Search Results (30,093)

Computational fluid dynamics methods range from Reynolds-Averaged Navier–Stokes (RANS) simulations to Direct Numerical Simulations (DNSs). RANS offers low computational cost at the expense of accuracy, while DNS provides high accuracy but at a prohibitive cost. The aim of this study is to evaluate whether multi-fidelity neural networks can learn a corrective mapping from RANS to DNS for a small canonical dataset and to determine how training-set composition and model architecture influence generalisation across geometries. In this study, multi-fidelity neural networks for data fusion between low-fidelity RANS and high-fidelity DNS data were applied to turbulent flow (Re = 5600) over parameterised periodic hills, defined by a geometry parameter characterising the steepness ratio. The inputs to the models were the coordinates and the corresponding RANS velocity components, and the outputs were the DNS velocity components, with data from both fidelities mapped onto the same mesh. Both a single-branch and a two-branch architecture were considered. Generalisability was assessed within a small dataset of five periodic hills defined by different values of the geometry parameter $\alpha$ (0.5, 0.8, 1.0, 1.2, 1.5). Both model architectures were trained on data from different combinations of the geometry parameter to evaluate interpolation and extrapolation capabilities. Both networks successfully corrected RANS flow fields for unseen geometries in interpolation regimes. When interpolating, the single-branch architecture achieved more than a 69% reduction in error, while the two-branch architecture achieved more than a 60% reduction, with both improving key flow features such as recirculation zones and jet structures. A key finding is that the single-branch architecture consistently outperformed the two-branch formulation, particularly in low-data regimes. The results show that multi-fidelity neural networks can improve RANS predictions using small datasets and simple inputs, provided that the training set spans the relevant geometric space. As the model does not require the geometry parameter as an explicit input, it is applicable to geometries lacking straightforward parameterisation. The demonstrated advantage of the single-branch architecture highlights the importance of architectural simplicity when training data is limited. Full article

On edge devices, convolutional neural network (CNN) inference is bottlenecked mainly by memory bandwidth, owing to the frequent memory accesses to feature maps and parameters. To address this challenge, we propose a memory-efficient hardware accelerator for depthwise separable convolution that minimizes off-chip memory traffic and parameter storage. The proposed architecture employs three key techniques: (1) a 64-bit run-length coding (RLC) packet compression that exploits feature-map sparsity after ReLU, (2) a mixed-precision scheme that represents feature maps and weights at different precisions, and (3) separated depthwise and pointwise convolution units. In particular, feature maps are transferred in RLC-compressed form, which reduces the amount of data exchanged with the host. The compressed data are decoded row by row, so the on-chip buffers hold only the rows required for computation rather than a complete feature map. In software simulation on ImageNet, the mixed-precision scheme reduced the parameter storage by 49.22% at a cost of 6.24 percentage points (pp) in Top-1 accuracy, and the RLC reduced the data by up to 56.07% in the deeper layers. Implemented on a Xilinx ZCU-104 FPGA, the proposed accelerator performs the depthwise separable convolution with a small number of logic resources and on-chip memory, confirming its feasibility for resource-constrained edge devices. Full article

Cardiovascular Disease (CVD) is a prevailing issue across the world. It requires an accurate method for diagnosing in order to treat the disease effectively. The use of machine learning (ML) techniques has gained popularity for diagnosing CVDs. Some existing ML models have demonstrated good results using the process of manual hyperparameter tuning. But manual hyperparameter tuning is a very tedious task that may cause fluctuations in diagnosis accuracy. To construct a PSO-tuned classifier that provides stable diagnosis for CVD, this paper proposes a robust hybrid framework named PSO-PNN using a particle swarm optimization algorithm for fine-tuning the value of the Smoothing Parameter ($\sigma$) of a probabilistic neural network (PNN). PSO-PNN could be employed as an effective classifier for cardiovascular disease. To validate the effectiveness of PSO-PNN, a dataset obtained from UCI Cleveland database was utilized containing 303 patients’ data. The results indicated that the proposed PSO-PNN framework improved the baseline PNN performance; obtained an accuracy of 91.3%, an ROC-AUC value of 94.8%, and a recall rate of 95.24%; and showed stable predictive performance across 30 independent PSO executions. Additionally, SHAP was applied as a post hoc explainability method to interpret the trained PSO-PNN predictions and identify the contribution of input features. The proposed framework may support future decision-support applications for CVD prediction; however, further external validation is required before practical clinical use can be considered. Full article

