DIYALA JOURNAL OF ENGINEERING SCIENCES

The escalating dependence on artificial intelligence (AI) within the healthcare sector presents significant challenges pertaining to data privacy, regulatory adherence, and the pragmatic implementation of predictive models. This review meticulously examines the amalgamation of Federated Learning (FL) and Differential Privacy (DP) as a prospective framework to mitigate these issues within decentralized healthcare infrastructures. We conduct a comprehensive analysis of extant FL-DP frameworks, concentrating on their capacity to safeguard privacy, uphold performance standards, and function efficiently in real-time clinical settings. The review encompasses architectural advancements, edge computing methodologies, adaptive privacy budgets, and the contributions of blockchain and the Internet of Medical Things (IoMT) in facilitating secure data interchange. Comparative assessments and case studies are synthesized to evaluate model precision, scalability, and conformity with regulatory mandates. Notwithstanding significant advancements, we delineate critical deficiencies, including ethical dilemmas, algorithmic equity, data disparity, and obstacles to deployment. Our contributions consist of a benchmarking framework, the delineation of unresolved research inquiries, and actionable insights for the formulation of secure, just, and scalable FL-DP systems within the healthcare domain. This paper delineates a strategic framework for prospective research and the execution of privacy-preserving AI within clinical practice. The outcomes highlight significant potential for real-world clinical implementation, fostering enhanced patient care, supporting regulatory compliance, and enabling scalable, privacy-preserving AI adoption across diverse healthcare environments.

The continuous scaling of MOSFETs to meet the demands of high-performance devices has intensified challenges such as short-channel effects (SCEs), which degrade device reliability and efficiency. To address these issues, this study investigates the influence of SiO₂ buried layer thickness and vertical positioning on the electrical performance of 20 nm n-MOSFETs incorporating a high-k dielectric gate stack. Using TCAD Silvaco ATLAS, the research systematically evaluates how buried oxide thickness (10 nm to 50 nm) and depth (30 nm to direct contact with the channel) affect key device parameters, including threshold voltage (Vth), drain-induced barrier lowering (DIBL), breakdown voltage, ON current (Ion), leakage current (Ioff), and the Ion/Ioff ratio. Results show that increasing the buried layer thickness increases Ioff from 8.51 × 10⁻¹¹ A/m (without buried layer) to 3.09 × 10-10 A/m at 50 nm, while Ion slightly increases to 2.3 × 10⁻³ A/m. Although this reduces the Ion/Ioff ratio from 2.37 × 10⁷ to 7.54 × 10⁶, it also significantly improves breakdown voltage from 85.09 V to 167.4 V at 10 nm thickness. Notably, the breakdown voltage decreases to 76.26 V at 50 nm. In terms of vertical positioning, placing a 10 nm SiO₂ buried layer in direct contact with the channel yields the highest breakdown voltage of 1186.7 V and the lowest Ioff of 5.15 × 10⁻¹¹ A/m, with an Ion/Ioff ratio of 3.62 × 10⁷. These results highlight that both the thickness and position of the buried oxide layer play a crucial role in suppressing SCEs and enhancing breakdown robustness.

Many The treatment of soft clay soil at different temperatures substantially influences its settlement properties. Increased temperatures can accelerate the consolidation of soft clay, resulting in alterations in settlement behavior. Estimating the settlement of soft clay soil beneath building foundations is a fundamental aspect of geotechnical engineering design. Traditionally, engineers rely on empirical formulas and classical consolidation theories, such as Terzaghi’s one-dimensional consolidation model, to predict settlement behavior. While these conventional methods provide useful approximations, they often struggle to capture the highly nonlinear and site-specific characteristics of soft clay soils. Recently, artificial intelligence (AI) models, particularly Deep Neural Networks (DNN) and other machine learning algorithms, have emerged as powerful tools for forecasting settlement more accurately. This study applied one of intelligent technique by using deep learning artificial Settlement model (SDNN) model to forecast underground settlement of soft clay soil treated with temperature. DNN model adopted 365 data set, input data parameters such as time period, various of distance between the heaters, temperature and load applied. The outcomes indicate that the SDNN model effectively forecasted ground settlement, showing similar results between actual and predicted values. The performance of the SDNN network model was notably impressive, achieving a mean absolute error (MAE) of 4.3356 % and a mean squared error (MSE) of 0.4494%. The artificial model also demonstrated strong efficiency and a favorable variance calculation coefficient.

