Differentiable physics in scientific machine learning: towards fast and accurate surrogates for fluid dynamics

Machine-learned surrogate models are creating a paradigm shift in engineering design by speeding up CFD calculations to near real-time. After training on computationally-expensive CFD calculations, these surrogate models provide fast approximate diagnostics for unseen examples, enabling rapid testing of new designs.

Surrogate modelling is essentially a method of interpolating between CFD model outputs, which can be extremely challenging if the solution space is structurally complex. This project seeks to explore how including information about the gradients of the CFD outputs with respect to the input parameters, rather than simply the outputs themselves, can improve the surrogate. By including more information about the solution space, the interpolations should improve and the associated surrogate models become more accurate.

Algorithmic differentiation (AD) allows the differentiation of complex pieces of software, such as CFD, with respect to their input parameters. RWTH has developed software routines that are themselves differentiable, as well as tools to support application of AD. The student will develop their own surrogate models based on these tools to explore the effect of including gradient data, initially based on simple examples, but with a view to implementation on complex datasets.