Master Theses

Simulating Structural Dynamics with Graph Neural Networks

Supervisors: Gregory Duthé (), Kostas Vlachas (), Prof. Dr. Eleni Chatzi

Content
Graph Neural Networks (GNNs) have recently emerged as a powerful way to learn the dynamics of a physical process. Using standard meshing techniques, it is easy to decompose a physical problem, such as the dynamical simulation of a structure into a discrete graph, which a GNN can then operate on. In this thesis, we investigate the use of GNNs as alternative structural dynamics simulators. We will explore operations on graphs (incl. message passing) and their equivalence to sub-structuring and condensation schemes. It will also be important to include as many physics-based inductive biases as possible.

The student(s) will have the following objectives:

• Review the existing methods and literature
• Become familiar with the PyTorch Geometric framework
• Implement a first prototype, based on existing work from the literature
• Formulate methods for sub-structuring/condensation schemes which exploit the expressivity of GNNs
• Demonstrate the method on a simple case study
  

Suggested Courses: 101-0157-01L Structural Dynamics and Vibration Problems

Suggested Competencies: Python, PyTorch (or equivalent ML framework)

Digital twins – a small-scale demonstration

Supervisors: Dr. Yves Reuland (), Dr. Panagiotis Martakis, Prof. Dr. Eleni Chatzi

Content
Many existing structures are nearing or passed already the service life they were designed for. With increasing traffic loads, the evolution of design codes, and material ageing engineers are tasked with assisting challenging decisions related to extending the lifetime of existing structures. The evaluation of existing bridges relies on computational models. Such representations of real structures, often termed as digital twins, should reflect real behavior as closely as possible and therefore, ideally involves monitoring data for calibration. Unvalidated modelling assumptions may lead to unrealistic evaluations, due to various sources of uncertainty and systematic errors:
• When real safety is lower than the model prediction, the assessment is insufficient and possibly unsafe.
• In case the real safety is higher, the assessment is overconservative and translates to unnecessary and expensive retrofit activities

With the use of a simple experimental structure, the entire chain of structural health monitoring, from measuring, through model calibration, until behavior prognosis will be addressed and validated. Such a holistic approach introduces the student into the necessary steps to develop a robust and accurate digital twin of an existing structure. Master thesis on monitoring-driven digital twinning of structures allows students to increase and deepen their knowledge of key aspects of structural dynamics, structural health monitoring, numerical modelling, uncertainties, and assessment of existing structures; through following the following steps:
- Design a monitoring campaign for a simple experimental specimen
- Develop numerical models of the experimental specimen
- Calibrate the numerical model based on the vibration monitoring of the specimen
- Estimate the uncertainties pertaining to the measurement and the model
- Study the influence of the amount and location of sensors on the model calibration
- Predict the structural behavior under new loading scenarios

Suggested Courses: Seismic Design

Finite-Element Updating Using Physics-Informed Neural Networks

Supervisors: Dr. Marcus Haywood-Alexander Marcus (), Prof. Dr. Eleni Chatzi

 

Content

Many high-value structural systems (such as bridges or wind turbines) monitor vibration data as a means of detecting faults within the structure or system. Modal characteristics of such systems can act as direct indicators of changes in local physics properties, and finite-element-model updating provides an approach to identify these properties on a finer resolution. With the explosion of data availability and data-driven machine-learning techniques, these have been used to drive such complex models. However, they require a large amount of data to adequately capture the model of modal characteristics to physical properties. Physics-informed are a popular new class of physics-enhanced machine learning algorithms, which utilise known physics in the form of ordinary/partial differential equations to promote the machine learning algorithm towards physically-informed solutions. By including the physical-properties as learnable parameters in training, PINNs can be used in a system-identification task.

