Improving confidence on CFD by deep learning
Résumé
Computational fluid dynamics (CFD) has been successfully applied in various industrial areas. However, the accuracy of the CFD result depends on not only the software but also the user's comprehension of the simulated flow phenomena. Compared with a CFD beginner, an expert with sufficient knowledge of both the simulated phenomenon and the computational tools can generally set a proper numerical and model parameters of the simulation, such as the turbulence model, mesh refinement, etc., which finally leads to a more accurate result. On the other hand, Deep Learning algorithms trained on validated cases can give the users confidence on the CFD results by helping them to properly configure the CFD simulation with regard to the corresponding flow phenomenon. To achieve this goal, the flow phenomena inside the simulation domain shall first be correctly identified in order to propose adequate and optimal configurations. In this paper, using carefully selected indicators of vortex location as an input dataset, a Graph Convolutional Neural Network (GCNN) algorithm is proposed to accurately detect the presence of vortices. The results show that the GCNN algorithm trained on limited cases can successfully detect the vortex location in new cases which are not included in the dataset. The datasets are generated in parallel using the code saturne open source CFD code.
Fichier principal
Improving_confidence_on_CFD_by_deep_learning__ParCFD2022__2_.pdf (881.04 Ko)
Télécharger le fichier
Origine | Fichiers produits par l'(les) auteur(s) |
---|