With the rapid development of artificial intelligence technology, image design feature enhancement plays a crucial role in computer vision, image processing, and multimedia applications. Traditional feature enhancement methods often have limitations when dealing with complex and changeable image data. Therefore, the study proposes an innovative MN-T (Multimodal Neural Network enhanced by Transformer) strategy to overcome these challenges. The MN-T strategy combines the attention mechanism of Transformer with the cross-modal learning capabilities of multimodal neural networks to more accurately capture and enhance key features in images. The Transformer attention mechanism enables MN-T to efficiently process global and local information in images, while the cross-modal learning capability of multimodal neural networks further enhances its ability to understand and express image features. The results show that MN-T strategy has excellent performance in processing different image data. Compared with the existing image design feature enhancement methods, MN-T strategy has achieved significant improvement in the root mean square error (RMSE), mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and other key performance indicators. The research results of this paper not only provide a new idea and method for image design feature enhancement methods, but also provide a new theoretical support and experimental basis for the research of related fields. Full article

Design of experiments (DOE) offers a powerful, systematic framework for optimizing nanoparticle (NP) formulations by replacing inefficient one-factor-at-a-time (OFAT) methods. By enabling the simultaneous evaluation of multiple variables, DOE uncovers critical factor interactions and identifies true global optima—critical for quality-by-design approaches. Despite its potential for systematic innovation, DOE remains underutilized in nanomedicine due to its perceived complexity; this review provides a practical roadmap to bridge the gap between statistical theory and robust NP optimization. It provides a practical overview of DOE concepts, including factor selection, design choice, graphical interpretation of results (perturbation/contour plots), model validation (regression analysis and ANOVA), and numerical optimization via desirability function (D). Common pitfalls and best-practice strategies are discussed to support reliable model building and decision-making. A practical case study on poly(lactic-co-glycolic acid) (PLGA) NPs illustrates a multistage workflow: utilizing Taguchi screening to isolate key factors, followed by central composite design (CCD), for precise surface mapping. Numerical optimization using Design-Expert^® software maximized EE% (highest importance) within size/zeta ranges, yielding optimal conditions (5 mg drug amount, 4 mL aqueous volume; D = 0.961). Confirmation runs (EE 41.2%, NP size 124 nm, zeta potential −15 mV) validated predictions (EE 47.6%, NP size 133 nm, zeta potential −17.2 mV), confirming model reliability. Ultimately, by bridging conceptual foundations with practical implementation, this review aims to encourage broader adoption of DOE, particularly among emerging formulation scientists, and serves as a roadmap to accelerate scalable NP development, fostering data-driven innovation and improving efficiency in nanomedicine research. Moreover, future integration of artificial intelligence (AI) and artificial neural networks (ANNs) with DOE will drive a predictive, data-driven approach to NP optimization—accelerating robust, scalable, and regulatory-ready nanomedicine development with fewer experiments. Full article

Road traffic surveys require sensing systems that can be deployed rapidly without modifying the road surface or requiring a permanent power connection. This paper presents the design, embedded implementation, and evaluation of a low-power roadside magnetic sensor that performs vehicle-event detection and classification directly at the edge. The sensing node integrates two RM3100 three-axis magnetometers (PNI Sensor, Santa Rosa, CA, USA) with an NXP MK22FN512VLH12 microcontroller (NXP Semiconductors N.V., Eindhoven, The Netherlands) based on a 120 MHz Arm Cortex-M4F core with 512 kB Flash and 128 kB SRAM. Magnetic-field data are acquired at 250 Hz and processed locally using baseline removal, low-pass filtering, signal-energy calculation, and peak-based event detection. Detected magnetic signatures are classified using an integer-quantised one-dimensional convolutional neural network implemented directly on the microcontroller. The model processes four synchronised 512-sample channels representing the three magnetic-field axes and their combined signal energy. Model development was supported by approximately 50,000 annotated events obtained from 36 h of real-world traffic measurements at eight locations. The selected model achieved an overall classification accuracy of 91.1% for the considered operational categories. The implemented network requires 288,128 multiply–accumulate operations per inference, while its quantised weights and biases occupy approximately 23 kB of Flash memory. Complete three-axis event signatures are stored locally for subsequent verification, whereas only the timestamp and predicted vehicle category are transmitted through the wireless interface. Based on the capacity of the applied LiFePO₄ battery and the estimated consumption of the implemented hardware, the expected autonomous operating period is approximately 41 days. The results demonstrate the feasibility of integrating magnetic sensing, embedded signal processing, and Edge AI on a conventional resource-constrained Cortex-M4 platform for non-invasive road traffic monitoring. Full article

