Editing genes has become much easier since the discovery of CRISPRCas proteins such as Cas9. Cas9 can be directed to cut almost any location on the genome by supplying it with a short guide RNA (gRNA) with a nucleotide sequence that matches the DNA target. With tools like Cas9, scientists can study disease, improve crops and develop therapies. Despite the apparent simplicity of Cas9based gene editing, it remains challenging to predict if a chosen gRNA causes Cas9 to cut DNA at the intended location, and if it risks cutting other locations too. A common strategy for Cas9 activity prediction is to collect large datasets describing which gRNAs cause which edits in a particular cell, and to train machine learning models on the observed patterns. However, such empirical models rapidly lose predictive power when applied to other cellular contexts or other experimental conditions. An alternative approach is to isolate Cas9, gRNA, and DNA from a cell, measure their dynamics with high precision, and thus learn about the molecular interactions that govern Cas9 activity in any context. With this molecular knowledge, one can formulate rational and general predictions about activity across cells and experiments.

This thesis uses biophysical modeling to translate understanding of Cas9’s molecular mechanisms to prediction of its genome editing activity in cells. Chapter 1 introduces CRISPR, outlining its origin as an immune system in bacteria and archaea, its transformation into a DNA editing toolkit, and the wide range of genetic engineering applications it has enabled. It discusses the main challenges to successful CRISPR applications, and explains how biophysical models can support gRNA selection and other design choices. Chapter 2 describes mathematically how Cas9 dynamics differ between cellfree experiments (in vitro) and living cells (in vivo) due to differences in Cas9 availability and the processes of target search and recognition. By connecting the two settings, the framework allows the findings of later chapters to be generalized across experimental and cellular conditions. Chapter 3 introduces CRISPRzip, a model of Cas9 target recognition that integrates the physics of gRNADNA interactions, and thus explains the variation in activity across gRNA and DNA sequences. Moreover, CRISPRzip incorporates the effects of Cas9 concentration and DNA twisting (supercoiling) on how Cas9 binds and cuts DNA. Because CRISPRzip adapts to gRNA sequence and to environment conditions like concentration and supercoiling, it provides a flexible basis for gRNA design even when the application setting deviates from the experiments used for training. Chapter 4 presents an experimental approach to quantify the effects of DNA supercoiling across thousands of DNA sequences. While these results can largely be explained by CRISPRzip, they also reveal a complex interplay between supercoiling and sequence that calls for further study. Chapter 5 extends the molecular description of Cas9 to the scale of the full genome. Using analytical models of target search and numerical predictions from CRISPRzip, it shows that prediction of ontarget and major offtarget activity does not require explicit consideration of the surrounding genome sequence under typical conditions. Also, it formulates a theory that connects different Cas9 delivery strategies and shows how to balance precision and efficiency in gene editing applications. Chapter 6 concludes the thesis by summarizing the main results and presenting an outlook towards biophysicssupported CRISPR activity prediction in cells
... Read more

Europe has a long and rich history of theatre development, closely linked to the growth of cities and urban conditions. The shape of the theatre and the behaviour of the audience changed over time. Still, the interaction between the city and the stage in most theatres has remained virtually unchanged since the 19^th century. The theatre serves only as a place for performance, catering to a specific audience. This often means the theatre building remains closed to the public for most of the day, opening only a few hours before the performance. Today, there is a desire to make theatre more accessible and an important part of the city's daily public life. 

The site for the new theatre in Delft is in a complex, layered environment. Urban redevelopment of large apartment blocks began in the 1970s but never fully filled the demolished granular urban tissue of the old city. The new theatre bridges the difference in scale between new construction in the west and traditional, undemolished houses in the east by cascading and reducing its height from west to east. By aligning the stage tower to the axis of the existing street, the building creates a new and recognisable landmark for the city. By lifting the main auditorium to the first floor, a big public area is created on the ground floor. This space can be used throughout the day as a public meeting place, for leisure activities, independently from the theatre. 

The result is a big urban gesture solving existing urban problems while also creating valuable indoor public space that can accommodate different events and activities.

