Decentralized learning (DL) is an emerging paradigm of collaborative machine learning that enables nodes in a network to train models collectively without sharing their raw data or relying on a central server. This paper introduces Zip-DL, a privacy-aware DL algorithm that leverages correlated noise to achieve robust privacy against local adversaries while ensuring efficient convergence at low communication costs. By progressively neutralizing the noise added during distributed averaging, Zip-DL combines strong privacy guarantees with high model accuracy. Its design requires only one communication round per gradient descent iteration, significantly reducing communication overhead compared to competitors. We establish theoretical bounds on both convergence speed and privacy guarantees. Moreover, extensive experiments demonstrating Zip-DL's practical applicability make it outperform state-of-the-art methods in the accuracy vs. vulnerability trade-off. Specifically, Zip-DL (i) reduces membership-inference attack success rates by up to 35% compared to baseline DL, (ii) decreases attack efficacy by up to 13% compared to competitors offering similar utility, and (iii) achieves up to 59% higher accuracy to completely nullify a basic attack scenario, compared to a state-of-the-art privacy-preserving approach under the same threat model. These results position Zip-DL as a practical and efficient solution for privacy-preserving decentralized learning in real-world applications.

Band offsets at semiconductor-oxide interfaces are determined through a scheme based on hybrid density functionals, which incorporate a fraction alpha of Hartree-Fock exchange. For each bulk component, the fraction alpha is tuned to reproduce the experimental band gap, and the conduction and valence band edges are then located with respect to a reference level. The lineup of the bulk reference levels is determined through an interface calculation, and shown to be almost independent of the fraction alpha. Application of this scheme to the Si-SiO2, SiC-SiO2, and Si-HfO2 interfaces yields excellent agreement with experiment.

Using density-functional calculations, we investigate the possibility of ammonia synthesis at supported Fe nanoclusters along catalytic routes closely resembling those in biological nitrogen fixation. To achieve similar catalytic conditions as at the active site of the enzyme nitrogenase, the clusters are passivated with either S or N atoms. From calculated potential-energy profiles for the N-2 hydrogenation, we find that low-temperature synthesis of ammonia is viable at the clusters passivated by N atoms due to the strong binding energy of the N-2 molecule in the initial adsorption step.

When described through a plane-wave basis set, the inclusion of exact nonlocal exchange in hybrid functionals gives rise to a singularity, which slows down the convergence with the density of sampled k points in reciprocal space. In this work, we investigate to what extent the treatment of the singularity through the use of an auxiliary function is effective for k-point samplings of limited density, in comparison to analogous calculations performed with semilocal density functionals. Our analysis applies, for instance, to calculations in which the Brillouin zone is sampled at the sole Gamma point, as often occurs in the study of surfaces, interfaces, and defects or in molecular-dynamics simulations. In the adopted formulation, the treatment of the singularity results in the addition of a correction term to the total energy. The energy eigenvalue spectrum is affected by a downwards shift in the energy eigenvalues of the occupied states, while those of the unoccupied states remain unaffected. Analogous corrections also speed up the convergence of screened exchange interactions despite the absence of a proper singularity. Focusing first on neutral systems, both finite and extended, we show that the account of the singularity corrections bears convergence properties which are quantitatively similar to those observed with semilocal density functionals. We emphasize that this is not the case for uncorrected energies, particularly for elongated simulation cells for which qualitatively different trends are found. We then consider differences between total energies of systems differing by their charge state. For systems involving localized electron states, such as ionization potentials and electron affinities of molecular systems or charge transition levels of point defects, the proper account of the singularity correction yields convergence properties which are similar to those of neutral systems. In the case of extended systems, such energy differences provide an alternative way to determine the band edges, but are found to converge more slowly with simulation cells than in corresponding semilocal functionals because of the exchange self-interaction associated to the extra charge.

Animals, including humans, interact with the external environment primarily through motion. Replicating their motor control skills in artificial embodied agents is a major objective of artificial intelligence research. This thesis presents a collection of studies focused on different aspects of artificial embodied intelligence, linked by one common underlying research question: the learnability of human-level motor control policies. Using biologically realistic computational models of the human musculoskeletal system, we can study motor skill learning and adaptation in simulation with unprecedented detail and efficiency. Advanced biomechanical simulators allow us to train policies that face the same hurdles as animals and humans, namely, dealing with a complex, high-dimensional system such as the human body. Throughout this thesis, we addressed the problems of adaptation and skill acquisition. In particular, we considered which inductive biases enable policy networks to deal with variable body shapes and adapt a locomotion strategy in real time. To this end, we devised DMAP, a policy network implementing principles of biological motor control, to facilitate extracting a representation of the agent's body from sensory input. In the second part of the thesis, we shifted our focus to motor control via muscle actuators. Through curriculum learning, a framework in which an agent faces progressively more difficult tasks, we trained policies to control realistic models of the human arm and perform dexterous object manipulation. These policies conquered the first two NeurIPS MyoChallenges, demonstrating for the first time that artificial neural networks can control a model of the human arm and rotate, grasp and throw objects. Solving these complex problems required innovating over existing learning algorithms. We devised Lattice, an exploration strategy to deal with the complexity of musculoskeletal environments with tens of degrees of freedom and actuators. Pairing Lattice with reinforcement and imitation learning, we obtained policies achieving human-level object manipulation and locomotion. We used these policies to compare the number of control dimensions that are necessary to perform different tasks, finding that they exceed previous estimations and that the low-dimensional control spaces transfer poorly across tasks. Finally, we distilled all the policies trained in these works into Arnold, a single, generalist motor control policy, with the flexibility to deal with multiple parts of the human body. Overall, our work introduced new training and analysis techniques that will facilitate future studies of the human motor control system, where data from human subjects can be complemented by analysis in simulation.