This study investigates the compressive performance of fiber-reinforced polymer (FRP) rebar-reinforced Seawater and Sea Sand Concrete (SSC) columns through an integrated approach combining finite element analysis, theoretical derivation, and machine learning. Finite element models were developed to quantify the influence of key parameters on the ultimate bearing capacity and lateral deflection. The results indicate that the compressive capacity decreases significantly with increasing eccentricity and slenderness ratio. Columns reinforced with steel rebars demonstrated superior load-bearing and anti-lateral displacement capabilities compared to their FRP-reinforced counterparts. A theoretical formula for predicting the compressive capacity was derived; however, it systematically overpredicted the experimental measurements by approximately 36%. To develop data-driven predictive models for the ultimate load capacity of FRP–SSC columns, four machine learning models, backpropagation neural network (BPNN), bootstrap aggregating BPNN (Bagging-BP), genetic algorithm-optimized BPNN (GA-BP), and gradient boosting regression trees (GBRT), were employed. Using sectional dimension, concrete strength, reinforcement parameters, eccentricity, and slenderness ratio as inputs, the validation sets of the models achieved R-values of 0.942, 0.918, 0.933, and 0.990, respectively. Feature importance analysis based on SHAP identified eccentricity as the most influential parameter. Results from this work can help to understand the behavior of FRP–SSC columns under compression. Full article

Gears serve as core transmission components, and their wear evolution directly affects equipment stability and service life under long-duration complex loading. Especially under complex loading and long-term service conditions, the tooth surface topography undergoes continuous evolution. However, traditional wear prediction methods based on physical models or empirical formulas have significant limitations in addressing nonlinear problems involving multiple coupled variables. This study proposes a deep neural network (DNN)-based method for gear wear prediction. Geometric parameters, loading conditions, and surface topography characteristics are integrated as model inputs to enable point-by-point prediction of tooth-profile wear. Experimental results demonstrate that the proposed model achieves excellent predictive performance in the mild-wear regime, with a mean absolute error (MAE) below 2.5 × 10⁻⁴ mm, a root mean square error (RMSE) below 5.0 × 10⁻⁴ mm, and R² values ranging from 0.92 to 0.99. The model also achieves satisfactory prediction accuracy at previously unseen measurement positions and for previously unseen superfinished gear samples. The proposed DNN effectively learns implicit wear-evolution patterns from experimental data and exhibits strong generalization capability, providing a practical approach for gear health monitoring and predictive maintenance. Full article

Predicting customer churn in the telecom sector is critical for retaining subscribers, maintaining brand reputation, and staying ahead of competitors. Losing customers not only reduces revenue but can also weaken long-term market position in a highly competitive industry. While machine learning has been widely used to address this challenge, most traditional approaches depend on centralizing customer data. This raises major concerns about user privacy, data ownership, and compliance with strict regulations such as GDPR. These challenges make it difficult for businesses to fully utilize customer data while safeguarding sensitive information. In this paper, we investigate a privacy-preserving approach to churn prediction that combines federated learning (FL) with differential privacy (DP). Rather than collecting all customer data in a single repository, the investigated framework enables multiple clients to collaboratively train a deep neural network while maintaining data locality during the federated training process. To further enhance privacy protection, we employ Differentially Private Stochastic Gradient Descent (DP-SGD) and add controlled noise to model updates, reducing the possibility of inferring individual data contributions. This work systematically evaluates how different privacy levels, expressed through ε and δ, influence model performance under simulated non-IID client distributions. The experiments analyze the privacy–utility trade-off using multiple evaluation metrics and compare the results with centralized and non-private federated-learning approaches. The findings show that the investigated framework maintains competitive predictive performance across a range of privacy budgets while demonstrating a clear privacy–utility trade-off. Very strict privacy budgets result in substantial performance degradation, particularly for smaller and more imbalanced datasets, whereas moderate privacy budgets maintain competitive predictive performance with limited degradation. This study highlights the potential of privacy-preserving federated learning for practical distributed analytics applications where protecting sensitive data is essential. Full article

