Learning non-linear aerodynamics with probabilistic models
by
https://psich.zoom.us/j/68312125271?pwd=RnN3eDBVSzVsWVdmU1ljZE5oUmRlQT09 Meeting-ID: 683 1212 5271 Kenncode: 593061
OHSA/B17
Abstract:
Understanding airflow dynamics and predicting the forces it exerts on an airplane is crucial for future applications such as autonomous flying agents. Challenges are that algorithms need to be accurate and robust to unseen data. In a recent project, we analyzed, how well we can predict the forces acting on a pitching airfoil subject to a flow in a wind tunnel. First I will show how a custom hidden Markov model can be used, to provide accurate predictions of the lift and the pitching moment, based on a minimal number of pressure sensors. As a by-product, it also separates the airflow into interpretable states, which could be used for detecting when airflow is close to being disrupted. In the second part of the talk, we will see, how a combination of a state-space model and a neural network can be used to learn the dynamics of an airfoil. We will see, that the model can accurately predict the pressure field and forces on the airfoil when provided only with trajectories that lie outside the training data. The model also provides an interpretable latent space, where the predicted dynamics can be analyzed.
Laboratory for Simulation and Modelling
SDSC Hub at PSI