ELLIS Excellence Fellowships: apply now for a project!

High-dimensional Bayesian Optimization (in collaboration with ASMPT)

Supervisor(s) at ASMPT 
Name: Richard van der Stam
Email: richard.van.der.stam [at] asmpt.com
Website: https://alsi.semi.asmpt.com/en/about-alsi/center-of-competency-nl/ 
Company department (if applicable): R&D AI/Vision Manager, ASMPT Center of Competency the Netherlands

Contact and/or supervisor(s) at Radboud University

Name: Mahyar Shahsavari
Email: mahyar.shahsavari [at] donders.ru.nl
Website: https://www.ru.nl/en/people/shahsavari-m
Department: Artificial Intelligence / Donders Institute

Faculty: Social Sciences

Project Description

Background/Motivation:

In complex manufacturing processes, such as semiconductor production, the development of a new machine or process may take experienced engineers several weeks to several years. Machine learning methods can help accelerate this cycle by using machine data to guide R&D teams toward machine improvements or optimal process parameters. Bayesian optimization is particularly attractive because it can optimize expensive black-box objective functions in a sample-efficient manner. However, many real-world industrial applications involve a large number of controllable variables, resulting in high-dimensional optimization problems where conventional methods struggle due to the curse of dimensionality and limited experimental budgets. This motivates the need for high-dimensional Bayesian optimization methods that can efficiently explore complex parameter spaces while remaining practical for industrial use.

Research Question/Goal:

We are interested in exploring various ways of applying modern High-dimensional Bayesian optimization algorithms to real-world problems.

Method:

Main methodological directions for high-dimensional Bayesian optimization include:

• Embedding / dimensionality reduction methods: Project the high-dimensional search space into a lower-dimensional latent space where optimization is more efficient [1].
• Structured or sparse variable selection methods: Identify the most relevant variables or interactions and focus optimization on those dimensions [2].
• Local or trust-region methods: Restrict optimization to promising local regions, avoiding global search over the full high-dimensional space [3].

Student Requirements:

Experience with/ Interest in: Bayesian Optimization, Process Optimization, Gaussian Processes

Expected time frame: 6 Months

Literature:

1. Saves, Paul. "High-dimensional Bayesian optimization using both random and linear embeddings." AIAA Journal (2025).

2. Eriksson, David, and Martin Jankowiak. "High-dimensional Bayesian optimization with sparse axis-aligned subspaces." Uncertainty in artificial intelligence. PMLR, (2021).

3. Eriksson, David, et al. "Scalable global optimization via local Bayesian optimization." Advances in neural information processing systems 32 (2019).

Model Compression of Pre-Trained Encoders (in collaboration with ASMPT) 

Supervisor(s) at ASMPT 
Name: Richard van der Stam
Email: richard.van.der.stam [at] asmpt.com
Website: https://alsi.semi.asmpt.com/en/about-alsi/center-of-competency-nl/ 
Company department: R&D AI/Vision Manager, ASMPT Center of Competency the Netherlands

Contact and/or supervisor(s) at Radboud University

Name: Mahyar Shahsavari
Email: mahyar.shahsavari [at] donders.ru.nl
Website: https://www.ru.nl/en/people/shahsavari-m
Department: Artificial Intelligence / Donders Institute
Faculty: Social Sciences

Project Description

Background/Motivation:

In the semiconductor industry, high production speed is a critical value proposition for chip manufacturers. This puts additional speed requirements on industrial AI models, such as segmentation or defect detection models. 

In this project we focus specifically on pre-trained foundation models available in the literature. These models are pre-trained on large quantities of data, but usually the models are large and slow. Therefore, we would like to compress these models while maintaining the similar model performance. Another important aspect is that these models are usually trained on large quantities of data that we do not have access to. Therefore, model compression is not straightforward as we must use our own or publicly available data to do model compression.

Research Question/Goal:

How can we compress these large vision foundation models to fit our strict speed requirements while maintaining similar performance?

Method:

We want to focus on self-supervised learning methods, such as DINOv3 [1]. These authors themselves propose model compression techniques which they apply but all work on the assumption that the pre-training dataset is available. Secondly, methods exist that adjust do domain adaptation of pre-trained weights [2]. 

Student requirements

Experience with / interest in: Computer Vision, Deep Learning, Self-Supervised Learning, Model Compression

Expected time frame: 6 months

Literature (max. 5)

1. Siméoni, Oriane, Huy V. Vo, Maximilian Seitzer, et al. “DINOv3.” arXiv:2508.10104. Preprint, arXiv, August 13, 2025. https://doi.org/10.48550/arXiv.2508.10104.

2. Pariza, Valentinos, Mohammadreza Salehi, Gertjan Burghouts, Francesco Locatello, and Yuki M. Asano. “NeCo: Improving DINOv2’s Spatial Representations in 19 GPU Hours with Patch Neighbor Consistency.” arXiv:2408.11054. Preprint, arXiv, August 20, 2024. http://arxiv.org/abs/2408.11054.

Visual (and non-visual) Path Planning (in collaboration with ASMPT)

Supervisor(s) at ASMPT 
Name: Richard van der Stam
Email: richard.van.der.stam [at] asmpt.com
Website: https://alsi.semi.asmpt.com/en/about-alsi/center-of-competency-nl/ 
Company department: R&D AI/Vision Manager, ASMPT Center of Competency the Netherlands

Contact and/or supervisor(s) at Radboud University

Name: Mahyar Shahsavari
Email: mahyar.shahsavari [at] donders.ru.nl
Website: https://www.ru.nl/en/people/shahsavari-m
Department: Artificial Intelligence / Donders Institute
Faculty: Social Sciences

Project Description

Background/Motivation:

In the semiconductor industry, we deal with a large variety of visual inspection tasks. For example, the detection of particles in the manufacturing process, inspection of the bonding process to verify no mistakes are made during or after bonding and setting up the software on the machine such that all the vision algorithms work properly. These tasks in large happen during the setup of the machine in the production line and are very labour intensive. 

We want to explore different path planning techniques to help us inspect and automate these tasks by path planning. Examples include: During the setup of the machine these methods can steer the process, or during the bonding of dies these methods can predict the outcome of the process.

Research Question/Goal:
What methods can we apply to automate the setup of our machines or predict the outcome of bonding processes by predicting next steps in the process.

Method:

For the methods we want to focus on self-supervised learning methods as recent advancements in these methods are very promising. These methods produce encoders that can be re-use without any or much finetuning.

There are different methods which re-use these pre-trained encoders and repurpose them for path planning directly on the latent space [2, 3]. These types of methods are also called world models [4] and use a predictor that predicts the next state for path planning. In many cases they use image encoders to represent the visual input [5], however video encoders are also becoming more and more powerful [1] to potentially replace image encoders.

Student requirements

Experience with / interest in: Computer Vision, Deep Learning, Self-Supervised Learning

Expected time frame: 6 months

Literature (max. 5)

1. Mur-Labadia, Lorenzo, Matthew Muckley, Amir Bar, et al. “V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning.” arXiv:2603.14482. Preprint, arXiv, March 17, 2026. https://doi.org/10.48550/arXiv.2603.14482.

2. Zhou, Gaoyue, Hengkai Pan, Yann LeCun, and Lerrel Pinto. “DINO-WM: World Models on Pre-Trained Visual Features Enable Zero-Shot Planning.” arXiv:2411.04983. Preprint, arXiv, February 1, 2025. https://doi.org/10.48550/arXiv.2411.04983.

3. Wang, Ying, Oumayma Bounou, Gaoyue Zhou, et al. “Temporal Straightening for Latent Planning.” arXiv:2603.12231. Preprint, arXiv, March 12, 2026. https://doi.org/10.48550/arXiv.2603.12231.

4. Maes, Lucas, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. “LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels.” arXiv:2603.19312. Version 1. Preprint, arXiv, March 13, 2026. https://doi.org/10.48550/arXiv.2603.19312.

5. Siméoni, Oriane, Huy V. Vo, Maximilian Seitzer, et al. “DINOv3.” arXiv:2508.10104. Preprint, arXiv, August 13, 2025. https://doi.org/10.48550/arXiv.2508.10104.

