ALMUSTAFA ENGINEERING JOURNAL

This work employs an analysis-and-simulation technique to quantitatively analyze the performance improvements of 4×4 MIMO systems, a key element in with multiple transmit and receive antennas; created through three typical adaptive beamforming approaches: Least Mean Squares (LMS), Recursive Least Squares (RLS) and Minimum Variance Distortionless Response (MVDR). This study particularly targets the problem of severe inter-channel interference and fast channel dynamics in dense urban channels (Rayleigh fading) as detailed below. This method includes two main approaches, which are to analytically develop the 4×4 MIMO mathematical model and analyze it in the first phase, while silence shows comprehensive simulation in a MATLAB environment investigating key performance metrics such as Bit Error Rate (BER) versus Signal-to-Noise Ratio (SNR). An analysis is performed for comparing convergence rate, jamming cancellation capability and computational complexity. The main hypothesis is that MVDR will provide the best interference suppression and LMS will be preferable on account of low complexity, making RLS a compromise. The results reveal the optimum beamforming structure that achieves the desired performance-complexity trade-off for practical deployment in high-density 5G networks.

Modern communication networks have undergone rapid growth due to the increasing number of users and the expansion of advanced digital services such as 4G/5G networks, the Internet of Things (IoT), and cloud computing. This growth has led to increased network complexity, making communication infrastructures more vulnerable to technical and operational faults that negatively affect network reliability and Quality of Service (QoS). Traditional fault detection methods mainly rely on periodic monitoring and post-fault analysis, which often result in delayed detection and limited predictive capability.This research focuses on designing and implementing an intelligent system for early fault detection in communication networks using deep learning techniques. The proposed system analyzes network performance data, including latency, packet loss, throughput, and error rates, to identify abnormal patterns that indicate potential faults before their occurrence. The system architecture consists of integrated modules for data collection, data preprocessing, deep learning-based analysis, fault detection and prediction, and performance evaluation. Several deep learning models are employed in this study, including Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks. In addition, a hybrid model is proposed to combine the strengths of these architectures in order to enhance detection accuracy and reduce fault detection latency. The system is evaluated using simulated network data and assessed based on multiple performance metrics such as accuracy, precision, recall, F1-score, false alarm rate, and detection latency. The experimental results demonstrate that deep learning-based approaches significantly outperform traditional fault detection methods. In particular, the LSTM model shows superior performance in handling time-series network data, while the proposed hybrid system achieves the highest overall accuracy, the lowest false alarm rate, and the fastest fault detection time. The findings confirm the effectiveness of deep learning techniques in improving early fault detection, reducing network downtime, and enhancing overall network reliability.

The Finite Element Method (FEM) is one of the most important numerical methods used in solving partial differential equations (FDIs), with wide applications in engineering and science. This is due to its high flexibility in representing complex geometric domains and handling diverse boundary conditions. However, the accuracy of the resulting numerical solutions is directly affected by the nature of the numerical approximation, particularly the size of the elements, the order of the approximation functions, and the quality of the numerical network. This leads to the emergence of what is known as asymptotic error, especially in problems with complex behavior or multiscale problems. This research aims to improve finite element methods by integrating optimization techniques with adaptation strategies to reduce asymptotic error and achieve an effective balance between solution accuracy and computational cost. The study relies on a theoretical analysis of error sources in FEM, including approximation error, partitioning error, and pre- and post-error analysis, in addition to examining the concepts of stability and convergence. Various adaptation techniques were also reviewed, such as element size-dependent (h), approximation order-dependent (p), and mixed adaptation (hp). On the practical side, a proposed algorithm was developed that estimates the dimensional error and incorporates it into an optimal optimization model. This model defines an objective function representing the magnitude of the asymptotic error, subject to constraints related to the number of degrees of freedom and computational cost. The results showed that the proposed algorithm is capable of reducing the asymptotic error and improving the convergence rate compared to traditional strategies, while also allocating numerical resources more efficiently within the studied domain. The study confirms that integrating optimization with Finite Element Method (FEM) represents an effective methodological approach for developing numerical solutions and improving their reliability in engineering and scientific applications.

