ABSTRACT. Communication systems operate across multiple spatial and physical scales, from molecular interactions in biological environments to radio-frequency propagation in large-scale wireless networks. Each scale exhibits distinct information carriers, channel dynamics, and performance limits, yet emerging applications increasingly require integrated, environment-aware design across scales. This presentation introduces environment-aware connectivity as a unifying framework for sensing and communication systems spanning nanoscale, body-area, and vehicular domains.

At the nanoscale, molecular communication is governed by stochastic transport and biochemical reactions. Beyond classical Shannon-based metrics, the role of semantic information is discussed, with synthetic cells proposed as a platform to investigate how meaning emerges in physical communication processes. At the mesoscale, implantable neural interfaces for peripheral nerve injury rehabilitation are examined, including cuff-electrode technologies, biophysical modeling, machine-learning-based signal classification, data compression, and low-latency wireless transmission for closed-loop operation. At the macroscale, environment-aware localization of ground radio transmitters using unmanned aerial vehicles is presented as a case study in dynamic propagation conditions. Overall, the talk highlights how environmental awareness enables cross-scale communication paradigms bridging biological, biomedical, and wireless systems.

ABSTRACT. Returned goods management is a significant problem in supply chain logistics, where vehicles should dynamically plan optimal tour schedules with respect to time windows, priorities, and estimated demands. Traditional heuristics are fast, but not adaptive. In this paper, we discuss reinforcement learning (RL) strategies DQN, DDQN+PER, and multi-agent RL (MARL) against a baseline heuristic. Using a simulated reverse logistics environment and 100 training episodes per agent, results demonstrate RL-based agents surpass heuristics to minimalize distance/waste and lateness. DQN minimizes lateness and distance, and MARL minimizes multiple objectives. The paper emphasizes the potential of RL for dynamic and data-driven logistics optimization.

ABSTRACT. Maritime safety plays a vital role in protection of coastal belt and ocean resources. With the growth of maritime traffic, the need of surveillance is also increasing. Traditional surveillance methods like Radio Detection and Ranging (RADAR), Sound Navigation and Ranging (SONAR) are useful and used to detect small objects in low-visibility sea conditions. This research presents a real-time maritime surveillance using the You Only Look Once (YOLO) deep learning architecture. A custom dataset was developed, that contains the images of ships, unidentified surface and aerial objects whose images are captured in dynamic weather conditions. The proposed model was trained and optimized to deliver accurate, real-time detection, integrating easily with data feeds from coastal cameras, UAVs, and radar systems to enhance situational awareness at sea. Experimental results show promising performance, achieving 89.7% precision, 90.5% recall, and a mean average precision (mAP@0.5) of 91.2%, outperforming common detection models such as Faster R-CNN and SSD. These findings highlight the effectiveness of deep learning in supporting safer, more efficient maritime operations. The proposed framework provides a scalable solution for coastal monitoring, marine traffic management, and environmental observation, with future development aimed at sensor fusion and improved visibility performance in challenging sea conditions.

ABSTRACT. Many families have preserved photographs of their ancestors in the form of photo negatives so that they can introduce their children to their elders and have them always remain in their memories. But the technology to convert photo negatives into images is now obsolete. It is needed to provide application based solution to convert these photo negatives into digital images so that people can preserve these assets. The "SMART IMAGESHOP" is an application that addresses the challenges of navigating vast image databases and transforming negative images into high-definition masterpieces. With advanced search functionalities, privacy assurances, and image enhancement capabilities, the application offers a groundbreaking solution. In today's data-rich world, efficient image classification and retrieval are critical. This application utilizes smart algorithms and advanced search mechanisms to enable users to easily navigate extensive databases while ensuring privacy by granting access only to the uploader. The innovation extends to image enhancement, empowering users to convert negative images to high-definition quality using their mobile phone's camera. This process involves sophisticated algorithms capable of producing vibrant and flawless images, eliminating the need for traditional, time-consuming methods, and reducing costs. With a user-friendly interface, cross-platform compatibility, and a focus on user comfort, the "SMART IMAGESHOP" project stands as a pioneering solution in image management and enhancement, revolutionizing how individuals interact with visual data.