Wire EDM plays a vital role in the precision machining of hard-to-cut materials, but its efficiency depends on the optimal selection of parameters. The influence of machining parameters on WEDM quality for Stainless Steel 202. This study integrates Taguchi’s L9 orthogonal design with machine learning (ML) to optimise and predict surface roughness (SR) outcomes. ANOVA revealed peak current as having a significant impact on machining quality, with a moderate non-significant effect from pulse on time; wire speed and pulse off time had minimal effect. Increased peak current and pulse on time result in higher discharge energy, which generates deeper craters on the workpiece surface, thereby leading to increased surface roughness. To boost predictive accuracy, three ML models—Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Machine (SVM)— were evaluated by using k-fold cross-validation in addition to the conventional 80/20 train-test split. RF achieved the highest prediction accuracy (R² = 0.931), followed by ANN (R² = 0.918) and SVM (R² = 0.810). This approach minimises experimental efforts and enhances machining precision. The findings suggest that combining statistical tools with ML can streamline WEDM processes, improve surface quality, and reduce defects. Future work may focus on real-time control systems, hybrid optimisation, and deep learning models for further improvement

The primary factor influencing road performance is pavement deterioration. Pavement cracking, a prevalent form of road deterioration, is a significant challenge in road maintenance. This paper proposes a method utilizing deep convolutional neural network models for precise crack detection, segmentation, and geometric parameter calculation in pavement crack identification. The system operates through three primary stages: Commencement, crack identification employs YOLOv10, a rapid and efficient object detection model. Secondly, crack segmentation employs a modified Unet 3+ variant known as Residual-Attention UNet 3+, which effectively distinguishes crack pixels from the background by utilizing attention mechanisms and residual connections to enhance accuracy. Finally, crack quantification, wherein the system computes the crack's geometric parameters, including width, length, angle, and orientation. We assessed performance using two datasets: SUT-Crack, a publicly accessible dataset, and IRD-Crack, a new real-world dataset compiled by the authors from roads in Diyala, Iraq, with diverse lighting conditions and surface complexities. The suggested technique attained an accuracy of 98.96% on the SUT-Crack dataset. It showed superior performance on the IRD-Crack dataset under actual situations, therefore validating its efficacy and generalization capability. This method offers a pragmatic and computationally efficient instrument for monitoring pavement cracks and can facilitate road repair choices.

This study aimed to evaluate the efficiency of biochar (BC) produced from Chara algae via semi-anaerobic pyrolysis as an adsorbent for removing lead (Pb²⁺) and copper (Cu²⁺) ions from aqueous solutions in a batch system. The preparatory process involved washing the algal biomass with tap water and deionized water, followed by a two-stage drying process: solar drying (48 h) and then thermal drying (75°C, 24 h). Dried sample was then ground into fine particles (250–400 μm) and subjected to a semi-anaerobic heat treatment at temperatures between 400 and 600°C for 50 min to adjust the physicochemical properties of the biochar. The adsorption performance was studied under a range of operating conditions, including a range of pH (2–6), contact time (10–150 min), initial heavy metal ion concentrations (20–100 mg/L), temperatures ranging from 25 to 80°C, and various adsorbent concentrations (0.1–3 g/L). The adsorbent was characterized using advanced analytical techniques. Fourier transform infrared spectroscopy (FTIR) was used to identify the functional groups active in adsorption, scanning electron microscopy (SEM) to analyze the surface structure, and energy-dispersive X-ray spectroscopy (EDS) to determine the elemental composition and confirm metal adsorption. Under best conditions (pH 6, dosage 1.0 g/L), the maximum adsorption capacities recorded were 322 mg/g for Pb²⁺ and 196 mg/g for Cu²⁺. The dynamic adsorption studies further confirmed the removing heavy metals from aqueous media, indicating its promise for scalable water treatment. applications.