The aim of this thesis would be to study the application of combining finite-element models with physics-informed neural networks to estimate local physical properties of a dynamic structure. In this thesis, the student will be expected to:

• Review the current state-of-the-art in literature
• Become familiar with physics-informed neural networks, and apply them to simple problems
• Develop a method for combining finite element models and physics-informed neural networks (FE-PINNs)
• Study the use of FE-PINNs in the context of novelty detection via model-updating, applying them to data from simulated or experimental data

Suggested Courses: Structural Identification and Health Monitoring, course on machine learning methods

Suggested Competencies: Python/MATLAB

Improving the design of rotating parts of wind turbines through metamaterial technology

Supervisors: Dr. Vasilis Dertimanis (), Prof. Dr. Eleni Chatzi

Industry Partners: Dr. Luca Sangiuliano, Phononic Vibes

Content

In the context of novel lightweight materials for noise and vibration reduction, Phononic Vibes Srl (https://phononicvibes.com/) is researching, designing and producing metamaterial technology. Metamaterials are often periodic structures, consisting of a host structure with embedded or added resonant structures, leading to superior vibro-acoustic attenuation performance in some tunable frequency ranges, called stop bands. In the following videos, the metamaterial concept is applied structure-borne (https://www.youtube.com/watch?v=DUee93HcPVQ&t) and airborne noise problems.

The resonant structures often have some small features that control their tuned frequency and the dynamic behavior of the entire metamaterial system. Once these resonant structures are applied to rotating systems, the stop band behaviour can be modulated according to the rotation speed to the entire system thanks to the presence of inertial forces (ref. 1). This interesting phenomenon can give rise to new methods to control the dynamic behavior of rotating systems and to reduce the unwanted dynamics.

Rotating parts of wind-turbines are often subjected to extreme dynamic conditions. To ensure the long service life and optimal design of rotating parts of wind-turbines, thus preventing failures and expensive maintenances, the dynamic behaviour of these parts must be thoroughly designed and controlled.

The goal of this master thesis is then to investigate the potential of metamaterial technology applied to the rotating parts of wind turbines to control and preserve their dynamic behaviour. In particular, the student will explore how to exploit the modulation effect given by the inertial forces to tune the metamaterial system to reduce critical dynamic conditions. Analytical and numerical models will be developed by the student to study the physics and to assess the metamaterial performance applied to such systems. The aim is to propose design guidelines to achieve a robust design solution. 

This thesis is in collaboration with Phononic Vibes Srl and is mainly simulation-based.

Suggested Competencies: Python/MATLAB

Advanced Data-Driven Methods for Infrastructure Monitoring and Maintenance Optimization

Supervisors: Dr. Charikleia Stoura, Prof. Dr. Eleni Chatzi 

Industry Supervisor: Dr. Giulia Aguzzi, KISTLER Instrumente AG

Content

Effective monitoring of civil infrastructure is vital for ensuring long-term safety, serviceability, and cost-efficiency. In recent years, the proliferation of multi-sensor systems and the availability of high-fidelity vibration, acceleration, and Weigh-In-Motion (WIM) data have opened new avenues for developing intelligent monitoring frameworks. These advancements present an opportunity to transition from periodic inspections to predictive maintenance regimes, enhancing asset management strategies in transport infrastructure.

This Master’s thesis aims to develop and benchmark modern AI-assisted methodologies for the analysis of real-world monitoring datasets from bridges and rail infrastructure, with data access supported by KISTLER. The work will focus on synthesizing information from various sources, including vibration signals, tilt measurements, and WIM data, to derive actionable insights into the condition and behavior of structures under operational loading.

The scope includes:

• Developing pipelines for data fusion and pre-processing of multi-sensor time series.
• Performing comparative analyses of existing Structural Health Monitoring (SHM) methods and fault detection strategies.
• Evaluating the added value of WIM data in augmenting structural response interpretation and load estimation.
• Exploring machine learning and signal decomposition techniques (e.g., modal analysis, dictionary learning, supervised/unsupervised ML) for enhanced pattern recognition.
• Investigating tilt estimation through acceleration measurements, with potential use in low-cost sensor deployment scenarios.
• Applying insights to optimize rail surface monitoring and recommend maintenance strategies.
  

