Electromagnetic Time Reversal (EMTR) features the bounded phase property, a well-established and reliable method for fault location in two-conductor transmission lines. Under the EMTR-based framework for mismatched media, this property asserts that the phase of the overall forward and backward transfer function remains strictly confined at the true fault location. Extending this property to multi-conductor lines is challenging because the transfer function ceases to be a scalar and instead becomes a matrix. Nonetheless, in this paper, we demonstrate that the bounded phase property applies to all eigenvalues of the transfer-function matrix, thereby providing a clear pathway for its application to multi-conductor lines.

Large image generation and vision models, combined with differentiable rendering technologies, have become powerful tools for generating paths that can be drawn or painted by a robot. However, these tools often overlook the intrinsic physicality of the human drawing/writing act, which is usually executed with skillful hand/arm gestures. Taking this into account is important for the visual aesthetics of the results and for the development of closer and more intuitive artist-robot collaboration scenarios. We present a method that bridges this gap by enabling gradient-based optimization of natural human-like motions guided by cost functions defined in image space. To this end, we use the sigma-lognormal model of human hand/arm movements, with an adaptation that enables its use in conjunction with a differentiable vector graphics (DiffVG) renderer. We demonstrate how this pipeline can be used to generate feasible trajectories for a robot by combining image-driven objectives with a minimum-time smoothing criterion. We demonstrate applications with generation and robotic reproduction of synthetic graffiti as well as image abstraction.

Evaluating explainability techniques, such as SHAP and LRP, in the context of audio deepfake detection is challenging due to lack of clear ground truth annotations. In the cases when we are able to obtain the ground truth, we find that these methods struggle to provide accurate explanations. In this work*, we propose a novel data-driven approach to identify artifact regions in deepfake audio. We consider paired real and vocoded audio, and use the difference in time-frequency representation as the ground-truth explanation. The difference signal then serves as a supervision to train a diffusion model to expose the deepfake artifacts in a given vocoded audio. Experimental results on the VocV4 and LibriSeVoc datasets demonstrate that our method outperforms traditional explainability techniques, both qualitatively and quantitatively.

Automatic speech recognition (ASR) systems struggle with dysarthric speech due to high inter-speaker variability and slow speaking rates. To address this, we explore dysarthric-to-healthy speech conversion for improved ASR performance. Our approach extends the Rhythm and Voice (RnV) conversion framework by introducing a syllable-based rhythm modeling method suited for dysarthric speech. We assess its impact on ASR by training LF-MMI models and fine-tuning Whisper on converted speech. Experiments on the Torgo corpus reveal that LF-MMI achieves significant word error rate reductions, especially for more severe cases of dysarthria, while fine-tuning Whisper on converted data has minimal effect on its performance. These results highlight the potential of unsupervised rhythm and voice conversion for dysarthric ASR. Code available at: https://github.com/idiap/RnV.

Liquid cooling in thermal management systems has been implemented mostly through the use of thermal interface materials, which introduce in in-series parasitic thermal resistances and reduced thermal performance. To achieve the full potential of liquid cooling, cooling structures should be fabricated directly in-chip, enabling effective heat transfer near the junction. In this work, we investigate an integrated cooling module encompassing enhanced fin structures (diamond-shaped and square pin fins) with an impinging jet manifold. Low thermal resistance of 0.1 cm2.K/W was achieved, while consuming a record low pumping power of 40μ W/mm2 using 25 μm channel width. Our approach was compared with bulky traditional fan cooling and no cooling cases, and resulted in 6.6x and 62 x reduction in thermal resistance, respectively. The cooling structure was directly fabricated at the back side of a GaN power IC which resulted in a thermal resistance of 1 K/W at 10 mW pumping power, 6.6 K/W thermal resistance was realized using fan cooling. Furthermore, 37 % increase in dynamic on-resistance (Ron) was reported at 10 W power dissipation using fan cooling, which was reduced to 12 % when using our approach at 15 W power dissipation.

The continuous rise in atmospheric carbon dioxide (CO2) level poses a major environmental and technological challenge, driving the urgent need for scalable CO2 utilization strategies. This thesis explores functional molecules based on main group elements as a versatile platform for sustainable CO2 conversion and energy device stabilization. In chapter 1, the scientific motivation and current limitations of CO2 capture and utilization are discussed, with emphasis on catalytic valorization approaches such as the cycloaddition of CO2 to epoxides and hydrogenation to value-added products. The role of main-group element chemistry and Frustrated Lewis Pairs (FLPs) in overcoming the challenges associated with transition metal catalysts is introduced, alongside the potential of functional molecular additives in stabilizing next-generation photovoltaic devices such as perovskite solar cells (PSCs). In Chapter 2, a metal-free catalytic system comprising thermally activated silica gel and tetrabutylammonium iodide (TBAI) is developed for the cycloaddition of CO2 to epoxides. This catalyst enables the production of cyclic carbonates with high yields. A continuous-flow reactor was designed and optimized for industrial integration, operating efficiently under mild conditions and compatible with modeled industrial flue gas, with life cycle and economic assessments supporting its practical viability and sustainability. Chapter 3 introduces a Hammett parameter-guided strategy for designing triarylborane Lewis acids to enhance FLPs catalysis. By tuning electronic properties, a family of boranes with improved Lewis acidity was synthesized, achieving much higher turnover numbers for catalytic CO2 hydrogenation under metal-free conditions. Building on this, Chapter 4 reports the synthesis of a covalent triazine framework (B_CTF) embedding both triarylborane (Lewis acid) and triazine (Lewis base) units within a porous network. While the intrinsic FLPs activity of this B_CTF was limited by weak basicity, post-functionalization with an iridium complex restored catalytic activity, providing a proof-of-concept for heterogeneous FLP platforms. Chapter 5 expands the application of boron-based molecules beyond catalysis into energy materials. Functional triarylboranes and boronic acids were applied as additives in perovskite PSCs, either as passivating interfacial layers or precursor additives. These boron compounds stabilized the formamidinium cation and led to improvements in both device efficiency and long-term thermal stability. In summary, this work advances functional molecule design for CO2 valorization and energy technologies, offering scalable, metal-free approaches for tackling key challenges in the transition to a low-carbon future.