This study presents a hybrid and interpretable modeling framework that integrates the Mori–Tanaka micromechanical model with artificial neural networks (ANNs) to predict the elastic modulus of bamboo-reinforced polypropylene composites. A synthetic dataset was created encompassing bamboo fiber volume fractions from 5% to 25%, enabling the ANN to generalize proficiently across diverse reinforcement setups. The ideal network architecture (2–15–1) attained superior predictive performance, with mean squared errors under 20 and regression coefficients surpassing 0.98, so validating the model's accuracy and robustness. To guarantee reliability, the model was evaluated on intermediate components not encountered during training, exhibiting consistent performance and resilience to overfitting. The interpretability of the black-box AI model was improved via sensitivity analysis and SHAP (Shapley Additive Explanations), which revealed that bamboo modulus was the primary factor affecting composite stiffness, contributing around 72% of predictive influence, whereas polypropylene accounted for 28%. These findings correspond with micromechanical theory and offer insights into material design methodologies. Combining physics-based modeling with artificial intelligence improves the accuracy of predictions and helps engineers make smart choices during the early stages of bio-composite development. This research enhances sustainable material innovation by offering a transparent, efficient, and scalable modeling tool suitable for comprehensive mechanical property forecasts and practical composite design.

Plant diseases can cause severe damage to crops and lead to food shortages and financial losses for both farmers and the agricultural sector. Detecting these diseases early is essential for protecting crops, increasing agricultural productivity, and ensuring food security. This paper introduces a new intelligent edge computing framework that provides a cost-effective, portable, and energy-efficient solution for deep learning-based automated plant disease detection. Unlike cloud-dependent systems, the proposed framework operates independently of an internet connection, making it ideal for real-time field deployment. It employs the NVIDIA Jetson Nano as an edge computing device and incorporates an Android-based interface for user interaction. The system utilizes a convolutional neural network (CNN) for feature extraction, followed by a deep classification network to identify plant diseases. Plant images are captured by a smartphone and transmitted to the Jetson Nano over a local WiFi network using the KDE Connect application for processing. After classification, the image with the predicted disease category is sent back to the smartphone using FTP for user display. The framework was trained and evaluated on the PlantVillage dataset involving 38 disease categories for 14 different types of healthy and diseased crop leaves, achieving a maximum accuracy of 99.1%. This efficient and practical system demonstrates the potential of edge AI in precision agriculture by enabling on-device disease diagnosis without relying on cloud computing infrastructure.

This research aims to evaluate various coatings on titanium alloy substrates to increase corrosion resistance and surface quality. This involves coating the substrates with layers of Hydroxyapatite (HAP), Chitosan (CS), and Magnesium oxide (MgO) using the EPD technique. The cell voltage value was set to 5 V, while the coating time was set to 5, 10, and 15 min. SEM was used to measure the thickness of the cross-sectional coating layers, and it was also used to show the morphology of the composite coating. From the results 10-minute deposition time has been established as the most optimal to produce an even, thick, uniform, and dense coating with good mechanical properties. The thickness more than doubles with increases in time with the coat almost tripling at 15 minutes and 5 minutes respectively. EDX analysis showed that the elemental compositions of calcium, oxygen, and phosphorus increased after 15 minutes, indicating a dense coating. The research findings indicate that the adhesion force transitions to a marginally positive value of 0.05 nN, suggesting a weak attachment rather than a robust bond. Additionally, Young's modulus shows a significant decline to 3.93 MPa, which signifies a reduction in the stiffness of the material. The surface stiffness decreases to 0.19 nN/nm between 10 and 15 minutes, indicating a loss of rigidity. Although longer coating times may increase thickness, they can also reduce functionality and raise defect risks. This reinforces that electrophoretic deposition (EPD) effectively improves the alloy's surface quality for biomedical applications.

