Fast flow prediction along airfoils operating at transonic conditions using Machine Learning

Ρεκούμης, Κωνσταντίνος; Rekoumis, Konstantinos

dc.contributor.author	Ρεκούμης, Κωνσταντίνος	el
dc.contributor.author	Rekoumis, Konstantinos	en
dc.date.accessioned	2023-03-03T10:02:08Z
dc.date.available	2023-03-03T10:02:08Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/57197
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.24895
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc/3.0/gr/	*
dc.subject	Airfoils	en
dc.subject	Machine Learning	en
dc.subject	CFD	en
dc.subject	Optimization	en
dc.subject	CNN	en
dc.subject	Μηχανική Μάθηση	el
dc.subject	Αεροτομές	el
dc.subject	Διηχητικές ροές	el
dc.subject	Βελτιστοποιήση	el
dc.subject	Υπολογιστιή Ρευστομηχανική	el
dc.title	Fast flow prediction along airfoils operating at transonic conditions using Machine Learning	en
dc.title	Γρήγορη Πρόβλεψη διηχητικής ροής γύρω από αεροτομή με την χρήση τεχνικών Μηχανικής Μάθησης	el
heal.type	bachelorThesis
heal.classification	Computational Fluid Dynamics (CFD)	en
heal.classification	Machine Learning	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2022-02
heal.abstract	This Thesis’s main focus is how we can couple Fluid Dynamics with Artificial Intelligence. This work is not an original idea as many other scientists have already explored this field, yet it is fascinating to further research and experiment as it shows immense potential. Based on the work of Hui et al., a system of Convolutional Neural Networks will be created where each Network would be responsible for predicting the Coefficient of Pressure distribution for an airfoil’s two sides (top and bottom sides). The airfoil of choice is the RAE-2822, operating under subsonic conditions close to the Critical Mach number. This way we can also study whether the Neural Network can predict the formation of sonic waves, phenomena with great mathematical and physical interest because of their extremely non-linear behavior. Also, an interesting concept used by Hui et al. is utilizing the Signed Dis- tance Function to colorize the input image’s pixels. Signed Distance Function enables us to describe more complex geometry with less image resolution. It achieves that by colorizing each pixel according to the distance information between the pixel’s center and its nearest geometry point. That leads to packing more information into fewer pixels by utilizing almost the entirety of available pixels. To train and test the Neural Networks, the RAE-2822 is randomly uni- formly deformed for deformation percentages ∈ [−20, 20] %, with 1000 spec- imens being used for training and 500 for testing. Each variant’s Cp distri- bution was calculated using the CFD solver MaPFlow. Then the Cp distri- bution’s values were extracted at specific length intervals for each side, thus creating a file storing these values for each side. This process, along with the creation of the SDF formatted images, constitute the data generation process. Then this data were used to train the Neural Network. After training the Networks, we study their accuracy in predicting the Cp distribution of previously unknown specimens and repeat the process for 4000 Epochs. Finally, the Networks’ error convergence per Epoch history would be studied both for the training and the testing sets. Also, the time per Epoch data as long as the single case prediction time data would be studied to validate the Networks’ prediction speed. Following the verification of the Neural Networks’ accuracy, we will study the influence of the minibatch size in the training-testing precision and speed. Also, we will examine the Networks’ precision in predicting Cp Distribution for airfoils out of the original training and testing range. Finally, a short and simplistic application of the trained Networks’ will be presented; a sim- ple geometry optimizer, whose purpose is to maximize the Lift capacity by optimizing the original RAE-2822 geometry for specific Free-flow conditions.	en
heal.advisorName	Παπαδάκης, Γεώργιος	el
heal.advisorName	Παπαλάμπρου, Γεώργιος	el
heal.committeeMemberName	Παπαδάκης, Γεώργιος	el
heal.committeeMemberName	Παπαλάμπρου, Γεώργιος	el
heal.committeeMemberName	Γρηγορόπουλος, Γρηγόρης	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ναυπηγών Μηχανολόγων Μηχανικών. Τομέας Μελέτης Πλοίου και Θαλάσσιων Μεταφορών	el
heal.academicPublisherID	ntua
heal.numberOfPages	131 σ.	el
heal.fullTextAvailability	false