National Technical University of Athens (NTUA)
AI-enhanced multiscale finite element methods for forward and inverse uncertainty quantification problems in structural mechanics
Abstract
dc:descriptionOver recent decades, there has been growing interest in high-performance materials tailored for complex engineering applications. By modifying material structures at fine scales, exceptional properties such as enhanced mechanical strength, improved thermal conductivity, and novel optical features can be achieved. To address the time-consuming and costly experimentation on these materials, several computational techniques have been developed. Among them, the multiscale computational homogenization method via the well-established FE^2 algorithm has gained significant attention. Despite its computational intensity, this algorithm is favored for its ability to reliably predict the complex macroscopic behavior of multiscale material systems due to non-linear phenomena at finer scales. However, identifying the parameters that characterize material behavior at fine scales still remains a nontrivial undertaking. This thesis presents a cost-efficient framework using machine learning strategies for implementing the computational homogenization modeling approach on multi-query analyses investigating fine-scale parameters. Novel computational methodologies are proposed for accurate and efficient forward and inverse uncertainty quantification analyses on multiscale material systems and are validated through real-world case studies. First, this thesis presents a strategy for performing Bayesian inference on microscale material properties using experimental observations from the visible structure. To tackle the computational load of repeated FE^2 analyses, a feed-forward neural network (FFNN) is used to emulate material behavior affected by microstructural parameters. This is achieved by training the FFNN on a dataset from offline representative volume element (RVE) solutions. Next, the thesis generalizes the FE^2 algorithm by employing a sequence of FFNNs to represent different scales in the multiscale system, with each FFNN learning the constitutive law of its corresponding length scale. This results in a FFNN that emulates macroscopic behavior by incorporating mechanisms from each finer scale. Based on this scheme, the thesis, subsequently, proposes a methodology to identify optimal typologies of nanocomposite materials for desirable structural responses under uncertain conditions. Finally, a hierarchical Bayesian framework is introduced to utilize disjoint experimental measurements in multiscale material systems for joint parameter inference. This framework integrates experimental data from different scales and material compositions to yield informed parameters for future model predictions.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Pyrialakos, Stefanos-Christos
- Πυριαλάκος, Στέφανος-Χρήστος
Subjects
dc:subject × 18- Ανάλυση υλικών σε πολλαπλές κλίμακες
- Ποσοτικοποίηση αβεβαιότητας
- Υποκατάστατη μοντελοποίηση
- Μπεϋζιανή επικαιροποίηση
- Μηχανική μάθηση
- Νανοσύνθετα υλικά
- Multiscale material modeling
- Uncertainty quantification
- Surrogate modeling
- Bayesian inference
- Machine learning
- Nanocomposite materials
- Επιστήμη Πολιτικού Μηχανικού
- Επιστήμες Μηχανικού και Τεχνολογία
- Δομοστατική
- Civil Engineering
- Engineering and Technology
- Structural Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57574