High-temperature power conversion modules (DC-DC converters, inverters, etc.) have enormous potential in extreme environment applications, including automotive, aerospace, geothermal, nuclear, and well logging. Power-to-volume and power-to-weight ratios of these modules can be significantly improved by employing Silicon Carbide (SiC) based power switches (MOSFET or JFET). Wide bandgap material such as SiC is capable of much higher temperature operation than conventional Silicon based power devices. For successful realization of such high temperature power conversion modules, associated control electronics also need to perform at high temperature. This paper presents a Silicon-on-Insulator (SOI) based high-temperature, high-voltage gate driver integrated circuit (IC) with improved peak output current drive over previous work as well as an on-chip low-power temperature sensor. This driver IC has been primarily designed for automotive applications where the under hood temperature can reach 200°C. This new gate driver prototype has been designed and implemented in a 0.8-micron, 2-poly, and 3metal Bipolar-CMOS-DMOS (BCD) on SOI process and has been successfully tested up to 200oC ambient temperature driving a SiC MOSFET or a SiC normally-ON JFET. In this design, the peak output current capability of the driver is 5 A and is thus capable of driving several power switches connected in parallel. An ultra low-power on-chip temperature supervisory circuit has also been integrated into the die to safeguard the driver circuit against excessive die temperature (≥ 220°C). This approach utilizes the increased diode leakage current at higher temperature to monitor the die temperature. Up to 200°C, the power consumption of the proposed temperature sensor circuit is below 10 μW.

The sale of hybrid electric vehicles (HEVs) has increased 10 fold from the year 2001 to 2008 [1]. Thus, high temperature electronics for HEV applications are desired in the engine compartment , power train, and brakes where the ambient temperature normally exceeds 150°C. Power converters (i.e. DC-DC converter, DC-AC inverter) inside the HEVs require Gate-Driver ICs to control the power switches. A Gate-Driver IC needs a step-down voltage regulator to convert the unregulated high input DC voltage (V DDH) to a regulated nominal CMOS voltage (i.e. 5 V), this step-down voltage regulator will supply voltage to the low-side buffer (pre-driver) and other digital and analog circuits inside Gate-Driver ICs. A linear voltage regulator is employed to accomplish this task; however, very few publications on high temperature voltage regulators are available. This research presents a high temperature linear voltage regulator designed and fabricated on a commercially available 0.8-µm BCD-on-SOI process. SOI processes offer 3 orders of magnitude smaller junction leakage current than bulk-CMOS processes at temperatures beyond 150°C. In addition, a pole swap compensation technique is utilized to achieve stability over a wide range (4 decades) of load current. The error amplifier inside the regulator is designed using an inversion coefficient methodology, and a temperature stable current reference is used to bias the error amplifier. The linear regulator provides an output voltage of 5.3 V at room temperature and can supply a maximum load current of 200 mA.

In the evaluation of high-value edible oils, increasing attention is devoted to non-destructive techniques capable of supporting authenticity assessment and quality control. Variations in the thermal behaviour of oils under controlled excitation can provide indirect but informative indicators of compositional changes and adulteration. In this study, an infrared (IR) thermography-based approach combined with deep learning is proposed for the analysis of tomato seed oil (TSO) and sunflower oil (SO). Thermographic sequences acquired during repeatable heating–cooling cycles were processed as spatiotemporal thermal signals reflecting heat absorption and dissipation dynamics. A convolutional neural network–long short-term memory (CNN–LSTM) model was employed to classify TSO, SO, and adulterated oil samples. The proposed system achieved an overall classification accuracy of 95.1% and a macro F1score of 94.8% in three-class classification. The results confirm that thermographic fingerprints capture stability-related differences induced by oil composition and blending, highlighting the potential of IR thermography coupled with CNN-based modelling as a rapid, non-contact decision support tool for edible oil authenticity screening.