ABSTRACT. The world has advanced technologically, with new inventions emerging that are better than the ones that came before. One such invention is the smartphone, which is used daily by almost everyone for purposes such as taking pictures, browsing the internet, texting others, etc. Typing text is one of the things that people use most frequently, and it can take a long time to write every word. Word prediction is used to make people's lives easier by displaying the word even before they finish it. It also helps to avoid vocabulary errors and increase productivity at work. This is categorized as natural language processing. In, this paper we talk about the next word prediction models that can help to predict the next word and compare them to find the model with higher accuracy. The models are Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU) and Bi-Direction LSTM. The libraries used are NumPy, keras, TensorFlow, nltk, regex, pickle and matplotlib.

ABSTRACT. In this work, a semi-blind approach for beamforming in a multi-access point (AP) association scenario for IoT networks is presented. For this purpose, the signal-to-leakage-plus-noise ratio (SLNR) based leakage probability is derived in closed form using the recently developed technique of statistical characterization of ratio of Indefinite Quadratic Forms (IQFs). The closed form expression for the leakage probability is thereby utilized in designing optimized semi blind transmit beamforming. An additional advantage of the proposed approach is its ability to facilitate the analysis of multi-AP association task in the presence of leakage. Theoretical findings are validated through Monte Carlo simulations. Furthermore, the results demonstrate that the proposed semi-blind beamforming strategy within a multi-AP association framework for IoT devices can significantly improve the network performance.

ABSTRACT. A predictive and slack-aware resource allocation framework for 6G Mobile Edge Computing systems is designed and presented in this paper. This framework integrates short- term workload forecasting, proactive admission control, and joint optimization of communication and computation resources under uncertainty. A stochastic optimization problem with slack margins is formulated to tolerate prediction errors, solved via an adaptable two-stage approach: predictive reservation using linear programming and real-time fine-tuning with convex optimization. Evaluated in a comprehensive 6G-MEC scenario with 10 edge servers, 500 users, and task classes (URLLC, eMBB, mMTC), the framework reduces average latency by 32.0% (from 48.7 ms to 33.1 ms), increases resource utilization by 43.6% (from 65.8% to 94.5%), boosts admission rate by 30.0% (from 0.70 to 0.91), and cuts latency violations by 60.0% (from 14.5% to 5.8%) compared to baselines. Robust under ±15% prediction errors, it enhances reliability and efficiency for 6G deployments.

ABSTRACT. The modern process of textile manufacturing needs automated defect detection of fabrics in the manufacturing process because the manual inspection is slow, inconsistent and subject to error. The latest advances in the deep learning industry have improved fabric inspection, however, the systems based on the Convolutional Neural Networks (CNNs) can give discontinuous and noisy segmentation outputs in reaction to illumination, texture, and fabric motion variations. The paper introduces a Quantum-Inspired Optimization-based Convolutional Neural Network (QIO-CNN) that will be applicable to overcome these limitations and assist in identifying defects in knitted fabrics in real-time. The suggested architecture is built upon CNN and QIO module to enhance pixel-level predictions using a Quadratic Unconstrained Binary Optimization (QUBO) formulation. The QIO module provides spatial smoothness and coherence between adjacent pixels and is a global filter on CNN-generated probability maps. The proposed model is tested on MATLAB/Simulink to evaluate the real-time feasibility of the proposed model on the industrial circular knitting data. The results of the experiments have revealed that the proposed approach has been able to reinforce intersection over union (IoU) and F1-score by an average of 4-5 percent relative to classical CNN approach, and at the same time, reduces false alarm rates by half. Visual inspection confirms the existence of more continuous defects and greater stability in changing lighting conditions. The results also validate the fact that QIO enhanced the spatial accuracy of deep learning results without increasing the complexity of the network.

