Search Results (2,205)

Deep learning has improved precipitation estimation and forecasting, but high predictive skill does not by itself ensure physically credible behavior, robustness under distribution shift, or reliable uncertainty. This critical narrative review examines physics-informed deep learning across quantitative precipitation estimation, radar and satellite nowcasting, short-range forecasting, numerical weather prediction post-processing, and selected Earth-system applications. We use a two-dimensional analytical framework—form of physical knowledge and point of model integration—to compare mechanisms that are often grouped under the same label but provide different levels of physical guarantee. Across the literature, useful physical information is strongly task- and scale-dependent. Physically meaningful inputs and transport-aware structures are most convincing when they represent processes that are both observable and dominant over the forecast horizon. Soft equation or consistency losses can reduce violations, but their effect depends on constraint validity, weighting, data quality, and precipitation regime. Hard parameterizations and output projections provide stronger guarantees for specified relations, although this does not necessarily translate into better precipitation forecasts when those relations are incomplete or scale-mismatched. Evidence for cross-region robustness, extreme-event reliability, and operational maturity remains comparatively limited. A key evidence gap is whether these benefits persist under matched ablations and transfer across regions, sensors, and precipitation regimes. Full article

Bayesian wavelet regression offers a mathematically grounded framework for modeling neuro-retinal signals that exhibit substantial noise, high dimensionality, and complex variations across temporal and spatial scales. Such characteristics are inherent to diverse neuroscience and vision science modalities, including electrophysiological recordings, neuroimaging, and retinal imaging, where accurate interpretation depends not only on effective signal representation but also on reliable uncertainty estimation and biologically meaningful feature identification. By combining wavelet-based multiresolution analysis with Bayesian probabilistic inference, these approaches provide adaptive mechanisms for separating informative structures from noise while maintaining clinically relevant signal characteristics. This review examines the mathematical principles and recent methodological advances in Bayesian wavelet regression, focusing on sparse signal representation, Bayesian regularization, and uncertainty-aware estimation. Particular emphasis is placed on sparsity-inducing prior formulations, including spike-and-slab, moment, and inverse moment priors, which enhance coefficient selection, improve parameter identifiability, and facilitate the recovery of informative multiscale patterns. In addition, Bayesian inference frameworks, including Markov Chain Monte Carlo, Hamiltonian Monte Carlo, No-U-Turn Sampling, and Variational Bayes, are analyzed with respect to posterior estimation accuracy, computational requirements, and practical applicability. The literature synthesis indicates that Bayesian wavelet approaches support multiscale representation, sparse coefficient selection, and probabilistic uncertainty characterization, with prior and inference choices involving trade-offs in sparsity, interpretability, estimation, and computational efficiency. These applications span neural signal analysis, neuroimaging, retinal imaging, and neuro-ophthalmology, while multimodal retina–brain integration and clinically validated uncertainty-aware modeling remain less developed. Extending beyond single-modality analysis, this review investigates hierarchical Bayesian models for multimodal brain–retina integration and compares Bayesian wavelet regression with Fourier methods, multitaper approaches, empirical mode decomposition, Gaussian processes, and deep learning. Using a hybrid systematic and narrative review approach guided by PRISMA principles, this work identifies challenges in computational scalability, cross-scale dependency modeling, multimodal uncertainty propagation, prior sensitivity, and clinical validation, and discusses emerging directions including Bayesian deep learning, graph wavelets, neural wavelet representations, foundation models, and real-time inference toward transparent, uncertainty-aware, and clinically meaningful neuro-retinal intelligence systems. In total, 237 selected studies are synthesized across methodological and application domains. Full article

Mountain–hill–plain transition zones combine terrain sensitivity, agricultural land use, and construction pressure and therefore require spatially differentiated management. This study develops a pressure-oriented framework that distinguishes composite ecological vulnerability from soil-erosion pressure and translates their multi-period co-occurrence into management demand. Using multi-source raster data for 2000, 2005, 2010, 2015, and 2020 on a common 30 m reference grid, we constructed an ecological vulnerability index (EVI) from 14 ecological, environmental, and human-activity indicators using a sensitivity–resilience–pressure framework, combined analytic hierarchy process (AHP)–entropy weighting, and unified natural breaks. Soil erosion intensity was estimated with the Revised Universal Soil Loss Equation (RUSLE) using digital elevation model (DEM)-derived flow accumulation, a capped effective slope length, and a fractional vegetation cover (FVC)-based cover-management factor. The EVI and RUSLE dimensions were integrated through a two-dimensional pressure matrix, a Coupled Pressure Index (CPI), and an Integrated Management Demand Index (IMDI). High vulnerability was concentrated in northern, urban-fringe, and selected agricultural areas, whereas moderate-or-above erosion was concentrated on slopes, in gullies, and on hilly farmland; areas subject to both pressures were spatially selective. The largest EVI weights were assigned to mean annual precipitation (0.3841), mean annual temperature (0.1727), ecosystem type (0.0820), biological abundance (0.0810), and built-up land ratio (0.0746). Non-vulnerable and slightly vulnerable areas increased from 40.62% to 64.27%, whereas highly and extremely vulnerable areas decreased from 38.98% to 23.29%. Moderate-or-above soil erosion accounted for 13.56% in 2020. The five-zone comprehensive management classification comprised stable conservation (40.19%), soil and water conservation priority (13.73%), ecological vulnerability regulation (27.94%), integrated management priority (1.48%), and transition management (16.66%). By linking ecological diagnosis, erosion-pressure identification, persistence analysis, and county-level zoning, the framework provides an interpretable spatial basis for prioritizing ecological conservation, soil and water conservation, and land-use management in mountain–hill–plain transition regions. Full article

