Deep neural networks and their differentiation for use in
gradient-based optimization in single- and multi-phase flows

Κεφαλούκου, Ειρήνη Σωτηρία; Kefaloukou, Eirini Sotiria

dc.contributor.author	Κεφαλούκου, Ειρήνη Σωτηρία	el
dc.contributor.author	Kefaloukou, Eirini Sotiria	en
dc.date.accessioned	2026-01-26T08:43:35Z
dc.date.available	2026-01-26T08:43:35Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/63263
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.30958
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	Βελτιστοποίηση	el
dc.subject	Νευρωνικά Δίκτυα	el
dc.subject	Υπολογιστική Ρευστοδυναμική	el
dc.subject	Optimization	en
dc.subject	Deep Neural Networks	en
dc.subject	Computational Fluid Dynamics	en
dc.title	Deep neural networks and their differentiation for use in gradient-based optimization in single- and multi-phase flows	en
heal.type	bachelorThesis
heal.classification	Optimization	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2025-07-07
heal.abstract	Target of this diploma thesis is the use of Deep Neural Networks (DNNs) in gradient-based Shape Optimization (ShpO), as low-cost surrogates of the primal and adjoint computations, thus reducing the optimization overall cost. The DNNs are trained on databases containing both the objective function values and their corresponding Sensitivity Derivatives. The networks architecture is inspired by the notion of the Hermite polynomials, since besides the objective function, incorporate gradients in their training process. The computation of gradients is accomplished by differentiating the networks outputs with respect to their inputs, using automatic differentiation in reverse mode. The accuracy of the computed gradients is verified against reference values of the adjoint method. Improving the networks generalization capabilities and reducing the cost of constructing their database is also investigated. Primary goal of this diploma thesis is the integration of the Hermite-DNNs in the gradient-based ShpO, so as to provide an approximation to the objective function values and derivatives. During the ShpO, the DNN-optimized solutions are reevaluated on the Computational Fluid Dynamics code. If necessary, this design is incorporated in the database and the networks are re-trained. All implementations are related to ShpO studies in single- and multi-phase flows. Two turbomachinery applications are presented, the first concerns a turbine blade-airfoil (single-phase turbulent flow), and the second a compressor blade-airfoil (single-phase turbulent flow). The proposed gradient-based optimization algorithm is implemented also in the design of an isolated airfoil (single-phase transitional flow) and a hemispherical-cylinder body (two-phase cavitating flow).	en
heal.advisorName	Γιαννάκογλου, Κυριάκος	el
heal.committeeMemberName	Μαθιουδάκης, Κωνσταντίνος	el
heal.committeeMemberName	Αρετάκης, Νικόλαος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Μηχανολόγων Μηχανικών. Τομέας Ρευστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	115 σ.	el
heal.fullTextAvailability	false