Optimisation and Prediction of Computational Fluid Dynamic Mesh using Evolutionary Algorithms and Neural Network Surrogate Models

This research aims to use Evolutionary Algorithms (EA) to optimise Computational Fluid Dynamic (CFD) mesh for a turbulent jet. The Star-CD CFD package is used to construct, solve and post process the mesh and the Covariance Matrix Adaption Evolutionary Strategy (CMA-ES) algorithm is used to optimise the mesh. A Recurrent Neural Network (RNN) is also trained to predict converged CFD results, from un-converged data, aiming to reduce computation time when CFD simulations are needed for optimisation of either the CFD mesh or design of turbulent jets.

Results from a mesh optimisation loop were not positive, so attention was focused on training the RNN to predict converged CFD results and preliminary findings from this work have been encouraging.

In this talk we present the motivation, method, results and conclusions of all the work undertaken to date, as well as our future plans and ways in which the research can be further explored.