Predicting the kinetic pathways of biomolecular systems at all-atom resolution is crucial for understanding protein function and drug efficacy, yet this task is hindered by the immense computational cost of conventional molecular dynamics (MD) simulations. While deep learning has revolutionized static structure prediction and equilibrium ensemble sampling, simulating the kinetics of conformational transitions remains a critical challenge. We introduce BioKinema, a physically grounded generative model that predicts continuous-time, all-atom biomolecular trajectories at a fraction of the cost of traditional simulations. In particular, BioKinema utilizes a scalable diffusion architecture with temporal attention mechanisms derived from Langevin dynamics. It employs a hierarchical forecasting-and-interpolation strategy to overcome the error accumulation that often plagues long-horizon generation. Through extensive validation, we demonstrate that BioKinema generates physically stable and dynamically accurate trajectories suitable for rigorous downstream analysis. The model captures key conformational transitions related to protein function. For protein-ligand complex systems, it successfully elucidates mechanisms such as induced-fit conformational changes and allosteric responses. Furthermore, BioKinema leverages enhanced sampling data to predict rare kinetic events, emerging as a powerful tool for estimating ligand unbinding pathways. Collectively, these results establish BioKinema as a robust alternative to MD that bridges the gap between static structure and dynamic function, enabling high-throughput exploration of the kinetic landscape for structural biology and drug discovery.

In this paper, we present a novel, high-resolution method, referred to as HRTR, for localizing underground objects. HRTR is based on a combination of the Time Reversal (TR) and Multiple Signal Classification (MUSIC) algorithms, and can be readily integrated with conventional ground-penetrating radar (GPR) systems without requiring any additional hardware. The proposed method offers significant advantages, particularly in achieving higher resolution, which enhances the ability to distinguish ground surface reflections and detect shallowly buried objects—challenges often encountered with conventional methods. The theoretical foundation of the proposed method is validated through numerical simulations using gprMax, as well as through experimental measurements from laboratory and field tests. The performance of HRTR is compared with conventional GPR methods, focusing on resolution improvements. Both simulations and experimental results demonstrate that HRTR produces clearer, sharper images with enhanced resolution. Unlike classical TR-MUSIC, the proposed HRTR method can be applied directly to conventional GPR measurements without the need for additional hardware or intensive computation.Moreover, it operates with just one antenna in monostatic mode or two in bistatic mode, avoiding the multiple-antenna requirement of TR-MUSIC. Furthermore, the proposed method enables the detection of deeply buried objects by using low-frequency signals for greater penetration while preserving spatial resolution. A graphical user interface was also developed and made available on GitHub for applying the proposed method to GPR A- and B-scans.

In addition to impressive performance, vision transformers have demonstrated remarkable abilities to encode information they were not trained to extract. For example, this information can be used to perform segmentation or single-view depth estimation even though the networks were only trained for image recognition. We show that a similar phenomenon occurs when explicitly training transformers for semantic segmentation in a supervised manner for a set of categories: Once trained, they provide valuable information even about categories absent from the training set. This information can be used to segment objects from these never-seen-before classes in domains as varied as road obstacles, aircraft parked at a terminal, lunar rocks, and maritime hazards.

In solid-state transformers (SSTs), as a promising conversion technology that integrates high-frequency galvanic isolation, a robust bidirectional power flow operation is essential, considering numerous SST real-world applications. Although an open-loop-operated resonant converter is suitable for the galvanic isolation stage in SSTs, a power reversal method (PRM) is required to diagnose the power direction transition and enable the current flow in the inverted direction by altering the modulation between two converter bridges, a process also known as active bridge switchover (ABS). In double-stage SSTs, in which the resonant conversion stage is combined with an active front-end (AFE) stage, the dynamic interactions of the two stages under power reversal are crucial for SST performance. Furthermore, to achieve a robust PRM in this special case, investigating the interactive behavior of the AFE closed-loop controller with the open-loop LLC converters in the detection and postdetection process is vital. Consequently, this article analyzes the operating principles of the LLC resonant converter in a two-stage SST under power flow reversal initially. Subsequently, a new PRM is proposed based on the unique interactive behavior of two stages in SST under power reversal. In this method, detection is based on monitoring the primary and secondary DC voltage slope rates of the LLC stages that exhibit opposite trends in the SST, and ABS is smoothly secured to avoid unwanted transients. This method does not rely on any additional or high-resolution sensors, while robust and fast switchover detection is achieved, which prevents false alarms caused by load and grid transients. The validity of the proposed method is established by experimental test results.