World Models for Industrial Use Cases (in collaboration with ASMPT) 

Supervisor(s) at ASMPT 
Name: Richard van der Stam
Email: richard.van.der.stam [at] asmpt.com
Website: https://alsi.semi.asmpt.com/en/about-alsi/center-of-competency-nl/ 
Company department: R&D AI/Vision Manager, ASMPT Center of Competency the Netherlands

Contact and/or supervisor(s) at Radboud University

Name: Mahyar Shahsavari
Email: mahyar.shahsavari [at] donders.ru.nl
Website: https://www.ru.nl/en/people/shahsavari-m
Department: Artificial Intelligence / Donders Institute
Faculty: Social Sciences

Project Description

Background/Motivation:

World models represent a recent trend in the AI literature, in which systems shift from reactive data processing to proactive planning and causal reasoning by explicitly simulating how the environment will develop over time given a system’s outputs. Among others, these concepts allow self-driving cars to better estimate safety risks, and allow longer-term planning for control tasks such as in reinforcement learning. A competent world model can also generate synthetic training data for other downstream tasks via ‘imaginative’ what-if reasoning, and this capability increases its sample efficiency relative to model-free alternatives.

In the semiconductor industry, AI applications range from visual defect inspection to data-driven machine control tasks. Unlike in tasks like self-driving cars or game-playing agents, however, it is much harder to gather data at scale in our industry, among others due to cost and IP concerns. Simulations are also not trivial to create for tasks occurring at the micro- or even nanoscale. The combination of higher sample efficiency, the promise of enhanced robustness, and better long-term planning, makes world models appealing to our industry as well.

Research Question/Goal:

How well do world models transfer to real-world use cases in the semiconductor industry?

Method:

We will begin the project by evaluating prominent methods in the literature, such as Dreamer [1][2], on benchmark open-source control tasks such as a simulated industrial robot arm. We plan to compare it to a more standard model-free reinforcement learning approach, such as PPO[4].

After our initial investigation to familiarise ourselves with the topic, in discussion with the student, we will deepen the topic towards a more specific research question. Options include:

• Focusing on a specific industry-relevant property of the relevant methods, such as their sample efficiency or safety/robustness
• Training agents based on offline data collected by a different agent, vs. the online ‘learning by doing’ approach
• Finding industrial use cases of world models beyond direct use inside a RL/control task

Student requirements

Experience with / interest in: Deep Learning, Reinforcement Learning

Expected time frame: 6 months

Literature (max. 5)

1. Hafner, D., Lillicrap, T., Ba, J., & Norouzi, M. (2019). Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603.

2. Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2023). Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104.

3. Zhou, G., Pan, H., LeCun, Y., & Pinto, L. (2024). DINO-WM: World models on pre-trained visual features enable zero-shot planning. arXiv preprint arXiv:2411.04983.

4. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.

Visual Anomaly Detection

Supervisor(s) at ASMPT 
Name: Richard van der Stam
Email: richard.van.der.stam [at] asmpt.com
Website: https://alsi.semi.asmpt.com/en/about-alsi/center-of-competency-nl/ 
Company department (if applicable): R&D AI/Vision Manager, ASMPT Center of Competency the Netherlands

Contact and/or supervisor(s) at Radboud University

Name: Mahyar Shahsavari
Email: mahyar.shahsavari [at] donders.ru.nl
Website: https://www.ru.nl/en/people/shahsavari-m
Department: Artificial Intelligence / Donders Institute
Faculty: Social Sciences
 

Project Description

Background/Motivation:

In the manufacturing of semiconductor devices, it is important to automatically detect any problems that could hinder the production process. One of these problems is foreign material blocking the bonding process. Therefore, we want to make use of visual anomaly detection methods that can automatically detect this foreign material. Over time many machine settings potentially drift, such as the lighting or temperature. 

We want to explore ways to improve our anomaly detection, by for example adjust online for these drifts, or make our vision encoders more robust to drifts in the machine. For example, we could create a method that models the lighting drift during the process such that the model accounts for this drift. We could also capture the directions relating to lighting and orthogonalize those, such that our features are robust to lighting changes.

Research Question/Goal:

What methods can we apply to make our anomaly detection methods more robust to various types of drifts we measure during the manufacturing process?

Method:

The methods such as PatchCore and SubspaceAD are good examples of anomaly detection methods that are relevant to our domain. Both depend on pre-trained encoders to do anomaly detection. We are interested in using these pre-trained encoders to model the measured drift over time, or find ways of using these encoders more effectively to improve our anomaly detection [3].

Student requirements

Experience with / interest in: Computer Vision, Deep Learning, Anomaly Detection
 

Expected time frame: 6 months

Literature (max. 5)

1. Roth, Karsten, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. “Towards Total Recall in Industrial Anomaly Detection.” arXiv:2106.08265. Version 2. Preprint, arXiv, May 5, 2022. http://arxiv.org/abs/2106.08265.

2. Lendering, Camile, Erkut Akdag, and Egor Bondarev. “SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling.” arXiv:2602.23013. Preprint, arXiv, February 26, 2026. https://doi.org/10.48550/arXiv.2602.23013.

3. Assran, Mido, Adrien Bardes, David Fan, et al. “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.” arXiv:2506.09985. Preprint, arXiv, June 11, 2025. https://doi.org/10.48550/arXiv.2506.09985.

Learning Geographic Routing Constraints for Electricity Grid Expansion with Graph Neural Networks (in collaboration with Alliander)

Supervisors 

Name: Yuliya Shapovalova, Gerson Foks
Email: yuliya.shapovalova [at] ru.nl (yuliya[dot]shapovalova[at]ru[dot]nl), gerson.foks [at] alliander.com
Website:
Department: Data Science, Institute for Computing and Information Sciences
Faculty: Faculty of Science

Project Description

Background/Motivation: Alliander, one of the Netherlands' primary distribution system operators, manages over 90,000 km of electricity grid and is facing a growing backlog of grid connection requests, with waiting times of up to ten years. Meeting the energy transition requires rapid physical expansion of the grid, but expansion planning today relies heavily on manual expert knowledge to navigate geographic constraints: cables should follow roads, avoid protected nature areas (Natura2000), and cannot cross open water without justification. These constraints are partly codified and partly tacit, embedded in decades of engineering decisions visible only in the existing cable topology.

Recent work within this group (Cambier van Nooten et al., 2025) showed that a GIN-inspired GNN can perform n-1 contingency analysis on Alliander's medium-voltage grid roughly 1,000× faster than the mathematical-optimisation baseline while generalising to unseen grid structures. That work, however, treats the grid as a purely topological object: geography enters only implicitly through existing cable placements. The present thesis extends the foundation from reliability assessment to geographically-constrained expansion planning, asking whether a GNN can learn the geographic feasibility rules that shape where new cables can actually go.

Research Question/Goal:

Main question: Can a graph neural network learn geographic routing constraints for electricity-grid expansion directly from existing grid topology and geospatial context layers, and use them to propose feasible routes for new connections?

Three sub-questions structure the work: 

1. Representation. How should heterogeneous geographic context (roads, Natura2000 areas, waterways) be encoded in a spatial graph representation of the medium-voltage grid? Possibly to be compared 1) rasterised node features, 2) auxiliary constraint-graph layers in a heterogeneous-GNN setup, and 3) coordinate-aware positional encodings (Klemmer et al., 2023). 
2. Constraint recovery. Does a GNN trained on existing cable routes implicitly learn which geographic features act as routing constraints, and does this generalise to regions held out during training? Existing placements encode tacit engineering rules; this sub-question tests whether those rules can be recovered from data alone. 
3. Route proposal. Can the learned representation propose feasible routes for new cable connections, and how do these routes compare against a) a least-cost-path raster baseline as used in traditional GIS transmission planning (Monteiro et al., 2005; Shu et al., 2012) and b) Alliander domain experts on held-out real connection cases?

Method:

Data. Grid topology from Alliander: nodes with (x, y) coordinates connected by lines composed of cable segments and connectors (also with (x, y)). Geographic context layers (Dutch road network (NWB), Natura2000 polygons, and waterways) aligned to the same coordinate system.