Spin waves (SWs) are collective excitations of spins in magnetically ordered materials. The quantized form of a SW is known as a magnon. SWs excited in ferrimagnetic insulators such as yttrium iron garnet (YIG) can transmit angular momentum over distances up to millimeters and avoid dissipative charge transport. They exhibit frequencies of several GHz, close to and beyond the clock frequencies of the state-of-the-art processors making them attractive candidates for an efficient data processing. Magnonics is a research field that explores the excitation, propagation, detection and manipulation of SWs. Recently, magnetization switching of Ni81Fe19 (permalloy or Py) elements induced by SWs was demonstrated enabling the realization of all SW-based computing devices. In this thesis, we address the SW-assisted magnetization reversal or switching of Py magnetic nanostripes as narrow as 50 nm. We excited propagating and interfering SWs having wavelengths between 7 microns down to 80 nm and studied in detail their interaction with nanomagnetic magnon memory bits (NMMBs). We assessed the ability of various magnon modes propagating in 100 nm-thick-YIG film to reverse arrays of Py nanostripes having widths down to 50 nm placed 25 microns away from the SW excitation antenna using all electrical spin wave spectroscopy. Propagating SWs reduced the critical switching fields of all the nanostripe arrays across the studied combinations of nanostripe widths and periods. Propagating ultrashort SWs having a wavelength of 101 nm reversed nanostripes at a precessional power Pprec of 10.6 nW, which is 6.5 times lower than the Pprec needed in case of 7 micron-long SWs. Secondly, we demonstrated the magnetization reversal of differently shaped 200 nm-wide and 800 nm-long Py NMMBs under SW excitation using magnetic force microscopy and micromagnetic simulations. Elliptical NMMBs reversed their magnetization at 10 dB lower SW excitation power than the rectangular NMMBs. We observed changes in the transmission spectra of short-wavelength plane-wave SWs after propagation over 100 microns beneath an array of over 5000 NMMBs programmed in various magnetic states. Simulations revealed that the reversal of an NMMB proceeded via the creation and propagation of vortices. The simulated switching time by vortex movement was below 4.2 ns. We observed a stochastic nature of the magnetization reversal. It was mainly controlled by the incubation delay before nucleating inhomogeneous spin textures that led to the vortex creation. We estimated that the energy needed to switch several elliptical NMMBs could reach down to 2 fJ considering the SW precessional power from experiments and the switching time obtained from simulations. Our microfocus Brillouin light scattering experiments demonstrated that NMMBs were selectively switched using SW interference. The NMMBs imprinted the SW interference pattern with unswitched NMMBs denoting the expected node positions. Our studies with NMMB switching assisted by SW pulses corroborated the stochasticity observed in the simulations and showed that the SW pulse period and the number of switched NMMBs followed an inverse relation. Lastly, X-ray imaging enabled us to study the microscopic interaction of short-wavelength SWs with NMMBs. We observed large phase delays for SWs propagating underneath the NMMBs leading to interference. Our observations are pivotal in realizing efficient all-magnon-based nanoscale memory devices and neural network.

One of the main milestones of the EUROfusion roadmap to fusion is the demonstrator reactor DEMO. Like ITER, the EU DEMO will be a large-scale tokamak with superconducting toroidal field (TF) coils to generate a strong magnetic field for plasma confinement. Even though no voltage should arise across the superconducting TF coils during DC operation, a large voltage develops during a fast safety discharge, due to the rapid exponential decay of the current. This discharge process is necessary to protect the coil in case of a quench (sudden loss of superconductivity). Considering the magnets in a large-scale tokamak like DEMO store approximately 10 GJ of energy, the maximum discharge voltage of one of these coils can reach tens of kilovolts, posing significant challenges. In the vacuum environment where the TF coils operate, such a high voltage requires exceptional insulation of the magnet to prevent Paschen breakdown, which could cause severe damage to the coils. This risk can be substantially mitigated by reducing the discharge voltage of a single TF coil below 5 kV. To reach this goal, a high-current Nb3Sn winding pack (WP) was designed for the TF coils of the EU DEMO. The proposed WP has an operating of ~105 kA, increased from the EU DEMO baseline of 66 kA. This higher operating current allows to reduce the coil inductance by reducing the number of turns from 226 in the nominal design to 142, while maintaining the total ampere-turn in the magnet. Due to the strong dependency of the discharge voltage on the operating current and the number of turns, the proposed design has a resulting discharge voltage of 4.23 kV, much smaller than the one of the nominal design 6.7 kV. Furthermore, the proposed high-current WP design of the DEMO TF coils also makes use of the react&wind (RW) technique and layer winding, allowing for grading, which ensures the optimal use of steel and Nb3Sn in each layer of the winding pack. The work conducted during this thesis proves that the proposed high-current WP results in a radial gain of almost 400 mm compared to the reference WP design. Before this thesis, the highest current conductor proposed for the EU DEMO operated at ~82 kA. Since this is the first time that a conductor exceeding 100 kA is considered for the EU DEMO, a full-scale conductor prototype, named RW4, was developed and tested in SULTAN. This test was used to validate its design principles by performing a DC and an AC characterization of the cable-in-conduit-conductor (CICC) foreseen for the first layer of the TF WP. The DC characterization of the RW4 conductor prototype was affected by sudden quenches. A significant part of this thesis is devoted to a detailed evaluation of the causes behind this unstable behavior. An in-depth study of the AC losses confirmed that the proposed RW design developed at SPC allows to design of low-loss CICCs.

