Monolithic integration of III-V compounds onto silicon (Si) substrates utilizing nano-ridge (NR) engineering holds great potential for developing optoelectronic devices that leverage the well-established CMOS process technology. To realize optimized processes that ensure high-quality growth of such novel three-dimensional structures, where feature sizes are in the nanometer range, reliable and rapid wafer-level monitoring is imperative. In this work, we develop an automated and scalable deep learning framework for intricate defect classification in NR structures using scanning electron microscopy images. The proposed method seeks to replace conventional, labor-intensive and error-prone manual defect inspection, significantly reducing turnaround time (TAT) and costs while improving classification accuracy. Furthermore, a graphical user interface (GUI) has been developed to enable efficient deployment in a 300 mm CMOS production environment.

Speech Command Recognition (SCR) has many applications in smart home systems, voice-controlled robots, and voice assistants. The modern SCR systems employ deep learning models trained on speech command datasets. Nowadays, the Google Speech Commands (GSC) dataset serves as a benchmark for training and evaluating SCR models, but it only includes English commands. In this work, we introduce an equivalent of the GSC for the Kazakh, Tatar, and Russian languages. We collected the dataset via a Telegram bot in real-world settings. The dataset includes the same 35 commands as GSC and contains 3,623 utterances for Kazakh, 3,547 for Tatar, and 1,625 for Russian. We trained and evaluated both monolingual and multilingual SCR models. The experimental results show that the multilingual model outperforms the monolingual models with an accuracy of 98.57%, 99.33%, 96.86%, and 97.33% on the test sets of the Kazakh, Tatar, Russian datasets and GSC, respectively. Furthermore, we deployed the model on an NVIDIA Jetson Orin NX 16GB, achieving an average inference time of 0.008 seconds. This performance makes it suitable for real-time applications. We have made the dataset, source code, and pretrained models publicly available to encourage further research in this area.

In the context of urbanization, engaging citizens in monitoring city data is crucial. Our study presents a novel framework that gamifies data visualization in smart cities to heighten public involvement. By incorporating game design principles into the representation of data, we have developed an interactive platform that facilitates community participation and invites feedback on urban issues. This transformation from conventional data visualization to an engaging experience empowers citizens to explore urban datasets dynamically. Elements such as point scoring, challenges, and leaderboards are integral in motivating user interactions and fostering active involvement. Utilizing real-time data captured from city sensors, the gamified format allows citizens to engage meaningfully in discussions around urban planning and resource management. Pilot studies were conducted across various cities to assess how this gamification influences participation rates and enhances the quality of public discourse. Feedback indicated that this approach significantly promotes a more informed citizenry and encourages residents to take ownership and responsibility within their urban environments. This development marks a significant step toward integrating citizen input into future urban management strategies.

The increasing number, severity, and sophistication of cybersecurity threats complicate the task of cybersecurity teams, and therefore, cybersecurity analytics is increasingly becoming an essential tool at their disposal to exploit new big data sources, which can be immensely valuable. Thus, there is a critical need for a comprehensive understanding of cybersecurity analytics in the enterprise context. This systematic literature review, leveraging multiple databases and sources pertinent to cybersecurity and business domains, aims at providing valuable insights into the evolution, adoption, techniques, models, data sources, challenges, and future research opportunities concerning enterprise cybersecurity analytics. Employing the PRISMA methodology, the review synthesizes existing literature to address six research questions, delivering a holistic understanding of the cybersecurity analytics landscape. The findings result in ten observations encapsulating various aspects of cybersecurity analytics. These observations underscore the driving factors and technological shifts influencing the adoption of cybersecurity analytics, emphasize the necessity of a holistic approach, reveal the widespread adoption across different sectors, highlight the diverse analysis techniques and data types in use, and elucidate the challenges faced by enterprises. The study concludes that while implementing cybersecurity analytics poses significant challenges, it also presents considerable opportunities for improving enterprises' cybersecurity situation. Finally, this study unveils the research gaps in enterprise cybersecurity analytics that need further exploration.

