A modeling paradigm is developed to augment predictive models of turbulenceby effectively utilizing limited data generated from physical experiments. Thekey components of our approach involve inverse modeling to infer the spatialdistribution of model discrepancies, and, machine learning to reconstructdiscrepancy information from a large number of inverse problems into correctivemodel forms. We apply the methodology to turbulent flows over airfoilsinvolving flow separation. Model augmentations are developed for the SpalartAllmaras (SA) model using adjoint-based full field inference on experimentallymeasured lift coefficient data. When these model forms are reconstructed usingneural networks (NN) and embedded within a standard solver, we show that muchimproved predictions in lift can be obtained for geometries and flow conditionsthat were not used to train the model. The NN-augmented SA model also predictssurface pressures extremely well. Portability of this approach is demonstratedby confirming that predictive improvements are preserved when the augmentationis embedded in a different commercial finite-element solver. The broader visionis that by incorporating data that can reveal the form of the innate modeldiscrepancy, the applicability of data-driven turbulence models can be extendedto more general flows.
展开▼