DREAM-Turbulence

Realistic aerodynamic configurations, both low-speed (wind turbines) and high-speed (aircraft), are dominated by highly anisotropic 3D turbulent flows, often with strong separation and curvature effects, whose accurate prediction still requires substantial improvements in turbulence modeling. Only the differential Reynolds stress (or higher order) equations contain a sufficient number of degrees of freedom to mathematically characterize these complex flows. The complete tensorial representations of the terms to be closed require the calibration of a large number of variable coefficients that non-linearly depend on local flow features. Recent developments of data-driven models are using machine learning and high-fidelity numerical or experimental databases to adequately formulate the unclosed terms without resorting to the restrictive simplifying assumptions and canonical flow configurations often adopted in classical turbulence modeling approaches. However, data-driven models are still in their infancy and struggle to generalize beyond the narrow class of flows for which they were trained due to poor choice of input features, violation of invariances, and lack of extensive high-fidelity reference datasets and improper use during training. To overcome these shortcomings and to promote disruptive change towards a new generation of efficient and generalizable machine-learning-based anisotropic turbulence models, the DREAM-Turbulence project will bring together a team of experts in both physics-based and data-driven turbulence modeling in France and Germany, who will work in close collaboration to produce and exploit high-quality data, improve tensorial representations, and develop data-efficient learning algorithms. The project also includes the generation of high-fidelity direct numerical simulation (DNS) data for 3D separated flow and the assessment of the proposed transparently interpretable modeling enhancements against an extended set of test-cases.