Machine learning for optimal mesh quality

Engineering companies are heavily reliant on computational simulation to help design and validate their products. During the last two years, AD&S have invested a lot of effort into the efficient production of large datasets (typically consisting of 10 000 - 100 000 computations). Through a series of methods (chimera sweeps, multi-fidelity, improved solver) the costs of such large datasets have been reduced by up to two orders of magnitude compared to the original codes.

Of course, the mesh spacing used (coarse and fine RANS for the multi-fidelity), play an important role in the computational effort. A scheme that is based on an adjoint solver for analysing the utilized meshes did not produce the desired solution. To date, AD&S have not been able to identify an automatic tool for optimizing the spacing distribution of the meshes used, and mostly rely on experience and "gut-feel" of highly-skilled individuals when defining the spacing distribution function.

The main aim of this project is to determine how machine learning can be used to capture engineering knowledge embedded within previous analyses and exploit it for construction of spacing distribution function for future products. The method will utilise the available wealth of modelling data to identify a relation between the mesh spacing distribution function and the level of required accuracy of the solution.

Neural network technique will be utilised to train a point distribution function that relate the mesh characteristic such as element type, spacing, stretching and gradation, to the geometric features and flow conditions such as angle of attack, Mach number and Reynolds number. The trained model will enable the determination of an optimal spacing distribution that "guarantees" a discretization error lower than a user-specified value for a specified level of solution fidelity.