Data-Driven Techniques to Enhance and Supplement Computational Fluid Dynamics Prediction Capabilities

Abstract

Computational fluid dynamics simulations and in particular Reynolds-averaged Navier-Stokes simulations are the backbone of modern aerodynamic analysis and have become a crucial part of aircraft design and analysis. However, when looking at load predictions or moving towards the borders of the flight envelope, existing numerical simulation capabilities are still limited in terms of accuracy, robustness and/or performance. Instead, adding more knowledge through data-driven techniques offers the possibility to tackle some of these weaknesses. Within this work, two different routes will be presented that are currently pursued to enhance and supplement numerical simulation capabilities through data-driven techniques. First, reduced order modeling is introduced in which either a deep neural network architecture or a manifold-learning technique which is combined with a residual minimization strategy are employed to rapidly predict aerodynamic responses throughout the flight envelope. An industrial-relevant aircraft test case is used to demonstrate the performance of the proposed methods. Secondly, an insight into ongoing work on data-driven turbulence modeling is given. The herein presented approach concentrates on the structural uncertainty of the turbulence model transport equations and augments an existing model with a data-based, machine-learned term. Besides a brief description also some demonstration results for an aerodynamically relevant test case are shown.