A document by Georgios Mylonopoulos. Click on the document to view its contents.

Temperature (𝑻)-dependent MOSFET noise is studied by using a custom-built broadband measurement system. Several n-and p-MOSFETs fabricated by technology nodes of 130 nm, 65 nm, and 22 nm are characterized. Wideband (20 kHz-500 MHz) noise spectroscopy clearly reveals that high-𝑻 (> 100 K) random telegraph noise (RTN), originating from relatively deep states, varies from device to device, whereas the low-𝑻 (< 30 K) counterpart exhibits much weaker variability and originates from very shallow states. Variable-𝑻 𝑪-𝑽 measurements of MOSFETs and p-n junction diodes suggest that these shallow states could emerge throughout the bulk Si at low-𝑻.

The traditional approaches in material research and hardware design are insufficient to address the evolving Operation and Maintenance (O&M) demands in contemporary power electronics. Overengineering and data acquisition practices lead to unsustainable costs and reduced profit margins. Digital Twins (DTs), defined as real-time simulation models of physical systems, emerge as promising solutions to meet stringent O&M requirements. In power electronics, DTs offer significant potential in thermal management, crucial for control performance, safety, and system lifespan. This paper aims to analyze the development of computationally efficient and high-fidelity DTs tailored for power electronics applications, emphasizing their predictive reliability. To achieve this goal, the proposed physics-based approach is enhanced by integrating Data-Driven Artificial Intelligence (AI)based techniques. The predictive reliability of the DTs produced through this workflow is then experimentally validated against a power electronic converter designed for induction heating applications. Additionally, the feasibility of real-time execution is demonstrated, affirming the practical applicability of the developed DTs.

Surface electromyography (sEMG) signals are electrical signals released by muscles during movement, which can directly reflect the muscle conditions during various actions. When a series of continuous static actions are connected along the temporal axis, a sequential action is formed, which is more aligned with people's intuitive understanding of real-life movements. The signals acquired during sequential actions are known as sequential sEMG signals, including an additional dimension of sequence, embodying richer features compared to static sEMG signals. However, existing methods show inadequate utilization of the signals' sequential characteristics. Addressing these gaps, this paper introduces the Spatio-Temporal Feature Extraction Network (STFEN), which includes a Sequential Feature Analysis Module based on static-sequential knowledge transfer, and a Spatial Feature Analysis Module based on dynamic graph networks to analyze the internal relationships between the leads. The effectiveness of STFEN is tested on both modified publicly available datasets and on our acquired Arabic Digit Sequential Electromyography (ADSE) dataset. The results show that STFEN outperforms existing models in recognizing sequential sEMG signals. Experiments have confirmed the reliability and wide applicability of STFEN in analyzing complex muscle activities. Furthermore, this work also suggests STFEN's potential benefits in rehabilitation medicine, particularly for stroke recovery, and shows promising future applications.

Language can disclose a lot about a person's emotional condition, social standing, and even personality. In the proposed research, the authors looked at the words that are most helpful in determining if a person is depressed or not. Additionally, the authors tried to include a variety of dataset features to see if they could help us form a more accurate judgement. The authors obtained the data from the social networking site Reddit. It includes erratic messages and comments made by a group of depressed people. Given that it performed the best, the authors used logistic regression using the Term Frequency-Inverse Document Frequency (TFIDF) for the proposed model. One of the factors that contributed to a marginal speed improvement was the average time between two consecutive postings or comments. In F1-score, the proposed model fared better than other models utilising the same dataset. The diagnosis of mental disease is turning into a significant issue as people's understanding of the value of mental health grows. Many psychiatrists found it challenging to make a diagnosis of mental disease in a patient due to the complexity of each mental disorder, making it impossible to start the patient on the right course of treatment before it was too late. However, the significance of integrating into people's daily life has produced a setting in which further knowledge about a patient's mental illness may be acquired.

Much has been written about the representation of noisy linear 2-ports. Here we present a theory of noisy Nports. We show how in the general case there are (2N)!/(N !) 2 equivalent representations and give the transformations relating them. We also discuss singular cases in which some of the transformations are not possible as well as how to measure the noise properties of an N-port.

