Open Master's Thesis Positions
On this page you will find a selection of possible Master Thesis opportunities, some notified to us directly by the research groups of MEST Tutors and some listed on the SiROP database.
This list is not exhaustive, other Thesis projects might exist, please check the respective listings of Departments and research groups you are particularly interested in.
See also Internship opportunities.
Projects directly supplied by MEST Tutors
Master projects:
- Exploring global vs. local learning of products to inform smarter innovation policy at the Energy and Technology Policy Group
Projects from the SiROP Database
ETH Zurich uses SiROP to publish and search scientific projects. Here is a selection of projects currently available which may be suitable for MEST students. For more information visit external page sirop.org.
Safe and adaptive learning-based MPC: is it even possible?
Model Predictive Control (MPC) excels at handling complex constraints, but its real-world success relies heavily on highly accurate system models and meticulously tuned cost functions. While learning-based MPC addresses this by adapting to changing environments online, it notoriously lacks the reliability required for practical deployment, risking system instability. This project aims to bridge the gap between adaptability and safety by leveraging differentiable MPC, a cutting-edge framework that treats optimization solvers as differentiable functions. By utilizing gradient-based techniques to tune cost functions and constraints, you will develop a control scheme that learns dynamically while providing rigorous safety guarantees.
Show details
Evaluating Incentive-Driven Demand Response for Residents
Residential occupants shape both how much energy households use and when the grid is stressed. Demand-side management lets households shift flexible activities (showering, laundry, dishwashing) into windows of high renewable availability, cutting peaks, costs, and CO₂. But residents don't respond to grid signals on their own: it depends on how clearly information reaches them, how easy the action is, and whether it's rewarded. This thesis investigates how day-ahead building-simulation results can be turned into concrete, well-timed recommendations for residents at Empa's NEST living-lab, and whether modest incentives produce a measurable, lasting shift in behaviour.
Show details
Dynamic Performance Certification of Grid-Forming Converters Under Large Grid Disturbances
As renewable energy penetration increases, power grids are becoming weaker and less voltage-stiff, creating new stability challenges for grid-connected converters. Grid-forming converters offer a promising solution by supporting voltage, frequency, inertia, and fault response during major grid disturbances. However, their ability to deliver these services quickly and reliably depends strongly on converter control settings and grid conditions. This thesis will investigate the large-disturbance dynamic performance of grid-forming converters beyond conventional case-by-case simulations and simplified voltage-source models. The student will develop analytical performance indicators that capture the influence of individual control loops, grid strength, and key control parameters. Particular attention will be given to the tradeoff between fast dynamic response and system stability. The proposed analytical certification method will be validated in MATLAB/Simulink and applied to the design of improved grid-forming control strategies. The project offers an opportunity to work on a highly relevant topic in renewable-energy integration, combining power-system theory, converter control, analytical modelling, and simulation.
Show details
Image-based prediction of material constructions for urban building energy simulation
Urban Building Energy Models (UBEMs) often require detailed inputs on building envelopes to produce accurate outputs. However, this information is typically neither stored nor readily accessible. This thesis aims to augment energy models with material constructions predicted from visual data and building register data using deep learning techniques. The project will focus on a case study in Zurich in collaboration with the City of Zurich.
Show details
Image-based prediction of window-to-wall ratio for urban building energy modeling
Urban Building Energy Models (UBEMs) often require detailed inputs on building envelopes to produce accurate outputs. However, this information is typically neither stored nor readily accessible. This thesis aims to augment energy models with WWR information obtained from visual data using deep learning techniques. The project will focus on a case study in Zurich in collaboration with the City of Zurich.
Show details
Self-consuming Generative Models: a Gaussian Analysis
Modern generative models are increasingly trained on web-scale datasets that contain content produced by earlier models. When this process is repeated over several generations, errors in the synthetic component can accumulate and the learned distribution can drift away from the original data distribution. This phenomenon, known as model collapse, can appear as bias in the mean, shrinking variance, or loss of low-probability modes. This thesis studies a concrete intervention: at each training generation, a curator may inject a limited amount of verified real data. The goal is to decide when and how much real data to inject so that model drift remains small over a finite horizon while respecting a data-acquisition budget. The project combines stochastic optimal control, optimal transport, probability, and numerical experimentation.
Show details
Research Internship - Machine learning and digital twins
Our research group works on increasing the efficiency of robotics systems in real-world deployment by
enabling quick data collection, calibration, and policy training while ensuring safety and efficiency. For
this, we develop novel learning-based control and policy optimization techniques.
We're looking for a skilled machine learning (ML) engineer to develop cutting-edge AI algorithms for
digital twin applications in real-world industrial systems. The digital twin applications will include process
monitoring, estimation, analysis, optimization, and control policy development. The project requires
hands-on expertise in ML, time-series forecasting, modelling, and control of industrial plants.
