WP 4: Parallel Efficiency and Multi-fidelity Optimisation

Objectives

• Implement adjoint-based r- and h-refinement methods suitable for unsteady industrial flows
• Port current adjoint solvers to GPU and benchmark performance
• Demonstrate adaptive and HPC methods on industrial adjoint cases.

Methodology

The mesh adaptation chain at WUT based on re-meshing using non-Euclidean metrics will be extended to unsteady flows. The hierarchical adaptation toolkit of QMUL will be parallelised integrated with an unsteady solver. The adaptation algorithms will be applied to unsteady flow cases and used for unsteady adjoint computations in WP5.

The unsteady adjoint in-house code "puma" of NTUA will be GPU-enabled. The currently non-GPU modules necessary for OpenFOAM and its adjoint will be ported to GPU. Runs on the nGPU platforms atNTUA, and the GPU clusters at RWTH and QMUL will be performed and the achieved speed-up will be quantified.

ESR 5 at ESI will extend the discrete prototype to support one shot methods with support for large geometric changes. The developments in WP4 are needed for the very large industrial cases of WP5.

Lead: WUT, participants: NTUA, ENGS, ESI, QMUL [ months 1-36]

Participating Research Fellows

• Jan Hückelheim (ESR 1 at QMUL): Application of AD for unsteady flows (WPs 3, 4, 5)
• Mateusz Gugala (ESR 2 at QMUL): Multi-fidelity unsteady optimisation (WPs 3, 4, 5)
• George Eleftheriou (ESR 5 at ESI): Discrete Adjoint Solvers and Parallelisation (WPs 1, 3, 4)
• Sheikh Razibul Islam (ESR 14 at WUT): Adaptation for unsteady adjoint optimisation (WPs 3, 4, 5)