Skip to Main content Skip to Navigation
Theses

Stratégies d'optimisation multi-objectifs par métamodèles de krigeage et de cokrigeage direct avec enrichissement multi-points : Application à la conception 1D d'un fan de refroidissement automobile

Abstract : The first design phases of a turbomachine, such as an automotive cooling fan, traditionally rely on low-fidelity models with low response times.An inverse method makes it possible, for example, to achieve a 3D geometry most suited to a nominal operating point and a vein geometry that the designer has specified. Although inherently dedicated to design, an inverse method has the major drawback of relying on the designer's know-how and on a trial-and-error approach in order to obtain the most efficient configuration.By coupling a direct method, dedicated to the performance analysis of any 3D geometry, with a multi-objective optimization algorithm, a more exhaustive and automated scan of the parameter space is performed and obtaining Pareto-optimal configurations is easier. Nevertheless, such a direct optimization strategy requires an intensive sampling of the latter space and is therefore no longer possible in more advanced design phases where evaluations are carried out by a more time-consuming high-fidelity model such as a Reynolds-Averaged Navier-Stokes (RANS) solver.Multi-objective Bayesian optimization strategies, embedding a metamodel and an infill criterion and performing a more efficient and parsimonious sampling of the parameter space, have thus been studied. In order to reduce their overall cost and to facilitate their development and their comparison, the sampled fan configurations in these optimization processes remain evaluated here through a direct method partly developed during this thesis.In addition to the criteria, this direct method can also evaluate the derivatives of the criteria with respect to design parameters for any sampled configuration. Therefore, the Bayesian optimization processes performed here are either based on kriging or cokriging metamodels. Their performances are compared on several fan shape optimization problems and the performances obtained by direct optimization processes are used as benchmarks. Moreover, the parameter space defining these optimization problems is not constrained in such a way that only feasible configurations can be evaluated. Indeed, like a RANS model, the direct method used here may sometimes not converge and fail to provide the criteria of some configurations.The goal of this thesis is thus to prepare for Bayesian optimization processesbased on evaluations from a RANS solver for which large non-feasible areas can be observed in the parameter space. If this solver comes with a sensitivity analysis method such as the adjoint, derivatives can be computed and the Bayesian optimization strategies based on cokriging highlighted in this thesis may prove to be interesting.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03267547
Contributor : Abes Star :  Contact Connect in order to contact the contributor
Submitted on : Tuesday, June 22, 2021 - 3:15:12 PM
Last modification on : Tuesday, September 14, 2021 - 10:20:08 AM
Long-term archiving on: : Thursday, September 23, 2021 - 6:47:06 PM

File

TH_T2743_mbuisson.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03267547, version 1
`

Citation

Martin Buisson. Stratégies d'optimisation multi-objectifs par métamodèles de krigeage et de cokrigeage direct avec enrichissement multi-points : Application à la conception 1D d'un fan de refroidissement automobile. Autre. Université de Lyon, 2021. Français. ⟨NNT : 2021LYSEC005⟩. ⟨tel-03267547⟩

Share

Metrics

Record views

46

Files downloads

54