Bergamot essential oil (BEO) is a high-value product, often involving synthetic metabolites. Hereafter is presented a preliminary study for the development of a device which can detect and identify natural and synthetic BEO. Fourier transform infrared attenuated total reflectance (FTIR-ATR) spectroscopy was used to identify the chemical fingerprint of the BEO which combined with principal component analysis (PCA), and Random Forest (RF) model can easily classify among BEO samples. Synthetic BEO often contains ethanol and 2-(2-hydroxypropoxy)-1- propanol (DPG) as diluents for artificial metabolites, absent in natural BEO. An adsorption peak at 1340 cm(-1), absent in natural samples, corresponds to C-H bending of alcohols (e.g. ethanol, DPG) or symmetric deformation of methyl (CH3) groups associated with artificial esters (e.g. linalyl acetate). Also, an absorption band between 3600-3100 cm(-1) confirmed the presence of DPG and synthetic ethanol. The classification model, trained on 260 samples using PCA, was validated through five -fold cross-validation (CV) and external testing. RF yielded an accuracy of 0.73 +/- 0.16, with a sensitivity and specificity of 0.96 +/- 0.02. Based on these results a concept based of non-dispersive infrared sensor (NDIR) has been developed for the development of a compact device including the classification model for the rapid detection of adulterated BEO.

This study explores single-event transient (SET) phenomena in enhancement-mode [Formula: see text]-GaN/AlGaN/GaN high electron mobility transistors (HEMTs) using technology computer-aided design (TCAD) simulations. A two-dimensional (2D) device model is developed, incorporating essential physical models for carrier mobility, recombination, impact ionization, and heavy-ion strike generation. The influence of gate voltage, drain voltage, and linear energy transfer (LET) values on transient drain current behavior is systematically analyzed. Simulation results show that SET effects are significantly more severe under off-state conditions, where the depletion of the two-dimensional electron gas (2DEG) channel leads to increased charge collection through leakage paths. Peak transient currents are observed to increase with both higher drain voltages and higher LET values, indicating a strong dependence of device vulnerability on operating conditions and ionizing radiation energy. These findings are consistent with trends reported in experimental studies and provide important design insights for improving the radiation tolerance of GaN-based devices. Overall, this work offers a theoretical foundation for optimizing GaN HEMTs to achieve enhanced performance and reliability in space and other radiation-intensive environments.

Single-photon avalanche diodes (SPADs) belong to a family of avalanche photodiodes (APDs) with single-photon detection capability that operate above the breakdown voltage (i.e., Geiger mode). Design and technology constraints, such as dark current, photon detection probability, and power dissipation, impose inherent device limitations on avalanche photodiodes. Moreover, after the detection of a photon, SPADs require dead time for avalanche quenching and recharge before they can detect another photon. The reduction in dead time results in higher efficiency for photon detection in high-frequency applications. In this work, an electronic interface, based on the pole-zero compensation technique for reducing dead time, was investigated. A nanosecond pulse generator was designed and fabricated to generate pulses of comparable voltage to an avalanche transistor. The quenching time constant (τq) is not affected by the compensation capacitance variation, while an increase of about 30% in the τq is related to the properties of the specific op-amp used in the design. Conversely, the recovery time was observed to be strongly influenced by the compensation capacitance. Reductions in the recovery time, from 927.3 ns down to 57.6 ns and 9.8 ns, were observed when varying the compensation capacitance in the range of 5–0.1 pF. The experimental results from an SPAD combined with an electronic interface based on an avalanche transistor are in strong accordance, providing similar output pulses to those of an illuminated SPAD.

Spiking neural networks (SNNs) emulate biological neurons transmitting information through discrete spikes or pulses of activity. SNNs found extensive application in neuromorphic computing due to efficient cognitive-aware computation, event-driven processing, and robustness to noise and faults. Numerous applications in pattern recognition, sensory processing, computer vision, etc., require complicated networks for enhanced performance and reliability. Realizing such complex SNNs on circuits poses significant challenges due to limitations in scaling up a single array of fully connected neurons within hardware resource and energy constraints. Contemporary neuromorphic chips incorporate small crossbars with time-multiplexed interconnects to address this issue. However, designing robust and efficient neuromorphic systems based on small crossbars is impractical. Therefore, strategic segmentation and positioning of SNNs are crucial for efficiently mapping them onto neuromorphic circuits. This work proposes the hill climbing optimization method to ef-fectively translate SNNs onto neuromorphic circuits, minimizing the spike communication on the interconnect.