In the past 20 years, improvements in night-time light (NTL) remote sensing have spurred a resurgence of interest in the mapping of human economic activity. Nevertheless, the full potential of NTL data for urban research is constrained by a relatively coarse spatial resolution and the blooming effect. Downscaling NTL data is a potential solution, aiming to obtain fine-resolution nocturnal data with high accuracy. Most existing remotely sensed image fusion techniques were developed for optical remote sensing images taken during the day. When NTL images are compared to optical images, they exhibit a greater quantity of dark (low value) pixels, higher levels of background noise, and a more obvious blooming effect. In this paper, we proposed a spatially non-stationary, geostatistical-based downscaling technique (random forest area-to-point Kriging) to downscale NTL data while accounting explicitly for the point spread function, thus, dealing with the blooming effect specific to NTL data. We compared several image fusion algorithms for downscaling while reducing the blooming effect. Numerical experiments on two megacities showed that downscaling was improved both numerically and visually by taking the PSF into consideration. During the RF regression, the R 2 increased and the RMSE decreased for both study regions, when accounting for the PSF. For the ATPK-based residual part, considering the PSF led to increased accuracy of prediction. The suggested methodology has the potential to increase the detail and accuracy of the NTL data available for modeling socioeconomic phenomena at the city scale, with wide potential for application in future socioeconomic research.

In the wake of the COVID-19 pandemic, efficiently allocating ICU resources for critical patients has become crucial, especially for those with chronic conditions. This study harnesses machine learning (ML) to forecast ICU admissions among COVID-19 patients in Kuwait, analyzing a dataset of 4399 patients to identify pivotal predictors for ICU needs. Employing cross-validation and Synthetic Minority Over-sampling Technique (SMOTE) to tackle data imbalance, the predictive variables were refined using backward feature selection with logistic regression and evaluated model interpretability with Shapley additive explanations (SHAP). The Support Vector Machine (SVM) model outperformed other models with an area under the curve (AUC) of 0.91, and the Extra Tree (ET) model showed better performance with an accuracy of 96.42%. Critical predictors included demographics, clinical outcomes like shortness of breath, elevated d-dimer levels, and abnormal chest X-rays. This research not only underscores the potential of ML in critical healthcare decision-making during pandemics but also highlights its role in discovery science, suggesting broader applications in healthcare and other scientific domains. The study advances medical informatics by integrating ML with healthcare, offering insights into disease dynamics and improving resource allocation strategies.

Background: Traditional myoelectric controls of prostheses for transhumeral amputees fail to provide intuitive coordination of the necessary degrees of freedom. Building upon promising advances in movement-based controls and computer vision, we have previously demonstrated that reconstructing the distal joints based on Artificial Neural Network (ANN) predictions, while knowing the shoulder posture and the movement goal (i.e., position and orientation of the targeted object), enables participants to position and orient an avatar hand to grasp objects scattered throughout a wide workspace with performances comparable to that of a natural arm. However, this previous control involved rapid and unintended prosthesis movements that resulted from sudden changes in the ANN predictions at each modification of the movement goal, rendering its use impractical for real-life scenarios. Methods: Here, we designed and tested novel methods to eliminate this abrupt change, based on an angular trajectory, determined from the speed of stump movement and the gap remaining between the current and the ’goal’ distal configurations. Two methods are presented to define this ’goal’ configuration, either relying solely on the movement goal or also taking into account the current shoulder posture. These new controls are tested offline and online (i.e., involving participants-in-the-loop) and compared to performances obtained with a Natural control as with our previously proposed control (i.e., without angular trajectory). Results: Despite a slight increase in movement time, potentially due to the increased trajectory curvature observed offline, the new controls allowed twelve non-disabled participants and six participants with transhumeral limb loss to reach objects at various positions and orientations in a virtual reality set-up without prior training. Furthermore, while valid participants rated the usability lower and the workload higher with our controls compared to the Natural control,  participants with upper limb disabilities did not, indicating their genuine interest in our novel controls. Conclusions: The good performances achieved with those controls, particularly when the current shoulder posture was accounted for, represent a necessary step towards applications in real-world scenarios. Notably, the good ratings by participants with upper limb disability highlight the potential acceptability and effectiveness of those controls for our target population.

