MSc Defences Spring 2025

Name of student(s)  

Valdemar Skou Knudsen

Study Programme  

Mathematics

Title  

Differentiable simulation and integration of neural networks in Richards equation-based modelling

Abstract  

In this thesis, we consider the Richards equation for modelling water
movement in unsaturated soils. An expository discussion and derivation
of the standard base-line numerical scheme comprises the first chapter,
with a focus on test scenarios and theoretical properties useful in model
validation. We then turn to a shorter chapter on the theoretical framework
behind exciting new developments in differentiable physics and integration of neural networks in process-based modelling. A novel minimization
problem combining a "solver-in-the-loop" approach and absolute constants
in the model is defined for the Richards equation. Further experimentation on this model is postponed for future work, due to computational
constraints.

Supervisor(s)  

Oswin Krause

External examiner(s)  

Anders Stockmarr

Date and time  

13.03.2025 10:00 TBC

Room  

TBD

Name of student(s)  

Betül Sert

Study Programme  

Health and informatics

Title  

Optimization of patient-reported outcomes (PRO) for multiple sclerosis patients: Analysis of user experiences and development of improvement proposals for the existing PRO-platform

Abstract  

Background: Patient-reported outcomes (PRO) hold great potential for improving disease management and clinical consultations for multiple sclerosis (MS) patients. However, usability challenges within the current PRO system limit patient engagement and adoption. This study
investigates MS patients' experiences and identifies key barriers to usability, providing concrete recommendations for improving the accessibility and integration of the existing PRO platform.
Methods: A qualitative research approach was applied, including semi-structured interviews, usability testing, and an analysis of the PRO system’s information infrastructure. Data were analyzed using Installed Base theory to assess how previous technological and organizational choices constrain
system flexibility. Additionally, usability and infrastructure development strategies such as Flexible Generification and Growing were evaluated to explore incremental improvements to PRO.
Results: The study highlights three major challenges: (1) Limited accessibility, where patients experience a complex login process involving a lengthy registration procedure and two-factor authentication, impeding engagement, (2) Lack of perceived value, as patients report that their PRO
data is rarely used in consultations by the doctors, reducing motivation for continued use, and (3) The need for iterative development, as balancing flexibility and standardization is necessary to prevent system fragmentation and technical debt.
Conclusion: To enhance PRO usability, the study recommends simplifying the login process through MitID or biometric authentication, improving integration with clinical workflows and existing health platforms, and adopting an iterative development approach with user involvement. A holistic approach is essential, combining technological optimization, organizational implementation, and user-centered design to ensure PRO becomes a valuable tool for both MS patients and healthcare
professionals.

Supervisor(s)  

Erling Carl Havn

External examiner(s)  

Jens Pedersen

Date and time  

18.03.2025 15:00-16:00

Room  

DIKU UP1-2-0-04

Name of student(s)  

Annike Kjølby Kristensen

Study Programme  

IT and cognition

Title  

Analysis of age-related differences in depression using structural connectomes

Abstract  

This project will segment MRI scans of children and analyze the structural differences and structural connectomes of depressed children by utilizing network graphs. This project will then compare the connectomes with those of depressed adults to identify age-related differences in brain structure associated with depression. This project seeks to answer following questions: What specific brain regions and connections are most affected by depression in children compared to adults? How do the structural connectomes of depressed children differ from those of depressed adults? Are there any identifiable patterns in brain connectivity that can be linked to age-related differences in depression?

Supervisor(s)  

Melanie Ganz-Benjaminsen

External examiner(s)  

Kristoffer Hougaard Madsen

Date and time  

20.03.2025 15:00-16:00

Room  

DIKU UP1-2-0-04

Name of student(s)  

Nikolaj Deichmann

Study Programme  

Statistics Thesis

Title  

Conditional Variational Flow Matching for generating chemical graph structures with desired properties

Abstract  

Machine learning (ML) has become a powerful tool in material discovery, particularly in the generation of molecular structures using graphbased models. While existing methods have successfully generated smaller
molecules, scaling to complex nanomaterials remains challenging. Monometallic oxides, with their relevance in batteries and catalytic applications, exemplify this difficulty due to their large configurational space.
This thesis explores an inverse materials discovery approach, where
desired properties guide the generation of suitable structures. Specifically, we integrate X-ray diffraction (XRD) signals, which are widely used
for characterizing crystalline materials, directly into the generative process. To achieve this, we leverage Conditional Variational Flow Matching (CVFM), a method that provides flexible and adaptive generative modelling through learned probability flow trajectories, enabling the conditional generation of nanomaterials aligned with data. By incorporating structural ”fingerprints” from XRD into the generation pipeline, our method aims to enhance the structural consistency of computationally designed materials, providing a more informed approach to nanomaterial generation. Our experiment find that combining Gaussian flow for node trajectories together with a categorical variant of flow on the edges are best able to leverage the information encoded by the XRD signal but that prediction of additional structural parameters are necessary in order to generate real meaningful structures.