Utilizing 4.5 fb(-1) of e(+)e(-) annihilation data collected with the BESIII detector at the BEPCII collider at center-of-mass energies between 4.600 and 4.699 GeV, the first observation of the singly Cabibbo-suppressed decay Lambda(+)(c) -> p pi(0) is presented, with a statistical significance of 5.4 sigma. The ratio of the branching fractions of Lambda(+)(c) -> p pi(0) and Lambda(+)(c) -> p eta is measured as B(Lambda(+)(c) -> p pi(0))/B(Lambda(+)(c) -> p eta) = (0.120 +/- 0.026(stat) +/- 0.007(syst)). This result resolves the longstanding discrepancy between earlier experimental searches, providing both a decisive conclusion and valuable input for QCD-inspired theoretical models. A sophisticated deep learning approach using a Transformer-based architecture is employed to distinguish the signal from the prevalent hadronic backgrounds, complemented by thorough validation and systematic uncertainty quantification.

At the microscopic level, the behaviour of a solid is governed by the intricate interplay between electrons, the lattice, and, in some cases, spin degrees of freedom. From these interactions emerge collective excitations involving charge, lattice, and spin dynamics, known respectively as plasmons, phonons, and magnons. These quasiparticles and their mutual couplings can only be described within a quantum mechanical framework, from which arise a variety of exotic phases of matter. Light-matter interaction offers a means to perturb these phases and drive them out of equilibrium. By analysing the resulting dynamics of the collective modes, we can gain insight into their couplings and energy exchange mechanisms. Since electrons can transfer energy and momentum through inelastic scattering to collective modes, ultrafast electron diffraction (UED) provides a powerful tool to directly observe their out-of-equilibrium. This thesis investigates the ultrafast dynamics of collective excitations in quantum materials using time-resolved electron diffraction. It focuses on a selection of systems that host intriguing phases of matter, aiming to unravel the underlying interactions between their collective modes. In chapter 1, I provide a brief overview of the development of our understanding of physical behaviour in solids, from single-particle descriptions to the emergence of collective modes. I introduce the fundamental principles of light-matter interaction that form the basis for studying the dynamics of collective modes and their couplings in ultrafast science. In chapter 2, I introduce the fundamental principles of electron diffraction, distinguishing elastic and inelastic scattering processes. By combining this technique with a pump-probe scheme, I demonstrate how UED can resolve the dynamics of collective modes in both time and momentum space. I then describe in detail the technical implementation of the UED setup, including a newly developed acquisition method that enhances the signal-to-noise ratio by nearly an order of magnitude. In chapter 3, I investigate the coupling between electrons, plasmons, and phonons in graphite. I use UED to reveal strong electron-phonon interactions involving two distinct lattice vibration modes. Under excitation with two different pump photon energies, graphite exhibits subtle dichotomies in the relaxation pathways of the photoexcited carriers. The resulting modulation of the phonon mode populations directly influences plasmonic dispersion. In chapter 4, I discuss the challenges associated with developing a microscopic description of high-temperature superconductors. Focusing on the prototypical cuprate Bi2Sr2CaCu2O8+x (Bi-2212), I investigate its out-of-equilibrium response under photoexcitation. In particular, I highlight the dynamics of Cooper-pair recombination in the superconducting phase and contrast them with those observed in the normal state. These measurements point toward a possible microscopic scenario underlying Cooper-pair formation in cuprate superconductors. In chapter 5, I introduce a theoretical phase of matter in which excitons constitute the ground state. The search for experimental evidence of this phase motivates the investigation of Ta2NiSe5, a material that undergoes a coupled structural and electronic phase transition. Upon photoexcitation with low and high fluence, two distinct dynamical regimes emerge, suggesting the coexistence of structurally driven and adiabatic phase transition.

Efficient optimization of organic reactions remains a central challenge in synthetic chemistry, often hindered by the vast combinatorial space of possible conditions and the limitations of traditional trial-and-error approaches. In this thesis, reaction optimization in homogeneous catalysis is advanced by integrating statistical and computational methodologies across four pillars: data collection and curation, molecular and reaction representation, robust modeling and model selection, and experimental design with a focus on Bayesian optimization. Emphasis is placed on strategies for low-data regimes, interpretability, and systematic decision-making under uncertainty, with outcomes demonstrated on representative case studies.

