Phd Studentship: AI-Enhanced Hybrid and Multi-Fidelity Large-Eddy Simulation for Efficient Prediction of Turbulent Flows

Numerical simulation is central to modern engineering design, enabling detailed analysis of complex fluid dynamic behaviour in aerospace, automotive, and energy systems. In industrial computational fluid dynamics (CFD), Reynolds Averaged Navier–Stokes (RANS) and Unsteady RANS (URANS) methods remain widely used because of their relatively low computational cost. However, their reliance on turbulence models limits accuracy for flows dominated by complex, unsteady turbulence. Large Eddy Simulation (LES) offers a higher fidelity alternative by resolving the large, energy containing turbulent structures while modelling only the smaller subgrid scale motions. This provides significantly improved predictive capability, but at a very high computational cost. High Reynolds number industrial LES may require billions of grid points and millions of timesteps, leading to runtimes of weeks even on advanced high performance computing systems.

Hybrid URANS–LES approaches aim to combine the efficiency of URANS with the accuracy of LES, making them attractive for practical engineering applications. Detached Eddy Simulation (DES) is widely used for high Reynolds number flows, applying URANS near walls and LES in separated or free shear regions. Delayed Detached Eddy Simulation (DDES) further refines this by basing the switching criterion on flow features rather than grid spacing, improving robustness across different geometries and mesh designs.

A promising research direction is the development of AI informed selection criteria to accelerate convergence and optimise computational workflows for multi fidelity URANS/LES simulations. Machine learning models could help determine the most appropriate turbulence modelling strategy—ranging from RANS to hybrid methods to full LES—based on evolving flow features. Such approaches may also support advanced meshing strategies, automated interrogation of flow history data, and code level optimisation to exploit emerging supercomputer architectures

Since LES accuracy depends strongly on the modelling of unresolved scales, the project will investigate how data driven subgrid scale models trained on high fidelity datasets can improve physical consistency and predictive performance. These models must respect the governing equations while capturing complex turbulence dynamics more effectively than traditional closures

The developed AI enhanced methodologies will be evaluated using representative turbulent flow configurations relevant to industrial applications, including external aerodynamic flows and internal turbomachinery environments. These test cases will allow assessment of how machine learning augmented hybrid and LES approaches influence accuracy, efficiency, and robustness compared with conventional methods.

The expected outcome is a suite of new hybrid AI–LES techniques capable of significantly improving the efficiency and predictive capability of high fidelity turbulence simulations. Such advances align with VECTA’s ambition to deliver transformational performance improvements—potentially exceeding 100× speedups—that cannot be achieved through hardware or algorithmic developments alone. Beyond aerospace, the resulting methods will be applicable to a wide range of scientific and engineering domains involving complex multiscale fluid flows.

This PhD forms part of the Virtual Exascale Calculations Transform Aviation (VECTA) EPSRC Strategic Prosperity Partnership, building on the ASiMoV programme to advance the modelling capabilities required for virtual design and certification of next generation aeroengines. VECTA brings together leading academic institutions and industry partners, including Rolls Royce, to develop transformative simulation technologies for complex engineering systems. The partnership aims to deliver step change improvements in predictive capability and computational performance, enabling high fidelity, physics based simulations to play a central role in future engineering design.

Advert information