Reduced order models and machine learning in analysis and optimum design of structures

Καλλίωρας, Νικόλαος

dc.contributor.author	Καλλίωρας, Νικόλαος	el
dc.date.accessioned	2020-07-24T07:46:28Z
dc.date.issued	2020-07-24
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/50967
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.18665
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	Μοντέλα μειωμένης τάξης	el
dc.subject	Μηχανική μάθηση	el
dc.subject	Ανάλυση κατασκευών	el
dc.subject	Βελτιστοποίηση	el
dc.subject	Βαθιά νευρωνικά δίκτυα	el
dc.subject	Reduced order models	en
dc.subject	Machine learning	en
dc.subject	Structural analysis	en
dc.subject	Βελτιστοποίηση τοπολογίας	el
dc.subject	Optimization	en
dc.subject	Deep Neural Networks	en
dc.subject	Topology optimization	en
dc.title	Reduced order models and machine learning in analysis and optimum design of structures	en
dc.contributor.department	Δομοστατικής	el
heal.type	doctoralThesis
heal.classification	Δομοστατική	el
heal.dateAvailable	2021-07-23T21:00:00Z
heal.language	en
heal.access	embargo
heal.recordProvider	ntua	el
heal.publicationDate	2019-04-18
heal.abstract	The main scope of the current dissertation is to contribute in the field of optimal structural analysis and design through developing and implementing new calculus methods by combining exact and approximation methods. As the size and complexity of analysis and design models is continuously increasing and the increase in the available computational power is not analogous, it is inevitable to investigate the exploitation of soft computing techniques. Exact methods have the ability to radically reduce the objective function of the problem in a relatively small number of iterations but lack the ability to overcome local minima. Actually, in real-life problems, it is more probable that the solution provided by exact methods will be a local minima than the global one. Additionally, the computational cost of calculating first and second-order derivatives or even the inability to calculate them resulted to examining the use of soft computing methods. In a short historical review, it can be witnessed that the use of soft computing techniques grew dramatically from 1970 to 1990. Applications of such methods in real life problems were performed often and attracted significant research interest. Derivative-free methods, because of the randomness included, have the ability to overcome local minima and, practically, guaranty finding the global minima if there is no limitation to population of iterative solutions. The main drawback of these methods is that a significant amount of iterations is almost always necessary. Due to the fact that objective function calculation also became computationally heavy in the following years, the necessity of multiple calculations, inevitable in soft computing, was an important drawback. The above resulted to a drop of interest in soft computing as models became too complex for these methods. In the last decade, research on modern soft computing methods drew a lot of attention, mainly because of the excessive amount of data generated, but not exploited, by companies offering on-line services. That led to generating new methods, able to handle large amounts of data and extremely complex problem definitions. For example, deep neural networks have been a major scientific breakthrough in modern research. The scope of this dissertation is to examine and apply modern soft computing methods in the fields of optimal analysis and design of structures. Some of the most computationally heavy problems dealt in structural design and analysis are the ones that include many solutions of the equation: {P} = [K] ∗ {U} as inversing the stiffness matrix is rather time and computational resources consuming. To deal with this issue, techniques of reducing the size of the stiffness matrix, or minimizing the necessary iterative solutions’ population or a combination of the previous two must be used. The core of this thesis is focused on providing new solutions in model order reduction and iterations population reduction in problems as the previously described ones. Additionally, parallel processing techniques are also examined in the solutions provided. Finally, a method for applying machine learning in the field of generative design is also presented.	en
heal.advisorName	Λαγαρός, Νικόλαος	el
heal.committeeMemberName	Γιαννάκογλου, Κυριάκος	el
heal.committeeMemberName	Κουμούσης, Βλάσης	el
heal.committeeMemberName	Προβατίδης, Χρήστος	el
heal.committeeMemberName	Πλεύρης, Ευάγγελος	el
heal.committeeMemberName	Χαρμπής, Δήμος	el
heal.committeeMemberName	Τσιάτας, Γεώργιος	el
heal.academicPublisher	Σχολή Πολιτικών Μηχανικών	el
heal.academicPublisherID	ntua
heal.numberOfPages	262 σ.	el
heal.fullTextAvailability	true