ABSTRACT. Pakistan’s agriculture faces collective hazards from climate abnormality, comprising floods, droughts, and random rainfall, which excessively distress smallholder farmers. Outdated decision-making procedures are unsatisfactory, which generates a requirement for adaptive, data-driven results. This review studies Adaptive Decision Support Systems (ADSS) from a Human–Computer Interaction (HCI) tactic, emphasizing how AI-driven analytics, real-time environmental statistics, and user-centered interfaces can advance climate-resilient cultivation. Confirmations from international and native case interpretations specify that mobile-first, multilingual, and involved ADSS designs with response loops progress usefulness, optimization, and farmer management. Little digital knowledge, restricted infrastructure, and acceptance obstacles are foremost trials of Pakistan; these can be talked through government-sponsored digital knowledge platforms, restricted mobile/voice advisories, and amalgamation of ADSS with extension facilities. Upcoming study should focus on moral and understandable AI, collective co-design, cognitive-load–aware interfaces, and association with countrywide digital agriculture bionetworks. Appropriately designed ADSS can be used as a transformative apparatus to reinforce food safety, reassure maintainable farming, and shape resilience in contradiction of climate dangers in Pakistan.

ABSTRACT. The size of large language models have been increased massively and this makes it harder to run them in a resource-constrained environments. To address this issue, many optimizations techniques have emergered with pruning LLMs being one of them. Structured pruning, a branch of pruning category, compresses LLMs by removing entire groups of parameters structurally with the requirement of further finetuning or pretraining to recover the performance when required. To the best of our knowledge, this is the first in-depth survey paper solely focused on the structured pruning of large language models. Twenty-five papers in this category were examined, out of which 18 represent widely recognized methods and 7 are recent approaches. To organize these methods, a four-dimensional taxonomy based on pruning granularity, importance estimation, recovery strategy, and pruning schedule is introduced. The analysis shows clear trends: the dominance of component-level pruning, gradient-based importance scores, parameter-efficient tuning, and one-shot pruning across the four dimensions, respectively. Finally, the emerging trends, challenges, and future directions are discussed. This survey aims to guide researchers and engineers seeking efficient ways to bring large language models to resource-constrained environments.

ABSTRACT. Accurate precipitation prediction from climate models is crucial for regional climate adaptation and water resource management. However, coarse-resolution CMIP6 models struggle to capture the spatial and temporal patterns of precipitation occurrence, particularly in topographically complex regions like Pakistan. Traditional forecast methods have limited success in predicting rain occurrence due to the highly skewed distribution and spatiotemporal dependencies of precipitation. We develop a machine learning framework using engineered spatiotemporal features and threshold optimization to predict daily precipitation occurrence (≥1mm/day) from CMIP6 data over Pakistan. Our feature engineering incorporates temporal lags, spatial gradients, rolling statistics, and cyclical temporal encodings. We train and evaluate the model using ERA5 reanalysis as reference data over a 10-year period, comparing against Random Forest baseline and raw CMIP6 predictions. The model achieves a Critical Success Index (CSI) of 0.342, 3% higher than that of Random Forest (CSI: 0.332). It is consolidated from literature that CSI in ranges 0.3-0.4 is considered modest, with study and region specific contributions. The framework is computationally efficient and generalizable to other regions and climate variables

ABSTRACT. This paper reviews the integration of chatbots and conversational AI in education, highlighting their role in enabling personalized learning, providing 24/7 support, and increasing student engagement. While these AI systems offer significant benefits by adapting to individual needs and reducing learning stigma, their adoption is hampered by persistent challenges, including privacy concerns, algorithmic bias, and contextual limitations. The paper synthesizes current applications, discusses ethical implications, and identifies future trends, concluding that the path forward requires a focused effort on developing transparent, adaptive solutions and providing comprehensive educator training.