The thesis will discuss how small objects such as bacteria and viruses can be manipulated on-chip using light via Photonic crystals (PhCs). In particular, PhCs could be used in the context of novel fast antimicrobial susceptibility tests (ASTs). We will see how point-like defects in the lattice structures form resonant cavities. The supported optical mode of these cavities is extremely spatially confined, thus generating large field gradients that allow the trapping of small polarisable objects (10 - 5000 nm) in their proximity. In contrast to classical optical tweezers, photonic crystal cavities do not require a bulk lens system for focusing and are, therefore, not limited by diffraction. Furthermore, when an object is trapped, the resonance is modified according to the refractive index and size of the object. Thus, they are true sensors capable of autonomously trapping their own analytes. The PhCs will be integrated into a microfluidic device to direct suspended particles to the cavity location for trapping. The so-formed optofluidic chip is mounted on an end-fire transmission setup to excite the cavity mode and read out its perturbation. The thesis defines the first steps towards more versatile and faster AST using photonic crystal cavities. Indeed, the rising threat of AMR is pushing the community to develop new techniques to overcome the long time to results challenge of currently accepted AST techniques. In this work, we exploit resonant 2D photonic crystal cavities to trap and sense single bacteria when healthy or exposed to antibiotics or bacteriophages. The technique allows the real-time monitoring of bacterium-antimicrobial interactions independently from bacterial growth.

Computational simulations of quantum molecular dynamics are indispensable for understanding and predicting phenomena in physical chemistry. This thesis explores molecular processes where multiple electronic states play a role. A significant part of this work is dedicated to studying charge migration, which is initiated by ionizing a molecule into a superposition of electronic states. The decoherence induced by the nuclei is often neglected, leading to predictions that could be challenging to observe experimentally. Here, we use semiclassical dynamics to include nuclear motion and propose an algorithm to efficiently identify molecules with long-lasting charge migration triggered by valence ionization. We highlight several previously unexplored molecules that exhibit long-lasting charge migration, suggesting them as promising candidates for future experimental studies.

Applications in attochemistry are based on molecules in which electronic motion can be controlled long enough to direct chemical reactivity. Advancing our understanding of the fundamental principles behind electronic decoherence due to electron-nuclear correlations is essential for designing such compounds. To address this, we compare charge migration in a series of structurally similar organic molecules of increasing size and flexibility. Surprisingly, extending the carbon skeleton in propynal prolongs the duration of charge migration. Semiclassical dynamics provides valuable insight into the mechanisms of decoherence. In particular, the overall decoherence can be decomposed into contributions from individual nuclear vibrations without additional calculations, enabling the identification of the normal modes responsible for the observed prolonged coherence.

The use of semiclassical methods neglecting nonadiabatic effects to describe dynamics involving multiple electronic states is limited to exceptional cases where the coupling between these states is negligible. In general, more advanced propagation methods are required to account for nonadiabatic effects. We develop and apply two different new mixed quantum-classical methods, based on Ehrenfest dynamics, that partially capture nuclear quantum effects without requiring the propagation of independent trajectories.

First, we present applications of the thawed Gaussian Ehrenfest dynamics, a method that combines single-trajectory Ehrenfest dynamics with the thawed Gaussian wavepacket dynamics. We highlight a significant limitation shared by other single-trajectory mean-field methods preventing electronic population transfer induced by conical intersections between electronic states belonging to different irreducible representations. In contrast, the thawed Gaussian Ehrenfest dynamics provide a good qualitative picture of the dynamics in the vicinity of conical intersections between electronic states of the same symmetry.

Second, we introduce SPEED, a variation of multitrajectory Ehrenfest dynamics, where all trajectories are propagated using a common time-dependent quadratic effective potential in the diabatic representation. This approach is equivalent to multitrajectory Ehrenfest dynamics when the diabatic potential energy surfaces and couplings are at most quadratic. We applied SPEED to study nonadiabatic dynamics involving conical intersections, atomic chemisorption on solid surfaces, and charge transfer between molecules, demonstrating its capability to efficiently and accurately describe various types of systems.