Model. The starting point is the GIN-inspired architecture of Cambier van Nooten et al. (2025), extended along two axes:

1. Spatial awareness. Nodes receive coordinate-aware embeddings via PE-GNN-style positional encoders (Klemmer et al., 2023) so that the model can relate topologically distant assets that are geographically close.
2. Heterogeneous context. Geographic constraint layers are injected either as per-node features (rasterised nearest-distance-to-feature), as an auxiliary context graph connected to grid nodes, or as heterogeneous edge attributes in the spirit of De Jonge et al. (2025).

Task formulation. Expansion is framed in two stages:

1. Edge feasibility scoring: for any candidate node pair, predict whether a direct connection is geographically feasible. Trained on existing cable segments as positive examples, with spatially matched negatives (e.g., candidate segments that would cross Natura2000 or open water), following the SEAL subgraph-labelling framework (Zhang & Chen, 2018).
2. Route proposal: given a source and target, construct a path. Baseline approach: greedy/beam search over the learned edge-feasibility scorer. Stretch approach: a reinforcement-learning agent over a heterogeneous multi-graph with action masking, following MetroGNN (Su et al., 2024), which is the closest published analog in the transport-infrastructure domain.

Student requirements

Experience with / interest in: Experience with / interest in: Python and PyTorch (or PyTorch Geometric / DGL); basic graph theory and deep learning; some exposure to geospatial data (GeoPandas, shapely, rasterio) is a plus but can be picked up; interest in the energy transition and applied ML.

Type of project (master thesis, internship (only for companies)): Master thesis (Alliander–Radboud collaboration).

Literature (max 5)

1. Cambier van Nooten, C., van de Poll, T., Füllhase, S., Heres, J., Heskes, T., & Shapovalova, Y. (2025). Graph neural networks for assessing the reliability of the medium-voltage grid. Applied Energy, 384, 125401.

2. Klemmer, K., Safir, N., & Neill, D. B. (2023). Positional Encoder Graph Neural Networks for Geographic Data. AISTATS 2023. arXiv:2111.10144.

3. Zhang, M., & Chen, Y. (2018). Link Prediction Based on Graph Neural Networks (SEAL). NeurIPS 2018. arXiv:1802.09691.

4. Su, H., Zheng, Y., Ding, J., Jin, D., & Li, Y. (2024). MetroGNN: Metro Network Expansion with Reinforcement Learning. Companion Proceedings of the ACM Web Conference 2024. arXiv:2403.09197.

5. De Jonge, M., Viebahn, J., & Marot, A. (2025). Generalizable Graph Neural Networks for Robust Power Grid Topology Control. arXiv:2501.07186.

Generative Models for Medium-Voltage Subgrids Synthetic Data Augmentation for Distribution-Grid AI (in collaboration with Alliander)

Supervisors 

Name: Yuliya Shapovalova, Gerson Foks
Email: yuliya.shapovalova [at] ru.nl (yuliya[dot]shapovalova[at]ru[dot]nl), gerson.foks [at] alliander.com
Department: Data Science, Institute for Computing and Information Sciences
Faculty: Faculty of Science
 

Project Description

Title: Generative Models for Medium-Voltage Subgrids: Synthetic Data Augmentation for Distribution-Grid AI

Alliander's medium-voltage grid contains tens of thousands of nodes and edges. For most modelling tasks it is sliced into roughly 180 feeder- or substation-level subgrids, each consisting of one substation and multiple middenspanningsruimtes (MS rooms). That sample size is the core bottleneck for data-driven methods: modern graph neural networks are typically trained on datasets one to three orders of magnitude larger, and recent empirical work on synthetic distribution-grid generation reports that even well-tuned graph VAEs struggle on hundreds-to-thousands of realistic feeders, producing disconnected components and repeated motifs on the more complex open benchmarks (Exploring VGAEs for Distribution Grid Data Generation, 2025).

Two things change if we can generate representative synthetic subgrids. First, downstream AI tasks (n-1 reliability assessment (Cambier van Nooten et al., 2025), load-flow surrogates, state estimation, data correction) become tractable with larger, better-mixed training sets. Second, because real grid topology is confidential, a high-quality generator opens a realistic sharing channel with academic partners: students and research groups can work on representative data without access to the real network. An internal intern project at Alliander made an early attempt at generative subgrid modelling but was constrained by data availability; the field has moved substantially since (DiGress and diffusion-based graph generators in 2023–2024; dedicated power-grid generators such as FeederGAN and DeepGDL), so a focused revisit is well-timed.

Research Question/Goal:

Main question: Can a deep generative model, trained on Alliander's 180 real subgrids (augmented with rule-based synthetic data), produce medium-voltage subgrids that are simultaneously 1) statistically representative of the real population, 2) diverse rather than memorised, and 3) useful as augmentation data for downstream grid-AI tasks?

Three sub-questions structure the work: 

1. Fidelity and diversity in the small-data regime. Which class of graph generative model, autoregressive (GraphRNN-style), one-shot adversarial (FeederGAN-style), latent-variable (VGAE / DeepGDL), or denoising-diffusion (DiGress), best balances statistical fidelity against sample diversity when trained on ~180 real subgrids plus rule-based synthetic data? Pretraining on the rule-based generator and fine-tuning on real data is tested as a transfer strategy.
2. Domain constraints. How much does adding power-grid-specific structural priors to the loss or architecture (radiality (tree structure per feeder), connectivity, expected average node degree, substation counts) improve sample validity, compared to a purely data-driven generator? Approaches include soft penalties, post-hoc rejection sampling, and constrained decoders (as in constrained VAEs for molecule design).
3. Downstream usefulness. Does augmenting the training set with generated subgrids improve performance on a concrete downstream task, specifically n-1 contingency prediction using the GIN architecture of Cambier van Nooten et al. (2025) as the target, and at what mixing ratio of real to synthetic data? This is the decisive test: a generator that produces pretty-looking graphs but does not help downstream tasks has not succeeded.

Optional stretch directions, picked in consultation with the supervisor after the baseline results: conditional generation (size, substation type, region), inclusion of node and edge attributes (coordinates, cable type, lengths) alongside topology, and joint generation of topology and operating conditions following PowerGrow (2025).

Method:

Data. Roughly 180 real Alliander medium-voltage subgrids (one substation + multiple MS rooms per subgrid) and an internal rule-based generator that produces less-representative synthetic subgrids in unlimited quantity.

Models (baseline to advanced). Following the taxonomy of Guo & Zhao (2022):

1. Autoregressive baseline: GraphRNN (You et al., 2018). Well-understood reference point, scales to hundreds of nodes, widely-used benchmark metrics.
2. Adversarial baseline tailored to this domain: FeederGAN (Liang et al., 2021), the closest published work, designed precisely for distribution feeders.
3. Latent-variable baseline: VGAE in the DeepGDL / 2025-VGAE-study tradition, which is a natural starting point when sample counts are small.
4. Advanced model: DiGress (Vignac et al., 2023), discrete denoising diffusion, current state-of-the-art on non-molecular graph generation, supporting categorical node and edge attributes and conditional generation.

Training strategy. The rule-based generator is used for 1) pretraining with domain adaptation to real data, and 2) creating a larger controlled benchmark on which the models can be debugged before being trained on the scarce real set. Cross-validation uses leave-out groups of subgrids rather than random splits.

Evaluation. Three complementary layers:

1. Distributional fidelity: Maximum Mean Discrepancy on degree, clustering, orbit, and spectral distributions (the GraphRNN benchmark), plus MMD on edge-attribute and node-attribute distributions.
2. Domain validity: rate of generated subgrids that are connected, radial, have plausible feeder depth, and match expected substation/MS-room cardinalities. This is the "did it generate a believable grid?" check.
3. Downstream utility: train the n-1 GIN from Cambier van Nooten et al. (2025) on real-only vs. real + synthetic data; measure held-out prediction accuracy as a function of the mixing ratio. This is the evaluation that the generative-modelling literature often skips and that gives the thesis a crisp, decision-relevant result.
4. Diversity vs. memorisation: nearest-neighbour distance in graph-edit-distance or WL-kernel space between each generated sample and its nearest training sample, to verify the model is not copying the 180 real graphs.