Adversarial examples undermine the robustness of deep neural networks (DNNs), posing a major challenge to deep learning. Although existing research has focused on external representations, such as shifts in feature distributions, internal dependencies between features, essential for model decisions, are often overlooked. Adversarial examples disrupt these dependencies, but quantifying such changes in high-dimensional feature spaces remains computationally challenging and underexplored. To address these challenges, we propose the Copula-based Adversarial Feature Index (CAFI), a novel metric that quantifies feature dependencies in the copula space. Copulas enable explicit modeling of feature interdependencies, overcoming limitations of methods that treat features as independent. By leveraging the degrees of freedom (dofs) in copula space, CAFI quantifies how adversarial perturbations destabilize these feature dependencies. Robust models maintain stable dependencies under attack, while non-robust models exhibit significant disruption. Beyond its diagnostic utility, CAFI also serves as a regularization tool to enhance adversarial training. Experimental results in multiple datasets show that incorporating CAFI regularization improves robustness by up to 8.61% against strong attacks like FWA, while preserving competitive accuracy on benign examples. The code is available at https://github.com/lynneliu-creator/CAFI.

The increasing complexity and frequency of cyber threats necessitate the adoption of advanced machine learning models for threat mitigation. However, the opacity of these models poses challenges in trust, interpretability, and accountability. Explainable AI (XAI) has emerged as a critical solution to enhance transparency in cybersecurity applications. This paper explores the role of XAI in improving trust and interpretability in machine-learned threat detection and mitigation models. By integrating explain ability techniques such as SHAP, LIME, and counterfactual explanations, security analysts can better understand model decisions, identify biases, and ensure compliance with regulatory frameworks. Furthermore, the study evaluates the trade-offs between model performance and explain ability, emphasizing the need for a balanced approach to maintaining accuracy while improving transparency. Case studies demonstrate how XAI techniques have been successfully applied to intrusion detection systems, malware classification, and anomaly detection. The findings highlight that explain ability not only fosters trust among stakeholders but also enhances security decision-making by providing actionable insights. Ultimately, integrating XAI in cybersecurity can bridge the gap between AI-driven automation and human expertise, ensuring robust and accountable threat mitigation strategies.

In the era of personalized communication, effective marketing email campaigns hinge on the ability to predict and tailor content to individual recipients’ preferences. This paper explores innovative approaches to text prediction by leveraging idiolect and sociolect modeling derived from social media interactions, specifically Twitter. By analyzing the linguistic patterns of individual users (idiolect) and their social circles (sociolect), we aim to enhance the accuracy of word predictions, thus saving keystrokes and improving message relevance. Our methodology incorporates context-sensitive and context-insensitive prediction modules trained on user-generated tweets, evaluating their performance through keystroke savings metrics. The findings highlight the potential of personalized text prediction in optimizing marketing strategies, ultimately fostering more engaging and effective email communications. This research contributes to the field of digital marketing by providing insights into how linguistic modeling can drive personalization in automated messaging systems.

Future wireless networks are anticipated to support a wide range of human-centric intelligent services and applications that are data-intensive and resource-consuming, requiring low latency and high reliability. This motivates the adoption of a novel paradigm called semantic-aware communication, seen as a revolutionary approach with the potential to transform the design and operation of wireless communication systems. Optimization problems involving unmanned aerial vehicle (UAV) mobility and non-orthogonal multiple access (NOMA), where multiple symbols are simultaneously transmitted and detected, are particularly difficult due to their NP-hard nature. Furthermore, conventional communication networks have focused on accurate message reception without considering the semantic meaning of messages. Next-generation networks can be enriched by incorporating semantic, context-aware reasoning into their design. In this paper, our objective is to maximize the overall sum-rate performance by optimizing the power allocation based on UAV's mobility in a UAV-based NOMA network. Although classical optimization algorithms have shown impressive results over the years, their underlying complexity makes them difficult to implement in practical scenarios. Additionally, current machine learning models are limited to training results that do not leverage accumulated self-organized knowledge for future use, resulting in limited generalizability to new situations. To resolve these challenges, we propose a novel data-driven, semantic-aware framework called the Active Semantic Generalized Dynamic Bayesian Network (Active-SGDBN). The framework integrates expert knowledge (exhaustive search optimization), active inference (inspired by cognitive neuroscience), and semantic learning to develop a more intelligent, adaptable, and efficient resource allocation scheme. Simulation results have validated the effectiveness of the proposed approach in reaching a near-optimal solution and surpassing other benchmark schemes in terms of achievable sum rate.