... Read more

The rapid integration of renewable energy is making long-term power system planning more important and more computationally demanding. Generation Expansion Planning (GEP) is a key problem in energy planning because it determines where and how much to invest in generation capacity. As renewables grow, GEP must be solved across many scenarios, demand forecasts, and technology options, making the speed of solving it a key bottleneck for timely policy analysis. GEP is commonly formulated as a large-scale mixed-integer linear program and is often solved using Benders Decomposition, where investment decisions are optimized in a master problem and the Economic Dispatch (ED) problems are solved as subproblems, but this remains computationally expensive at scale. Recent work proposed learning the dual solution of the ED subproblem with a machine-learning surrogate to generate approximate Benders cuts during an inexact phase, followed by exact refinement to recover optimality guarantees. The runtime benefit of this framework, however, depends on whether the learned surrogate produces cuts that are useful within the Benders algorithm, which has not been studied empirically. This thesis investigates how machine-learning dual surrogates can be more effectively designed and integrated into Benders Decomposition to accelerate GEP. Three aspects are studied. First, a constraint-aware data generation strategy is proposed to construct surrogate training instances that better reflect the investment decisions encountered during Benders iterations. Second, limitations of self-supervised dual learning are diagnosed. To address these limitations, this thesis proposes Heuristic Soft Label Augmentation (SLA), which adds cheap, structure-aware auxiliary guidance for learning ED shadow-price classes without requiring exact dual labels. Third, the thesis studies how each subproblem’s dual solution should be converted into Benders cuts, comparing single-cut, full multi-cut, and a proposed K-means clustered cut-aggregation strategy and analysing their effect on overall runtime. The results show that constraint-aware data generation improves the usefulness of surrogate-generated cuts during the inexact phase, and that SLA reduces the dual surrogate’s optimality gap and can lead to faster GEP solve time. Cut management has the largest effect: the proposed K-means clustered multi-cut strategy achieves the largest runtime improvement in the tested setting by preserving timestep-level dual information while limiting master-problem growth. Combining the three improvements, ML-enhanced Benders solves GEP 3 times faster than both exact Benders and a direct solver on the tested instances. Overall, the thesis shows that accelerating MLenhanced Benders requires improving the surrogate’s integration across the pipeline: training-data alignment, dual learning, and cut management, with cut management the dominant contributor at the tested scale.
... Read more

Conflict analysis has become one of the key techniques behind the success of modern Constraint Programming (CP) solvers. In Lazy Clause Generation (LCG), conflicts are analysed to derive learned constraints, known as nogoods, which help the solver avoid revisiting infeasible regions of the search space. While learned nogoods can substantially improve search efficiency, maintaining large nogood databases introduces computational and memory overheads, making periodic removal of low-quality nogoods essential. Existing approaches largely rely on heuristics adopted from SAT solving, yet their comparative performance in the CP setting has not been systematically studied.

This thesis investigates the impact of nogood quality metrics and database reduction strategies on the performance of CP solvers. We analyse eight nogood quality metrics, including both established measures such as activity and Literal Block Distance (LBD), as well as new CP-specific metrics. We analyse their correlation with nogood usefulness, defined in terms of propagation behaviour across progressively stricter notions of utility. We find that all metrics can, to some extent, predict the usefulness of nogoods, and we identify four metrics for further study.

We then propose and evaluate a set of nogood management schemes that combine these metrics in different ways. Experiments across hundreds of optimisation and satisfaction instances from the MiniZinc Challenge show that a scheme retaining nogoods with either low LBD or high activity achieves the fewest conflicts and outperforms the default management strategy of the state-of-the-art solver Pumpkin. For anytime behaviour, incorporating the number of variables enables faster discovery of high-quality solutions. We further demonstrate that periodic database reductions are necessary, but that the choice of database size parameters involves a trade-off between finding solutions quickly and proving optimality. Finally, we find that differences in solver performance cannot be fully explained by schemes' ability to remove nogoods that cause few propagations and retain the active ones, suggesting avenues for future research. ... Read more

Detecting money laundering in financial transaction data is a task where graph neural networks (GNNs) have shown strong potential. Such data is naturally represented as a directed multigraph, since two accounts, each represented as a node, may exchange many separate payments, each forming a distinct edge with its own amount, currency, and timestamp. Preserving these parallel edges, rather than collapsing them into a single connection, retains the fine-grained structure that allows for distinguishing laundering behaviour from ordinary activity. Yet these models also introduce a new vulnerability, as an adversary could manipulate the transaction graph to alter the neighbourhood of a suspicious account such that the GNN misclassifies it as benign. Existing adversarial robustness research operates on the adjacency matrix, which records at most one edge per node pair and therefore cannot represent the parallel transactions between two accounts that this task depends on. Multigraph GNNs therefore lack both a framework for evaluating robustness under structural perturbations and defences against such perturbations.

This thesis extends adversarial robustness analysis to multigraph GNNs through three contributions. First, it reformulates GNN message passing and attack optimisation over the incidence matrix instead of the adjacency matrix, yielding the first gradient-based structural attack that retains multi-edge structure. Second, it introduces unnoticeability loss terms that constrain perturbations to maintain the graph's statistical fingerprint, including the frequency of characteristic patterns such as short transaction cycles, keeping the attack statistically plausible and unnoticeable at the macro level. Third, it scales the framework to large networks with projected randomised block coordinate descent. On the IBM synthetic anti-money laundering dataset, learned attacks substantially reduce detection accuracy compared to non-learnable perturbations, and adversarial training recovers robustness, showing that multigraph GNNs are both vulnerable to structural manipulation and defensible against it. ... Read more