Student requirements

Experience with / interest in: Python and PyTorch; solid grounding in deep learning and at least one family of deep generative models (VAE, GAN, or diffusion); willingness to learn Graph Neural Networks and a modest amount of power-systems vocabulary on the fly. Prior knowledge of the electrical grid is not required, as the project abstracts away detailed electrical calculations and focuses on topology and attributes.

Type of project (master thesis, internship (only for companies)): Master thesis (Alliander–Radboud collaboration).

Literature (max 5)

1. Guo, X., & Zhao, L. (2022). A Systematic Survey on Deep Generative Models for Graph Generation. IEEE Transactions on Pattern Analysis and Machine Intelligence. arXiv:2007.06686.

2. Liang, M., Meng, Y., Wang, J., Lubkeman, D. L., & Lu, N. (2021). FeederGAN: Synthetic Feeder Generation via Deep Graph Adversarial Nets. IEEE Transactions on Smart Grid, 12(2), 1163–1173.

3. Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V., & Frossard, P. (2023). DiGress: Discrete Denoising Diffusion for Graph Generation. ICLR 2023. arXiv:2209.14734.

4. You, J., Ying, R., Ren, X., Hamilton, W. L., & Leskovec, J. (2018). GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models. ICML 2018. arXiv:1802.08773.

5. Cambier van Nooten, C., van de Poll, T., Füllhase, S., Heres, J., Heskes, T., & Shapovalova, Y. (2025). Graph neural networks for assessing the reliability of the medium-voltage grid. Applied Energy, 384, 125401.

Active Bayesian quadrature for automated model comparison

Supervisors 
Name:                  Max Hinne
Email:                    max.hinne [at] donders.ru.nl (max[dot]hinne[at]donders[dot]ru[dot]nl) 
Website:              https://www.ru.nl/personen/hinne-m 
Department:      MLNC
Faculty:                Social Science

Project Description

Background/Motivation:

Bayesian automated scientific discovery attempts to learn the structure of a model from data [1]. Such automated scientific discovery procedures have resulted in breakthroughs in, for example, genomics, molecular biology, computational psychology, and materials science. 

Under the hood, this means the procedure evaluates a large number of Bayesian models, and estimates the probabilities of these models under the data. This is very computationally intensive, and we need clever approximation techniques to make this scalable to larger model spaces and data sets.

In this project, you contribute to this development. You will explore the idea of active Bayesian quadrature [2] for automated model comparison. In a nutshell, Bayesian quadrature estimates the fit of a single model with the smallest number of model evaluations possible. Here, we try to instead find the smallest number of model evaluations to find the *relative difference between a pair of models*. Ideally, by smartly selecting where in parameter space to evaluate likelihood of the models, we can determine that model 1 is more likely than model 2, without fully exploring either model’s parameter space. 

Research Question/Goal: How can Bayesian quadrature be adapted to find the minimum number of likelihood evaluations for accurate Bayesian model comparison?

Method:

The approach starts by implementing an existing Bayesian quadrature baseline [3] followed by extending this idea to estimating the Bayes factor, that encodes the relative fit of two models. We evaluate the new approach by comparing with other model fit estimation procedures that are already implemented in our efficient bamojax (‘Bayesian modelling in JAX’) library [4], and see how the approach performs in terms of the accuracy in Bayes factor estimates and the number of required model evaluations.

The resulting method can be used in Bayesian automated scientific discovery across all empirical domains where many models must be compared efficiently.

Student requirements

Experience with / interest in:

• A strong interest in Bayesian computational modelling and scientific discovery.
• Experience with efficient implementation of AI algorithms.
• Prior experience with modern coding frameworks such as JAX or PyTorch are a plus.

Type of project: Master thesis

Expected time frame: Approx. 45 ECTS (depending on Master thesis requirements of the programme the student is in)

Literature (max 5)

1. Hey T, Butler K, Jackson S and Thiyagalingam J, Machine learning and big scientific data. Phil. Trans. R. Soc. 2020;A.37820190054 http://doi.org/10.1098/rsta.2019.0054.

2. Osborne, M. A. et al.Active learning of model evidence using Bayesian quadrature. Advances in Neural Information Processing Systems 1, 46–54 (2012). https://proceedings.neurips.cc/paper/2012/hash/6364d3f0f495b6ab9dcf8d3b5c6e0b01-Abstract.html

3. Chai, H. & Garnett, R. Improving Quadrature for Constrained Integrands. in Proceedings of Machine Learning Research vol. 89 (Naha, Okinawa, Japan, 2019). https://proceedings.mlr.press/v89/chai19a.html

4. Hinne, M. Bamojax: Bayesian Modelling with JAX. Journal of Open Source Software, 2025;10(114), 8642, https://doi.org/10.21105/joss.08642. 

AI-driven scheduling application for IC/MC nurses

Supervisors

Technical supervision
Name: Silvan Quax (member ELLIS Unit Nijmegen)
Email: silvan.quax [at] radboudumc.nl (silvan[dot]quax[at]radboudumc[dot]nl)
Website: https://rtc.diagnijmegen.nl/
Department: Department of Medical Imaging (Radboudumc)

Faculty: Faculty of Medical Sciences

Clinical supervision

Name: Laura Ros

Email: laura.ros [at] radboudumc.nl (laura[dot]ros[at]radboudumc[dot]nl)

Department: Intensive Care C1b (Radboudumc)

Name: Femke Heutink

Email: femke.heutink [at] radboudumc.nl (femke[dot]heutink[at]radboudumc[dot]nl)

Department: Intensive Care C1b (Radboudumc)

Project Description

Background/Motivation:

In the high-stakes environment of Intensive Care (IC) and Medium Care (MC) units, the roster is more than just a calendar. It is the backbone of patient safety and staff well-being. However, the current process for creating these schedules has reached a breaking point. What was once a manageable administrative task has evolved into a complex mathematical puzzle that exceeds human cognitive limits.

The difficulty lies in the sheer scale and the web of interdependencies within the IC/MC units. Every shift represents a collision of competing requirements: legal labor laws, diverse contract types, varying clinical experience levels, and personal availability. When a single piece of information—such as a late absence or a shift change—is introduced, it creates a domino effect across the whole roster. Because these variables are often submitted late or incomplete, the planning office is forced into a cycle of reactive ‘firefighting’ that results in an unstable schedule that requires constant, manual revisions.

The main problem is that manual scheduling can find a functional roster, but it rarely finds an optimal or fair one. In a unit of this size, there are trillions of ways to assign shifts. A human planner, overwhelmed by the volume of data, naturally relies on any solution that works rather than the one that is best for the team. As highlighted by Hamster (2025), manual scheduling inherently leads to inequality. Without an algorithmic solution, certain units or individuals systematically shoulder a heavier burden of socially demanding shifts, such as nights and weekends. This imbalance directly erodes staff vitality, fuels frustration, and compromises the long-term sustainability of the workforce.

To move beyond these manual limitations, this project proposes the development of an artificial intelligence-based (AI) scheduling system. The goal is to create a model capable of processing the many interdependencies, while balancing the institutional needs with individual preferences. It has already been demonstrated by Hamster (2025) that automated scheduling has potential, and this has been recognized by nurses as well (Gerlach, 2025). The transition to an automated, algorithmic approach is not merely a matter of convenience. It’s necessary to ensure a fair, predictable, and future-proof workplace for those providing critical care.

Research Question/Goal:
The primary goal is to develop an AI model that generates fair, predictable, and sustainable rosters while accounting for staff preferences and staffing requirements. Key research questions include:

• To what extent can an AI model combine individual preferences with staffing requirements without violating hard constraints (e.g., labor laws)?
• Which optimization techniques provide the best balance between fair distribution and overall schedule quality?
• How do nurses perceive the fairness and workload of AI-generated schedules compared to manual methods?
• Is the developed model technically scalable and reproducible for other medical departments?

Method:

Data and constraint formalization

This research will utilize one or more years of scheduling data from the Radboudumc IC/MC units to develop and validate a scheduling model. The technical framework begins with the formalization of constraints, the distinction between hard constraints (e.g., working hours act and mandatory staffing ratios) and soft constraints (e.g., individual preferences). The model must ensure legal compliance while maximizing staff satisfaction. This technical framework is consistent with solutions proposed by Hamster (2025) and Zhang et al. (2025). 