Automated pulmonary auscultation demands rapid and reliable anomaly detection. Contemporary deep learning frameworks frequently face deployment barriers due to their reliance on memory-intensive Transformer mechanisms and high-cost processing hardware. Addressing this limitation, the present study introduces the Neuro Scatter Mamba Kolmogorov–Arnold Neural Network (NSMK-Net), a lightweight, Transformer-free architecture. The model integrates 1D Wavelet Scattering, bi-directional Selective State Space Models (Mamba), and Kolmogorov–Arnold Networks (KAN). By substituting quadratic self-attention with continuous-time differential discretization, the framework achieves very good computational efficiency under severe hardware constraints. Model optimization followed an eco-friendly “Green-AI” methodology, successfully converging on a standard 4 GB VRAM graphics unit. Regarding real-world deployment, the finalized architecture can be considered as a possible candidate for future “Edge-AI” applications, because it requires only 0.34 MB of parameter storage (89,342 parameters) and executes inference in approximately 48 milliseconds per respiratory cycle. Evaluated on the SPRSound dataset, the proposed model achieved a cycle-level accuracy of 84.67% (Macro-F1: 0.48). When tested under the strict official 60/40 partition of the ICBHI 2017 dataset, the network delivered a global accuracy of 41.56% (Macro-F1: 0.31) alongside an official reported ICBHI Score of 49.38%. These metrics indicate a stable detection capability when processing highly imbalanced clinical data. By replacing fixed activation nodes with learnable edge non-linearities and utilizing linear sequence memory, this new structural approach reduces the dependency on high-end hardware for medical acoustic processing. Full article

Quality control in green coffee bean production is critical for food safety and economic sustainability. A persistent gap in existing automated inspection systems is their inability to capture both sides of each bean, which leads to systematic under-detection of surface defects. This work presents three contributions: (i) a novel mechatronic apparatus that mechanically guarantees dual-sided imaging of every bean, (ii) a public 12-class dataset of green coffee bean defects, and (iii) an embedded, real-time inspection pipeline validated on low-cost hardware. The apparatus sequentially presents each bean, from a standard 350 g sample, to two 16-megapixel cameras under controlled LED illumination. A dataset of 9600 images spanning 12 classes (11 defects and 1 normal) was generated from expert-classified samples and enriched through data augmentation. Four convolutional neural network (CNN) architectures, VGG-16, VGG-19, ResNet-50 and YOLOv8, were trained and benchmarked using precision, recall, F1-score and mean average precision. YOLOv8 achieved the best overall performance, with a precision of 97.4%, a recall of 99.6%, an F1-score of 0.930 and a mean average precision of 96.5%, outperforming VGG-16 (accuracy 86.07%), VGG-19 (accuracy 67.03%) and ResNet-50 (accuracy 87.76%). Dual-sided acquisition raised mean per-class detection accuracy from 0.727 to 0.908, a relative gain of 25.7% over an equivalent single-sided configuration. Deployed in real-time “track” mode on a Raspberry Pi 4, the system simultaneously classifies defects and counts beans by category, processing a 350 g sample in approximately 38 min. Combining mechanical innovation with lightweight deep learning enables practical, scalable, and cost-effective quality control for laboratories specialized in coffee analysis. Full article

Graph neural networks remain challenging to apply to traffic forecasting, often failing to consistently outperform simple baselines. We observe that the strongest baseline is regime-dependent: last-value persistence dominates at short horizons on smooth freeway data, while time-of-day climatology prevails at longer horizons and on bursty arterial networks. Rather than treating this as a limitation, we propose a learnable approach that automatically selects the optimal baseline. Our method anchors predictions to a learned, per-horizon convex blend of persistence and climatology, introducing only twelve scalar parameters. This learned anchor recovers whichever baseline is locally most effective, allowing the model to focus on capturing residual variations that neither baseline captures. We evaluate on six public benchmarks (METR-LA, PEMS-BAY, PEMS03/04/07/08) spanning traffic speed and flow data under standard 70/10/20 chronological splits with masked evaluation metrics and holiday-aware climatology. Our anchored temporal models consistently beat both baseline methods on the 12-step average across all datasets, and outperform at every horizon on five of the six benchmarks. When integrated into two strong architectures (STID and Graph WaveNet), the anchor yields substantial gains at long horizons where climatology is most informative. Notably, within our lightweight framework, learned spatial graph components do not improve accuracy and can slightly degrade performance, a finding we analyze and discuss. Full article

