Stabilization Strategies for a Variational MultiScale Method coupled with an Artificial Neural Network

Stabilization Strategies for a Variational MultiScale Method coupled with an Artificial Neural Network

Rajampeta, Sharath (TU Delft Aerospace Engineering)

Hulshoff, S.J. (mentor)

Delft University of Technology

Aerospace Engineering

2021-10-08

The use of Large Eddy Simulations(LES) in industry is steadily increasing as the complexity of the problems are increasing. However there remains a potential for significant improvements in LES. The Variational Multiscale Method(VMM) introduced by Hughes et al. (1998) provides a more consistent framework for turbulence modelling. The use of an Artificial Neural Network (ANN) to construct complex flow models in LES is becoming more prevalent. To mirror these developments in LES, attempts in the TU delft by Robijns (2019), Janssens (2019) and Pusuluri (2020) to couple the ANN to the VMM have met with moderate success but with several limitations. The strategies employed by these authors to stabilize the VMM-ANN simulations had drawbacks like their very nonphysical nature, requiring huge amounts of data, loss in the accuracy of the ANN, the empirical nature of the strategy etc. New strategies are devised in this thesis whose motivation have a physical and mathematical basis. They aim to address several drawbacks of the methods the previous authors used. These strategies were tested individually on several 1D burgers formulations and conclusions
were drawn. Finally a combination of strategies was found that stabilized a very complex VMM-ANN problem.

variational multiscale method
Artifical neural networks
Stabilisation

http://resolver.tudelft.nl/uuid:2114d9a4-7333-47bd-9793-77721a8303b0

Student theses

master thesis

© 2021 Sharath Rajampeta