Machine learning model

Mixed Integer Linear Programs (MILP) are commonly used due to their capacity to guarantee mathematically optimal solutions under strict constraints (Hamster, 2025). However, these traditional models frequently struggle with scalability and flexibility. As the complexity of the IC/MC unit increases (e.g., increase in staff), the computational time required for MILP solvers grows exponentially. This renders them too rigid to efficiently handle late absences, sudden roster disruptions, or dynamic operational changes. AI and Machine Learning (ML) techniques offer a solution to these limitations by enabling rapid, adaptive, and data-driven scheduling. Methods such as reinforcement learning (Nagayoshi & Tamaki, 2022) and machine learning-augmented constraint programming (Ben Said & Mouhoub, 2026) can process vast amounts of historical roster data to implicitly learn complex staff preferences, hidden constraints, and fairness patterns that are otherwise difficult to hard-code. Furthermore, AI systems excel at sequential decision-making, allowing planners to dynamically adjust and optimize schedules in real-time without needing to recalculate the entire mathematical puzzle from scratch.

Evaluation

The model's performance will be evaluated with a retrospective analysis. The algorithm will be used to reconstruct schedules for specific historical periods, using the exact staff availability and constraints present at those times. These generated rosters will then be benchmarked against the original manual schedules using quantitative metrics for fairness and preference fulfillment. Furthermore, the model will be stress-tested with several scenarios—such as simulated peak holiday demands—to evaluate its resilience and ability to maintain an equitable distribution of labor. This comparative analysis will demonstrate the model’s capacity to provide a more stable and transparent scheduling logic than current manual practices.

Student requirements

Experience with / interest in:

• Knowledge of ML-based optimization algorithms and/or mathematical modeling
• Programming experience with Python and popular data science libraries (e.g., pandas)
• Interest in healthcare logistics and operational management

Type of project: 
Master thesis

Expected time frame: 

The expected time frame for the project is 6 months. Before the project starts, we make sure that (1) data already has been extracted and anonymized; and (2) an overview of the input variables and hard- and soft constraints are available for the student.

Literature

1. Hamster, M. (2025). The Rhythm of the Roster. Master thesis, University of Twente. Available at: https://purl.utwente.nl/essays/108598
2. Gerlach, M., Renggli, F. J., Bieri, J. S., Sariyar, M., & Golz, C. (2025). Exploring nurse perspectives on AI-based shift scheduling for fairness, transparency, and work-life balance. BMC nursing, 24(1), 1161. 
3. Zhang, S., Tang, P. M., & Lau, H. C. (2025). A Nurse Staffing and Scheduling Problem with Bounded Flexibility and Demand Uncertainty. arXiv preprint arXiv:2505.22124.
4. Nagayoshi, M., & Tamaki, H. (2022). An Approach of Exchanging Work Shifts Using Reinforcement Learning on a Constructive Nurse Scheduling System. Journal of Robotics, Networking and Artificial Life, 9(2), 154-158.
5. Ben Said, A., & Mouhoub, M. (2026). Machine Learning and Constraint Programming for Efficient Healthcare Scheduling. International Journal of Software Engineering and Knowledge Engineering, 1-32.

Adaptable nonlinear physical devices for efficient AI

Supervisors 
Name: Prof. dr. Marcel van Gerven
Email: marcel.vangerven [at] donders.ru.nl
Website: https://www.ru.nl/personen/gerven-m-van
Department: Machine Learning and Neural Computation
Faculty: FSW
Cosupervisor: Wilfred van der Wiel, BRAIN, UT

Project Description

Are you an AI/Physics student interested in performing groundbreaking research at the intersection of machine learning, physics and engineering that may help reduce the energy consumption of AI systems? Then this exciting project is for you.

Research Question/Goal: In this project you will work on training networks of (simulated) nonlinear silicon devices that may one day replace standard compute components. You will focus on training these networks to solve various AI tasks more efficiently.

Method: RNPUs are devices that can learn nonlinear functions by setting external parameters [1-2]. By coupling these devices, AI models can be taught to solve real world problems. The question is how to optimally learn these parameters.

In this project you will examine new approaches (e.g. based on genetic programming [3], Bayesian optimization [4] or perturbative learning [5]) to effectively train networks consisting of nonlinear processing elements. You will examine how to train neural networks and/or dynamical systems consisting of these components to solve challenging AI problems (e.g. speech processing or control of simulated humanoid robots).

The approach will be fully tested in simulations (https://github.com/BraiNEdarwin) but the aim is to make the approach suitable for use in the lab in the long term. This project entails a collaboration between Prof. van Gerven (RU) and Prof. van der Wiel (UT).

Student requirements

Experience with / interest in: Machine learning, Physics-inspired methods, Excellent programming and math skills

Type of project: Master thesis

Expected time frame: Start as soon as possible for a 6-9 month MSc project (funded maximally 6 months by the ELLIS unit)

Literature (max 5)

1.    Chen, Tao, Jeroen van Gelder, Bram van de Ven, et al. “Classification with a Disordered Dopant-Atom Network in Silicon.” Nature 577, no. 7790 (2020): 341–45. https://doi.org/10.1038/s41586-019-1901-0.
2.    Escudero, Manuel, Mohamadreza Zolfagharinejad, Sjoerd van den Belt, Nikolaos Alachiotis, and Wilfred G. van der Wiel. “Physical Analog Kolmogorov-Arnold Networks Based on Reconfigurable Nonlinear-Processing Units.” arXiv:2602.07518. Preprint, arXiv, February 7, 2026. https://doi.org/10.48550/arXiv.2602.07518.
3.    Vries, Sigur de, Sander Wessel Keemink, and Marcel Van Gerven. “Kozax: Flexible and Scalable Genetic Programming in JAX.” Proceedings of the Genetic and Evolutionary Computation Conference Companion, July 14, 2025, 603–6. https://doi.org/10.1145/3712255.3726681.
4.    https://bayesoptbook.com/
5.    Dalm, Sander, Marcel van Gerven, and Nasir Ahmad. “Effective Learning with Node Perturbation in Deep Neural Networks.” ArXiv Preprint ArXiv:2310.00965v2 [Cs.LG], October 2023. http://arxiv.org/abs/2310.00965.
 

Bayesian Reinforcement Learning across Multiple Environments 

Supervisors 

Name: Maris Galesloot MSc & Prof. Dr. Nils Jansen 

Email: maris.galesloot [at] ru.nl, nils.jansen [at] ru.nl 

Website: https://ai-fm.org, https://marisgg.github.io, https://nilsjansen.org 

Department: Institute for Computing & Information Sciences (Dept. of Software Science) Faculty: Faculty of Science 

Project Description 

In reinforcement learning (RL), an agent interacts with an environment with unknown dynamics in  order to maximize cumulative reward (e.g., robotics control or video games) [1]. A central challenge  is the exploration–exploitation trade-off: the agent must gather information about the environment  while using existing knowledge to act effectively. In Bayesian RL, the agent maintains and updates a  posterior distribution over environment dynamics, allowing principled exploration via uncertainty aware decision making [2]. 

In many applications, an agent must operate across multiple related environments corresponding to  different tasks. These tasks may share structure but differ in their underlying dynamics or reward  functions. Moreover, the task identity may not be directly observed, so the agent must  simultaneously infer the current task while learning to act optimally. Exploiting such shared structure can significantly improve sample efficiency compared to learning each task independently. 

Research Question/Goal: This project studies Bayesian approaches to multi-task RL [3] in settings where the underlying task is  latent. We consider environments drawn from a small set (or mixture) of possible Markov decision  processes (MDPs), where the agent does not observe which environment it currently faces. The goal  is to design and analyze Bayesian RL methods that maintain beliefs over both environment dynamics and task identity, enabling efficient transfer of information across tasks while balancing exploration  and exploitation. 

