Open student projects

Reconstructing time-varying graph signals is a critical challenge in graph machine learning and graph signal processing, with various applications such as missing data imputation in sensor networks and time-series forecasting. Effectively addressing these tasks requires accurately capturing the spatio-temporal information inherent in these graph signals. Current methods, however, often rely exclusively on either graph signal processing methods or graph neural networks, limiting their ability to fully leverage the strengths of both approaches. This project aims to develop a new model for time-varying graph signal reconstruction, integrating the inductive bias from graph signal processing to graph neural networks.

Requirements:
1. Experience in Python and PyTorch
2. Familiar with graph machine learning and/or graph signal processing.

Keywords: Graph neural networks, graph signal processing, time-series reconstruction.

You can apply via SiROP or directly contact Zepeng Zhang.

This project explores how we can use physical and engineering knowledge about structures to better simulate them using neural networks. For some years, graph neural networks (GNNs) have been state-of-the-art methods for surrogates of finite element method (FEM) simulations. While there have been some approaches that use physics-informed loss functions, what the current models lack is the incorporation of physical knowledge directly into the model architecture. This is what we will do here; we will go one step beyond. The student will start from a state-of-the-art graph neural network and apply a physics/engineering inductive bias to improve the performance of the simulator and make it more stable. The goal is to evaluate stable rollouts and simulate dynamic oscillations of simple geometric structures, like composite plates.

Python programming and machine learning knowledge are required. Additionally, familiarity with GNNs or FEM is recommended. You can apply directly by contacting Kevin Steiner.

This project aims to develop an agent-based framework for fault diagnosis and root cause analysis in industrial systems. It focuses on overcoming the limitations of supervised approaches by enabling unsupervised fault diagnosis without relying on labeled failure data.

The approach combines Large Language Model (LLM) reasoning with digital twins to allow interactive and iterative analysis of system behavior. By leveraging counterfactual reasoning, the agent can simulate alternative operating conditions to better identify and isolate the origin of faults. This enables a shift from passive anomaly detection toward more interpretable and actionable diagnosis. Ultimately, the project seeks to advance autonomous, data-efficient, and explainable diagnostic systems.

Keywords: Fault Diagnosis, Root Cause Analysis, Unsupervised Learning, Large Language Models, Digital Twin.

You contact Amaury Wei or apply via IS-Academia (project #14319).

Having interpretable and stable simulations is important for industrial applications. Nowadays, machine learning methods for simulating material deformation can produce results on par with finite element method (FEM) simulators. However, the information we recover is often incomplete and bound to the parameters tracked in the dataset eventhough we could infer the missing information with known material laws. This is what we will do in this project; we will incorporate material laws into our neural network. The student will start from a state-of-the-art graph neural network and incorporate the materials constituitive relations into the model itself to accurately predict the material stress. The goal is to evaluate this physics-informed approach with respect to learning the stress as a distinct learning objective.

Python programming and machine learning knowledge are required. Additionally, familiarity with GNNs or FEM is recommended. You can apply directly by contacting Kevin Steiner.

• Background in Computer Science, Data Science, Electrical Engineering, Aerospace Engineering, Robotics, Mechanical Engineering, or a related field
• Strong programming skills in Python
• Familiarity with machine learning, data analysis, and signal processing
• Interest in anomaly detection, fault diagnosis, or predictive analytics
• Strong problem-solving and communication skills
• Knowledge of Graph Neural Networks (GNNs) is a plus.

The success of tabular foundation models across a wide range of real-world prediction tasks has demonstrated the potential of large-scale synthetic pre-training to improve performance on real-world tabular data. This project will investigate how such models can be pre-trained for prognostics and health management (PHM), where labeled failure data are scarce, degradation trajectories are only partially observed, and operating conditions vary substantially, using synthetic data generated from structural causal models (SCMs).

The central hypothesis is that synthetic pre-training data can encode PHM-specific biases, such as degradation trajectories, sensor noise, missingness, operating regimes, and failure-time structure. The student will design synthetic tabular datasets that reflect realistic PHM assumptions and use them to pre-train or adapt a tabular foundation model. The resulting models will be evaluated using our extensive PHM benchmarking framework across downstream tasks such as fault detection, remaining useful life estimation, and condition classification. A key research question is whether PHM-informed synthetic pre-training improves generalization compared with generic tabular pre-training or task-specific training only. The expected outcome is an empirical study and prototype pipeline for PHM-oriented tabular foundation model training.

Strong Python programming skills and a solid foundation in machine learning are required. Familiarity with structural causal models, transformer architectures, distributed multi GPU training, and Kubernetes is advantageous. You can apply directly by contacting Raffael Theiler.