This research presents a comprehensive investigation and optimization of the Pt/AlN Schottky Barrier diode (SBD) using technology computer-aided design (TCAD) modeling. The study explores the electrical characteristics of AlN SBDs with various metal contacts, including Aluminum (Al), Silver (Ag), Tungsten (W), Gold (Au), Nickel (Ni), and Platinum (Pt). Through the comparative analyses of different metal/AlN Schottky contacts, the Pt/AlN structure emerges as the most promising due to its superior barrier height and lower leakage current. At [Formula: see text]K, the diode demonstrates a barrier height of 2.72[Formula: see text]V, a nearly ideal leakage current of 0.046[Formula: see text]pA, and a breakdown voltage of 363[Formula: see text]V. The research extends to examining the temperature-dependent electrical behavior of Pt/AlN Schottky diodes, particularly for high-power and high-temperature applications. Analysis carried out across temperatures ranging from [Formula: see text]K to [Formula: see text]K reveals a trend of increasing ON resistance and consistently lower leakage current with rising temperature. Importantly, the study indicates that the impact of temperature on the barrier height and breakdown voltage of the diode is negligible, thus rendering it suitable for high-temperature operation. Leveraging the unique properties of AlN as an ultra-wide bandgap material within the III-V compound semiconductor family, this research provides valuable insights into the potential applications of Pt/AlN Schottky contact. The study highlights that the Pt/AlN Schottky contact is effective not only for high-power, high-temperature SBDs but also as superior metal/semiconductor gate contacts for field-effect transistors (FETs). Their suitability is attributed to their ability to handle high voltages, minimize reverse leakage current, and demonstrate improved thermal stability.

Background: Obstructive sleep apnea is a sleep disorder that is linked to many health complications and can even be lethal in its severe form. Overnight polysomnography is the gold standard for diagnosing apnea, which is expensive, time-consuming, and requires manual analysis by a sleep expert. Artificial intelligence (AI)-embedded wearable device as a portable and less intrusive monitoring system is a highly desired alternative to polysomnography. However, AI models often require substantial storage capacity and computational power for edge inference which makes it a challenging task to implement the models in hardware with memory and power constraints. Methods: This study demonstrates the implementation of depth-wise separable convolution (DSC) as a resource-efficient alternative to spatial convolution (SC) for real-time detection of apneic activity. Single lead electrocardiogram (ECG) and oxygen saturation (SpO2) signals were acquired from the PhysioNet databank. Using each type of convolution, three different models were developed using ECG, SpO2, and model fusion. For both types of convolutions, the fusion models outperformed the models built on individual signals across all the performance metrics. Results: Although the SC-based fusion model performed the best, the DSC-based fusion model was 9.4, 1.85, and 11.3 times more energy efficient than SC-based ECG, SpO2, and fusion models, respectively. Furthermore, the accuracy, precision, and specificity yielded by the DSC-based fusion model were comparable to those of the SC-based individual models (~95%, ~94%, and ~94%, respectively). Conclusions: DSC is commonly used in mobile vision tasks, but its potential in clinical applications for 1-D signals remains unexplored. While SC-based models outperform DSC in accuracy, the DSC-based model offers a more energy-efficient solution with acceptable performance, making it suitable for AI-embedded apnea detection systems.