ABSTRACT. Predictive maintenance will be one of the mainstays of modern industrial systems, driven by developments in the Internet of Things, digital twin, and other intelligent sensing technologies. This paper presents a sensor-based predictive maintenance framework with a lightweight digital twin model for a conveyor belt system. The configuration includes load cells, current sensors, temperature sensors, and rotary encoders to track load fluctuations, motor health, thermal behavior, and movement of the belt. In real time, the data obtained from these sensors are processed using Python and mapped onto a 3D digital twin model created in Unity/Blender, which allows dynamic visualization of system performance, offering early detection of anomaly conditions. The PdM process begins with gathering sensor data that is analyzed against pre-defined thresholds, indicating anomalies such as increased consumption of current, overheating, abnormal load, or fluctuations in speed. The maintenance score will be updated in real time, and when it reaches the critical threshold, alerts will be produced and recorded into a database for historical analysis. The dashboard presents a flowchart to show the status of systems and logic of decision making visually. In order to allow remote monitoring, the system will send alert messages and show real-time sensor data, anomaly alerts on a user-friendly dashboard. It also provides on-site visual feedback through an LCD and mirror setup within a physical display box. In contrast to the current state of digital twins, which require proprietary cloud technology, expensive sensor networks and integration hardware to run, the offered framework is implemented as a low-cost, lightweight solution and can be scaled to support specific conveyor systems, including small and medium-sized ones, while still allowing real-time anomaly detection and visualization of maintenance predictions. The proposed strategy is cost-effective and scalable for small to medium-sized industrial environments with real-time sensing, predictive maintenance. It further offers digital twin visualization through both virtual and physical displays.

ABSTRACT. PrintFusion is an innovative e-printing platform designed to revolutionize remote printing through smart and AI-driven technology. For users such as students facing tight deadlines, traditional brochure design and printing can be a lengthy process involving multiple redesigns and time-consuming vendor searches. PrintFusion streamlines this experience by providing AI-powered design recommendations to help users select optimal colors, fonts and layouts to achieve professional results without the need for trial and error. Beyond design support, PrintFusion simplifies vendor selection by offering AI-enhanced suggestions that consider price, quality and delivery speed to ensure that users receive high-quality prints efficiently. By addressing key market gaps such as limited customization and fragmented workflows, our PrintFusion delivers a smarter, more seamless and user-friendly printing solution. This all-in-one platform empowers users with convenience, advanced customization and reliable vendor matching thus reflecting latest advances in AI-driven design and automation

ABSTRACT. Traditional industry performs their operations with conventional technologies such as mechanical machines, standalone computers, and human resources. The Industrial Revolution changed industrial operations with the deployment of emerging technologies and fewer human resources for industrial operations and management. In the smart industry, the industrial Internet of Things (IIoT) is used for data collection, processing, and analysis. IIoT is also used for automation, remote monitoring, and communication of devices with each other. The smart industry uses various devices, servers, applications, protocols, and continuous connectivity with the internet; its attack surface has increased due to multiple threats and vulnerabilities. IIoT is exposed to various cyber-attacks such as ransomware, cross-site scripting, Man in the middle, sniffing, and spoofing. Intrusion detection systems (IDS) have been used as technical control for the cyber defense of digital assets and services. IDS has been used with misuse and anomaly-based methods. Misuse methods are not practicable for threat mitigation in smart industry due to their limitation, such as the detection of known attacks which are stored in signature database; however, emerging attacks, such as ransomware or DDoS use various strategies to initiate cyber-attacks on systems. Deep learning (DL) based IDS are more efficient due to its high-performance metrics. In this paper, we have discussed IIoT importance in the smart industry and cybersecurity challenges related to IIoT. Furthermore, we have briefly described IDS classification and IIoT datasets for the development of DL-based systems. Finally, we have provided literature on DL-based security solutions for threat mitigation in IIoT environments.

ABSTRACT. Globally, 25% of the GHG (Green House Gas) emissions come from transportation. Out of this quarter, 70% of the emissions come from on road vehicles. To mitigate this problem smart mobility and electric vehicles are necessary to adopt. This study combines the data from EV Data Explorer, International Energy Agency and the World Bank database, to study the regions of Europe and China in their efforts to implement EV and smart mobility to mitigate carbon emissions. In Europe, the number of EVs increased exponentially while the transport CO₂ emissions declined slightly, indicating partial decoupling. In China, EV stock rose rapidly while a simultaneous increase in emissions was observed as well, because of reliance on fossil-based electricity generation. The correlation analysis found a strong positive value for China and weak negative value for Europe. The findings demonstrate that simply adopting EVs will not support deep decarbonization, without reducing carbon emissions from power grid. This will be importantly supported by policies that integrate smart mobility and data-driven solutions. The experiences of Europe and China can provide points of reference for developing economies such as Pakistan.

