Potentials and Limitations of the A Priori Data-Augmentation of Turbulence Closure Models

Beschreibung

The CFD simulations which are commonly used in engineering design processes are based on the RANS equations and accompanying turbulence closure models. Structural and parametric deficiencies of such closure models, however, often lead to inaccurate flow field predictions. This thesis is concerned with the investigation of the extent to which replacing the closure coefficients of eddy viscosity models with functions of the local flow field, extracted from DNS data, can remedy their limitations. Specifically, the effects of such an a priori augmentation on the accuracy of the flow field predictions and the potential for developing more universal closures are assessed. A two-step augmentation procedure is proposed. First, given the DNS data, inverting the closure model reveals the optimal spatial distribution of its coefficients for a particular application. This case-specific knowledge can then be generalized by regressing functions capable of predicting those optimal values based on quantities that describe the local flow field. For the training cases considered, the errors in the predicted velocity fields can be reduced by up to 65 %. Furthermore, it is demonstrated that the augmented models provide more accurate predictions than the baseline for a wide range of Reynolds numbers and geometrically different test cases. However, the augmentations often turn out to be harmful when applied to test cases that are characterized by different flow phenomena than the training case. Using multiple training cases at a time to cope with this loss of universality is found to be ineffective due to a trade-off between accurately fitting the training data and the a posteriori model stability. In conclusion, a priori data-augmentation is suitable for developing highly specialized models that achieve the desired accuracy gains for classes phenomenologically similar, not too complex flows.

Autor*in

Hannes Mandler

Themen in »Potentials and Limitations of the A Priori Data-Augmentation of Turbulence Closure Models«

Data-Driven Turbulence Modeling NeuralSST Model Scientific Machine Learning

Stimmen zu »Potentials and Limitations of the A Priori Data-Augmentation of Turbulence Closure Models«

Details

Link teilen