This paper proposes two power-efficient digital classifier designs for a neural network-based sleep apnea (SA) detection system. The digital classifiers designed in this work are rectified linear unit (ReLU) and signum (sign), which are used in the hidden layer and the output layers, respectively of our proposed binarized neural network model (BNN). The BNN model yielded around 88% accuracy using the selected classifiers with balanced evaluation metrics. Studies of measurement results such as accuracy, power consumption, and comparison with other widely used classifiers such as hyperbolic tangent (tanh) and sigmoid were conducted on digital hardware using a general-purpose field-programmable gate array (FPGA) called Nexys Artix-7. By proposing our binarizing technique called Shift-Accumulate-based Binarized Neural Network (SABiNN) on the neural network model and using the stacked multiplexer design method with look-up-tables for both ReLU and sign classifiers, the power consumption rates of the selected classifiers were significantly reduced without compromising performance. The 4-hidden layer 2-(8-12-6-4)-1 BNN model consumed a maximum of 5 W of power with a thermal margin of 11.4 oC including a low resource utilization report. The proposed classifier designs demonstrate promising results in accurately modeling neural network models that enable SA detection, offering the potential for cost-effective and scalable healthcare solutions.

Obstructive sleep apnea is a sleep disorder that is linked with many health complications and severe form of apnea can even be lethal. Overnight polysomnography is the gold standard for diagnosing apnea, which is expensive, time-consuming, and requires manual analysis by a sleep expert. Recently, there have been numerous studies demonstrating the application of artificial intelligence to detect apnea in real time. But the majority of these studies apply data pre-processing and feature extraction techniques resulting in a longer inference time that makes the real-time detection system inefficient. This study proposes a single convolutional neural network architecture that can automatically extract spatial features and detect apnea from both electrocardiogram (ECG) and blood-oxygen saturation (SpO2) signals. Using segments of 10s, the network classified apnea with an accuracy of 94.2% and 96% for ECG and SpO2 respectively. Moreover, the overall performance of both models was consistent with an AUC score of 0.99.

Gallium oxide (Ga2O3) is a promising ultra-wide bandgap material offering a large bandgap ( >4.7 eV) and high critical electric fields. The increasing demand for electronic devices for high-power applications in electric automobiles, high-performance computing, green energy technologies, etc., requires higher voltages and currents with enhanced efficiency. Vertical transistors, such as fin-shaped field-effect transistors (FinFETs) have emerged to meet the growing need with improved current handling capabilities, reduced resistance, and enhanced thermal performance. However, to fully exploit the Ga2O3 power transistors, precise and reliable physics-driven models are crucial. Therefore, a comprehensive surface potential model has been developed in this work for a vertical Ga2O3 FinFET. The electric potential across the channel is explained by analyzing the two-dimensional (2D) Poisson equation employing parabolic approximation. Such a surface potential model is instrumental in determining the performance of the Ga2O3 FinFET as it affects the threshold voltage, the drain current, and fringing capacitance. Exploiting the surface potentials, a fringing capacitance model is derived which is crucial in analyzing the speed of the device in compact integrated circuits. In addition, statistical analysis of the Ga2O3 FinFET using the Monte Carlo simulation technique is performed to determine the leakage current fluctuation due to doping variations. The validation of the analytical model with experimental results confirms the effectiveness and prospects of the developed models in the rapid development and characterization of next-generation high-performance vertical Ga2O3 power transistors.

Type 1 diabetes (T1D) presents an escalating health challenge requiring precise glycemic control within management protocols. The need for energy-efficient wearable medical-grade devices and biosensors, notably continuous glucose monitoring (CGM), facilitates real-time monitoring and augments interest in blood glucose prediction. However, existing research predominantly concentrates on short-term blood glucose forecasting, overlooking the critical need for accurate long-term prediction. This work addresses this gap by exploring the feasibility and efficacy of long-term blood glucose prediction (60, 120, and 240 minutes) leveraging deep multitask learning (DMTL). In addition, long-short-term memory (LSTM) based recurrent neural network (RNN) architecture has been explored within the shared layers of the DMTL framework. The predictive model effectively captures blood glucose trends, achieving a root mean square error (RMSE) of 56.66 +/- 6.96 mg/dL at the 240-minute (four-hour) mark while maintaining 90.36 +/- 4.22% predictions within clinically benign zones in the Clarke error grid. These predictions offer a strategic advantage in long-term care planning, facilitating proactive interventions to stabilize glycemic levels through insulin therapy and dietary intake adjustments. Furthermore, predictive analytics hold significant promise in advancing next-generation AI-enabled automated self-management assistive tools, including artificial pancreas and predictive low-glucose suspend systems.