In response to the evolving demands of microelectromechanical system (MEMS) integration, aiming to achieve high-performance electronic products, innovative interconnect solutions are becoming essential. The interconnects must possess key features, including the capability for miniaturization, low processing temperatures, and the ability to maintain a stable microstructure with optimized electrical, mechanical, and thermomechanical properties. To meet these demands, this study designed a novel Cu-Sn-based solid-liquid interdiffusion (SLID) interconnect. The study involved the implementation of a metallization stack, incorporating Co as a layer to interact with low-temperature Cu-Sn-In SLID and form intermetallic compounds (IMCs). Since Cu6(Sn,In)5 forms at a lower bonding temperature than other phases commonly observed in the Cu-Sn-In system, the study aimed to develop interconnects comprising a stable single-phase (Cu,Co)6(Sn,In)5. Bonding conditions were established for the Cu-Sn-In/Co system and the Cu-Sn/Co system as a reference. Thorough assessments of their thermomechanical reliability were conducted through hightemperature storage (HTS), thermal shock (TS), and tensile tests. The Cu-Sn-In/Co system emerged as a reliable low-temperature solution with the following key attributes: 1) a reduced bonding temperature compared to the Cu-Sn SLID interconnects, 2) the absence of the Cu3Sn phase and resulting void-free interconnects, and 3) high thermomechanical reliability, particularly following thermal annealing after bonding.

This study introduces the Subspace Rotation Algorithm (SRA), an innovative gradientfree method designed to discover the global optimal weight matrix. The SRA consists of two fundamental algorithms: the Left Subspace Rotation Algorithm (LSRA) and the Right Subspace Rotation Algorithm (RSRA). The combination of LSRA and RSRA, in two formats, LSRA-RSRA and RSRA-LSRA, can harness the advantages of both individual algorithms, resulting in enhanced performance. Our observations reveal that shallow and wide Multilayer Perceptrons (MLP), trained using RSRA-LSRA, achieve higher training accuracy compared to the Backpropagation (BP) algorithm. Moreover, when combining SRA with the BP algorithm, a remarkable impact on training MLP models is observed. Our experiments demonstrate that the BP algorithm may become trapped in local optima, while RSRA-LSRA, though capable of escaping local optima, may not fully realize its potential when the number of hidden nodes is limited. The synergy of RSRA-LSRA and the BP algorithm allows for the utilization of the advantages from both approaches, achieving optimal MLP performance with fewer hidden nodes.

While grid-forming (GFM) technology has gained increased attention for voltage-source converters (VSC) connected to low-inertial AC systems, there exists a gap in addressing their fault ride-through (FRT) capability under large grid disturbances. Specifically, the challenge lies in resolving the resynchronization issue of GFM-VSCs under current-limitation mode without relying on a phase-locked loop (PLL). To bridge this gap, this article proposes a novel approach called direct current-synchronization control (DCSC), which directly regulates the VSC current for synchronization. The validity of DCSC is substantiated by establishing equivalent relationships between current and power, as well as power and phase angle. Thus, achieving synchronization by controlling the VSC phase angle is made possible through the direct control of the VSC current. The stability boundary of DCSC is theoretically analyzed, concluding that DCSC has the same stability boundary as power synchronization control (PSC) with reactive power regulation in continued normal operations but a 90-degree stability boundary under large grid disturbances, irrespective of voltage magnitude. Additionally, boundary conditions for system stability assessment during large grid disturbances are established, and a control gain self-adaptability (CGSA) scheme is introduced to accelerate resynchronization after faults. Therefore, the DCSC scheme exhibits identical control dynamics to PSC with reactive power regulation in continued normal operations, while offering enhanced FRT performance under large grid disturbances. Experimental results validate the theoretical findings, affirming the effectiveness of the proposed control method.