ABSTRACT. As customers’ reviews become more essential in informing people about the quality of the product, there is a pressing need to analyze customers’ sentiments and opinions regarding their experience. This study aims to develop an opinion mining framework to assess and analyze customer reviews regarding the quality and services of the product. By using machine learning and natural language processing techniques, this research explores extracting and classifying customer opinions from reviews and converting them into the count of positive, negative, and neutral opinions. A big dataset of fine food product reviews containing 568454 reviews was collected from kaggle.com. In this study, three machine learning algorithms, namely Naïve Bayes, Support Vector Machines, and Logistic Regression, were employed with Term Frequency- Inverse Document Frequency (TF-IDF) and Bag of Words vectorizers to analyze the sentiments. The classification models were compared using evaluation parameters. The findings of this study provide that SVM with TF-IDF outperformed the other models with an accuracy of 72.3% on the big dataset.

ABSTRACT. Poor diet and lifestyle issues are promoting gastric problems and abdominal diseases. Such problems, although they have varying causes, depend on the type of food being consumed. Hence, to study the effect of food on gastric activity, the gastric mobility signals first need to be monitored, for which Electrogastrography (EGG) signals are being used to analyze the electrical activity of the stomach. However, the food-specific classification has not yet been performed. This paper presents a novel approach to classify the Pre-Meal, Junk-Triggered, and Dairy-Induced Gastric Responses using EGG signals. For this three-class problem, six different machine learning (ML) classifiers have been used, and the highest train accuracy of 92.95% and Test accuracy of 92.44% with the bagging k-NN classifier has been achieved. Further optimization techniques could be applied in the future to enhance the evaluation scores. Moreover, these models could be deployed in real-time gastric monitoring systems.

ABSTRACT. This contribution presents an algebraic precoding and decoding scheme based on the COSMIC (Connectivity- Oriented Sensing Method for Imaging and Communication) approach. COSMIC takes advantage of the limited radar field of view to embed information while preserving antenna orthog- onality. The result is joint high-resolution radar imaging and reliable data transmission. Simulations confirm the advantages of COSMIC over state-of-the-art Integrated Sensing and Communi- cation waveforms in sidelobe suppression and spectral efficiency. Index Terms—6G, Integrated Communication and Imaging,

ABSTRACT. The advent of Noisy Intermediate-Scale Quantum (NISQ) machines has brought new opportunities for expanding machine learning abilities with the help of quantum computing. Herein, a complete hybrid quantum-classical learning system is introduced that incorporates classical preprocessing, quantum feature extraction, and classical machine learning in a strategic way to surpass the performance of existing systems for the tasks of complex pattern recognition. We perform large-scale experiments with two popular benchmark datasets: the Iris flower dataset and the Wisconsin Breast Cancer Diagnostic dataset. Our framework shows consistently improving accuracy of 2-7% over stark classical methods without sacrificing practicality on state-of-the-art NISQ hardware. With rigorous ablation studies and cross-validation with cutting-edge approaches, we conclude that the best task distribution between traditional and quantum processors is key to near-term quantum advantage. Our introduced framework effectively achieves a quantum circuit depth reduction of 40% using smart classical preprocessing along with improving classification performance using quantum-augmented feature representation. Our implementation offers both hardware-based execution with Qiskit as well as fast classical simulation modes, so it is available for researchers with limited quantum computing capabilities.

ABSTRACT. Personality detection from human text has gained research attention in recent years. Using deep learning and natural language processing, it has applications in career pre- diction and human-computer interaction. Our survey introduces a new taxonomy that divides deep learning approaches into six model groups. These include Transformer, RNN, CNN, Graph, Hybrid and Custom models. Based on our analysis from 22 studies, we give a comparison of features and dataset usage. Our review shows that transformer-based models achieve state-of- the-art accuracy but they require high computational resources. The main challenges in personality detection have been limited datasets and high computational demands, which need special attention. This survey serves as a reference for researchers in choosing suitable personality prediction methodologies