Show details
Thermal and Structural Analysis of 3DP Building Components
Additive manufacturing has shown great potential in the building industry, enabling the fabrication of complex, customisable geometries that were previously impossible with conventional fabrication. Combining the advancement of this technology with ceramics, a fundamental building material with exceptional properties of durability, thermal stability and strength, functional properties of building components can be enhanced. This research investigates the potential of optimising the thermal and structural properties of ceramic 3D-printed components through a parametric design workflow that explores parameters of form, thickness and infill geometry.
Show details
Organic 3D-Printable Paste Development for Ceramic DIW
Clay can be shaped and formed through a wide range of techniques, and a subsequent sintering process transforms it into ceramic, a stiff and durable material well-suited to building components. In Direct Ink Writing (DIW), clay must retain a certain water content to enable extrudability for printing. After printing, the part passes through a drying stage in which water is lost, causing shrinkage, and is then sintered at high temperatures. The shrinkage and warping that occur during drying and sintering introduce dimensional deviations, which are problematic for the tolerances required in building components. Therefore, this research develops an organic composite paste to be co-extruded during the 3D-printing process, which aims to reduce the drying and sintering shrinkage of the resultant ceramic component.
Show details
Physics-Consistent Lifelong Learning and Adaptive Control for Energy-Efficient Ovens
Modern household ovens are expected to provide high cooking performance while minimizing energy consumption throughout their lifetime. However, the thermal characteristics of an oven gradually change due to aging effects such as insulation degradation, heating element wear, and sensor drift, causing conventional factory-calibrated models and controllers to lose accuracy over time. This thesis investigates a novel lifelong learning framework that enables ovens to continuously adapt to these changes by combining physics-consistent machine learning with adaptive control techniques. The project will develop thermodynamic models that incorporate known physical principles, employ Bayesian optimization for automatic calibration, and use online system identification to continuously update the model using operational data. The learned models will then be integrated into adaptive control algorithms that automatically adjust controller parameters to maintain temperature regulation accuracy while improving long-term energy efficiency. The proposed methods will be validated using simulation, industrial datasets, and available hardware at V-ZUG, with the possibility of deployment on a real oven. This project provides an opportunity to work at the intersection of control theory, machine learning, system identification, and industrial automation on a real-world problem with significant practical impact.
Show details
The frequency safety control of large-scale power electronics-dominated power systems
The increasing penetration of inverter-based renewable energy resources is transforming the frequency regulation paradigm of modern power systems. Reduced system inertia, distributed frequency-support resources, and emerging spatial frequency variations pose significant challenges to maintaining secure system operation. This project develops a distributed safety-critical control framework for low-inertia power systems based on Control Barrier Functions. The proposed approach aims to guarantee nodal frequency safety by ensuring that all bus frequencies remain within prescribed operational limits throughout transient evolution. By exploiting the sparse network structure of power systems, global frequency-safety requirements are decomposed into local safety conditions that can be enforced through distributed control actions and limited information exchange. The resulting framework combines rigorous safety guarantees with scalable real-time implementation, providing a promising solution for frequency-safe operation of future power-electronics-dominated power systems.
Show details
Probabilistic Forecasting for Predictive Maintenance
Predictive maintenance asks a simple question with high stakes: when should a machine be serviced before it actually breaks? Getting the answer wrong is expensive either way, since unplanned failures cause downtime while servicing too early wastes parts and labor. Despite decades of work, the problem is far from solved. Two difficulties stand out. First, sensor signals span very different timescales, with wear accumulating over months but the most informative measurements living at high frequency. Second, any maintenance decision is taken under genuine uncertainty about the future, and a safe decision requires that uncertainty to be made explicit. This thesis explores how to combine modern long-context sequence models, such as Transformers and state-space models like Mamba, with probabilistic forecasting techniques, to produce calibrated predictions of future machine health that can drive safer and more efficient maintenance decisions on standard public datasets.
Show details
Welfare based evaluation of autonomous vehicle integration in Zurich
The City of Zurich has identified several metric groups for assessing autonomous vehicle (AV) integration, including allocation of urban mobility space and time, safety, environment and livability, traffic performance and reliability, and equity and access. This project asks how these metrics should be turned into a clear evaluation framework for policy analysis.
Show details
Meta-Learning transformer policies for control of non-linear systems
What if, instead of designing a controller for one specific plant, we trained a single policy that already knows how to control an entire family of plants? This is the promise of meta-learning for dynamical systems: a policy is trained once, offline, on a randomized class of simulated systems, and is then deployed zero-shot on the real hardware, with no on-system tuning. Recent work has shown that this idea can be realized with Transformer architectures through a mechanism called In-Context Learning (ICL), and has been validated on a physical brushless DC motor.
This thesis builds on that line of work and proposes a decoupled architecture that explicitly separates system identification from control inside the policy. The framework will be developed in simulation, extended to nonlinear plants and generic reference trajectories, and validated experimentally in our labs.
