Colloquium: David Papp (FPP)

31 juli 2018 12:30 - Locatie: Lecture room B, Faculty of Aerospace Engineering, Kluyverweg 1, Delft.

Prediction of unsteady nonlinear aerodynamic loads using deep convolutional neural networks

Accurate modelling of unsteady nonlinear aerodynamics is essential for modern combat aircraft. However, conventional aerodynamic models lack the required fidelity while sophisticated Computational Fluid Dynamics (CFD) simulations are impractical in many cases. In the present work a Convolutional Neural Network (CNN) based surrogate model is introduced computing motion-induced surface pressure distributions and integral aerodynamic loads. Pressure distributions are predicted from geometry descriptions and motion history using a deep encoding-decoding CNN architecture. Integral forces and moments are derived from predicted pressure fields by an auxiliary encoding convolutional network. To train the model the flow characteristics are inferred from CFD simulations. After optimization, the model can replace CFD solvers within predefined flow and flight conditions. Experiments are conducted on the MULDICON test configuration of NATO. Results show good correlation with CFD reference data throughout wide ranges of inputs. Additionally, exploiting GPU-acceleration, the surrogate offers significant reduction in computation time.