The rapid evolution of AI-driven language technologies has inadvertently widened the gap between high-resource and marginalised languages. Despite significant progress in AI-driven translation for high-resource languages, low-resource languages remain underrepresented due to limited data, a lack of benchmarks, and evaluation challenges. This study presents a comprehensive systematic review of machine translation for low-resource languages, focusing on advances in neural machine translation (NMT) and large language models (LLMs) between 2017 and 2025. Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, 63 studies were selected from the 1696 articles in the Scopus, Web of Science, and Google Scholar databases. The review identifies five dominant methodological approaches: data augmentation, back-translation, transfer learning, pre-training, and parameter-efficient fine-tuning. The findings reveal that model performance is highly dependent on resource availability: transformer-based NMT excels in moderate data settings, while LLMs demonstrate promising zero-shot and few-shot capabilities in extremely low-resource scenarios. Hybrid NMT–LLM approaches emerge as a particularly effective paradigm. The study also highlights critical challenges, including the absence of standardised benchmarks, over-reliance on inadequate evaluation metrics such as Bilingual Evaluation Understudy (BLEU), limited human evaluation, and significant geographic and linguistic underrepresentation. Additionally, ethical concerns related to bias, cultural representation, and community engagement are increasingly relevant. The findings contribute to advancing inclusive and equitable AI-driven language technologies. Full article

Dichromatic optical interferometry provides rich optical information for lubricant film-thickness measurement. However, experimental data are typically limited to a small number of discrete operating conditions, making it difficult to learn the nonlinear relationship among entrainment speed, dichromatic interference images, and lubricant film thickness. To address this limitation, a Dual-Wavelength Physics-Guided Graph Neural Network (DW-PG-GNN) is proposed for intermediate lubrication-state characterization from sparse experimental observations. Rather than directly regressing film thickness from interference images, the proposed framework learns the nonlinear relationship among entrainment speed, dichromatic interference images, and lubricant film thickness through a unified physics-guided learning framework. Specifically, graph representations capture the structural continuity and cross-wavelength coupling characteristics of dichromatic interference fringes, while a physics-guided residual learning strategy embeds analytical film-thickness priors from classical elastohydrodynamic lubrication (EHL) theory into an implicit neural representation for intermediate state prediction. A differentiable interference renderer further constrains the consistency between predicted film thickness and reconstructed dichromatic interference images, ensuring optical and physical consistency throughout the learning process. Validation based on sparse experimental datasets acquired from a ball-on-disc EHL rig demonstrates reliable average film-thickness prediction and interference-image reconstruction under the investigated operating conditions. The prediction errors remain within 2.2% for trained conditions and 8.67% for unseen entrainment-speed interpolation conditions. Full article

Estimating conditional average treatment effects (CATEs) in high-dimensional causal inference problems remains challenging because complex nonlinear relationships, heterogeneous feature distributions, and dependence among covariates can limit the effectiveness of conventional machine learning approaches. To address this challenge, we propose a copula-enhanced neural learning framework that integrates empirical copula transformations, manifold-based feature augmentation, structured treatment–covariate interaction representations, and deep neural networks for flexible CATE estimation. The empirical copula transformation does not introduce additional dependence information; instead, it provides a rank-based feature representation that normalizes marginal distributions, reduces sensitivity to heterogeneous feature scales and extreme observations, and offers a dependence-aware representation for subsequent learning. The proposed framework is evaluated through Monte Carlo simulations under diverse data-generating mechanisms and a real-world application using the Criteo uplift dataset. The simulation study examines the contribution of individual model components through ablation experiments and compares the proposed approach with established causal learning methods. Results demonstrate that the proposed framework achieves competitive CATE estimation accuracy while providing stable policy evaluation based on Inverse Propensity Scoring (IPS) and Doubly Robust (DR) estimators. In the Criteo application, the proposed method exhibits predictive performance comparable to conventional neural-network approaches while producing more stable Doubly Robust policy value estimates. These findings suggest that copula-based feature representations combined with deep learning provide a flexible approach for heterogeneous treatment effect estimation, particularly in high-dimensional settings with complex covariate dependence. The benefits of the proposed framework depend on data characteristics, including sample size, dimension, dependence structure, and treatment assignment mechanisms. Full article