Lakewater microorganisms secrete proteases which contribute to the turnover of dissolved organic matter and the degradation of peptidic contaminants. However, little is known about the identities and substrate specificities of these proteases. Herein, we sought to characterize the global proteolytic fingerprint of the extracellular proteases present in Lake Geneva, the largest freshwater body in Central Europe. Using Multiplex Substrate Profiling by Mass Spectrometry (MSP-MS), we identified preferred enzymatic cleavage next to positively charged and certain nonpolar amino acids, while cleavage next to negatively charged residues was disfavored. Specifically, the detected dominant cleavage sites were surrounded by arginine and lysine, consistent with a trypsin-like substrate specificity. This pattern was conserved across seasons and water depths and was shared with two other Swiss lakes. In contrast, we observed variability in the numbers and types of less prevalent cleavage sites across samples, suggesting that the degree of heterogeneity in proteolytic substrate specificity varies spatially and temporally. Using class-specific inhibitors, we found that serine and metalloproteases contribute to both exo-and endoproteolytic activity in lakewater. Our findings expand our understanding of protein stability in lake ecosystems and may be used to predict the fate of peptidic contaminants in the environment.

AI assistants, such as ChatGPT, are being increasingly used by students in higher education institutions. While these tools provide opportunities for improved teaching and education, they also pose significant challenges for assessment and learning outcomes. We conceptualize these challenges through the lens of vulnerability, the potential for university assessments and learning outcomes to be impacted by student use of generative AI. We investigate the potential scale of this vulnerability by measuring the degree to which AI assistants can complete assessment questions in standard university-level Science, Technology, Engineering, and Mathematics (STEM) courses. Specifically, we compile a dataset of textual assessment questions from 50 courses at the École polytechnique fédérale de Lausanne (EPFL) and evaluate whether two AI assistants, GPT-3.5 and GPT-4 can adequately answer these questions. We use eight prompting strategies to produce responses and find that GPT-4 answers an average of 65.8% of questions correctly, and can even produce the correct answer across at least one prompting strategy for 85.1% of questions. When grouping courses in our dataset by degree program, these systems already pass the nonproject assessments of large numbers of core courses in various degree programs, posing risks to higher education accreditation that will be amplified as these models improve. Our results call for revising program-level assessment design in higher education in light of advances in generative AI.

Ultrafast transient spectroscopy enables disambiguation between integral and contaminant chlorophyll (Chl) in photosynthetic protein complexes.

Photocatalytic water splitting has drawn considerable attention as an alternative way to produce renewable energy rather than relying on fossil fuels. Although extensively studied from both engineering and scientific perspectives, this process still has some issues that need to be addressed. Since the first reported photocatalytic water splitting by titanium dioxide, this material remains one of the most promising photocatalysts, due to its suitable band gap and band-edge positions. However, predicting both of these properties is a challenging task for existing computational methods. Density functional theory --- a workhorse of materials science more generally --- lacks the ability to obtain the right electronic gap as well as excitation energies. It is not just the band gap that we need to get right: we need to capture band alignment and even more complicated physics such as charge carriers and polaronic states. All of these properties can pose a challenge to a semi-local DFT. Higher levels of theory such as GW or hybrid functional provide better results, but at increased computational cost and/or scaling. Therefore, here I present novel computational strategies based on Koopmans spectral functionals with the idea to break new ground when it comes to light-driven reactions. These orbital-density dependent functionals are proven to be reliable when it comes to calculating the spectral properties of materials, such as the ionization potentials and electron affinities of molecules, as well as the band gaps of solids. Furthermore, they predict reliable trends when it comes to studying band alignments with and without accounting for solvation.

In this thesis, I present the main advantage of this method over standard DFT in addressing the spectral properties and give some ideas regarding the potential use of this method in studying the role of polarons in photocatalytic processes. The research is focused on the prototypical photocatalyst TiO2. Overall, this work demonstrates the capabilities of Koopmans functionals in terms of modeling photocatalytic materials, bridging the accuracy and computational cost. While this method seems promising for spectral properties, challenges remain in modeling polaron localization. Addressing these limitations --- in particular, the role of screening parameters in systems energetics and alternative electronic minimization strategies --- will be a key focus of future work.
More broadly, this thesis opens new avenues for modeling photocatalytic materials with implications for solar fuels and energy conversion.

