AutoEncoders for Aerodynamic Predictions

Explore the application of autoencoders in aerodynamic predictions during this 1-hour 5-minute talk by Gianluca Iaccarino from Stanford University. Delve into the workings of autoencoders, their trustworthiness, and their use in predicting flow past wing geometries. Learn about non-linear compression techniques for creating low-dimensional latent representations of data and their relation to physical inputs. Discover how careful dataset construction can lead to interpretable latent variables for both attached and separated flow conditions. Examine the impact of uncertainties in autoencoder architecture, hyperparameters, and training data on predictions. Compare this approach to Gaussian Process regression and linear compression strategies, highlighting its advantages in extracting useful information on prediction uncertainty. Explore the balance between model uncertainties and variabilities from uncertain operating conditions when establishing prediction confidence. Conclude with a discussion on incorporating multi-fidelity data in autoencoder training. Engage in a Q&A session covering topics such as handling noisy data, computational costs, dynamic stall predictions, and the application of autoencoders in 3D, unsteady, turbulent flow scenarios.

DDPS | “AutoEncoders for Aerodynamic Predictions”