PhD Seminar Series: “Making Aerodynamics Great Again (But Smaller): Low-Dimensional, Data-Efficient Surrogates”

Abstract: 

The increasing computational cost of high-fidelity Computational Fluid Dynamics simulations has led to a growing need for efficient surrogate modelling techniques in aerodynamics. This seminar explores the integration of manifold learning and neural networks to construct high-accuracy surrogate models that capture the underlying structure of aerodynamic data while significantly reducing computational requirements.
We begin by addressing dimensionality reduction as a key enabler of surrogate modelling. Isometric Feature Mapping (Isomap) and β-variational autoencoders (β-VAE) are employed to uncover a latent space where aerodynamic flow fields exhibit smooth variations, enabling efficient interpolation. By mapping high-dimensional aerodynamic fields onto this manifold and applying Gaussian Process Regression or neural network-based interpolation, we reconstruct aerodynamic quantities with minimal information loss, providing fast and reliable predictions.
Additionally, we discuss how multi-fidelity modelling can further enhance surrogate performance. By combining low- and high-fidelity data using multi-fidelity Gaussian Process Regression (MF-GPR) and Gappy POD, we demonstrate how surrogate models can increase accuracy while reducing the amount of high-fidelity data required.
Join the seminar to explore ideas on how data-driven surrogate modelling and multi-fidelity strategies are reshaping the way we approach aerodynamic predictions, making high-fidelity insights more accessible and computationally feasible.