Computer simulations of flows are extremely important in many areas − whether in the design of modern pipelines or in automotive development. Such simulations can be used, for example, to calculate air streams around vehicles or blood flow in human veins without having to perform expensive real-life experiments. Flow simulations are now routine in many industries. However, it can still take several days to perform the required calculations. Within this Proof-of-Concept grant, Prof. Nils Thuerey is working on methods to fundamentally improve such simulations. The Proof of Concept Grant for his project, entitled “dataFlow: a Data-driven Fluid Flow Solving Platform”, will help to refine a deep-learning algorithm that was the outcome of Prof. Thuerey’s earlier ERC Starting Grant project “realFlow”. The algorithm could make it possible to perform flow simulations on normal desktop PCs within just a few seconds. For this method, neural networks are trained to recognize recurring patterns in simulations. Once those patterns have been saved to memory, the data-driven model is able to generate new results very efficiently.
