulation costs by two orders of magnitude and I/O by as much as three orders of magnitude. Consequently, industrial development processes are potentially accelerated by years. The combination of classical, well-established fluid mechanics techniques and cutting-edge AI technology, its ability to fully scale on today's supercomputers, and its cloud-based n:m on-the-fly training make it a very powerful tool for industrial transformation and even ready for the upcoming exascale machine JUPITER.