E-Dissertationen

Predictive Safety in Reinforcement Learning

2026-06-11T06:20:20Z

Predictive Safety in Reinforcement Learning Elnagdi, Murad Autonomous robots operating in unstructured environments require control policies that achieve their tasks reliably while respecting safety constraints. While Reinforcement Learning (RL) has emerged as a powerful data-driven paradigm for optimizing these policies through interaction, its application in robotics is hindered by three fundamental barriers during exploration: it is unstable under sparse rewards, since feedback signals are rare or delayed until the task is solved; it is unsafe, risking damage to the robot and its surroundings; and it is sample-inefficient, as it requires extensive interaction with the environment. This thesis addresses these challenges by establishing prediction as a framework for exploration. We combine model predictive control (MPC) and predicted safety signals to guide data collection, constrain risk, and preserve task performance during training and execution of RL agents. Our methodological progression begins by addressing the inefficiency of random exploration in sparse reward settings. We introduce a hybrid training framework that utilizes MPC as a synthetic expert to guide the agent through complex navigation tasks. This approach accelerates convergence by alternating planned trajectories with trial-and-error experience, yielding a lightweight policy for independent deployment. Transitioning to the safety-critical domain of multi-robot systems, we subsequently employ predictive control as a distributed safety filter. We develop a scalable behavior-based formation controller secured by distributed nonlinear MPC shields, which ensures collision-free training, faster convergence, and enables zero-shot transfer to larger teams and to physical hardware. However, relying on static safety shields often induces conservative behavior that hinders learning. To overcome this limitation, we propose a dynamic safety shield that utilizes a supervisor agent to adapt constraint parameters online. By tuning the shield's sensitivity to the environment, we reduce solver failures and prevent the conservative behaviors typical of static shields. Ultimately, to eliminate the computational bottleneck of running optimization solvers at runtime, we transfer these predictive principles into a modular action regulator. This learned mechanism uses cost critics to preemptively adjust actions based on predicted risk, decoupling safety enforcement from reward maximization. Collectively, these studies show that combining short-horizon planning with learned risk estimation makes RL safer and more sample-efficient without sacrificing task performance. The proposed methods are evaluated in extensive simulation, with the MPC-based frameworks further validated on physical robots to demonstrate their robustness under real-world uncertainties.

Characterising the broadband polarisation properties of extragalactic radio sources at 3 GHz

