We are happy to get Bachelor and Master students involved in our research. Here we present a collection of possible topics for Bachelor and Master theses, while courses are presented at the institute level. The below list of possible topics is not exclusive – feel free to contact us with own ideas!
Research can only make a difference if we talk about it. Therefore, we encourage and support all students to submit their reports and theses in the form of scientific publications. Some of our graduates have even successfully published in international journals or conference proceedings, as exemplified below:
In our knowledge base, we have collected information that might be become relevant for you while working on your project or thesis.
We speak better German, Spanish, Italian, Marathi, Hindi, Thai and Persian than this website suggests – feel free to contact us in our mother tongue!
Wind flow through streets and between buildings drives pedestrian comfort, ventilation, and heat stress in cities. Computational Fluid Dynamics (CFD) can simulate these flows accurately, but a single simulation can take hours to days. That is far too slow for interactive urban design or for optimization over many design variants. Therefore, surrogates are replacing the expensive CFDs in order to make optimization of space possible.
You will learn: Scientific machine learning, the basics of CFD and urban flow physics, working with large simulation datasets, evaluating surrogate models
Requirements: Solid Python skills and some experience with PyTorch (or willingness to learn it), basic knowledge of numerical methods or fluid mechanics (helpful but not required)
Related third-party projects: ReSpace!
Contact: Rebekka Walter Ich spreche Deutsch!
You can find parameters of standard materials in text books – or fit them to match your own data! The ability to distinguish material parameters and constitutive mechanisms strongly depends on the applied loading program. You will develop a general framework for designing loading histories that provide the most information about a predefined constitutive model library. The framework will initially be demonstrated for compressible generalized standard materials and can later be applied to constitutive model discovery.
You will learn: Constitutive modelling, parameter identifiability, sensitivity analysis, optimal experimental design, numerical optimization, and scientific programming in Python.
Requirements: Knowledge of continuum mechanics, Python programming skills, interest in numerical methods and optimization. Previous experience with machine learning is helpful but not required.
Related third-party projects: KIStoff
Contact: Atharva Potnis. I speak Marathi! मी मराठी बोलतो! I speak Hindi! मैं हिंदी बोलता हूँ !
The mechanical behavior of heterogeneous materials, such as composites, porous materials, and mechanical metamaterials, is strongly influenced by their underlying microstructures. Understanding these structure–property relationships is essential for predicting complex nonlinear phenomena, including plasticity, damage, buckling, and contact. However, high-fidelity numerical simulations are often computationally expensive, making the exploration of large microstructural design spaces challenging. This project focuses on developing machine-learning-based surrogate models, combined with active learning and uncertainty quantification, to accurately predict mechanical behavior using minimal simulation data.
You will learn: Scientific machine learning, computational mechanics, surrogate modeling of structure–property relationships, active learning, and uncertainty quantification.
Requirements: Python programming (PyTorch is a plus), basic understanding of machine learning fundamentals, fundamentals of solid mechanics and numerical methods.
Contact: Hooman Danesh I speak Persian! من فارسی صحبت میکنم! (Man Farsi Sohbat Mikonam!)
Designing materials with tailored mechanical properties is a central challenge in materials engineering. For example, the microstructures of composites and architected materials can be designed to achieve specific stiffness, strength, energy absorption, or nonlinear mechanical responses. However, identifying suitable microstructures among countless possible configurations is a challenging inverse problem, particularly when evaluating each design requires expensive numerical simulations. This project explores data-driven inverse design strategies, ranging from surrogate-assisted gradient-free optimization and Bayesian optimization to modern generative AI approaches, such as denoising diffusion models and flow matching, to efficiently discover and generate microstructures with desired mechanical properties.
You will learn: Inverse microstructure design, surrogate-assisted optimization, Bayesian optimization, gradient-free optimization, generative AI, diffusion models and flow matching.
Requirements: Python programming (PyTorch is a plus), basic understanding of machine learning fundamentals, fundamentals of solid mechanics and numerical methods.
Contact: Hooman Danesh I speak Persian!
You will extend the statFEM–EUCLID framework to identify plastic constitutive models from sparse and noisy displacement measurements. The statistical finite element method (statFEM) combines numerical predictions with experimental observations to reconstruct displacement fields and quantify uncertainty, while EUCLID enables automated constitutive model discovery. The aim is to extend this approach to path-dependent material behavior, accounting for irreversible deformation and loading history, and to validate it through numerical benchmarks under different loading conditions and measurement uncertainties.
You will learn: Computational plasticity, Bayesian data assimilation and uncertainty quantification, Constitutive model identification, Finite element modelling in Julia.