There are three primary obstacles to the efficient conversion of photocatalytic CO₂; slow charge transfer, low efficiency of light capture and large thermodynamic energy barriers at the solid-gas interface. To address these limitations, we have created a hybrid composite material that utilises Au as a dual-functional active material based on plasmonic-photothermal effects and has been developed via the chemical/fault engineering of MnOₓ/CN. Upon illumination with 1 Sun (82.4°C), the resonance frequency of Au is enhanced providing thermal/photonic contributions into the CN matrix. Operando diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) and computational density functional theory (DFT) studies employed the engineering of both defects in combination to produce an internal electric field that facilitates charge transfer via the S-scheme pathway and increases the lifetime of both positive (6.88 nS) and negative charges through optimising their redox potential. Together this mechanism decreases the activation energy barrier for the rate-limiting *COOH intermediate by almost 50%, giving rise to an energy barrier of +0.32 eV when under standard thermodynamic conditions (approx. 1-2 eV). The overall improvement in performance of this composite heterostructure (i.e., nanoscale morphology) resulted in superior catalytic and stability test results (CO=112.8 μmol/(gh), CH₄=18.52 μmol/(gh)) over 100 hours continuous operation, ultimately resulting in a change in the way we use both dual defect engineering with plasmonic management of thermal energy to overcome current thermodynamic barriers to solar-to-chemical energy conversion.

Real-time obstacle detection constitutes a foundational requirement for safe autonomous navigation in dynamic, unstructured environments. While deep learning architectures have achieved remarkable accuracy in visual perception tasks, their computational complexity, memory footprint, and power consumption frequently exceed the operational constraints of low-cost embedded platforms deployed on mobile robots. To address this critical limitation, this paper presents a lightweight, vision-based obstacle detection system specifically engineered for resource-constrained robotic applications. The proposed pipeline integrates computationally efficient handcrafted feature extraction—combining Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP)—with a hardware-optimized linear Support Vector Machine (SVM) classifier. The system operates entirely on CPU, requires no GPU acceleration, cloud offloading, or model parallelization, and is explicitly designed for deterministic execution on ARM-based embedded processors. We evaluate the framework on a custom-curated dataset of 8,500 RGB images captured across diverse indoor and outdoor environments under varying illumination conditions, surface reflectivity, and partial occlusion scenarios. Quantitative experiments demonstrate that the optimized model achieves 96.4% classification accuracy, 97.1% precision, 95.8% recall, and an F1-score of 0.964 on a strictly held-out test set. Crucially, the system sustains an average end-to-end inference latency of 8.2 milliseconds per frame on a Raspberry Pi 4 Model B operating at 1.5 GHz, with a peak memory footprint of 45.2 MB and a power draw of approximately 1.2 W during continuous operation. These metrics correspond to a sustainable throughput of ~122 frames per second, significantly exceeding the 30–60 Hz sampling rates required for reactive collision avoidance and path planning in mobile robotics. Our primary contribution is a mathematically grounded, hardware-aware vision pipeline that successfully reconciles high detection reliability with strict real-time processing constraints. By prioritizing linear computational complexity, cache-friendly memory access patterns, and margin-maximizing classification in a fused handcrafted feature space, the proposed system offers a practical, reproducible, and energy-efficient perception layer for low-power autonomous platforms operating in bandwidth-constrained or safety-critical scenarios where deep learning deployment remains infeasible.