2026-06-10T07:00:16Z

Characterising the broadband polarisation properties of extragalactic radio sources at 3 GHz Ranchod, Shilpa Magnetic fields are ubiquitous in the Universe and are fundamentally linked to galaxy evolution, influencing star-formation, cosmic-ray propagation and feedback mechanisms. The radio frequency regime, at centimetre wavelengths, is a powerful probe of cosmic magnetic fields through the detection of galaxies emitting linearly polarised synchrotron emission. The Faraday rotation measure (RM) of this emission can be used to study the properties of both the emitting source and the intervening magneto-ionic media (e.g. the Galactic interstellar medium). Large, dense samples of polarised extragalactic sources form RM grids that can be used to map the foreground magnetic field of the Milky Way. Over the past few years, the increasing sensitivity of modern interferometers has given new accessibility to the μJy radio sky, increasing the RM sky density. It is therefore key to establish a clear understanding of the faint polarised source population and the foreground effects that shape RM measurements. Broadband spectro-polarimetry provides an enhanced perspective on this, revealing Faraday complexities in the Stokes Q and U spectral behaviour, which trace turbulence or differential Faraday rotation along the line-of-sight. Through the projects presented in this thesis, I aim to address two important factors in improving our interpretation of RM grid experiments, (i) understanding both the observational biases and physical nature of Faraday complexity at low Galactic latitudes, (ii) characterising the extragalactic polarised source population.
Firstly, in analysing the SPASS/ATCA RM catalogue (Schnitzeler et. al., 2019), the most extensive broadband polarisation catalogue in the southern sky, we report a Galactic latitude dependence of Faraday complexity for polarised sources at |b| < 10°, with the degree of complexity increasing towards the Galactic plane. Through higher angular resolution (θ = 15′′) follow-up observations of 95 sources, we find that this trend is primarily driven by contamination from large-scale Galactic polarised emission in the SPASS/ATCA spectra, which we effectively filter out. We find 42% of the observed sources in our sample are Faraday complex, with an increased fraction of Faraday complex sources surrounding the spiral arm tangents and towards the Galactic centre. We constrain the scale of this complexity to < 2.4 pc, consistent with turbulent injection scales in the spiral arms. These results emphasise the importance of broadband spectro-polarimetric observations to fully characterise small-scale and/or turbulent structures in the Galactic magnetic field, and that it is essential to correctly filter contaminating polarised emission when interpreting foreground turbulence.
Secondly, we reprocess the VLA-COSMOS 3 GHz Large Project (Smolčić et. al., 2017), one of the deepest S-band continuum surveys (2.3 μJy beam⁻¹), for polarisation calibration and imaging. Here, we present the deepest polarised source count at 3 GHz to date, and the second deepest overall, returning an RM density of 42 deg⁻². We find that these source counts are consistent with those at the more typically-studied 1.4 GHz band, a combined effect of spectral index and depolarisation, which we attribute to differential Faraday rotation in the lobes of radio galaxies. Through the available multi-wavelength catalogues, we identify all polarised sources as radio galaxies (i.e. active galactic nuclei), and confirm that no star-forming galaxies are detected in polarisation. We place an upper limit on the density of polarised star-forming galaxies to be < 0.58 deg⁻², implying that surveys much deeper than 2.6 μJy beam⁻¹ will be required to readily probe this population, even at higher frequencies where Faraday depolarisation effects are less pronounced.
Finally, I present a demonstrator project with the recently installed MeerKAT S-band receivers, with a focus on extragalactic wide-field imaging. We present MeerKAT S-band observations of the DEEP2 field, the emptiest radio field in the southern sky. The total intensity source counts are consistent with those from the literature, and also show that only a fraction of integration time is required with MeerKAT for comparable results with legacy S-band surveys. Here, I also present the science goals and survey design for the MeerKAT+ S-band Legacy survey, a 3000 hour, full-Stokes survey of the southern sky at Dec ≤ −40°. This survey will be extremely versatile across Galactic, galaxy evolution, magnetism and transient science, and with an expected 10⁵ polarised source detections, will provide a higher frequency perspective on the RM grid of the southern sky.

Development of an automated RNA capture-SELEX for enriching modular RNA light-up sensors

2026-06-09T09:31:14Z

Development of an automated RNA capture-SELEX for enriching modular RNA light-up sensors Legen, Tjasa This thesis examines recent advances in selecting and functionalizing RNA aptamers for molecular sensing applications. Two related studies have been carried out: developing a robotic platform for automated RNA aptamer selection targeting small molecules and creating a modular approach to design fluorogenic RNA sensors. Together, these studies offer a framework that combines standardized aptamer discovery with adaptable sensor design.
The systematic evolution of ligands by exponential enrichment (SELEX) was adapted to a robotic platform, minimizing manual intervention and enhancing reproducibility. Traditional SELEX often involves immobilizing small-molecule targets, which can alter ligand properties. To address this, capture-SELEX was used, in which RNA libraries are immobilized via hybridization to captureoligodeoxynucleotides (ODNs), and unmodified ligands in solution facilitate the recovery of bound sequences. By optimizing the robotic system, a preferential immobilization strategy for the library was systematically investigated, resulting in more substantial enrichment of binding sequences. The platform completes up to twelve selection cycles in 72 hours and has successfully enriched aptamers for several small molecules, including neomycin B, theophylline, and riboflavin. Interaction analysis using fluorescence polarization and isothermal titration calorimetry confirmed specific binding properties of enriched aptamers, with affinities in the micromolar range. Although the current capture-SELEX protocol used on the robotic system has limitations in the affinity of the enriched aptamers compared to immobilization-based selections, it offers a standardized, high-throughput method for the rapid assessment of the utility of aptamers for small molecules.
In the second part of this work, the capture-SELEX strategy was extended to enable the selection of modular allosteric RNA sensors. RNA libraries were designed to couple ligand-binding aptamers to fluorogenic RNA scaffolds, enabling molecular recognition to be directly translated into fluorescence output. Using this library in capture-SELEX, aptamers that bind thiamine pyrophosphate (TPP) were identified. These aptamers were then fused to Broccoli and its red-shifted variant, Red Broccoli, to develop ligand-responsive sensors. By optimizing the linker and spacer regions, the signal-to-background ratio was improved. The selected sensors demonstrated specificity for TPP and thiamine monophosphate, with minimal response to thiamine or unrelated nucleotides. The modular design facilitated the easy swapping of fluorogenic domains and the adjustment of sensor properties, demonstrating its effectiveness for rapid and efficient development of RNA-based sensors.
Overall, this thesis establishes a scalable framework for RNA aptamer discovery and deployment. By integrating automated selection with modular sensor engineering, the presented approach provides a generalizable strategy for developing RNA-based sensing systems applicable to biosensing, synthetic biology, and molecular diagnostics.