Ischemic stroke is a leading cause of death and disability worldwide. It is an emergency condition in which a blood clot obstructs an artery in the brain, leading to progressive brain damage. In recent years, a minimally-invasive endovascular procedure - mechanical thrombectomy - has been developed for the extraction of the thrombus. The procedure has proved to be effective in a large number of cases, however, with variable delays resulting in variable clinical outcomes. The delay comes from multiple factors - difficult clot penetration, unadapted choice of thrombectomy technique, or wrong arterial branch in which the technique is performed. In this thesis, done in collaboration with the team of Prof. Paolo Machi from the University Hospitals of Geneva, we build foundation blocks for the development of an efficient and safe robotic mechanical thrombectomy. We tackle the challenges of clot penetration and identification, and we propose the development of a teleoperated endovascular robotic system, which aims to provide several advantages with respect to the conventional manual procedure - a safer and more ergonomic environment for the interventionalist, potential for strategies with shared control between human and robot, improved precision and improved force sensitivity.

We developed three endovascular robotic devices for guidewire navigation and one haptic leader device. We evaluated the haptic device in combination with one of the guidewire navigating robotic devices in an in vitro experiment with three experts in mechanical thrombectomy. Three robotic control modalities were tested, and even though several improvements are necessary particularly regarding the force feedback implementation, interventionalists recognized the system's potential to enhance not only mechanical thrombectomy but also a range of other neuroendovascular procedures.

We investigated different clot penetration strategies and analyzed the information that can be extracted from the measurement of the interaction forces between the guidewire and the clot. In particular, with a strategy of constant translational velocity, we demonstrated that the interaction forces depend on factors such as the volume of the clot, the stiffness of the clot, and the pressure with which the clot is impacted in the artery. We concluded that additional factors not investigated enough in this thesis, such as the fracture toughness of the clot, may be essential to fully explain the observed force profiles. We also developed an autonomous clot penetration strategy combining translational and rotational movements according to a force-based law. This strategy proved to be efficient in penetrating clots with different consistencies both in a straight arterial model and in a realistic vascular model.

Finally, we developed methods for the identification of the beginning and the end of the clot. We focused on a force-based method with a force measurement at the proximal end of the guidewire, but we also explored the possibility to implement a distal force measurement. Both measurements bring important information, with the distal measurement appearing very promising for the detection of the extremities. Additionally, we developed a method based on the trace of the guidewire during penetration, which is suitable for a penetration strategy employing rotation. In the future, this vision-based method can be extended to suit possibly other penetration strategies.

This thesis focuses on the analysis and reproduction of integral experiments for validating stainless-steel nuclear data, through two complementary programs conducted in the CROCUS zero-power reactor at EPFL. These programs are the semi-integral PETALE program, which includes criticality and transmission experiments with heavy reflectors, and the hybrid pile-oscillation program BLOOM. Both belong to HARVEST X, an EPFL coordinated initiative to reproduce, extend, and cross-validate experiments for stainless-steel and its Fe-Ni-Cr components. The main application concerns heavy reflectors in light-water reactors, with broader relevance to reactor pressure vessel fluence, Generation IV designs (especially fast reactors), and fusion systems. With PETALE, the aim is to deliver benchmark-quality results, including comprehensive covariance data and reduced risks of elemental compensation in alloys using pure elemental reflectors. A further goal is to provide feedback on modern nuclear data libraries, notably JEFF-3.3 and the newly released JEFF-4.0, and to prepare for data assimilation. To achieve these objectives, a new dosimetry analysis framework was developed, enabling the quantification of correlations among its 480 activation dosimetry measurements. This approach supports robust estimation of C/E ratios and their renormalization, including full covariance propagation, without ad hoc assumptions. High-resolution modelling with JEFF-3.3 and JEFF-4.0 confirms improved iron data in the fast neutron range. Trends observed in C/Es with JEFF 3.3 - front-to-back reflector differences of 5.7 % at ~2 MeV and 6.5 % at ~3.6 MeV - are reduced to below 1% with JEFF-4.0. Overall, stainless-steel performance improves as well, although the new chromium and nickel evaluations perform less favorably. These results highlight the need for careful treatment of alloys with reduced iron concentrations compared to typical light water reactor grades. Preliminary assimilation attempts with JEFF-3.3 covariance data suggest a 4-8% increase of the 56Fe inelastic cross section would improve agreement with fast-range observations. The BLOOM program was designed and carried out to complement PETALE, focusing on pile-oscillation experiments. Conducted in 2024, it involved 45 samples oscillated using a dedicated experimental channel, the SAFFRON array, and current mode-operated fission chambers. The measured reactivity worth have uncertainties as low as 0.015 pcm, the theoretical pile-noise limit. Local flux perturbations of ~10% (up to 30%) were measured with relative uncertainties of 0.1-1% (down to 0.05%). Simulations of 30 oscillations using Serpent2 and the black-body exact perturbation method yield complementary results to PETALE's transmission data. Chromium sample results improve with JEFF-4.0 (C/E are +1%, +3.5% previously), whereas nickel results slightly degrade (+1% bias vs. <0.5%). The next step of HARVEST-X involves the ongoing benchmarking of PETALE within ICSBEP. Additional experimental follow-ups are planned, notably the accelerator-based GRAPE campaign. Further oscillations and analysis are also in preparation for BLOOM, including validation using local signals, and higher reactor power oscillations to improve precision. Complete data assimilation studies for dedicated applications are still to be performed, in collaboration with the Paul Scherrer Institute.