The detection of a person's membership in a family is crucial in many cases, and the need for it becomes urgent during and after wars and natural disasters. While DNA testing and facial images are common methods for kinship detection, they have limitations in certain scenarios. This paper presents an innovative approach to intrafamily kinship classification using palm skin texture images, a modality not previously explored for this purpose. Unlike conventional palm biometric verification aimed at individual identification, this research investigates inherited skin texture patterns to establish familial relationships. A hand palm dataset was created from the MKH (Mosul Kinship Hand) images dataset. It contains 332 segmented and annotated skin images from 84 individuals across 15 families, with multiple samples per person. Two deep convolutional neural network pre-trained models (VGG16 and EfficientNetB0) were used independently for feature extraction via transfer learning. Based on these features, neural network classifiers were designed to detect the membership of an individual to one of the given families. The results were evaluated using the appropriate metrics and gave a test accuracy of 96% by using EfficientNetB0 and 79% from VGG16. The findings of this study suggest that hand palm images may offer a promising and practical preliminary approach to kinship detection. The dataset utilized in this study is accessible upon request from the author to support reproducibility and encourage further research.

Permanent magnet synchronous motors are widely used in industrial laboratories due to their high starting torque and efficiency compared to static induction machines. The growing interest can be explained by the development of permanent magnet magnetic materials. To evaluate the transient behavior of permanent magnet synchronous motors, a MATLAB model is considered based on the Parke transform to simplify the control strategy that implements voltage-to-frequency (V/f) ratio to maintain more stability during unsteady conditions and facilitate the analysis to track the machine action. Managing this transient phase is essential to prevent excessive inrush currents and ensure stability. This study rigorously examines the behavior of machines during two transition periods until they reach a steady state speed. The rise time (Tr) is the time it takes for a signal to change from (10 - 90) % of its final value. The transient speed duration of the first approach varies between 0.4 ms for the unloaded machine and performs longer duration time 0.7 ms under load conditions. The unloaded machine obviously responds faster than the loaded machine .The second, controlled approach has a limited duration time periods from 10.25 to 10.8 ms, with a rise time of Tr1 = 0.2ms and Tr2 = 2,5ms for both loads applied sequentially. The controlled machine demonstrates a shorter rise time which can be attributed to the same speed root forced to different load conditions. This approach highlighted the benefits of the inherent dc boost voltage to minimize transient effects in synchronous machine.

The increasing reliance on electronic voting systems introduces challenges in securing voter data and ensuring privacy. Traditional e-voting systems are prone to cyber threats, compromising the integrity of elections, including data leakage, tampering, and spoofing attacks, as well as the problem of data centralization that increases the risk of hacking and loss of electoral integrity. To address this, we propose a secure e-voting system that uses federated learning to enhance security and privacy. The system employs multiple Convolutional Neural Networks (CNNs) including VGG16, VGG19, ResNet18, MobileNetV2, a custom CNN, and a ResNet-VGG hybrid across six distributed sites. In order to verify that only eligible and registered voters may cast ballots, the proposed system uses a three-step biometric process that includes user ID identification, fingerprint matching for high security using unique biometric data to prevent identity fraud, and gender identification as an extra layer of verification to lower the risk of identity theft. Furthermore, the system trains models locally and only transfer trained weights to avoid centralized sensitive data. When compared to the Sokoto-Coventry fingerprint dataset, it achieved high rates of 99.75% for identity recognition, 99.90% for fingerprint recognition, and 99.97% for gender recognition. These results highlight the effectiveness of the proposed system in providing a secure, privacy-preserving, and scalable solution for e-voting.