The first pillar establishes practical guidance for building statistically robust, machine-learning-ready datasets, with ongoing challenges in streamlining these processes discussed in detail. The second pillar develops a reaction-agnostic featurization strategy for bidentate ligands and curates ligand datasets to enable interpretable and transferable modeling. The third pillar introduces ReaFS (reaction feature selection) as a systematic methodology for selecting informative features and validating multivariate linear regression models, with a focus on model stability and interpretability in low-sample regimes. The fourth pillar extends classical Bayesian optimization by incorporating experimental cost into the acquisition strategy, resulting in cost-informed Bayesian optimization (CIBO) that reduces expected expenditure relative to cost-agnostic policies while respecting practical constraints. Application of selected methodologies is demonstrated on reaction scope exploration and optimization, including azidofunctionalization and carboetherification reactions. The thesis concludes by outlining future directions for integrating these tools with broader chemical synthesis planning, highlighting the need for standardized practices, rigorous validation, and continued methodological innovation.

This study advances distributed sensing with optical fibres by proposing gas-mediated platforms that overcome two long-standing limitations of silica-based methods: the unavoidable cross-sensitivity between temperature and strain, and the loss of reliable temperature sensitivity at low temperatures. Three studies based on distributed sensing in gases are developed and experimentally validated. First, a threshold-temperature alarm is realised by filling a side air-hole optical fibre with carbon dioxide and interrogating it with conventional optical time-domain reflectometry. When local temperature falls below a pressure-set point, carbon dioxide condenses inside the holes and liquifies, sharply increasing local optical loss while preserving core guidance. The resulting change in the backscattered trace identifies the position of cold or hot spots with simple hardware, and the contrast can be tuned by selecting the optical wavelength. This provides a robust, multi-point, distributed event detector rather than a full thermometer. Then, to eliminate the temperature cross-sensitivity in the strain sensing based on phase-sensitive Rayleigh scattering, a solid-core photonic-crystal holey fibre is filled with a gas at controlled pressure so that the negative thermal response of the gas compensates the positive thermo-optic response of silica. Operating near this athermal point yields a strain-only readout while retaining strong coherent Rayleigh backscattering signal from the solid core. The analysis quantifies how the compensation depends on gas species, pressure, wavelength, and modal overlap, and demonstrates dominant strain sensitivity with greatly suppressed thermal drift. Finally, this study establishes an absolute, calibration-free thermometry based on stimulated Brillouin scattering in gas-filled hollow-core fibres. The Brillouin frequency shift, linewidth, and gain can be theoretically predicted based on the thermodynamic gas state equations at given temperature and pressure. Experiments with neon, argon, and nitrogen across 67-350 Kelvin confirm these predictions with high agreement (without calibration), showing a monotonic sensitivity increase with decreasing temperature, and demonstrate distributed temperature measurements with sub-Kelvin accuracy using the Brillouin echo scheme. Overall, this study presents the following: i) a simple distributed temperature alarm for threshold crossings; ii) a phase-sensitive Rayleigh system with intrinsic strain-only response; and iii) an absolute Brillouin thermometer that remains accurate and highly sensitive in the cryogenic regime. In summary, gas-mediated fibre sensors extend the scope of distributed sensing technology, transcending the inherent limitations of solid silica and offering tuneable physical properties through gas selection, pressure, and waveguide design.