ABSTRACT. Floating photovoltaics FPV systems are a renewable energy source that has been proposed as a potential solution, especially in areas with limited land and rising energy consumption. This review analyses the international opportunities for reservoir-based floating solar farms and measures their effects on aquatic ecosystems in recent literature (2018-2025) and in global FPV datasets. FPV systems have a cooling effect of water, which makes them 5% to 10% more efficient than land-based solar systems. The evaporation losses can be reduced by covering less than 15% to 20% of the total reservoir area without interfering with the ecological conditions. However, insufficient coverage can change thermal stratification and light penetration, dissolved oxygen dynamics, and nutrient cycling, which can affect aquatic biodiversity. Key gaps in ecological monitoring, material sustainability, and policy frameworks are found in developing regions. The results highlight the importance of standardizing environmental assessment procedures, combined with techno-environmental planning and policy development, to achieve the sustainable implementation of FPV systems. In general, reservoir-based FPV is an option that opens a new opportunity to increase clean energy production without adversely affecting water resources or their ecosystem reliability.

ABSTRACT. This study provides a comprehensive genomic and proteomic characterization of the Naegleria fowleri Karachi_NF001 strain, revealing critical insights into its pathogenic mechanisms, evolutionary lineage, and immunogenic potential. Phylogenetic analysis demonstrated a 90–100% sequence similarity with other N. fowleri strains, supporting conserved genetic lineage and regional adaptation. Eight Open Reading Frames (ORFs) were identified, with ORF135 being the longest at 6,633 bp and potentially encoding a protein of approximately 240 kDa. Conserved domain analysis revealed functionally significant domains, including nucleotide transferase-related (pfam10127, COG3641), cytochrome oxidase assembly (SURF1, Shy1), calcium signaling-related (pfam06449, pfam01365), and membrane interaction (MIR domain, pfam02815), all with E-values <1.0E-10, indicating strong functional conservation. The protein-protein interaction (PPI) network comprised 232 interactions among 51 proteins, with a high clustering coefficient of 0.751 and key hubs involved in protein kinase activity (e.g., D2V5S4_NAEGR) and cyclic AMP turnover. Functional enrichment highlighted Reactome pathway involvement in heme biosynthesis and iron metabolism, KEGG pathway activation in purine/pyrimidine metabolism, and STRING clusters related to phosphodiesterases and transporters. Immunogenicity prediction identified multiple peptides with potential vaccine value, including one with high HLA-DR7 (Class II) binding affinity of 0.87 and several with Class I affinities ranging from 0.25 to 0.34; these peptides were also non-toxic and non-allergenic with immunogenicity scores >0.6. These findings collectively advance the understanding of N. fowleri’s molecular biology, support future epitope-based vaccine development, and provide a reference for targeted therapeutics in the context of Primary Amoebic Meningoencephalitis (PAM).

ABSTRACT. Breast cancer has become one of the causes of mortality of women, which points to a crucial necessity of effective and precise diagnostic methods. This research study provides an in-depth comparison of supervised and unsupervised machine learning on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, which is comprised of 569 patient’s records with 30 diagnostic features. Multiple supervised machine learning models like Logistic Regression, support vector machine, k-Nearest Neighbors, random forest , gradient boosting, and XGBoost were evaluated along with some unsupervised machine learning models including K-means, Gaussian Mixture Model and Hierarchical clustering. Supervised ML models were evaluated using evaluation matrices such as Accuracy, Precision, Recall (Sensitivity), Specificity, F1-score and ROC-AUC, whereas unsupervised ML model were evaluated using metrics such as Silhouette, Calinski Harabasz, and Davies-Bouldin. Results showed superiority of supervised models compared to unsupervised models in predictive accuracy and diagnostic reliability. SVM was the most precise with the highest accuracy of 98.2%, whereas the F1 score and recall was the highest of 97.62% as compared to other Supervised ML models. Conversely, the unsupervised ML models like Hierarchical Clustering and Gaussian Mixture Models had aligned accuracies of about 90% supporting the claim that they could be used in exploring data and grouping the features in the absence of labels. When comparing each model using the analysis of feature ranking, it was found that worst area, mean concavity and worst radius are the most powerful predicting features of the malignancy. This study presented an evidence-based comparison of early breast cancer detection based on supervised and unsupervised ML models, focusing on ways to remain robust when only limited data are available and explainable in terms of their importance.