Υποκατάστατα μοντέλα βασισμένα στη μηχανική μάθηση για ποσοτικοποίηση αβεβαιοτήτων στην υπολογιστική ρευστοδυναμική

Bakis, Dionysios; Μπάκης, Διονύσιος

dc.contributor.author	Bakis, Dionysios	en
dc.contributor.author	Μπάκης, Διονύσιος	el
dc.date.accessioned	2025-02-14T11:08:50Z
dc.date.available	2025-02-14T11:08:50Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/61114
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.28810
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc/3.0/gr/	*
dc.subject	Machine Learning	en
dc.subject	Optimization	en
dc.subject	Uncertainty Quantification	en
dc.subject	Computational Fluid Dynamics	en
dc.subject	Μηχανική Μάθηση	el
dc.subject	Βελτιστοποίηση	el
dc.subject	Ποσοτικοποίηση Αβεβαιότητας	el
dc.subject	Υπολογιστική Ρευστοδυναμική	el
dc.title	Υποκατάστατα μοντέλα βασισμένα στη μηχανική μάθηση για ποσοτικοποίηση αβεβαιοτήτων στην υπολογιστική ρευστοδυναμική	el
dc.title	Machine learning-based surrogate models for uncertainty quantification in CFD	en
heal.type	bachelorThesis
heal.classification	Machine Learning	en
heal.classification	Optimization	en
heal.classification	Uncertainty Quantification	en
heal.classification	Computational Fluid Dynamics	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2024-09-20
heal.abstract	This Diploma Thesis is in the area of Aerodynamic Shape Optimization under Un certainties, and, in particular, it explores the adequacy of Feedforward Fully Con nected Deep Neural Networks as surrogates of the high-fidelity, yet very expensive, Computational Fluid Dynamics solver for the task of quantifying the effects of un certainty (Uncertainty Quantification, UQ). This is the key process in search for a well performing design that is less sensitive to the presence of uncertainties (Robust Design), instead of just the best-performing design point. Uncertainties must be taken into account whenever optimization is carried out. For example, in Mechanical Engineering, minor fluctuations in model parameters can yield sub-optimal performances. In conjunction with the evolution of computing systems, the use of methods that take uncertainties into account increases the reli ability of the outcome of an optimization process. In most cases, UQ requires the computation of a stochastic model’s mean and vari ance. Here, UQ is carried out using Monte Carlo and Polynomial Chaos Expansion coupled with the fluid solver or a surrogate model for two aerodynamic problems involving transitional flows: an isolated airfoil and an isolated wing. The flows are simulated with the in-house PUMA software of the Parallel CFD & Optimiza tion Unit solving the Reynolds-Averaged Navier-Stokes equations along with the Spalart-Allmaras turbulence model, and the γ − Re˜ θ transition model. Cases with uncertainties related to coefficients that appear in the γ − Re˜ θ transition model are studied. The UQ methods, in general, require repetitive calls to the analysis code, which renders them prohibitively expensive, especially, whenever CFD software is involved. For this reason, the development of surrogate models aims to accelerate the optimization process by greatly reducing the computational cost. On the other side, surrogate models perform computations of lower fidelity than those obtained from the expensive CFD tool. This trade-off is being investigated in the two aerodynamic problems. The DNNs’ hyperparameters are tuned manually and also, by using evolutionary algorithms. Lastly, the involvement of two very promising techniques (Stacking Ensemble and Feature Selection) is examined, aiming to check how the prediction accuracy can further be improved as well as how surrogate models with fewer input features perform compared to those used so far.	el
heal.advisorName	Giannakoglou, Kyriakos	en
heal.committeeMemberName	Aretakis, Nikolaos	en
heal.committeeMemberName	Mathioudakis, Konstantinos	en
heal.committeeMemberName	Giannakoglou, Kyriakos	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Μηχανολόγων Μηχανικών. Τομέας Ρευστών. Εργαστήριο Θερμικών Στροβιλομηχανών	el
heal.academicPublisherID	ntua
heal.numberOfPages	112 σ.	el
heal.fullTextAvailability	false