Geospatial information is embedded in both natural systems and human society, making it inherently multimodal and heterogeneous. It describes objects and events associated with geospatial locations on Earth, with various forms such as remote sensing imagery, ground-level imagery, and textual descriptions.

With the rise of geospatial artificial intelligence (AI), multimodal learning has become a key to integrating these diverse sources. However, the inherent heterogeneity across modalities poses significant challenges, constraining both the development and generalizability of geospatial AI systems. Considering the sparsity of aligned data across modalities, external knowledge offers a potential solution. By providing modality-agnostic semantics and priors and guiding the multimodal learning process, it plays a critical role in bridging modality gaps.

Despite efforts to incorporate external knowledge into multimodal systems, the current understanding of how external knowledge integration influences multimodal complementarity, alignment, and representation remains limited, particularly within geospatial contexts. To address these challenges, this dissertation investigates knowledge-aware multimodal systems that integrate diverse types of knowledge, including commonsense and geospatial knowledge, across multiple modalities with a focus on geospatial vision-language and geo-localization tasks.

For geospatial vision-language applications, this dissertation investigates the complementarity of external knowledge to multimodal learning. To improve cross-modal alignment, KTIR incorporates external commonsense knowledge to bridge semantic gaps between remote sensing imagery and textual descriptions by expanding the semantic scope of text. In addition, ConVQG integrates external knowledge into multimodal fusion by proposing a contrastive visual question generation pipeline that generates knowledge-enriched, image-grounded questions.

In the context of geo-localization, this dissertation develops generalizable representations for both retrieval-based and navigation-based approaches by incorporating external knowledge in the learning process. For retrieval-based geo-localization, ConGeo proposes a model-agnostic pipeline that integrates the location prior in the contrastive learning objective to achieve robust cross-modal alignment between ground and aerial views. Building on this alignment, GeoExplorer further addresses the generalizability of representation in active, navigation-based geo-localization and proposes a curiosity-driven reinforcement learning pipeline to emphasize comprehensive environment modeling and exploration for geo-localization.

Taken together, the dissertation highlights the multifaceted complementarity of external knowledge in geospatial multimodal information and introduces knowledge-aware systems that leverage such knowledge to improve cross-modal alignment and enhance the generalizability of geospatial representations. With these contributions, the dissertation aims to foster knowledge-enriched, robust, and practically applicable geospatial AI systems that leverage the distinctiveness of geospatial information and accommodate real-world complexity.

The electron-positron stage of the proposed Future Circular Collider (FCC-ee) aims to achieve unprecedented precision at large centre-of-mass energies, serving as a high-luminosity Higgs, top, and electroweak factory that goes beyond what was enabled by the 2012 Higgs-boson discovery at the Large Hadron Collider (LHC) at CERN. As a lepton collider, FCC-ee is strongly constrained by synchrotron radiation power across its four operating modes: the energy loss per turn must be replenished by the radio-frequency system while respecting the nominal 50 MW per-beam limit.

This thesis explores the use of nested magnets, which introduce a dipole component into the quadrupoles and sextupoles of the arc lattice to reduce the per-turn energy loss from synchrotron radiation. At the FCC-ee scale, this is a novel design choice that poses non-trivial optics challenges. We develop and compare several lattice families to address these issues, ultimately converging on a Common Layout Configuration that shares a unified design across the operational energies. The resulting optics deliver a robust reduction in synchrotron radiation losses, which can be traded for consumption power savings or, alternatively, converted into luminosity gains; these gains are confirmed for both of the currently considered optics, despite their very different interaction region designs and chromatic correction schemes. In addition to the production optics, we introduce a simplified ballistic optics, which turned out fundamental for the commissioning phase and for studies of errors under relaxed optics constraints. Finally, we assess the impact of systematic quadrupolar errors in the arc dipoles.