Reduced order models from CFD

Factorial design is a common approach to design-space investigation. More advanced techniques include Taguchi orthogonal arrays, Latin Hypercubes and others.

Massimiliano Molinari

The focus of this work is the use of modelling techniques to explore new design spaces, based on CFD simulations, while still keeping the computational cost manageable.

One can reduce the complexity of the original problem by building a ROM (reduced order model), which represents a reliable “surrogate” of the original, complex model. There are many Response Surface Methodology (RSM) techniques commonly used for this purpose in several industrial design areas; they combine Design of Experiments (DoE) techniques (e.g. factorial design and Taguchi orthogonal arrays) to efficiently select relevant points of the design space to build the ROM.

Comparison between old data currently used in design (by Lieblein, 1956) and the correlations obtained with CFD tools. By modelling the design space by means of ROM (reduced order model) techniques, we aim to explore the design space and to reveal the boundaries of reliability of the CFD correlations.

An optimisation algorithm can then be run using this model to quickly find a Pareto front and, essentially, an optimal solution. The key issue is the assessment of the range of applicability of the model and the suitability of different model types (e.g. polynomial or Neural Networks) for representing the problem in hand. Furthermore, an important question is how the limits of feasible design space can be determined using these models. We are also looking at ways of integrating historical simulation or experimental data into models.

« back