Artificial intelligence and supercomputing at the service of aerodynamics: from vehicle optimization to urban hazard mitigation

This thesis presents a set of methodologies for dimensionality reduction of turbulent flow data, with a focus on high-fidelity simulations of external aerodynamics in industrial contexts, such as the flow around simplified automobiles. These simulations, typically performed on unstructured meshes with hundreds of millions of degrees of freedom, require scalable tools for analysis and modeling. All developments have been implemented in pyLOM, an open source Python library designed to reduce dimensionality of data on the order of magnitude of terabytes.

The work progresses in four main stages. First, classical reduction techniques based on singular value decomposition (SVD), such as eigen orthogonal decomposition (POD), dynamic mode decomposition (DMD) and spectral POD (SPOD), have been adapted to high-performance computing by taking advantage of parallel QR factorization. This has allowed these algorithms to be applied to multi-terabyte datasets, such as the direct numerical simulation of flow in the Stanford diffuser. Secondly, a variational autoencoder (VAE) based on convolutional neural networks (CNN) has been developed for nonlinear dimensionality reduction. This strategy is able to successfully capture the temporal dynamics of the Windsor car rear pressure with only two latent variables. Both methodologies have been combined to create a new method called Geometry-Agnostic Variational-autoencoder Integration (GAVI), replacing the SVD step with a VAE that operates on QR-factored data. GAVI provides compact latent spaces without the need for structured meshes, achieving high energy recovery in several test cases. Finally, a transformer-based strategy is proposed to compensate for the energy loss in the reduced models. By learning the spatial distribution of the unresolved fluctuations, the accuracy of both the instantaneous fields and the root-mean-square value of the fluctuations is improved.