This thesis explores the application of artificial intelligence methods in digital health, clearly demonstrating the pathway from technological breakthroughs to practical health solutions in four critical areas: plant disease detection, food recognition, personalized glycemic response forecasting, and musculoskeletal modeling. Leveraging deep learning and reinforcement learning techniques, this work highlights the feasibility of practical, scalable AI-driven solutions to longstanding healthcare challenges.

First, a deep convolutional neural network trained on an extensive dataset successfully identifies plant diseases from leaf images, presenting an effective approach to smartphone-assisted agricultural diagnostics. Next, advanced instance segmentation models enable accurate food segmentation and recognition from real-world images, facilitating improved dietary assessments essential for nutritional epidemiology. Subsequently, Temporal-Fusion-Transformers accurately forecast individualized postprandial glycemic responses using continuous glucose monitoring data and nutritional information, highlighting personalized nutrition possibilities. Finally, deep reinforcement learning methods are utilized to synthesize physiologically accurate human movements within musculoskeletal simulations, demonstrating potential applications in biomechanics and rehabilitation.

The thesis further emphasizes the transformative potential of crowdsourced participatory research to accelerate AI innovations, reducing barriers to practical implementation and promoting rapid, tangible public health impacts.

This thesis addresses challenges in modeling and control of multi-agent systems for future urban mobility, where strategic competition, decentralized coordination, and uncertainty are inherent. Motivated by applications in electric ride-hailing markets and large-scale traffic monitoring, the work develops principled frameworks that combine tools from game theory, optimization, and learning-based control.

The first part investigates the management of electric ride-hailing fleets in competitive markets. We propose bi-level formulations for optimal electricity pricing, modeled as Stackelberg and Reverse Stackelberg games, and establish conditions for equilibrium existence and uniqueness. Building on these results, we design distributed algorithms with provable convergence to local Stackelberg equilibria, along with learning-based extensions using bandit methods and no-regret learning, which eliminate the need for explicit information sharing between operators and regulators. Beyond single-shot settings, we introduce a new class of multi-stage resource allocation games inspired by Tullock contests, modeling charging scheduling and fleet rebalancing decisions in systems where profitability depends on supply-demand balance. For this class, we prove the uniqueness of Nash equilibria, show that it generalizes receding-horizon and Blotto-type games, and derive analytical solutions in the latter case.

The second part focuses on cooperative traffic monitoring with fleets of aerial drones. We develop a Gaussian process-based framework that fuses historical and real-time data to assign adaptive monitoring priorities across urban regions. This framework is integrated with centralized path-planning via distribution matching and decentralized strategies based on scalable multi-agent reinforcement learning. To account for temporal dependencies, we investigate recurrent neural architectures for implementing shared drone policies in adaptive patrolling.

Taken together, the contributions of this thesis advance the theoretical and algorithmic foundations of multi-agent control under competition, cooperation, and uncertainty. While grounded in applications to smart mobility, the developed models and methods extend to broader domains where strategic interaction and data-driven adaptation are key to effective system management.

We study the behaviour of electromagnetic waves in terms of their propagation through time and space.

First, while time typically flows in only one direction for complex systems, the reciprocal nature of most wave systems allows for a form of time reversal. Indeed, for a wave f(t, r), its time-reversed counterpart f(-t, r) can also physically exist. Moreover, methods have been developed to create this time-reversed wave, enabling the application of time reversal to wave focusing and imaging. In this context, we record the direct-time wavefront f(t, r) emitted by an unknown source on a (ideally closed) surface, known as a time-reversal mirror. This recording is then time reversed and played back from the mirror, causing the resulting wave to converge back to the original source location. If the original source is absent, the wave continues propagating, slightly blurring the focus and exhibiting the diffraction patterns observed in classical optics.

A wide range of methods has been developed to enable wave imaging and focusing beyond the diffraction limit, achieving "super-resolution." We begin with a review of the theory, methods, and applications of super-resolution imaging and focusing in the radio-frequency range. We then contribute to the theoretical framework of time-reversal by, first, proposing a new convergence metric based on a combination of probability theory and electromagnetic energy density, and second, generalising the theory of the time-reversal cavity for nonreciprocal media, high-order multipole sources, and proposing a link between the dispersion relation of homogeneous media and the attainable resolution. We subsequently present three applications of time-reversal imaging for electromagnetic compatibility pre-compliance testing aimed at imaging the current from an electrostatic discharge: i) using a resonant metalens, ii) leveraging cable radiation, and iii) with a low-cost setup.

In the second part of the thesis, we study the spatial behaviour of complex electromagnetic sources. This behaviour can be described analytically by the multipole expansion, a widely used method in wave systems. Traditionally, this method projects complex field distributions onto a basis of spherical harmonics. The strong connection between this method and Green's function warrants, first, an analysis of the singular behaviour of this function in uniaxial media. We then present the novel Cartesian time-domain multipole expansion, along with a practical recursive implementation. We illustrate the method on time-reversal imaging of intricate sources. Additionally, we apply this expansion to model the time-domain fields radiated by, first, impulse-radiating antennas, and second, lightning. Finally, we discover that the Cartesian approach is essential in anisotropic media to describe novel degrees of freedom in wave propagation.