Requirements: Basic knowledge of continuum mechanics and finite element methods, Programming skills preferably in Julia or at least Python, Interest in inverse problems and material modelling.
Contact: Andrea Morini Parlo Italiano!
You will develop efficient neural operators to predict the history-dependent response of elastoplastic materials. You will explore combinations of physics-informed learning with reduced-order modeling to capture stress evolution and internal variables while maintaining equilibrium and constitutive consistency. Accuracy, computational efficiency, and generalization to unseen loading paths will be evaluated against finite-element simulations.
You will learn: Physics-informed operator learning, the fundamentals of plasticity, reduced-order modeling, and evaluating machine learning models against finite-element simulations
Requirements: Solid Python skills and some experience with PyTorch, basic knowledge of solid mechanics and numerical methods; familiarity with finite elements is helpful but not required
Contact: Hamidreza Eivazi I speak Persian!
You will investigate scalable transformer architectures for predicting physical fields on irregular meshes and point clouds. Local attention and multiscale communication will be explored to capture both fine-scale features and long-range interactions efficiently. The focus is on large point sets, memory and runtime efficiency, and generalization across mesh resolutions and geometries.
You will learn: Transformer architectures and attention mechanisms, learning on meshes and point clouds, working with large simulation datasets, evaluating accuracy and computational scalability
Requirements: Solid Python skills and basic experience with PyTorch, familiarity with neural networks and linear algebra; experience with scientific computing
Contact: Hamidreza Eivazi I speak Persian!
You will develop a two-stage probabilistic framework for identifying material parameters from sparse measurements. A generative model reconstructs full-field displacements and their uncertainty. An inference model then combines these reconstructions with geometry, loading conditions, and reference simulations pairing known parameters with displacement fields, supplied as in-context examples. The aim is to infer parameter distributions for new materials without retraining, while propagating reconstruction uncertainty through the identification process.
You will learn: Generative modeling with flow matching or diffusion models, in-context learning, inverse problems in solid mechanics, probabilistic parameter identification
Requirements: Solid Python skills and basic experience with PyTorch, familiarity with neural networks and basic probability and statistics; knowledge of solid mechanics or finite-element methods is helpful but not required
Contact: Hamidreza Eivazi I speak Persian! and Narunat Pantapalin I speak Thai ฉันพูดภาษาไทย!
The mechanical behavior of solid materials is typically described using constitutive models whose mathematical structure is prescribed in advance. However, selecting an appropriate model can be challenging, especially for materials exhibiting complex nonlinear behavior. Data-driven constitutive model discovery offers an alternative by identifying the structure of the model from data while ensuring consistency with mechanics principles. In this project, you will explore how continuum mechanics and sparse optimization techniques can be combined to discover physically admissible constitutive models.
You will learn: constitutive modeling, continuum mechanics, sparse optimization, model identification, analysis of experimental data, scientific programming.
Requirements: Knowledge of continuum mechanics or solid mechanics, Python programming skills, interest in numerical methods and optimization. Previous experience with machine learning is helpful but not required.
Contact: Jorge-Humberto Urrea-Quintero Hablo español!
Neural networks can approximate complex material behavior, but conventional architectures do not automatically satisfy fundamental physical principles. Physics-Augmented Neural Networks (PANNs) address this limitation by incorporating principles of continuum mechanics and thermodynamics directly into their architecture. In this project, you will investigate how physically constrained neural networks can be trained using mechanical data to obtain reliable constitutive models. You will explore how neural network architecture, physical constraints, and optimization influence the accuracy and consistency of the identified material response.
You will learn: Physics-augmented machine learning, neural network architectures, constitutive modeling, neural network training.
Requirements: Knowledge of solid mechanics, Python programming skills, interest in machine learning and numerical methods.
Contact: Jorge-Humberto Urrea-Quintero Hablo español!
Experimental measurements and numerical models are subject to uncertainties that can significantly affect predictions of material behavior. Quantifying these uncertainties is essential for assessing the reliability of computational models. In this project, you will investigate Bayesian approaches for identifying and updating constitutive models from data, as well as data assimilation methods such as the statistical finite element method (statFEM). Particular attention will be given to evaluating whether the quantified uncertainties are well calibrated and whether the resulting predictions are consistent with available observations.
You will learn: Uncertainty quantification, Bayesian inference, constitutive model calibration, data assimilation, uncertainty calibration and validation, scientific programming.
Requirements: Knowledge of solid mechanics, Python programming skills, interest in probability theory, statistics, and numerical methods. Previous experience on uncertainty quantification methods is desired.
Contact: Jorge-Humberto Urrea-Quintero Hablo español!