Riassunto analitico
This thesis makes a contribution to the field of computational fluid dynamics by means
of the development of an innovative data-driven turbulence modelling approach.
The work addresses the fundamental challenge of balancing computational efficiency
with simulation accuracy in turbulence modelling by combining elements from Reynolds-
Averaged Navier-Stokes (RANS) and Large Eddy Simulation (LES) methodologies. While
RANS models offer computational efficiency but limited accuracy in complex flows, and
LES provides higher fidelity at increased computational cost, this study presents a hybrid
framework that leverages the advantages of both approaches.
The present study builds upon and improves a previous research project at the Von Karman
Institute for Fluid Dynamics, where a novel data-driven technique was introduced
to obtain a scale-adaptive closure term.In the original model, an eight-input Artificial
Neural Network (ANN) was employed to correct the k-ω SST eddy viscosity definition,
exploiting a high-fidelity Direct Numerical Simulation (DNS) database of channel flow.
The methodology involves the application of spatial filtering operations on the DNS data
to emulate LES outcomes, and the subsequent calculation of the exact sub-grid viscosity,
the eddy viscosity according to the k-ω SST definition, and a selected set of parameters
to represent the flow and simulation set-up.
To ensure physical consistency, the sub-grid viscosity correction is constructed to satisfy
fundamental invariance properties, while the parameter set is systematically reduced to
a collection of dimensionless numbers through dimensional analysis. The ANN is then
trained to predict the correction factor as a function of these non-dimensional quantities,
allowing for accurate sub-grid value predictions across different mesh resolutions.
In this study, a thorough Shapley value analysis was employed to assess the contribution
of each input parameter to the model’s predictions. This analysis enabled the identification
and elimination of less significant variables. This optimisation process results in a
more efficient neural network while maintaining predictive accuracy.The enhanced turbulence
model undergoes an a priori validation analysis as well as an a posteriori one with
the OpenFOAM tool. To enhance the model’s generalization capabilities, the training
dataset is extended to include temporal boundary layer DNS data. This extension represents
a significant step toward developing a more universal turbulence model applicable
to diverse flow configurations.
|
