Markus Mrosek from Volkswagen had a secondment to International School for Advanced Studies (SISSA) in June 2022.
The purpose of the secondment is two-fold:
1) industrialize the previously developed multi-fidelity approach involving active subspaces (MF-AS) (https://arxiv.org/pdf/2110.14396.pdf);
2) investigate Reduced Order Models (ROM) based on residual minimization (ResMin-ROM) for
automotive vehicle aerodynamics with geometric parametrization.
The work implemented during the secondment includes:
- Training in the python code for creating the MF-AS models, including the packages
Emukit (https://emukit.github.io/) and ATHENA (https://github.com/mathLab/ATHENA)
- Comparison between MF-AS and Gaussian Process Regression in drag coefficient prediction accuracy on 11 industrial external aerodynamic datasets
- Implementation of an automatic hyperparameter optimization procedure for the MF-AS ResMin-ROM
- Investigation of the “monotonicity problem”: the objective function for the residual minimization – calculated as some norm of the residual fields – and the error of the ResMin-ROM with respect to the true solution do not always behave monotonically
- Study effect of increasing number of snapshots (more full order simulations and/or more snapshots during the convergence of each full order simulation) on the monotonicity problem
- Investigation of the weighting of velocity and pressure residuals