Optical network automation is a critical enabler of modern communication networks, enhancing efficiency, reliability, and scalability to meet escalating demands for highspeed, high-capacity data transmission. By automating tasks such as routing, spectral assignment, failure management, and resource optimization, automation minimizes operational costs while improving network performance. However, widespread adoption faces significant challenges, including organizational resistance, fragmented multi-vendor environments, and the need for high-quality training data in machine learning (ML) models. Physical impairments, financial constraints, regulatory compliance, vendor lock-in, and high computational demands for realtime operations further complicate implementation. Emerging technologies like digital twins and large language models (LLMs) offer promising solutions for network automation, predictive maintenance, and intelligent decision-making. However, LLMs face notable limitations, including a lack of understanding of signal degradation, fiber attenuation, and other physical network behaviors. Additionally, they struggle with persistent memory retention, causal inference, and multi-step planning, impacting their ability to diagnose network failures and optimize long-term performance. Digital twins help mitigate these constraints by providing real-time, dynamic simulations of the physical network, enabling LLMs to assimilate real-world conditions, improve reasoning, and enhance predictive analytics. Furthermore, advances in elastic optical networks (EONs), flexible grid systems, and federated learning support dynamic routing, real-time Quality of Transmission (QoT) evaluation, and predictive failure management while preserving data confidentiality. The synergy between digital twins, LLMs, and edge AI enables continuous monitoring, proactive reconfiguration, and autonomous network control, paving the way for self-optimizing, resilient next-generation optical networks. This paper explores both technical and non-technical challenges to automation, emphasizing how digital twins and LLMs can overcome energy constraints, interoperability issues, workforce reskilling challenges, and regulatory complexities. By addressing these barriers, these technologies promise a shift toward highly efficient, scalable, and sustainable optical networks, capable of meeting the growing demands of an increasingly datadriven world.

Neck pain or cervicalgia is the discomfort in or around the cervical spine which is the region of spine located just below the head. One of the most common causes for neck pain is poor posture. Excessive use of electronic devices like computers, cell phones, etc aggravates the neck posture. This research focused on the development of a device that monitors cervical spine posture and provides feedback to the user to help them maintain a good neck posture. The device uses three flex sensors to measure neck bending and an MPU6050 IMU sensor for measuring acceleration and rotation in the neck movements. It employs an ERM motor to provide haptic feedback to the user about their neck posture. HC-05 Bluetooth modules have been used for wireless communication and a LiPo battery powers the device. A Bluetooth adapter, which contains a HC-05 and a USB-to-TTL converter, transmits sensor data to a computer running the Python script that feeds this data to a Long Short Term Memory (LSTM) Network to predict the user’s neck posture type. The neck band containing the sensors has a diameter of 110 mm and height of 120 mm. It is made up of flexible Thermo Polyurethane and has a velcro for easy wear. An enclosure made up of Polylactic Acid with its dimensions as 56 mm x 56 mm x 58 mm houses all the components of the device including the vibration motor. The enclosure has a push button, ports for uploading code and charging the battery, and a hole for sensor wires coming from the neck band. The LSTM model was trained on a dataset containing 3400 data samples. The device uses Python to feed the sensor data to the trained LSTM model which classifies the neck postures into 2 categories, good and bad. The Python script also sends the posture type predictions back to the device via Bluetooth. Real-time sensor data for every user wearing the device is saved in CSV files along with the predicted posture labels. Based on the LSTM model predictions about the neck posture type received from the python script, the Arduino code running on the device controls the vibration motor and alerts the user with two vibrations for bad posture and no vibrations for good posture. The LSTM model’s classification performance was evaluated using accuracy. Over 100 epochs, validation loss decreased from 70% to 28.51%, and validation accuracy increased from 40.59% to 95%. The accuracy obtained on the test set was 82%. User testing with 20 participants showed that the device accurately distinguished between good and bad neck postures with an overall accuracy of 79%, which varied by activity: 81% at a desk, 82% while walking, and 78% while sitting in different postures. User feedback indicated 75% found the device comfortable for extended wear, 85% found the neck band to be non-intrusive, 65% participants could easily set it up, and 40% reported minor discomfort from the vibration motor. Additionally, 80% felt the device feedback helped them correct their neck posture with 75% finding the vibration alert effective, and 70% agreeing that the usage of the device improved their neck posture awareness.