Nonlinear convection-diffusion-reaction equations are fundamental to CFD, with applications ranging from environmental modeling to reactive flows. While the Finite Volume Method (FVM) is prized for its conservation properties and low cost, it struggles with steep gradients and discontinuities. Conversely, Physics-Informed Neural Networks (PINNs) offer mesh-free flexibility but suffer from slow convergence. This study introduces Learning-Enhanced Finite Volume Methods (LE-FVM), a hybrid framework that integrates FVM with state-of-the-art deep learning paradigms: PINNs, Finite Volume Graph Networks (FVGN), and Kolmogorov-Arnold Networks (KANs). By embedding ML models directly into the FVM workflow—handling flux calculations, adaptive mesh refinement, or acting as end-to-end solvers—the framework ensures solutions are strictly constrained by integral conservation laws. Key innovations include a "twice-message aggregation" mechanism for unstructured grids and an adaptive loss-weighting scheme based on the Neural Tangent Kernel (NTK) to balance PDE residuals and boundary conditions. Benchmark evaluations on 1D and 2D systems (Burgers’, Fisher’, Burgers–Huxley) demonstrated that LE-FVM provides high-fidelity solutions over 95% of the time. For the 1D Burgers’ equation ( Re=1 ), LE-FVM reduced maximum absolute error by 173x compared to the Galerkin Finite Element Method (GFEM). In unsteady 2D flows, FVGN achieved 77% higher accuracy and 56% lower training time than purely data-driven models. Furthermore, the framework exhibited exceptional generalization to novel geometries (airfoils and ellipses) without retraining. By blending deterministic conservation laws with stochastic learning, LE-FVM offers an order-of-magnitude speedup for parametric studies, setting a new bar for solving complex nonlinear PDEs in science and engineering.

The constant shifts in today’s global supply networks mean, more than ever, the necessity for custom, efficient, resilient, and adaptable processes in supply chain systems. It has been proven that fluctuating market volatility and customer demand, geopolitical tensions, inclement weather, and the implementation of many different digital technologies show that old ways of supply chain management do not cut it. This means organizations require next-generation systems, technologies, and processes that enable rapid, seamless decision-making and continuous operational excellence. The emergences of BDA has solidified itself as one of the most important, and vital, tools for organizations. which can refine processes, be more proactive and prepared for challenges, and hone their operational excellence. This paper focuses on the integration of BDA and AI in supply chain management and how it is being applied to enhance vital capabilities, including risk management, demand forecasting, sustainability, and collaborative decision making. AI, through predictive analytics, provides organizations with the frameworks to map out supply chain interdependencies, mitigate, and expose the hidden risks that lead to uncertainty. BDA allows these organizations to foster greater alignment with their supply chain partners. Moreover, this work examines the methods, frameworks, and insights proving the impact of intelligent analytics on visibility, agility, and overall operational performance enhancement of a supply chain. These analytics enable real-time visibility, predictive maintenance, and dynamic routing, as well as adaptive routing—all vital for successfully competing in rapidly changing markets. This synthesis further analyzes the challenges pertaining to the scalability of AI systems, the data quality and diversity, and the opacity of design issues of some machine learning algorithms. These challenges justify the uneven adoption of advanced analytics and call for more thorough guidance and benchmarking for the frameworks employed. This work addresses the challenges identified, concentrating on the real-time adaptability, sustainable process optimization, and cross-organizational data governance and human–AI collaboration as the most promising lines of development for supply chain intelligence. Exploring these will help balance the need for a responsive and resilient supply chain system.

Fractional differential equations' (FDEs) subject t'o non-local boundary conditions' a're fundamental t'o modeling a'nomalous' t'ransport, viscoelastic relaxation, a'nd systems' with long-range memory dependencies. t'he coexistence of fractional-order differential operators' a'nd integral/non-local boundary constraints' presents' significant computational challenges, primarily due t'o t'he global coupling of degrees' of freedom a'nd t'he loss' of spectral a'ccuracy near boundary layers. t'his' paper develops' a'nd rigorously a'nalyzes' a' hybrid spectral scheme t'hat couples' a' Chebyshev–Galerkin spatial discretization with a' graded convolution quadrature t'emporal integrator, a'ugmented by a' constraint-preserving projection t'echnique for t'he non-local boundary operator. t'he proposed framework reformulates' t'he continuous' problem into a' structured, well-conditioned a'lgebraic system while maintaining exact enforcement of t'he integral boundary condition. t'heoretical a'nalysis' establishes' unconditional stability a'nd derives' optimal L^2-error estimates' t'hat explicitly depend on t'he fractional order \alpha\in(0,1), solution regularity, a'nd polynomial degree N. Comprehensive numerical validation on linear, nonlinear, a'nd variable-order benchmark problems' demonstrates' exponential spatial convergence a'nd t'emporal convergence rates' of \mathcal{O}(\tau^{min(2,2-\alpha)}) on suitably graded meshes. Comparative benchmarks' a'gainst established finite-difference t'ime-stepping a'nd pure spectral collocation methods' confirm superior a'ccuracy-to-computational-cost ratios' a'nd enhanced robustness' under varying non-local parameters. t'he proposed hybrid methodology provides' a' reproducible, high-fidelity computational framework for t'he numerical simulation of constrained fractional dynamical systems.

