Presentation pdf 4.0MB

Parallelisation and Error Handling of Shape Optimisation for Expensive and Constrained Multi-Objective, Multi-Parameter CFD Problems

Existing Bayesian Optimisation methods generally explore the parameter space sequentially and those that have exploited parallelism are generally single objective. Here we utilise a penalised acquisition function in order to collect batches of expensive and constrained multi-objective, multi-parameter CFD runs. Latin-Hypercube sampling of the design space is performed asynchronously, whilst solution exploitation/exploration is done batch-synchronously. The reason for this is that sampling the design space does not depend on prior samples, whilst the exploitation/exploration phase uses a surrogate model which depends on prior samples in order to make predictions which maximise the acquisition function. The acquisition function must therefore be sequentially penalised in order for this to work, otherwise there is a risk of computational resources being expended on designs which are too similar to the detriment of global search. This leads to batch-sequential parallelisation.

The parallelised optimisation approach is evaluated with a CFD model, which is an expensive 3D Lagrangian parcel-tracking simulation, with one-way coupling to a converged Eulerian steady flow field. The shape optimisation approach is based on our recent work in this area [1]. Although block-structured mesh generation is accurate and convergence of the solution can be monitored, direct use of mesh generation in optimisation is challenging because it cannot be guaranteed that a mesh will be generated at all from the scripting process, or if it is, the mesh may be skewed. A second problem is that even if a mesh is successfully generated, it may not be of sufficient quality to allow the solution to converge. For both problems, the acquisition function can be penalised in a similar manner to the parallelisation strategy.

The work is performed using the open-source software OpenFOAM on the Isambard supercomputer held at the UK Met Office and funded by the EPSRC for the GW4 Alliance. It also extends the Bayesian Optimisation code IscaOpt [2] and the CFD suite generated previously [1]. Prior work using this software was supported by an EPSRC grant (reference number: EP/M017915/1) for the University of Exeter. The current project is supported by the EPSRC (10%) and Innovate UK (90%) on a KTP programme (reference number: 11477) with industrial application to hydrodynamic separators at Hydro International UK.

REFERENCES [1] S. Daniels, A. Rahat, R. Everson, G. Tabor, and J. Fieldsend (2018). A suite of computationally expensive shape optimisation problems using computational fluid dynamics. In: Parallel Problem Solving from Nature – PPSN XV, pp. 296–307. [2] A. Rahat, R. Everson, J. Fieldsend, Alternative infill strategies for expensive multi-objective optimisation. In: Proceedings of the Genetic and Evolutionary Computation Conference, pp. 873–880, ACM, 2017.