This interdisciplinary thesis connects experimental data analysis, system identification, and data-driven modeling with strong industrial relevance. Students will have the opportunity to work with cutting-edge sensing technologies and contribute toward the digital transformation of infrastructure maintenance.

Hybrid Modeling for Twinning and Structural
Health Monitoring in Engineered Systems

Supervisors: Dr. Konstantinos Vlachas (), Prof. Dr. Eleni Chatzi 

Content

Ensuring the safe, efficient, and resilient operation of structural assets throughout their life cycle requires the integration of sensing data with computational models that are both rapid to evaluate and reliable. This hybrid paradigm is particularly critical for systems subject to uncertainty, evolving conditions, or extreme events that may push their response beyond design assumptions. This thesis will focus on developing and validating hybrid representations that combine physics-based and data-driven models for structural response prediction, uncertainty quantification, and real-time model adaptation. The aim is to demonstrate the potential of such hybrid formulations in advancing digital twin frameworks for the optimized operation and maintenance of real-life structures. Depending on their background and interests, students may work on one or more of the following tasks:

Expected project outcomes

1. Development of reduced-order but reliable models for dynamic response prediction
2. Integration of sensing and vibration data with physics-based models
3. Implementation of uncertainty quantification and adaptive updating strategies
4. Real-time data assimilation and model adaptation under varying operating conditions
5. Experimental validation of hybrid digital twin approaches

Required Competencies

Structural mechanics/dynamics, Finite Element Modelling
Computer programming experience will be beneficial

Software Knowledge

• Proficient in Python (or MATLAB) coding

Exploring Applications of Physical Computing to
Hybrid Modeling for Twinning and StructuralCondition and Structural Health Monitoring

Supervisors: Dr. Marc Serra Garcia (AMOLF), Dr. Bart Van Damme (Empa), Dr. Andrea Bergamini (Empa),Prof. Dr. Eleni Chatzi (ETH)

Content

Advances in numerical
modeling, of linear and non-linear mechanical metamaterials, as well as machine
learning, are leading to an exciting new area of mechanics: Physical signal
processing. The fundamental principle in physical signal processing consists in
using the natural dynamical response of a Micro-Electromechanical System (MEMS)—for example, its resonance characteristics, to
detect specific acoustic signatures with near-zero power. As an example, recent
work has demonstrated the ability of a MEMS device to discriminate between
different spoken words presented to it, and correctly categorize them. The
implications of this capability are wide-ranging; from the development of
passive voice-activated devices that activate with specific keywords, guaranteeing the privacy of users, to the classification of complex acousticsignals based on specific fingerprint features.

This master thesis project, will build a bridge between physical signal processing and condition monitoring of composite structures by establishing the functional relationship between acoustic emission signals from damage event in composites and classification capabilities of MEMS physical computing devices.  The ultimate goal of this effort will be the fabrication of very low cost devices that can be distributed on structures as a ‘smart dust’ and will provide information about their state, if remotely interrogated.

Expected project outcomes

1. Literature survey of acoustic signals from typical damage events in composite structures, both in laboratory scale and real-life structures. 
2. Familiarize with and analyze actual acoustic emission data, as found for example on https://data.4tu.nl OR https://data.mendeley.com/datasets/svz439j8hb/1 to determine fundamental quantities such as duration, power spectral density to identify basic design features of such signals. 
3. Set up the environment for the in-silico training and validation of a mechanical device for the identification of available events.

If time permits:

1. Collect signals from tests on laboratory scale samples under controlled conditions representative of actual damage events and non-significant events (such as Hsu-Nielsen test) and pre-process them.

1. Train the network on collected data.

Required Competencies

Structural mechanics/dynamics, Finite Element Modelling, Computer programming experience will be beneficial

Software Knowledge

• Proficient in Python (or MATLAB) coding