Method: We will extend Posterior Sampling for Reinforcement Learning (PSRL) methods [4] to the multi-task  setting. In particular, we consider models where environments are drawn from a mixture of MDPs  with shared priors. The agent maintains a posterior over mixture components (tasks) and MDP  dynamics and samples a candidate environment from this posterior to guide exploration. The  approach will build on recent work on posterior sampling with mixture priors [5] and related multi task Bayesian RL models [3]. Instead of mixtures, we may also explore settings in which  environments are drawn based on a hidden latent variable. We evaluate performance in simulated  environments by comparing learning efficiency and regret against standard PSRL baselines. In the  project, you will be embedded within an active research group that specializes in handling  uncertainties revolving unknown environments. 

Student requirements 

Experience with / interest in: Reinforcement learning, Markov decision processes, machine learning

Type of project: Master thesis 

Expected time frame: 6 months 

Literature (max 5) 

1. Richard S. Sutton, Andrew G. Barto: Reinforcement learning - an introduction, 2nd Edition. MIT  Press (2018) 

2. Mohammad Ghavamzadeh, Shie Mannor, Joelle Pineau, Aviv Tamar: Bayesian Reinforcement  Learning: A Survey. Foundations and Trends in Machine Learning (2015) 

3. Aaron Wilson, Alan Fern, Soumya Ray, Prasad Tadepalli: Multi-task reinforcement learning: a  hierarchical Bayesian approach. ICML 2007 

4. Ian Osband, Daniel Russo, Benjamin Van Roy: (More) Efficient Reinforcement Learning via  Posterior Sampling. NIPS 2013 

5. Joey Hong, Branislav Kveton, Manzil Zaheer, Mohammad Ghavamzadeh, Craig Boutilier: Thompson Sampling with a Mixture Prior. AISTATS 2022

Hysteretic networks for optimal control

Supervisors

Name: Prof. dr. Marcel van Gerven
Email: marcel.vangerven [at] donders.ru.nl
Website: https://www.ru.nl/personen/gerven-m-van
Department: Machine Learning and Neural Computation
Faculty: FSW
Cosupervisor: Martin van Hecke, AMOLF

Project Description

Are you an AI/CS/Physics student interested in performing groundbreaking research on learning to control complex systems such as (simulated) robots? Then this project might be for you.

Research Question/Goal: It has been shown that networks of hysteretic elements can be designed so that they act as in-memory, sequential computing devices [1,2]. In this project, we will explore if such networks can be exploited for effective training of controllers. We will focus on controlling simulated (humanoid) robots [3]. The goal is to improve the efficiency and effectiveness of these controllers.

Method: Building on our work in genetic programming, we have already shown that complex systems can be effectively controlled [3-4]. Here, we explore if the incorporation of hysteretic elements improves the performance of controllers that require memory for effective control. 

You will test this in simulated robotics environments [5]. This project entails a collaboration between Prof. van Gerven (RU) and Prof. van Hecke (AMOLF). 

Student requirements

Experience with / interest in: 

Machine learning, Physics-inspired methods

Excellent programming and math skills

Type of project: Master thesis

Expected time frame: Start as soon as possible for a 6-9 month MSc project (funded maximally 6 months by the ELLIS unit)

Literature (max 5)

1. Liu, Jingran, Margot Teunisse, George Korovin and Martin van Hecke. “Controlled pathways and sequential information processing in serially coupled mechanical hysterons, PNAS 121 e2308414121 (2024) https://doi.org/10.1073/pnas.2308414121
2. Shohat, Dor and Martin van Hecke, Geometric Control and Memory in Networks of Hysteretic Elements, Phys. Rev. Lett. 134, 188201 (2025). https://doi.org/10.1103/PhysRevLett.134.188201
3. De Vries, Sigur, Sander Keemink, and Marcel Van Gerven. “Discovering Continuous-Time Memory-Based Symbolic Policies Using Genetic Programming.” Genetic Programming and Evolvable Machines 27, no. 1 (2026): 5. https://doi.org/10.1007/s10710-025-09530-9.
4. Vries, Sigur de, Sander Wessel Keemink, and Marcel Van Gerven. “Kozax: Flexible and Scalable Genetic Programming in JAX.” Proceedings of the Genetic and Evolutionary Computation Conference Companion, July 14, 2025, 603–6. https://doi.org/10.1145/3712255.3726681.

Tassa, Yuval, Yotam Doron, Alistair Muldal, et al. “DeepMind Control Suite.” arXiv:1801.00690. Preprint, arXiv, January 2, 2018. https://doi.org/10.48550/arXiv.1801.00690

Brain-Inspired Agentic AI: Spiking Neural Networks for  Autonomous, Adaptive Intelligence 

Contact and/or supervisor(s) at Radboud University 

Name: Mahyar Shahsavari 

Email: Mahyar.shahsavari [at] donders.ru.nl 

Website: https://www.ru.nl/en/people/shahsavari-m 

Department: Machine Learning and Neural Computing (MLNC) 
Department Faculty: Social Science 

Project Description 

Background/Motivation: 

Recent advances in Artificial Intelligence have led to the emergence of agentic systems capable of autonomous decision-making, planning, and adaptation. However, most existing  approaches rely on conventional deep learning models that are computationally expensive,  energy-intensive, and poorly suited for real-time, adaptive environments. 

In contrast, the brain operates using spiking neural networks (SNNs), which process  information through sparse, event-driven signals and support efficient temporal learning.  Neuromorphic computing systems inspired by this paradigm offer significant advantages in  energy efficiency, online learning, and robustness. Despite these benefits, their potential  for enabling agentic AI systems remains underexplored. 

This project is motivated by the need to bridge brain-inspired computation and  autonomous AI, investigating how SNNs can support adaptive, efficient, and scalable  agentic behavior. 

Research Question/Goal: 

• How can spiking neural networks be used to design autonomous and adaptive  agentic AI systems?  

• Which learning mechanisms (e.g., STDP, reinforcement learning, surrogate  gradients) best support decision-making and adaptation in SNN-based agents?  • Can neuromorphic approaches improve efficiency, robustness, and real-time  performance compared to conventional deep learning models? 

Method: 

The project will combine theoretical analysis, simulation, and experimental evaluation: 

1. Literature Review 

Review existing work on:  

• Spiking neural networks and neuromorphic computing  

• Agentic AI and reinforcement learning 

• Brain-inspired learning rules (e.g., Reward-based STDP, Local Forward Forward, SRNN)  

2. Model Design 

• Develop SNN-based agent architectures (e.g., spiking RL agents or decision making networks)  

• Explore biologically plausible learning rules and compare with surrogate  gradient methods  

3. Simulation & Implementation 

• Implement models using frameworks such as SnnTorch, SpikingJelly,  BindsNET 

• Evaluate in simple environments (e.g., control tasks, grid-world, or event based settings)  

4. Evaluation 

Measure performance in terms of:  

• Task success / reward  

• Energy efficiency (proxy metrics)  

• Adaptation and generalization  

• Robustness to noise  

5. Analysis 

• Compare SNN-based agents with ANN baselines  

• Identify strengths, limitations, and future research directions  

Student requirements 

Experience with / interest in: 

• Machine Learning / Deep Learning fundamentals  

• Python programming  

• Interest in neuromorphic computing, spiking neural networks, and Agentic AI • Basic understanding of neural networks and optimization  

Literature (max. 5) 

1. D. B. Acharya, K. Kuppan, and B. Divya, “Agentic AI: Autonomous intelligence for complex goals—A  comprehensive survey,” IEEE Access, 2025. 

2. M. K. K, S. Mehta.S, A. H. L, S. N, G. Jadhav and A. Mitra, "Neuromorphic-Driven Agentic AI for  Autonomous Decision-Making Systems," 2024 4th International Conference on Mobile Networks and  Wireless Communications (ICMNWC), Tumkuru, India, 2024, pp. 1-8, doi:10.1109/ICMNWC63764.2024.10872131. 

3. J. López-Randulfe and L. B. Larsen, “A multi-agent model for growing spiking neural networks,”  CoRR, vol. abs/2010.15045, 2020. [Online]. Available: https://arxiv.org/abs/2010.15045 

4. M. Karamimanesh, E. Abiri, M. Shahsavari, K. Hassanli, A. van Schaik, and J. Eshraghian, “Spiking  neural networks on FPGA: A survey of methodologies and recent advancements,” Neural Networks,  vol. 186, p. 107256, 2025. 