Show details
Multi-Fidelity Bayesian Optimization with Parallel Information Sources
Many engineering and machine-learning problems boil down to the same question: which settings make this system work best? Tuning a controller, picking the hyperparameters of a neural network, or finding a good policy for a robot all require optimizing a performance metric whose value can only be measured by a slow, costly, or risky experiment. Bayesian Optimization (BO) is the standard tool for this setting, but it traditionally relies on a single, expensive source of information, even when cheaper approximations (simulators, reduced-order models, short training runs) are readily available.
In this Master's thesis you will design and implement a new BO framework that can leverage many sources of information in parallel, each with its own accuracy and computational cost, and that learns online which sources are worth querying and which to drop. The work is explicitly scoped to lead to a co-authored publication at a top ML venue.
Show details
Preferential Bayesian Optimization with Cost Priors, Constraints, and LLM-Assisted Operator Interfaces
Almost every problem in modern control and machine learning eventually reduces to minimizing a cost function. In practice this cost combines several competing objectives — tracking accuracy, energy consumption, safety margins — whose relative weights are usually chosen by hand, with little principled justification. Preferential Bayesian Optimization (PBO) offers a more systematic alternative: rather than fixing the cost in advance, the system shows the user two candidate outcomes, asks which one is preferred, and progressively learns a surrogate of the user's true preferences from these comparisons. The same mechanism underlies the alignment of modern large language models (LLMs) to human feedback. This thesis extends PBO along three directions: incorporating structured prior knowledge about the cost components, handling constraints on individual objectives, and coupling the framework with an LLM-based conversational interface, translating natural-language instructions into structured updates to the optimizer.
Show details
Safe and Performance-Aware Reinforcement Learning for Legged Robots
Reinforcement learning has become a powerful tool for robotic locomotion, including legged robots, humanoids, and wheel-legged systems. Modern simulators such as NVIDIA Isaac Sim and Isaac Lab make it possible to train policies with algorithms such as proximal policy optimization (PPO) in high-fidelity, GPU-accelerated environments. However, learned policies often rely on reward engineering and may exhibit poor transient behavior, limited robustness, or safety violations under disturbances, actuator limits, and model mismatch. This thesis will investigate how tools from nonlinear control theory can be combined with modern reinforcement learning to improve the reliability of learned locomotion policies. In particular, the project will study prescribed performance control (PPC), control Lyapunov functions
(CLFs), and control barrier functions (CBFs) as mechanisms for reward shaping and, where feasible, lightweight online action filtering. The main application will be wheel-legged robotic locomotion, with possible extensions to wheeled-biped or humanoid robots in Isaac Sim /
Isaac Lab.
Show details
Automatic Calibration Procedure for a 5-Axis Robotic 3D Printer
Additive manufacturing systems are typically operated in an open-loop fashion, where both motion and material extrusion are precomputed offline. This makes the process sensitive to disturbances such as geometric misalignments, material variations, and changes in process conditions. Accurate geometric calibration, such as bed leveling and multi-axis alignment, is therefore an important prerequisite for reliable and repeatable printing. This is especially relevant for 5-axis additive manufacturing systems, where additional rotational axes introduce further calibration challenges. This project focuses on improving the reliability of a custom-built 5-axis 3D printer at IfA by developing automated calibration and startup procedures.
Show details
Characterisation of past vertical extension projects in Zurich
To achieve ecological sustainability, urban densification must shift from carbon-intensive demolition toward alternative methods, such as vertical extensions. However, the lack of systematic mapping of past projects hinders evidence-based scaling-up of this approach. This thesis aims to bridge this gap by developing a data-driven pipeline to identify, characterise, and evaluate past projects. Using computational tools, the workflow integrates open data and satellite imagery to extract features of past projects. The resulting structured, geocoded database and quantitative and qualitative assessment framework will ultimately facilitate a rigorous analysis of implementation potential and the strategic scaling-up of urban densification practices.
Show details
Evaluating the use of thermal indicators for indoor heat stress assessment
As heat events intensify and become more frequent, concerns about the health impacts of indoor heat stress are growing. Yet, most indoor overheating studies still rely on dry-bulb temperature or comfort indices, with no consensus on which indicators are appropriate and health-relevant in buildings. This project evaluates a small set of candidate indicators for typical Swiss apartments and offices by combining building simulations with selected thermophysiological model outputs.
Show details
Feedback control for the first Swiss local energy markets
How can one safely control an electricity grid with multiple selfish stakeholders such as electric vehicles owners and solar panel owners, in real time? This project investigates this question for Walenstadt, a Swiss town where "the grid of tomorrow" is currently being created. The goal is to test PRIME, a recently proposed feedback market mechanism that controls the grid by providing economic incentives to the stakeholders, to drive their decision-making and achieve coordination. The tests are performed in simulation on a realistic model, but tests in the real grid of Walenstadt are a possibility if the simulations are successful.
Show details
Conceptual design of decentralised data centres in various urban building energy systems
The master’s thesis will conceptually design how decentralised data centre can be integrated and utilised within different types of buildings. There is a strong emphasis on achieving a technical solution that is both redundant and sustainable. Different types of buildings with different energy demands will be modelled to outline the integration of data centres.
Show details