Recent advances in agentic AI have demonstrated its strong potential for coordinating multi-step reasoning, tool use, scientific information extraction, and self-improving decision making. This project will develop a self-improving agentic AI system that extracts structured technical knowledge from prognostics and health management (PHM) research papers and connects it to an evaluation framework.

The goal is to move beyond simple paper summarization toward an executable, framework-based representation of scientific knowledge. The student will assist in scaling up and existing multi-step agent workflow for reading PHM papers, extracting candidate knowledge, checking consistency, and storing the results in a structured representation. This knowledge will then be linked to benchmark tasks and evaluation criteria, enabling more systematic comparison of scientific PHM methods. A central research question is how reliably agentic AI can transform unstructured scientific literature into reusable evaluation knowledge. The expected outcome is a prototype literature-to-evaluation pipeline for self-improving PHM research agents.

Strong Python programming skills and a solid foundation in machine learning are required. Familiarity with agentic AI, large language models, containerization, and Kubernetes is advantageous. You can apply directly by contacting Raffael Theiler.

Accurately predicting the remaining useful life of a battery requires understanding how degradation evolves under changing operating conditions. Current loads, charging and discharging schedules, and rest periods vary throughout service, making future degradation dependent on how the battery will be used. This semester project aims to develop Neural Stochastic Differential Equations (Neural SDEs) that learn battery degradation dynamics and predict capacity evolution across successive cycles under specified future operating conditions.

The student will use NASA’s ProgPy battery simulator to generate lifetime trajectories under variable current load profiles and prescribed degradation laws. They will develop a model that infers battery health from current and voltage observations and learns how its degradation rate evolves. The model will then forecast capacity and remaining useful life under alternative future operating scenarios, such as a transition from mild to demanding use. A central research question is whether modelling stochastic degradation dynamics improves prediction under changing operating conditions compared with a deterministic Neural ODE model. Evaluation will focus on previously unseen operating schedules, including changes in load intensity, duration and ordering.

Requirements: Experience in Python and PyTorch and a solid foundation in machine learning are required. Familiarity with time-series modelling, dynamical systems or stochastic differential equations is advantageous.

You can apply directly by contacting Hugo Walter.

Recent advances in neural scene reconstruction, such as Neural Radiance Fields (NeRF) and Gaussian Splatting, have significantly improved the performance of downstream tasks, including novel view synthesis and geometric reconstruction. Building on these innovations, multi-modal approaches have been explored to incorporate additional scene attributes such as depth, surface normals, thermal data, and semantic information to enrich existing scene representations. However, current multi-modal methods often rely on a tightly coupled correspondence between RGB data and other modalities, which limits their applicability in uncontrolled, real-world scenarios.

Description:
This project aims to address the challenge of strong cross-modal dependencies in 3D reconstruction, with a particular focus on RGB and thermal imaging. The goal is to develop robust techniques that can establish correspondences between different modalities, or align image features in three-dimensional space, thereby enabling accurate reconstruction without the need for strict pixel-level alignment across modalities.

Objectives:
 – Conduct a comprehensive literature review on multi-modal 3D reconstruction techniques, with an emphasis on robustness and cross-modal integration
 – Investigate contrastive learning approaches for identifying correspondences between RGB and non-RGB modalities
 – Explore strategies for 3D-space feature alignment and rendering-based pose optimization
 – Extend the reconstruction framework to incorporate additional modalities, including semantic labels and inputs with weak or no geometric correlation

Keywords: Multi-modal fusion, Visual-based 3D reconstruction, thermal imaging

You can apply by directly contacting Chenghao Xu.

Most deep models are trained under a close-world assumption and hence do not possess knowledge of what they do not know, leading to over-confident and inaccurate predictions for unknown objects. Out-of-distribution (OOD) detection, which aims to identify unknown objects, has attracted much attention as it plays a vital role in a variety of applications such as industrial inspection, medical diagnosis, and autonomous driving. One of the primary challenges of effective OOD lies in the wide-existing domain shift in the real world. The testing dataset might have a significant distribution from the training data of deep learning models. Thus, it is even more difficult for the model to distinguish between the domain difference and the OOD objects. In this task, the student is expected to work on the OOD detection task with domain shift. Specifically, we focus on test-time domain adaptation, which focuses on OOD detection during inference time online. The potential application case can be found here: https://www.epfl.ch/labs/cvlab/data/road-anomaly/. Overall, the goal of this project is to improve the OOD performance on new domains through the use of test-time domain adaptation techniques. The student is expected to undertake the following tasks in this project: – explore current test-time domain adaptation methods – improve those methods and modify them for OOD detection.

You can apply via SiROP or directly contact Han Sun.