In this review paper, the structural behavior of reinforced geopolymer concrete (GPC) beams is analyzed based on the synthesis of data obtained experimentally, numerically, and comparatively in peer reviewed literature from 2014 till 2025. Geopolymer concrete (a material synthesized from industrial aluminosilicate byproducts such as fly ash (FA), metakaolin (MK) and ground granulated blast furnace slag (GGBS) by means of alkaline activation) has generated significant research interests as an eco-friendly replacement of OPC concrete. LCA studies reveal reduction in carbon emissions from 30% to 60% with respect to OPC concrete. In this paper, various important parameters that influence the behavior of GPC beams with respect to flexural performance, shear strength, ductility, bond strength, seismic resistance, fire resistance, and durability will be reviewed. Fibers are found to enhance the post crack behavior and fracture energy of the beams. Seismic behavior of GPC beam-column joints have been found to satisfy the design codes. Important knowledge gaps still remain in long term creep and shrinkage properties, cyclic behavior, residual strength after fire exposure and standardization of design code for GPC structures.

Despite the progress made by Wi-Fi 6 network, they still suffer from interference in dense environments. This research aims to enhance the channel selection process using machine learning algorithms to reduce the interference. Synthetic dataset generated to simulate dense wireless network using the network properties such as RSSI, noise level, channel width, channel utilization, the number of adjacent access points, and interference level. The Random Forest (RF), Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) as machine learning algorithms used to classify the channels and select the optimal channel. The evaluation results of algorithms performance show that Random Forest outperformed the other algorithms. Random Forest achieved high accuracy reached to 97.67%, also achieved high network performance thereby contributed to increase throughput by approximately 92% and reduce the delay by 41%, in addition to reduce packet loss rate by 78% compared with random channel selection method (Rand). The results also showed a direct relationship between improving classification accuracy and improving network performance metrics, confirming the effectiveness of using machine learning techniques in channel selection within dense wireless environments.

This survey presents a systematic and analytical review of deep learning algorithms for multilingual scene text detection and recognition. We critically analyze over 50 publications from 2013 to 2025, categorizing methods into regression-based, segmentation-based, and hybrid detection approaches, alongside CTC-based, attention-based, Transformer-based, and vision-language recognition models. Our analysis reveals that segmentation-based detectors such as DB++ achieve 87.9% F-measure on ICDAR2015, while vision-language models like CLIP4STR reach 94.1% average recognition accuracy. End-to-end methods with detection-recognition synergy, particularly DNTextSpotter, achieve 85.3% on Total-Text. We systematically evaluate 14 benchmark datasets across 10+ scripts, quantify annotation cost trade-offs, and identify critical gaps in low-resource language support. This survey provides comparative analysis through quantitative benchmarking, identifies key architectural trends, and outlines future research directions including unified multi-task frameworks, LLM integration, and efficient edge deployment.

In this paper, we introduce and study subclass ????(????,ɤ,????,????) of the punctured unit for meromorphic, univalent functions disk ????∗ which associated with linear operator. We obtain coefficient estimates, distortion and growth theorems. Also we discuss some results related with ????−neighborhoods and radii of starlikeness and convexity.