This paper presents an algorithm for a team of heterogeneous mobile robots to estimate and adaptively sample a stationary, isotropic, Gaussian process. An estimation framework is proposed to assimilate measurements from robots with differing sensing capabilities (i.e., measurement noise variance). To improve computational efficiency of the Gaussian process regression, the survey area is divided into regions that each use a common semivariogram matrix constructed with a truncated measurement set determined by an adaptive selector. As the mission proceeds, a Voronoi-based algorithm periodically partitions a density function-representing time-varying sampling priority-to identify high-value sampling locations. The path for each robot is modeled as an mechanical system: a sequence of masses (waypoints) are interconnected by springs and dampers and pulled towards Voronoi cell centroids. At each path planning cycle the robots are iteratively simulated with heterogeneous dynamics (e.g., speed, turn radius) following their respective waypoint paths as stiffness/damping parameters are adjusted to satisfy mission time constraints. Numerical simulations show that the proposed approach reduces mapping error when compared to non-adaptive lawnmower coverage of a survey area. The algorithm is demonstrated experimentally using two cooperating autonomous surface vessels to map the bathymetry in a section of Lake Norman near Charlotte, NC.

Machine Learning (ML) with distributed privacy preservation is growing in significance as it focuses on facilitating multi-party learning without requiring actual data sharing. This is especially helpful for companies that want to work together but are unable to do so because of ethical, regulatory, or budgetary constraints on sharing data. In order to address these issues, this study examines three privacy-preserving algorithms: regularized logistic regression with Differential Privacy (DP), stochastic gradient descent (SGD) with differentially private updates, and a distributed Lasso that distributes gradients among data centers. The study emphasizes the relationship between error rate and privacy through these algorithms. In order to improve error rates for large datasets, both DP algorithms modify their sensitivity dependent on the amount of data, highlighting the significance of training data volume in model performance in the study. Results demonstrate that using the SGD; error rate can be reduced by employing random projections in advance.

This paper investigates the efficacy of linear regression enhanced by gradient descent with momentum for predicting real-world outcomes. The introductory sections establish the significance of machine learning in driving sustainable innovations and detail the foundational aspects of linear regression, a statistical technique pivotal for modelling relationships between variables using a least squares approach. Enhanced with momentum-based gradient descent, these models achieve faster and more stable convergence, which is particularly beneficial in complex, real-life data scenarios. Through empirical analysis using the Boston and California housing datasets, we demonstrate that linear regression, when optimized, can effectively predict housing values with high R 2 scores, indicating robust predictive power across socioeconomic and geographic variables. Our findings underscore the model's utility as a forecasting tool in today's data-driven landscape. Future research directions include optimizing momentum coefficients and learning rates and potentially incorporating adaptive methods to enhance convergence efficiency. This study provides insights into the continuous improvements required in predictive analytics to maintain accuracy and reliability in diverse applications.

Our research explores the integration of Artificial Intelligence (AI) within mirrorless cameras to automate color grading and editing processes, utilizing Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs). The project aims to harness AI's potential to analyze and learn from a photographer's historical stylistic choices, enabling personalized image enhancement directly in-camera. Current DSLR and mirrorless cameras already leverage AI for various functions such as autofocus, noise reduction, and image stabilization. Building on these applications, my project seeks to add a layer of creative automation by training models on a photographer's edited images, thereby recognizing and applying their unique stylistic traits to new photos. Our approach aims to streamline the editing workflow, offer inspiration, and maintain artistic individuality, addressing concerns over the potential homogenization of creative expression in photography. By embedding this technology directly into camera hardware, it provides an innovative tool for photographers to instantly visualize and apply their preferred aesthetics, challenging the traditional post-processing workflow. Our research not only advances the technical capabilities of camera systems but also explores the balance between technological innovation and the preservation of artistic integrity in this age of photography.

A document by Alejandro de la Fuente . Click on the document to view its contents.

This paper explores the significance of Quality of Transmission (QoT) estimation in optical networks and highlights the increasing use of machine learning (ML) techniques to enhance QoT estimation accuracy. It presents a survey of literature in this area, categorizing studies into classification and regression algorithms. ML methods are shown to improve QoT estimation, mitigate nonlinearities, and optimize decisionmaking processes. Ultimately, these advancements reduce the reliance on conservative margins, maximize network capacity, and decrease infrastructure investment. Accurate and real-time QoT information is the foundation for efficient routing and spectral allocation (RSA) systems, it enables proactive failure management, facilitates network reconfiguration, provides inputs for optical line optimization and drives optical network automation.  This paper explores and provides insights into both traditional ML models techniques and large language models (LLMs) as applied to optical communications.