Intuitive Explainable Artificial Intelligence for Molecular Design

2026-06-09T08:00:56Z

Intuitive Explainable Artificial Intelligence for Molecular Design Lamens, Alec With the onset of artificial intelligence (AI), machine learning (ML) approaches have become increasingly embedded into the drug discovery pipeline with the aim of accelerating research and reducing cost. Exemplary applications include quantitative structure-property relationship modeling and de novo design of candidate molecules. However, the practical impact of these ML models has been limited by the opaque nature of their predictions, termed the black-box problem. The proliferation of large-scale public chemical datasets combined with advances in high-performance computing has driven a shift from traditional ML approaches toward more complex deep learning (DL) methods. While powerful, these models further exacerbate the inability to rationalize model outcomes. Therefore, the field of explainable AI (XAI) has attracted substantial interest to improve the transparency of model decisions. Given the wide range of available XAI approaches and lack of established practices, these methods are often applied without careful consideration for their suitability to the task at hand. This represents a problem in drug discovery where interpretable explanations are critical to guide experimental follow-up. To address this issue, this dissertation proposes the adaptation of established XAI approaches to operate on the level of chemical structure and facilitate interpretable ML for molecular design. The first study demonstrates the use of Shapley additive explanations (SHAP) to analyze feature sets driving the correct prediction of compound promiscuity. This is followed by a benchmarking study comparing several methods for the calculation of feature importance values, revealing inconsistencies across approaches that complicate the interpretation of feature-importance explanations. To support chemically meaningful interpretability, SHAP values are then combined with counterfactual reasoning, yielding the SHAP-CF methodology. Thereafter, a novel counterfactual method is introduced that generates candidate counterfactuals through the iterative recombination of molecular cores and substituents. This method is evaluated on a challenging multi-class kinase inhibitor prediction task, demonstrating its capacity to generate large numbers of counterfactuals, enabling in-depth analysis. Next, a closely related XAI concept, contrastive explanations, is adapted for molecular design by introducing rational chemical modifications to test compounds through the exchange of scaffolds or substituents with structural analogues, termed MolCE. This approach is shown to provide chemically relevant contrastive explanations that enable direct causal insights into the structural factors that drive model predictions. Finally, MolAnchor is introduced, a domain-adapted variant of the rule-based Anchors method. By conforming the principles underlying the Anchors framework to operate on retrosynthetically meaningful molecular fragments, MolAnchor produces interpretable fragment-level decision rules. A follow-up study showcases how the if-then formulation of the fragment rules is particularly complementary to causal reasoning. Taken together, the studies presented hereindemonstrate the utility of XAI methods that have been augmented with domain-specific information to provide chemically intuitive explanations for predictive models applied to molecular design.