This work comprehensively investigates the systematic integration of highly relevant life events and physiological parameters with continuous glucose monitoring (CGM) data to understand the synergy in glucoregulatory systems. Multitask learning (MTL) is employed in learning from numerous subjects while tailoring features in patient-specific layers. Systematic combinations of inputs are fed to the system providing personalized blood glucose level (BGL) predictions at multi-step prediction horizons (PHs) as output. Three cutting-edge long-short-term memory (LSTM) networks are adapted in the shared layers of MTL architecture. Moreover, PHs are varied with 30-minute intervals up to 120 minutes to identify the long-term effects of relevant input features on BGL prediction suggesting optimal deep learning (DL) architecture. The empirical result demonstrates that the most relevant features for BGL prediction are glucose, bolus insulin, and carbohydrate estimate from meals, while exercise and basal insulin rate have momentary effects. The best predictive root means square error (RMSE) achieved are 16.06 +/- 2.74, 30.89 +/- 4.31, 40.51 +/- 5.16, and 47.31 +/- 5.78 mg/dL for 30-, 60-, 90-, and 120-min PH, respectively, maintaining 94.06 +/- 3.08 % predictions in clinically safe zones (A+B) in Clarke error grid analysis (EGA). The insights learned from the experiments will assist in selecting appropriate DL models, features, and timelines based on specific needs, with significant promise in improving T1D management through better therapeutic and lifestyle modification.

Type I diabetes (T1D) is attributed to pancreatic beta cell impairment and insulin diminution, requiring multiple daily injections to maintain blood glucose homeostasis. Predicting human blood glucose concentration (BGC) is crucial in clinical decision support systems for diabetes control and closed-loop insulin delivery. However, the complex glucose behavior, inter-subject variability, and multi-factor dependence make the BGC predictions challenging, causing mathematical models and human experts to fail in achieving acceptable precision. This paper presents end-to-end multi-task learning (MTL) that accumulates learning from all subjects and fine-tunes it from the individual patient data for personalized predictive models. The MTL technique enables simultaneous training of relevant tasks, harnessing relevance (generalization) and variations in individual data (personalization). The proposed network architecture consists of shared layers and layers tailored to specific subjects. Deep long short-term memory networks (LSTMs) followed by dense layers form the shared layers and the subject-specific part consists of only dense layers. The model is trained and evaluated using the OhioT1DM clinical dataset. Results demonstrate average root means square error (RMSE) and mean absolute error (MAE) of 16.65, 10.76, and 32.19, 22.91 for 30- and 60-minute prediction horizon (PH), respectively. Moreover, the results are consistently robust, and the effectiveness is confirmed by benchmarking with different techniques.

The prevalence of sensors and sensor networks has resulted in the Internet of Things (IoT), transforming modern lifestyles. However, resource-constrained IoT edge devices require inventive sensing, computing, and wireless telemetry strategies to support the massive number of devices in an IoT network. Traditional telemetry and data encoding schemes for short-range sensor networks are limited by network bandwidth, data processing capability, and power constraints imposed by the battery energy density. To mitigate these challenges, an energy-efficient architecture is presented for analog orthogonal pulse (AOP) generation and AOP-based data encoding for high-density spectrum-efficient wireless telemetry. Unlike conventional digital pulse-based encoding, higher-order analog orthogonal pulses are used for spectrum-efficient data encoding. The orthogonal pulse generator contained a reduced number of functional blocks for energy efficiency. The MATLAB-Simulink package is used to design and simulate the proposed encoding architecture and, finally, embedded into a microcontroller unit. Test results show the successful generation of analog orthogonal pulses and pulse-based signal encoding.