5. Jerry Smit Architecting Autonomy: The Neuroscience Behind Agentic AI Systems, 2025,  https://medium.com/@jsmith0475/architecting-autonomy-the-neuroscience-behind-agentic-ai systems-29aab6d5d131

Efficiently learning mixed behaviours from data: Probabilistic circuits for mixtures of Markov models 

Supervisors 

Name: Maris Galesloot MSc & Prof. Dr. Nils Jansen 

Email: maris.galesloot [at] ru.nl, nils.jansen [at] ru.nl 

Website: https://ai-fm.org, https://marisgg.github.io, https://nilsjansen.org 

Department: Institute for Computing & Information Sciences (Dept. of Software Science) Faculty: FNWI 

Project Description 

In many domains, such as web search or user modelling, we use Markov chains to describe  stochastic processes (Puterman, 1994). However, a single model cannot capture all different  behaviours. Instead, the system is better represented as a mixture of Markov chains (Kausik et al.,  2023), where each component contributes according to its mixture weight. Learning such mixtures  from data is difficult because component identifiers may be hidden, transition dynamics may  overlap, and state spaces can be large. Factored state representations (Guestrin et al., 2003) help  manage complexity but also induce a blow-up in the state space. This motivates the need for a  representation that is both expressive and scalable. 

Research Question/Goal: This project investigates the use of probabilistic circuits (Choi et al., 2019), also known as sum product networks, (Melibari et al., 2016), as a compact and tractable representation for mixtures of  Markov chains in factored state spaces. Key questions are: How can mixtures of Markov chains be  encoded efficiently as probabilistic circuits? What structural assumptions can we exploit? How do  circuit-based representations compare to traditional approaches? 

Method: We will define how transition dynamics and mixture weights can be expressed using probabilistic  circuits, and identify factorizations (across features or components) that reduce circuit size. The  initial idea is to implement the circuits in Cirkit (https://github.com/april-tools/cirkit). In this project,  we can collaborate with our colleague Antonio Vergari from the University of Edinburgh, an  international expert in probabilistic machine learning, and probabilistic circuits in particular. 

Student requirements 

Experience with / interest in: Probabilistic machine learning, mixture models, Markov chains,  reinforcement learning 

Type of project: master thesis 

Expected time frame: 6 months 

Literature (max 5) 

1. Markov Decision Processes: Discrete Stochastic Dynamic Programming, 1st Edition, Martin L.  Puterman, 1994. 

2. C. Guestrin, D. Koller, R. Parr, and S. Venkataraman. Efficient solution algorithms for factored  MDPs. J. Artif. Intell. Res., 2003. 

3. C. Kausik, K. Tan, and A. Tewari. Learning mixtures of Varkov chains and MDPs. ICML, 2023. 4. M. Melibari, P. Poupart, P. Doshi, and G. Trimponias. Dynamic sum product networks for tractable  inference on sequence data. PGM, 2016. 

5. YooJung Choi, Antonio Vergari, Guy Van den Broeck, Probabilistic circuits: A unifying framework  for tractable probabilistic models,

Multi-phase AI for kidney cancer 

Supervisors  

Name: Sarah de Boer, Msc 

Email: sarah.deboer [at] radboudumc.nl 

Website: https://www.diagnijmegen.nl/people/sarah-de-boer/ 

Department: Beeldvorming, Diagnostic Image Analysis Group 

Faculty: Radboudumc 

Name: Alessa Hering, PhD 

Email: alessa.hering [at] radboudumc.nl 

Website: https://www.diagnijmegen.nl/people/alessa-hering/ 

Department: Beeldvorming, Diagnostic Image Analysis Group 

Faculty: Radboudumc 

Project Description 

Background/Motivation: 

Our team is developing an AI-based pipeline for segmenting and classifying renal masses (also called lesions) in computed tomography (CT) scans. For patients with suspicion of kidney  cancer, a multi-phase CT scan is commonly used to diagnose and stage the cancer. Multi-phase CT  refers to an imaging protocol where a contrast agent is given to the patient and at subsequent  timepoints a CT scan is made. The contrast agent goes through the body and depending on the time  point, different organs are enhancing. These enhancements allow radiologists to assess organs for  tumors or other findings. At the moment, the AI models we have built and are common in the  literature, only take one contrast-phase CT scan as input. We would like to investigate whether it is  beneficial for AI models to input multiple contrast phases. 

This project will be embedded in the COMFORT project (https://comfort-ai.eu) which aims to  develop robust and trustworthy multimodal AI systems to enhance clinical outcomes for prostate  and kidney cancer patients. Our goal is to create internationally and interdisciplinarily validated  decision support systems that improve clinical prognosis, patient stratification, and individualized  therapy options. 

Goal: Develop a method that incorporates multiple CT contrast phases for kidney cancer. 

Method: 

In recent literature^1,2,3,4,5, multiple methods for leveraging multi-phase CT in AI have been proposed.  Most of the work apply their strategies to liver cancer, as open data is available, or it is applied to  kidney cancer, but the models are unfortunately not publicly available. Therefore, the first step  would involve researching these methods and picking a promising approach. Next, the chosen  method needs to be built, likely from scratch. However, for baseline methods we have code  available for single phase models, which can be used to compare the developed method to. If time  allows, proposing a new strategy for improving the existing methods is possible. 

Student requirements 

Experience with / interest in: AI, deep learning, medical image analysis, building AI models from  scratch in Pytorch, strong mathematical foundation of deep learning methods 

Type of project (master thesis, internship (only for companies)): Master thesis 

Expected time frame: 6 months 

Literature (max 5) 

1. Uhm, K.-H., Jung, S.-W., Choi, M. H., Hong, S.-H., & Ko, S.-J. (2022). A Unified Multi-Phase CT  Synthesis and Classification Framework for Kidney Cancer Diagnosis With Incomplete Data. IEEE  Journal of Biomedical and Health Informatics, 26(12), 6093–6104.  

https://doi.org/10.1109/JBHI.2022.3219123 

2. Uhm, K.-H., Jung, S.-W., Choi, M. H., Shin, H.-K., Yoo, J.-I., Oh, S. W., Kim, J. Y., Kim, H. G., Lee, Y. J.,  Youn, S. Y., Hong, S.-H., & Ko, S.-J. (2021). Deep learning for end-to-end kidney cancer diagnosis on  multi-phase abdominal computed tomography. Npj Precision Oncology, 5(1), 54.  https://doi.org/10.1038/s41698-021-00195-y 

3. Dai, C., Xiong, Y., Zhu, P., Yao, L., Lin, J., Yao, J., Zhang, X., Huang, R., Wang, R., Hou, J., Wang, K.,  Shi, Z., Chen, F., Guo, J., Zeng, M., Zhou, J., & Wang, S. (2024). Deep Learning Assessment of Small  Renal Masses at Contrast-enhanced Multiphase CT. Radiology, 311(2), e232178.  https://doi.org/10.1148/radiol.232178 

4. Kim, D. W., Lee, G., Kim, S. Y., Ahn, G., Lee, J.-G., Lee, S. S., Kim, K. W., Park, S. H., Lee, Y. J., & Kim,  N. (2021). Deep learning–based algorithm to detect primary hepatic malignancy in multiphase CT of  patients at high risk for HCC. European Radiology, 31(9), 7047–7057.  

https://doi.org/10.1007/s00330-021-07803-2 

5. Zhang, Y., Peng, C., Peng, L., Huang, H., Tong, R., Lin, L., Li, J., Chen, Y.-W., Chen, Q., Hu, H., &  Peng, Z. (2021). Multi-phase Liver Tumor Segmentation with Spatial Aggregation and Uncertain  Region Inpainting (No. arXiv:2108.00911). arXiv. https://doi.org/10.48550/arXiv.2108.00911

Real-Time Stress Detection with Multimodal EEG–ECG  Fusion using Spiking Neural Networks 

Contact and/or supervisor(s) at Radboud University 

Name: Mahyar Shahsavari 

Email: Mahyar.shahsavari [at] donders.ru.nl 

Website: https://www.ru.nl/en/people/shahsavari-m 

Department: Machine Learning and Neural Computing (MLNC) Department Faculty (if known): Social Science 

Project Description 

Stress detection using physiological signals has become an important research area in  healthcare, human-computer interaction, and wearable technologies. Traditional machine  learning and deep learning approaches (e.g., CNNs, LSTMs) have shown promising results  when applied to electroencephalogram (EEG) [4] and electrocardiogram (ECG) [5] signals.  However, these methods often fail to fully exploit the temporal dynamics and event-based  nature of biological signals. 