The more human development is linked with nature, the greater will be the negative impact of any destruction of the environment caused by us. These technologies are not only allowing us to improve the efficiency of work processes, but also represent an effective means of achieving the UN Sustainable Development goals, like improving the quality of our lives, increasing resistance to climate change, protecting the environment, and reaching global equity, bringing us closer to "environmental intelligence" by 2030. All things considered, it is crucial to carry out the analysis regarding means to utilize AI to promote environmental sustainability. The approach was phased. Beginning with Data collection and statistical description. The documents used in this phase were from the Basra Environmental Directorate, the Iraqi Ministry of Environment, urban planning offices, and other sources. Descriptive statistics showed an improvement in important indicators; PM2.5 decreased from 55.0 to 40.5 µg/m³, green area coverage increased from 12% to 18%, daily waste increased slightly from 2500 to 3000 tons, and the smart projects index increased from 1 to 7. Pearson correlation revealed negative relationships between PM2.5 and green areas and smart projects. This means that more vegetation and AI interventions can reduce air pollution, though correlations are not always clear, and looked at AI uses in waste management, air monitoring, energy efficiency, and smart green projects, using data from 150 participants in three urban types. And the results indicated major growth in AI for waste management. Smart air monitoring coverage grew 305% from 10% to 60%, energy monitoring made steady progress, and smart green projects grew from 2 to 13 per year. Urban Sustainability Index increased from 35 to 83, 137% improvement, but some aspects still need review.The research provides a good basis for further research and modeling. And the need to include smart technology in urban planning in order to achieve quantitative changes in the quality of the environment, even in cities that witness rapid growth and pressure, such as the city of Basra, but the practical application may still face some difficulties in its application.

Digital image forgery has emerged as a significant issue, as numerous tools have been developed to alter images and advancements have been made in creating images with AI. Modern editing techniques, such as copy-and-move, splicing, object removal, and editing with AI help, can create images that look so real that they test traditional forensic methods. Convolutional neural networks (CNNs) are good at capturing local texture artefacts, but they have a hard time modelling long-range contextual connections. Transformer-based models, on the other hand, offer strong global representation learning but often create feature sets that are too complex and duplicate.
This article suggests a way to find fake digital images. It's called Hybrid Dung Beetle–Addax Optimisation with Attention-Based Siamese Network (HDBAO-ASN). The framework uses the Vision Transformer (ViT) and ConvNeXt architectures to get global and local features that work well together. Then, a mix of the Dung Beetle Optimiser (DBO) and the Addax Optimisation Algorithm (AOA) is used to pick out traits that make the data different and cut down on duplicates. After the features are optimised, they are put through an Attention-Based Siamese Network, which makes it easier to learn the similarities between suspicious picture areas and better classifies forgeries.
The experiments were done with the CASIA 2.0, MICC-F220, MICC-F600, and MICC-F2000 files. The average scores for accuracy, precision, recall and F1-score of the proposed method were 97.5%, 97.2%, 97.0% and 97.1%, respectively and AUC was 98.7%. The framework was also demonstrated to be invariant to JPEG compression and geometric transformations and manipulation by AI. The outcomes show that HDBAO-ASN is a good and dependable choice for modern picture forgery detection and multimedia forensic uses.

Off-grid renewable–hydrogen microgrids must jointly size components and dispatch in real time, but most previous works address these problems in isolation, provide no operational safety guarantees, and do not explicitly recover curtailed energy. This paper presents a bi-level co-optimization that couples a two-stage Bayesian (Tree-structured Parzen Estimator) sizing search with a Lyapunov-constrained soft actor-critic (SAC) dispatch policy, along with an execution-time safety shield and a curtailment-recovery reward that prices any recovered surplus as sold green hydrogen, applied to a PV–wind–battery–electrolyzer system. Compared to single-agent SAC, multi-agent SAC, Robust-SAC and a perfect-foresight MILP oracle for five random seeds, the proposed method is the cheapest and safest in all cases, reducing net annual cost by ≈4–5% versus strong reinforcement-learning baselines and constraint violations by ≈6×; the safety shield alone reduces violations by ≈4.9×, and Bayesian sizing reduces levelized hydrogen cost by ≈42% versus MILP sizing. The contributions are the integrated bi-level co-optimization, the shield-augmented Lyapunov dispatch, and the curtailment-recovery objective.