This paper presents an in-depth reliability analysis of neuromorphic circuits utilizing phase change memory (PCM). In-memory computing with PCM cells to store synaptic weights is exploited to address the performance gap between processing needs and memory constraints in machine learning computation. The integration of PCM for hardware realization of spiking neural networks (SNNs) enhances data processing and transmission capabilities. However, challenges remain in memristor-based crossbar architectures, including the effects of voltage drops and current fluctuations on the reliability of PCM. Hence, a reliability model accounting for process variations and temperature influences is proposed. The model effectively predicts the reliability of PCM in neuromorphic circuits, paving the way for more robust cognitive computing applications.

Obstructive sleep apnea is a sleep disorder that is linked with many health complications and severe form of apnea can even be lethal. Overnight polysomnography is the gold standard for diagnosing apnea, which is expensive, time-consuming, and requires manual analysis by a sleep expert. Recently, there have been numerous studies demonstrating the application of artificial intelligence to detect apnea in real time. But the majority of these studies apply data pre-processing and feature extraction techniques resulting in a longer inference time that makes the real-time detection system inefficient. This study proposes a single convolutional neural network architecture that can automatically extract spatial features and detect apnea from both electrocardiogram (ECG) and blood-oxygen saturation (SpO

Nailfold video-capillaroscopy (NVC) is a diagnostic technique that allows direct observation of microcirculation. NVC has gained significance for its non-invasive evaluation of subtle abnormalities, contributing to the diagnosis and monitoring of various pathologies. In recent years, technological advances have led to the development of new capillaroscopy techniques, as the Sidestream dark-field (SDF) capillaroscopy. It enables real-time visualization of microcirculation with high resolution and it is less susceptible to motion artifacts. A ring of light-emitting diodes (LEDs) provides structural and functional images of the microcirculation with high contrast. However, SDF can be susceptible to artifacts caused by pressure applied during imaging. These artifacts may distort the observed microvascular patterns, requiring careful interpretation by trained professionals. In this context, we developed an NVC based on the sidestream technique, employing off-the-shelf components, 3D printing and stacking algorithm. The aim is to provide a high quality, easy-to-fabricate NVC with reduced artifacts caused by the oil (immersion microscopy topology). The NCV is designed to be applied to the periungual area. Using a black 3D printed case compared to a white 3D case, the NVC enhances image quality by minimizing light glare. Furthermore, we conducted experiments using different illumination wavelengths and processing algorithms to evaluate the most effective image quality. The green LED crown produced superior images due to its absorption by hemoglobin. Our device provides further insights and advancement into early diagnosis and monitoring for Raynaud's phenomenon.

Impact localization plays a vital role in continuous monitoring of damage and fatigue testing in various healthcare applications. It is particularly important for assessing potential risk factors in spinal cord and orthopedic injuries, and in monitoring the health of athletes. Flexible sensor technology, especially the inkjet printing process on flexible substrates, is gaining significant interest due to its cost efficiency, mass production, simplicity, and environmental sustainability advantages. However, the efficiency of inkjet printed sensors is limited due to the constraints of their low-maintenance fabrication process. The integration of artificial intelligence with inkjet-printed sensors can address the limitations in their operational efficiency. In the field of artificial intelligence, Echo State Network is increasingly being implemented in a wide range of applications due to its computational simplicity. This work presents an Echo state network integrated tactile sensor. Echo state network reservoir is inkjet-printed on a polyethylene terephthalate film substrate and works as a tactile sensor. The readout layer of the Echo State Network is constructed with multilayer perceptrons followed by a majority voting layer. The tactile sensitivity is tested with a pencil impact experiment on different sensor surface areas. Our designed Echo State Network with an integrated tactile sensor accurately predicts the location of the pencil drop with a precision of 92.31%.