Applying machine-learning approaches to develop novel turbulence models
After several decades of research in the field of turbulence models, well-behaved wall bounded flows can be computed with reasonable accuracy using Reynolds averaged Navier-Stokes (RANS) models, the computationally least expensive form of turbulence simulation. However, when flow separation occurs (e.g. bluff bodies, sustained adverse-pressure gradients), the accuracy of RANS based models deteriorates considerably. In situations where separation occurs and the flow is dominated by large-scale anisotropic vortical structures, Large-Eddy simulation (LES) methods are the method of choice. Unfortunately, for complex geometries that feature wall bounded flow, LES is prohibitively expensive and, even though it can now be performed on the latest high-performance computers, will not permit parametric studies crucial in the design stage. In an attempt to address the above shortcomings of both RANS and LES models, so-called hybrid RANS/LES have been developed that promise to exploit the strength of each model. However, these methods are hard to derive in a mathematically rigorous manner and the currently used methods are based on certain assumptions or even formulated in an entirely ad-hoc manner. At the same time, even hybrid RANS/LES methods can be too costly in an industrial context.
The aim of this project is to continue developing a turbulence model creation framework pioneered at the University of Melbourne that relies on applying gene-expression programming algorithms to high-fidelity experimental or numerical data. Preliminary work on simple canonical flow configurations has shown good performance. However, it is now of key importance to adapt and refine the approach for considerably more complex flow configurations and for higher Reynolds numbers.
Leader: Richard Sandberg
Staff: Andrew Ooi
Optimisation of resources and infrastructure
drag reduction; fluid dynamics; numerical modelling; turbulence