Solar air heaters (SAHs) offer a practical and environmentally friendly method of harnessing solar energy. Enhancing the thermal performance of SAHs is critical, and various techniques, including surface modifications, have been explored. This experimental study investigates the influence of dimpled rib roughness on the thermal and hydraulic performance of a SAH with an equilateral triangular duct. The experiments were conducted over a range of operating conditions: relative roughness height (e/D) from 0.023 to 0.058, relative roughness pitch (L/e) from 3 to 9, relative roughness width (S/e) from 6 to 18, and Reynolds numbers from 6000 to 18000. The results indicate that the dimpled rib configuration significantly enhances heat transfer compared to smooth SAH. The maximum enhancement in the Nusselt number (Nu) and friction factor (f) is 5.78 and 2.26 times higher, respectively, than the smooth SAH. The highest performance enhancement criteria (PEC) value of 4.64 is achieved at e/D=0.058, L/e=3, and S/e=6. Additionally, empirical correlations for Nu and f were derived based on experimental data, providing valuable tools for SAH design optimization. This study demonstrates the potential of dimpled ribs to significantly improve SAH performance, making it a viable option for efficient solar energy utilization.

Deep beams transfer loads primarily through short, steep compression struts rather than classic flexural action; their response is highly sensitive to how forces enter the member and to the support layout. In tanks, transfer girders, and podium edges, midspan three-point loading is common, and in practice those loads are rarely equal. This study interrogates that everyday imbalance. Even when applied at midspan, unequal loading is a major driver of deep-beam response and capacity. By contrast, with a central, equal load, the compression-strut paths that carry force from the supports to the loading node(s) are typically symmetric. However, in the present study, the load inequality causes these struts to be asymmetrical. As a result, the strut carrying the larger load failed before the one carrying the smaller load. Therefore, the deep beam fails early. Twenty-one deep beam specimens were analyzed using SAP 2000 software, which is based on the well-known finite element method. Three patterns of load distribution between three concentrated load points were adopted: 33%-33%-33%, 50%-25%-25%, 25%-50%-25%, 67%-16.5%-16.5%, 16.5%-67%-16.5%, 75%-12.5%-12.5% and 12.5%-75%-12.5%. These load cases were studied using different concrete's compressive strength values of 20, 30, and 40 MPa. Based on these results, load capacity remained essentially unchanged, while midspan deflection and shear stresses decreased by 4.0–4.6% and 3.8–17%, respectively; in contrast, the maximum positive moments increased by 0.55–7%.

The current study highlights the importance of accurate temperature prediction in Iraq, a country facing economic challenges due to its hot, arid climate and increasing climate change effects. Conventional forecasting methods, such as statistical and shallow machine learning models, struggle to address the complex time-dependent characteristics of meteorological data. The present study proposes to improve the temperature forecasting of the three large cities in Iraq, i.e., Dohuk, Erbil, and Mosul, using the deep learning models that can learn both short- and seasonal weather trends. A meteorological dataset of 24 years (2000-2024) was created with five major characteristics, namely, temperature, wind speed, relative humidity, total precipitation, and surface pressure. The models to be used in the deep learning model were three, namely (Long Short-term memory (LSTM), Gated Recurrent Unit (GRU), and Artificial Neural Network (ANN). The metrics of performance were Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Squared Error (MSE), and R². The LSTM model performed the best in all the cities with RMSE values of 2.544, 2.366, and 2.323 and R² scores of 0.941, 0.948, and 0.952 in Dohuk, Erbil, and Mosul, respectively. The study confirms that LSTM is the most effective in modeling complex temporal dependencies in climatic time series, making it a significant contribution to understanding deep learning's application in weather forecasting in the Middle East. It suggests integrating AI-driven technology into the national meteorological system for climate-resistant decision-making in agricultural, water resource management, and urban development sectors.