Recent research highlights that Spiking Neural Networks (SNNs) are particularly well-suited  for processing temporal biosignals due to their biological plausibility and spike-based  computation [1]. Moreover, combining multiple physiological modalities such as EEG and  ECG has been shown to significantly improve stress detection [2]. 

Emerging works demonstrate the potential of multimodal SNNs for stress detection, achieving  improved accuracy and energy efficiency compared to conventional neural networks [3].  Additionally, SNNs enable real-time and low-power inference, making them ideal for wearable  and edge devices. 

Despite these advances, challenges remain: 

• Effective fusion of multimodal signals in SNN frameworks  

• Robust real-time inference 

• Generalization across subjects and conditions  

This project aims to address these gaps by designing a real-time multimodal SNN system for  stress detection. 

Research Question/Goal: 

Main Question: 

How can spiking neural networks effectively fuse multimodal EEG and ECG signals for  accurate and real-time stress detection? 

Sub-goals:

• Design a multimodal fusion strategy for EEG + ECG in SNNs  • Develop a low-latency, real-time SNN model  

• Compare SNN vs ANN-based models (CNN/LSTM)  

• Evaluate robustness across subjects and datasets  

• Investigate energy efficiency and neuromorphic deployment potential  

Method: 

1. Data Acquisition 

• Public datasets (e.g., Physionet EEG dataset, DEAP, WESAD, or similar multimodal stress  datasets)  

• Signals: EEG + ECG (optionally EDA for extension)  

2. Preprocessing 

• EEG: filtering, artifact removal, band extraction  

• ECG: HRV features, R-peak detection  

• Synchronization of multimodal signals  

3. Spike Encoding 

Convert signals into spikes using:  

• Rate coding  

• Temporal coding  

• Learnable encoding (advanced option)  

4. Model Architecture 

Design multimodal SNN architecture:  

• Separate encoders for EEG and ECG  

• Fusion strategies: early fusion (input-level), late fusion (decision-level), hybrid fusion (recommended, novel)  

• Possible extensions:  

• Attention-based SNN  

• Graph-based EEG modeling  

5. Training 

Surrogate gradient learning  

Compare with:  

• Choosing one of CNN / LSTM/Transformer baseline  

• Hybrid ANN→SNN conversion  

6. Evaluation 

Metrics:  

• Accuracy, F1-score  

• Latency (real-time capability)  

• Energy efficiency (if possible) 

7. Real-Time Setup (Optional but Strong for ELLIS) 

• Streaming inference simulation  

• Deployment on:  GPU / edge device / neuromorphic hardware (if available) 

Student requirements 

Experience with / interest in: 

• Machine Learning / Deep Learning  

• Signal processing (EEG, ECG preferred but not required)  

• Python (PyTorch, preferably SNN frameworks like Norse / SpikingJelly)  

• Interest in:  

• Neuromorphic computing  

• Brain-inspired AI  

• Healthcare AI  

Literature (max. 5) 

1. Park, S.S., Choi, YS. Spiking neural networks for physiological and speech signals: a review. Biomed. Eng. Lett. 14,  943–954 (2024). https://doi.org/10.1007/s13534-024-00404-0 

2. Huang Z, Guo B, Xu H, Ruan H, Guo D. Multimodal Spiking Neural Network With Generalized Distributive Law for  Biosignal and Sensory Fusion. IEEE Trans Biomed Eng. 2026 Jan 12;PP. doi: 10.1109/TBME.2026.3653109. 

3. Tan, C.; Ceballos, G.; Kasabov, N.; Puthanmadam Subramaniyam, N. FusionSense: Emotion Classification Using  Feature Fusion of Multimodal Data and Deep Learning in a Brain-Inspired Spiking Neural Network. Sensors 2020, 20,  5328. https://doi.org/10.3390/s20185328 

4. Joshi, A., Matharu, P.S., Malviya, L. et al. Advancing EEG based stress detection using spiking neural networks and  convolutional spiking neural networks. Sci Rep 15, 26267 (2025). https://doi.org/10.1038/s41598-025-10270-0 

5. Lee, Y. Abrupt Change Detection of ECG by Spiking Neural Networks: Policy-Aware Operating Points for Edge Level MI Screening. Appl. Sci. 2025, 15, 12210. https://doi.org/10.3390/app152212210

Inferring environmental structure from cell motility data

Supervisors 

Name:                                Inge Wortel
Email:                                 inge.wortel [at] ru.nl
Website:                            https://computational-immunology.org/, https://das.cs.ru.nl/ 
Department:                     Data Science, Institute for Computing and Information Sciences
Faculty:                              Faculty of Science

Project Description

Background/Motivation: Our immune cells continuously patrol the body to detect and eliminate infected or cancerous cells. To perform this task effectively, they must navigate complex environments, such as lymphoid tissues, inflamed regions and tumours, where their motion is defined by a combination of cell-intrinsic behaviour and cell-extrinsic physical constraints. A large body of research studies these processes by recording time-lapse videos of immune cells in tissues [1]. These recordings are typically obtained through fluorescence microscopy, in which specific cells are “labelled” and become visible against an otherwise black background. This means that these videos can be somewhat misleading: environmental structure is often not labelled and therefore not visible directly – even though it can strongly constrain the observed behaviour. 

An important question when studying these videos is if we can infer this environmental context from the observed trajectories, as we otherwise risk misinterpreting the observed cell behaviour. Work in other domains has shown this is possible in principle, for example in single-particle tracking [2], animal location data [3], or ship trajectories [4]. However, none of these methods are particularly suited to the specific challenges in cell migration videos: a combination of both persistent and highly stochastic movement, a limited observation window, and, relative to other applications, trajectories with a low number of data points [5]. In this project, you will work on a method for inferring environmental structure from trajectories that is specifically designed to deal with these challenges. 

Research Question/Goal: Can we infer the environmental obstacles that affect immune cell migration by analysing cell trajectories and/or video data?

Method: In this project, we will recast this problem as detecting points in space where the locally observed motility is unexpected from the “null distribution” of expected motility when we do not condition on a particular location. You will explore different approaches to doing this, e.g. data-driven approaches through bootstrapping from the available trajectories or model-driven approaches based on persistent random walks. You will use a combination of experimental cell motility videos and simulated data in this project. 

Student requirements

Experience with or interest in: analysis of trajectory or time-series data, stochastic processes or random walk models, basic knowledge of probability theory, computational biology

Type of project: Master thesis

Expected time frame: 6 months (max period funded by ELLIS), but the duration of the thesis project can extend beyond the funded period if required for the student thesis. 

Literature (max 5)

1. Pizzagalli, D.U., Carrillo-Barbera, P., Palladino, E. et al. (2024). Systematic analysis of immune cell motility leveraging Immunemap, an open intravital microscopy atlas. bioRxiv. https://doi.org/10.1101/2024.12.02.626343
2. Wu, HM., Lin, YH., Yen, TC. et al. Nanoscopic substructures of raft-mimetic liquid-ordered membrane domains revealed by high-speed single-particle tracking. Sci Rep 6, 20542 (2016). https://doi.org/10.1038/srep20542
3. Aarts, G., MacKenzie, M., McConnell, B., Fedak, M. and Matthiopoulos, J. (2008), Estimating space-use and habitat preference from wildlife telemetry data. Ecography, 31: 140-160. https://doi.org/10.1111/j.2007.0906-7590.05236.x
4. Wang, L., Chen, P., Chen, L., & Mou, J. (2021). Ship AIS Trajectory Clustering: An HDBSCAN-Based Approach. Journal of Marine Science and Engineering, 9(6), 566. https://doi.org/10.3390/jmse9060566
5. Schienstock, D and Mueller, S.N. Moving beyond velocity: Opportunities and challenges to quantify immune cell behavior. Immunol Rev.2022;306:123–136. https://doi.org/10.1111/imr.13038