Carbon fibre-reinforced polymers are increasingly employed in thick, load-bearing structures across several sectors. However, their potential is hindered by manufacturing-induced defects such as thermal overshoot, residual stresses, and uneven cure, which compromise performance, dimensional stability, and process reliability. Conventional mitigation strategies optimise curing cycles but become inefficient for thick parts due to excessively long cure times. This limitation calls for alternative manufacturing routes. This thesis investigates stepwise curing as a defect-mitigation strategy in thick carbon/epoxy composites, involving partial reaction of the thermoset matrix prior to deposition, then further reaction during tow/layer deposition and compaction, and finally post-curing. This concept is adapted both to prepreg lay-up and additive manufacturing by tow deposition. The objectives were thus to: assess the influence of stepwise curing on internal strain and warpage; quantify how tow-by-tow curing affects thermal overshoot, residual stress, and process time; and evaluate the effect of pre-curing and in-situ consolidation on the interlaminar performance of composites. An integrated experimental-numerical approach was adopted. Distributed fibre optic sensors enabled real-time strain monitoring during curing of thin and thick laminates. Thermo-mechanical simulations incorporating cure kinetics and evolving material properties were developed to predict the influence of pre-cure level, deposition speed, and heat input on stress generation, overshoot, and cure time. A proof-of-concept winding line was also built to manufacture curved laminates under controlled pre-cure and compaction conditions, followed by mechanical and microstructural characterisation. Experimental results on thin laminates showed that pre-curing plies up to 50% conversion reduced internal strain and warpage by 31% and 14%, corresponding to 94 µe and 11 MPa reductions compared to composites made from uncured plies. In thick laminates, pre-curing to 28% conversion lowered peak temperature and residual strain by 29% and 17% (166 µe and 24 MPa reductions). Numerical simulations confirmed that tow-by-tow stepwise curing could reduce the exotherm by up to 92% and through-thickness stress gradients by 65%, compared to batch curing, while enabling faster cycles. Finally, tow-wound curved laminates produced from partially pre-cured tows showed low void content < 2% and stable interlaminar performance with an average interlaminar shear strength of 56±4 MPa and stable fracture toughness around a pre-cracked energy release rate of 462±102 J m-2, confirming that pre-curing below gelation does not impair interlaminar properties. The findings demonstrate that the proposed manufacturing strategy effectively mitigates cure-induced defects in thick composites, providing experimental and numerical evidence of defect reduction via stepwise curing. The work establishes partial pre-curing as an effective method for controlling stress gradients, thermal overshoot, and residual strain, thereby enhancing dimensional accuracy, structural reliability, and manufacturing efficiency. This research opens opportunities for new reliable and automated composite manufacturing processes based on pre-cured components and highlights the potential of coupling real-time monitoring with predictive simulations for adaptive cure control and performance assessment throughout manufacturing and service life.

Simulating turbulence in the tokamak boundary, which is crucial for understanding and optimizing future fusion power plants performance, is computationally challenging due to the complex, multi-physics, and multi-scale nature of plasma dynamics. In the boundary region, plasma interacts with neutral particles recycled from the walls, leading to nonlinear phenomena that control confinement and heat exhaustion. Accurately capturing these coupled processes requires solving large-scale, stiff, and non-local problems. This thesis contributes to the development of efficient numerical algorithms based on low-rank linear algebra to accelerate plasmaâ neutral simulations. The key idea is to exploit the low-dimensional structure underlying the data and operators, which can often be approximated in reduced subspaces, thereby lowering memory and computational cost. The proposed methods are implemented in GBS, a three-dimensional, flux-driven turbulence code for self-consistent plasma and kinetic neutral modeling in the tokamak boundary. GBS solves the drift-reduced Braginskii equations coupled with a kinetic neutral model that is solved deterministically through integration along characteristics. The first contribution of this thesis introduces a subspace acceleration method for expediting the solution of sequences of large-scale linear systems, such as the ones arising from the numerical discretization of time-dependent partial differential equations coupled with algebraic constraints. By leveraging the history of previous solutions, accurate initial guesses for an iterative solver are constructed through reduced-order projection methods combined with randomized linear algebra techniques, drastically reducing the number of iterations needed for convergence. Theoretical analysis highlights that smooth temporal evolution ensures high approximation quality, while large-scale simulations of the plasma boundary demonstrate significant speed-up, through faster solution of the resultant linear systems. The second contribution concerns the acceleration of the solver used to simulate the kinetic neutral dynamics. When the neutral dynamics are described deterministically via the method of characteristics, the resulting integral formulation leads to dense linear systems which are too costly to assemble and solve in practice. This thesis employs hierarchical matrix approximations to represent the corresponding operators in a data-sparse format, leading to a computational complexity nearly linear in the number of unknowns, in contrast with a quadratic scaling of the dense approach. The hierarchical matrix solver achieves over 90% reduction in computation time and memory load, enabling high-resolution simulations of neutral dynamics consistent with the plasma grid. These developments advance the efficiency and accuracy of tokamak boundary simulations, enabling detailed and computationally tractable modeling of plasmaâ neutral interactions, which are essential for the design and operation of future fusion devices.