Supervisor(s)  

Raghavendra Selvan

External examiner(s)  

Anders Therkelsen

Date and time  

07.04.2025 14:00 - 15:00

Room  

SCI-DIKU-UP1-2-0-04 and 1-2-0-06

Name of student(s)  

Rikke Vesterbæk Nielsen

Study Programme  

Computer Science

Title  

Supporting Legal Compliance Through Transparent Process-aware Information Systems Enabling Adaptability To The Evovling Nature Of Legislation

Abstract  

Cybersecurity threats in digital supply chains pose increasing risks to critical infrastructure and business operations. To address this, the NIS2 Directive (Directive (EU) 2022/2555) mandates stricter risk management, incident reporting, and supply chain security requirements. However, organisations struggle to assess third-party risks and enforce compliance across complex supplier networks. This thesis investigates how process-aware compliance tools can support NIS2-aligned supply chain security.
It explores Dynamic Condition Response (DCR) Graphs as a modelling framework to structure supplier risk assessments, compliance workflows, and incident response mechanisms. Using the Danish GDPR compliance tool Lexoforms as a case study, the research demonstrates how existing solutions can be adapted to track, assess, and enforce security standards across supply chains. A Process-Aware Information System (PAIS) prototype is developed to model supplier interactions, automate compliance verification, and improve risk monitoring. The findings indicate that model-driven compliance tools improve regulatory transparency and cybersecurity governance, but ensuring adaptability to diverse supply chain structures remains a challenge. The study concludes that integrating AI-assisted monitoring and end-user development capabilities could further streamline supply chain compliance and proactive risk mitigation.

Supervisor(s)  

Thomas Troels Hildebrandt

External examiner(s)  

Søren Debois

Date and time  

03.04.2025 13:00 - 14:00

Room  

Sigurdsgade 41, lokale 2.03

Name of student(s)  

Víctor Escudero González and Gabriel Alexandre Araújo Ribeiro

Study Programme  

Computer Science

Title  

The Effects of Group Definition in Recommender Systems

Abstract  

A Group Recommender System, in contrast to an individual recommender system, recommends an item to a set of users. In activities that can be performed as a group (e.g watching a movie, travelling etc.), a recommender system can leverage a groups’ dynamics in order to make recommendations more accurately. With the increase of users in social media and streaming platforms, these systems are more relevant now than ever. However, a large majority of the research focuses on highly-specific situations, with pre-defined datasets that tend to not be scalable to real-life applications, thus making it difficult to assess the benefits of the "group" aspect in them.

This work aims to study the effects of dynamic group definition in recommender systems. We do so by researching the state-ofthe-art, integrating it to already-existing recommender models and testing it in real-life datasets.

Supervisor(s)  

Christina Lioma

External examiner(s)  

Troels Andreasen

Date and time  

22.04.2025 10:00 - 11:30

Room  

DIKU UP1 1-1-N116A

Name of student(s)  

Sofus Gerlow Tornbjerg

Study Programme  

Computer Science

Title  

Procedural Generation of Gears

Abstract  

This thesis introduces a comprehensive and versatile framework for procedurally generating gears, including spur, helical, bevel, and internal gears, with the aim of producing high quality watertight meshes that adhere to the gear geometry and mating conditions presented in this work. The procedural interfaces developed in this work greatly simplify the
creation of complex gear mechanisms by simplifying parameter selection and automating the verification that the mating conditions are valid. This includes a very simple interface for generating planetary gears. The generated meshes are validated through a series of rigid body simulations demonstrating that the generated gears behave as expected,
including backlash avoidance, sliding contact behavior, and transmission ratios. The procedural interfaces’ usefulness in real world applications is demonstrated by replicating an innovative six-stroke engine design by Porsche.

Supervisor(s)  

Kenny Erleben

External examiner(s)  

Jeppe Revall Frisvad

Date and time  

24.04.2025 9:00 - 11:00

Room  

Kenny Erlebens kontor

Name of student(s)  

Weronika Paulina Mlotkowska

Study Programme  

IT and Cognition

Title  

Exploration of the Cross-Modal Relationship Between Baroque Music and Paintings

Abstract  

Visual and audio arts have the power to evoke emotions in audience, reflecting the artist’s intended expression. This study explores the techniques used in painting and music to convey emotions, investigating whether visual and auditory arts share common features across emotional categories. It further examines potential correlations between image and audio features within each emotional group.

The analysis focuses on Baroque art as a case study, with feature analysis based on the techniques and characteristics of paintings and musical pieces as presented by art historians in the literature. While initial analyses revealed no statistically significant correlations between low-level audiovisual features, suggesting Baroque art’s emotional impact may reside in more complex interactions that cannot be identified in simple feature mapping. Furthermore, this research evaluates the emotional alignment between AI-generated music (produced using MusicGen, AudioLDM-2 and Riffusion models) and the Baroque paintings that inspired these musical compositions.

Supervisor(s)  

Serge Belongie

External examiner(s)  

Dan Witzner Hansen

Date and time  

30.04.2025 10:00 - 11:00

Room  

Pionercenter ØV3, Seminar Room