National Technical University of Athens (NTUA)
Reduced order models and machine learning in analysis and optimum design of structures
Abstract
dc:descriptionSince the first structures where designed by human, finding the optimal design was a major problem. This increased as structural domain size and complexity also increased. Analysis of structures and its computational cost depends on the size and complexity of the structural system. The use of approximation methods increased due to the increase of computational cost of exact methods. Work on gradient-free algorithms, surrogate modeling methods, machine learning algorithms, topology optimization and generative design is presented in this dissertation. Applications of the above and combinations of them in analysis and design of structures are also presented. A new metaheuristic algorithm is proposed and tested while an improved version of Harmony Search Algorithm (HS) (firstly introduced by Zong Woo Geem) is also presented. Additionally, the application of different deep neural networks on accelerating topology optimization procedure is examined and four approaches are presented. The first one, DL-TOP uses a deep belief network to predict the final result of Solid Isotropic Material with Penalization (SIMP) method according to a few initial iterations. The second solution, DL-SCALE, focused on model upgrading, uses a deep belief network to accelerate topology optimization by sequentially providing close-to-final results from less dense discretizations to dense ones. The third methodology, DLRM-TOP, again uses a deep belief network to predict the final result of a topology optimization applied on a dense discretization by providing as input the results of significantly less dense discretizations. The final proposed methodology, CN-TOP, focuses on model enhancing by using a deep convolutional neural network to increase the density of a topology optimization result mesh and reduce necessary computational time. The formulated combinations of topology optimization method SIMP and deep learning methods proposed are modified and used as a AI-driven design methodology, DzAIN. A user can almost instantly produce a number of designs without any drawing input needed, apart from the domain and loading and support conditions. The performance of the above methodologies clearly states that deep learning methods and soft-computing can be a very powerful tool in accelerating procedures in analysis and optimal design of structures while it is also efficiently applicable in Generative Design.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kallioras, Nikolaos
- Καλλιώρας, Νικόλαος
Subjects
dc:subject × 22- Μοντέλα μειωμένης τάξης
- Μηχανική μάθηση
- Βελτιστοποίηση
- Ανάλυση κατασκευών
- Σχεδιασμός κατασκευών
- Βαθιά νευρωνικά δίκτυα
- Βελτιστοποίηση τοπολογίας
- Reduced order models
- Machine learning
- Optimization
- Structural analysis
- Structural design
- Deep neural networks
- Topology optimization
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Natural Sciences
- Computer and Information Sciences
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/46064