Theater performers express emotions while acting through a variety of facial expressions and vocal techniques. To improve their skills or while preparing for performances, actors rehearse their scripts and practice their facial expressions and dialogue delivery. Computer vision tools can be utilized to detect emotions from facial expressions. Audio processing tools can be used to analyze vocal characteristics like tone, pitch, loudness, etc providing insights into the emotional delivery of actors through their speech. This research focuses on developing an application that combines computer vision and audio analysis for emotion recognition, aimed at helping actors enhance their skills and performances. The emotion detection tool utilizes deep learning models to predict emotions based on actors’ facial expressions and speech. The emotion conveyed by the user through facial expressions is predicted by utilizing Mediapipe’s FaceMesh model and Blendshapes. A pre-trained wav2vec2 based audio emotion classification model predicts the emotion conveyed in the user’s speech. The emotions detected in the user’s video and speech are logged into a PDF report which also contains a customized list of links of tutorials which a particular user could learn from to improve their acting skills. The proposed emotion recognition system was tested with 20 volunteers for detecting emotions in both video and audio data. 65% of the participants confirmed the system accurately identified their facial expressions, and 75% found it useful for assessing their dialogue delivery. Participants appreciated the immediate feedback and detailed PDF reports, which proved to be helpful for them to identify their mistakes and enhance their acting skills. This multimodal approach shows promise as a valuable resource for beginner actors trying to improve their skills.

This paper presents a Model-on-Demand (MoD) approach to system identification and its integration with a three-degree-of-freedom Kalman filter-based Model Predictive Control (3DoF-KF MPC) framework. MoD estimation represents a hybrid of local and global modeling techniques, judiciously formulated to take advantage of both while not being computationally demanding.  The 3DoF-KF MPC algorithm enables responses to setpoint changes and measured and unmeasured disturbances to be tuned intuitively and independently, thereby providing superior performance and ease of use over tuning with move suppression and error weights as done with conventional MPC algorithms. The algorithm proposed in this paper involves estimating MoD-based predictive models that are seamlessly integrated into  3DoF-KF MPC to generate control actions that vary with operating conditions. This results in notable performance enhancements in the context of both SISO and MIMO control compared to conventional ARX models.  Performance and robustness of the 3DoF-KF MoD MPC framework are demonstrated in this paper through two case studies involving (i) epidemic control of a variant of the widely used SISO Susceptible-Infected-Removed (SIR) model and (ii) a nonlinear, highly interactive MIMO Continuous Stirred Tank Reactor (CSTR) model. The second case study further provides guidelines for designing informative databases for effective MoD-based MIMO identification and implementing 3DoF-KF MPC-based control for a demanding class of systems.  Overall, this paper demonstrates technological and practical improvements in system identification and control of nonlinear SISO and MIMO systems through the synergistic integration of MoD estimation and 3DoF-KF MPC, providing an effective approach for operating complex nonlinear process systems.

Wildfire-related outages significantly impact transmission and distribution systems and have been explored in most research articles and reports. However, there are instances where actual wildfires events can also lead to generation outages. These can be due to various reasons including transmission-induced generator outages, environmental factors, personnel safety and evacuation. For hydropower resources, a generator outage can lead to flooding or other dam safety issues. Hydropower in the Pacific Northwest (PNW) is a major source of frequency response on the Western Interconnection, with Oregon and Washington producing one-third of US' hydropower. This study reviewed some recent wildfires to assess the short-term impacts of hydropower outages due to wildfires, investigated the operations and dispatch during such events, and how this information can help inform long term planning of a grid with an increasing instances of wildfires events.

