Saetta, Ettore (2023) Machine Learning in Aerodynamics: Clustering and Autoencoders for CFD Applications. [Tesi di dottorato]

Anteprima

Testo Saetta_Ettore_36.pdf Download (46MB) | Anteprima

Abstract

A hot topic in all scientific fields over the last decade has been Machine Learning (ML), a branch of artificial intelligence. Fluid Mechanics has also been exploring the great potential of ML technology. Aerodynamic analyses traditionally involve Experimental Fluid Dynamics (EFD), Computational Fluid Dynamics (CFD), and semi-empirical models for preliminary design. In the context of CFD, the rapid growth of computational resources and development of new technologies and algorithms, is progressively increasing the feasibility to perform high-fidelity CFD simulations, even for complex geometries. Nevertheless, the journey toward the practical implementation of high-fidelity CFD for real-world applications remains challenging. In this context ML holds the potential to revolutionize our approach to Fluid Dynamic analyses and significantly enhance their outcomes. Despite initial skepticism, ML has found successful integration in various real-world applications, ranging from autonomous vehicles to medical diagnosis. Also in the Fluid Dynamic framework, it proved to be very promising for a wide array of applications, including turbulence modeling, flow control, shape optimization, and flow field prediction. A fascinating aspect of combining ML with Fluid Mechanics is that Fluid Mechanics is grounded in well-established theories and first principles. The fusion of physical knowledge with data-driven methodologies can be mutually beneficial for both Fluid Mechanics and ML. As Fluid Mechanics continues to uncover the great potential of ML, various concerns and open questions remain. The main areas of focus include the interpretability of these models, the handling of extensive data acquired from numerical simulations and experiments, ensuring that these models yield physically meaningful results, and the uncertainty quantification of the ML outcomes. In this dissertation, ML algorithms are investigated within the context of CFD, with a particular focus on unsupervised algorithms. In a first part, a well-established clustering algorithm is repurposed for a novel application: physical-based domain decomposition. This is crucial for various applications, ranging from the shape optimization to the aerodynamic forces decomposition. The ML method demonstrated its ability to overcome the limitations of existing deterministic algorithms used to date. In a second part, autoencoders (unsupervised deep-learning algorithms) are employed for real-time predictions of flow fields around wing sections. The investigation on autoencoders, as presented here, begins with the assessment of their accuracy in predicting aerodynamic flow fields, particularly in strongly non-linear regimes. Subsequently, it addresses several challenges in ML techniques, including the interpretability of models, by performing extreme data compression, and the assessment of the uncertainty quantification of the autoencoder predictions. This part is pivotal for enabling the deployment of ML technologies in real-world systems. The ML algorithms investigated here are linked by a common underlying theme: the unsupervised extraction of field information and the inherent uncertainty embedded in this process. The applications proposed here should be intended as a proof of the significant impact that ML can have on aerodynamic analyses in the near future, unlocking the possibility to perform high-fidelity CFD simulations for industrial applications within reasonable timeframes and contained computational costs. Furthermore, it could open the door to real-time flow predictions with a satisfactory level of accuracy, enabling tasks like shape optimization, real-time flow control, and visualization using high-fidelity data.

Downloads

Actions (login required)

Modifica documento