Remote Sensing (RS) image captioning has traditionally relied on specialized models tailored to domainspecific tasks. The emergence of large vision-language models (VLMs) offers a promising alternative due to their versatility across tasks and domains. However, fine-tuning VLMs for specific applications introduces significant challenges, including computational overhead, overfitting risks, and reduced generalization capabilities. To address these limitations, we propose RAGCap a Retrieval-Augmented Generation framework that leverages pre-trained VLMs for RS captioning without the need for fine-tuning. Our approach employs a similarity-based retrieval model (SigLIP) to identify relevant image-caption pairs from the training set. These retrieved examples, along with the test image, are processed by a multi-image capable VLM (Qwen2VL) using a carefully designed prompt structure. This enables the model to generate captions that not only accurately describe the test image but also preserve the domain-specific style. Extensive evaluations on three RS benchmark datasets demonstrate that RAGCap achieves competitive performance compared to finetuned models while offering enhanced efficiency and generalization. Our framework provides a practical and scalable solution, maintaining the versatility of VLMs while effectively adapting to domain-specific requirements. Code will be available at: https://github.com/BigData-KSU/RAGCap

Artificial intelligence (AI) is a technology which is originating incremental advances in almost every known field, be it agriculture, faceto-face academic instruction, institution-level housekeeping and waste management, or sports. In disciplines which involve high-end technical and scientific applications such as space exploration, defense, software development, medicine and surgery, and automotive design, its current presence is evidenced to be even more natural. In this paper, we discuss how it can revolutionize enzyme engineering, particularly with regards to its application in the production of bioenergy. Introducing automation in enzyme engineering can empower development of bioenergy solutions by enabling precise, rapid, and cost-effective design of catalytic systems or enzymatic systems. We highlight the synergistic potential of AI, enzyme engineering and bioenergy generation to address the challenges of fossil fuel dependence and climate change. AI ingests vast datasets for training and utilizes large quantities of computational power for implementation. Its various potential and contemporary applications in bioenergy generation such as the enablement of in-silico simulations of directed evolution are succinctly described in this paper. We delve into key advancements that can be brought about in enzyme engineering through introduction of AI. Critical discussions are presented on the possible limitations of AI models, the trade-offs between stability and activity of enzymes engineered for bioenergy production, the bottlenecks associated with experimental validation, and the precautions needed to overcome regulatory, ethical, and interdisciplinary collaboration challenges. We imagine an optimistic outlook for intelligent enzyme engineering approaches in bioenergy generation. Through future directions, we emphasize the need for interdisciplinary collaboration, hybrid modeling approaches, quantum computing, federated learning, and robust datasets integrating omics information to push the boundaries of AI-driven enzyme engineering. The convergence of AI and enzyme engineering represents a paradigm shift in the renewable energy landscape. It provides unprecedented opportunities to create sustainable and scalable bioenergy solutions. Establishing connections between computational innovations and experimental biology can expedite the global transition toward a carbon-neutral energy future.

Magnet resonance imaging (MRI) using a permanent magnet array for polarization and encoding drastically simplifies the system hardware, enabling portability. The image reconstruction is model based, where a pre-measured magnetic field map is required and its accuracy is critical for imaging. However, the magnets inevitably decay and thus requires impractical regular measurements of the magnetic field map. This study propose usage of physics-guided neural network (PgNN) to selfcorrect the encoding magnetic field decay, eliminating the need for regular measurements of field maps and making model-based MRI imaging approach practical. PgNN is trained using reconstructed images and acquired signals of phantom scans with known ground truth images. The image reconstruction signal equation is integrated into PgNN as physics concepts to guide the training to generate a correction term for the decaying magnetic field. The loss function, derived from comparing the reconstructed image to the phantom ground truth, improve PgNN's performance. The corrected encoding magnetic field improves the reconstructed image's structural similarity index measure and its perspective from the human visual system. Owing to the physics guidance, the training time is below 20 minutes and only one set of training data is required. PgNN rapidly corrects the encoding field and improves the image quality without additional cost, replacing regular and repeated measurement of the encoding magnetic fields.

In this paper, we explore the integration of two groundbreaking technologies, reconfigurable intelligent surfaces (RISs) and orthogonal time frequency space (OTFS) modulation, with a focus on their use in multicast communications to improve high-speed wireless connectivity. We introduce a phase shift design algorithm for RIS-assisted OTFS systems to maximize the channel gain in the presence of both delay and Doppler spread. Furthermore, we propose a method that utilizes the majorization-minimization (MM) algorithm to jointly optimize beamforming and adjust the phase shifts of RIS elements in a multicast scenario. We validate the effectiveness of the proposed algorithms by numerical simulations. The results demonstrate significant performance improvements over existing benchmark approaches.