In recent years, reconfigurable intelligent surfaces (RISs) have been proposed as a promising disruptive technology for future wireless communication systems. RISs enable unprecedented dynamic and programmable control of the electromagnetic waves by integrating software-defined metasurfaces into wireless environments. When smartly configured with the phase shifts of incident signals, RIS systems have the potential to improve spectral efficiency, energy efficiency, coverage, security, and other wireless metrics without the need for additional transmit power or active radio frequency chains. However, optimizing RIS-assisted wireless networks is highly nontrivial, due to the high-dimensional search space, cascaded channel model, coupled design of active and passive beamforming, etc. Machine learning (ML), and in particular deep reinforcement learning (DRL), has shown great promise in addressing these challenges by providing intelligent, adaptive, and real-time resource allocation and control. In this survey, we present a comprehensive overview of the state-of-the-art Machine Learning (ML) empowered RISs, from the fundamentals to the various ML paradigms, including supervised learning, unsupervised learning and the DRL framework. We then discuss in details ML-based solutions for channel estimation, beamforming design, power allocation, and resource management in RIS-aided multiple access systems. We further survey recent advances in ML for mobile edge computing, federated learning, unmanned aerial vehicles (UAVs), and physical layer security with RISs. Finally, we discuss several open challenges and future directions to spur future research on ML-empowered RISs, including scalability, hardware impairments, and integration with future 6G wireless networks. Critically, we provide a substantive technical treatment of Explainable AI (XAI) for RIS scenarios, detailing how SHAP value attribution, Grad-CAM saliency mapping over RIS element indices, and Transformer attention maps can be applied to interpret black-box DRL policies and CNN-based models used for continuous phase-shift and beamforming optimisation, enabling operators to understand, trust, and debug ML-driven RIS control decisions. Full article

Indoor localization based on Received Signal Strength (RSS) fingerprinting remains a prominent paradigm owing to its economic viability and seamless integration with existing wireless infrastructure. Nevertheless, positioning accuracy is frequently compromised by spatiotemporal environmental dynamics, device heterogeneity, and high-variance noise fluctuations inherent in individual Reference Points (RPs). To overcome these limitations, this paper introduces an Improved Soft Rough set-based Covering (I-SRC) model underpinned by a rigorous three-stage localization architecture. Stage (i): Raw offline RSS measurements undergo advanced filtering and structural optimization to construct a robust, noise-resilient radio map. Stage (ii): A specialized SRC methodology is deployed to classify the training instances, effectively mitigating the high dimensionality of the RSS feature space while preserving critical spatial characteristics. Stage (iii): A high-fidelity online matching algorithm correlates real-time RSS vectors with the established offline database to estimate coordinates. Comprehensive evaluations across multiple benchmark datasets demonstrate that the proposed I-SRC framework achieves a robust classification accuracy of approximately 96.38% and 97.75% on the UJI-V.1 and UJI-V.2 datasets, respectively. Crucially, the model yields outstanding positioning precision, recording low Average Positioning Errors (APE) of 0.58 m and 0.64 m on the respective datasets, thereby significantly outperforming contemporary state-of-the-art fingerprinting baselines. These results confirm that the I-SRC framework offers an efficient, scalable, and highly accurate solution for complex indoor positioning environments. Full article

Remotely operated vehicles (ROVs) suffer from severe navigation trajectory optimization problems in variable water-depth environments, as near-wall hydrodynamic effects cause synchronous scaling drift of added mass and damping coefficients. This parameter variation leads to obvious model mismatch in traditional fixed-gain controllers and seriously deteriorates ROV trajectory tracking accuracy. To address the scale-type parameter mismatch issue, this paper proposes an adaptive trajectory tracking control strategy combining forgetting-factor recursive least squares (RLS) online identification and periodic linear quadratic regulator (LQR) gain scheduling. A closed-loop coupling framework is established to estimate the discrete state-space matrices of ROVs via the RLS algorithm, and the optimal feedback gains are updated every 50 sampling steps to adapt to time-varying hydrodynamic characteristics. Three typical water-depth scenarios with different parameter mismatch degrees are set up for sinusoidal trajectory tracking simulations, adopting PID and fixed-parameter MPC as comparison methods. The results indicate that the proposed method maintains comparable steady-state performance with fixed-parameter MPC under nominal conditions, and reduces the two-dimensional trajectory RMSE by 8.4% and 57.4% under moderate and severe parameter mismatch conditions, respectively. A critical mismatch threshold of fixed-parameter MPC compensation capability is also determined. This study provides a feasible technical reference for high-precision adaptive motion control of ROVs in variable-depth water environments. Full article

Background and Objectives: Chronic kidney disease (CKD) affects roughly 788 million adults worldwide according to the Global Burden of Disease (GBD) 2023 estimate. Extracellular vesicles (EVs) are promising sources of biomarkers and therapeutic carriers, while artificial intelligence (AI) offers tools for high-dimensional EV analysis and computational design. Materials and Methods: We conducted a scoping review using a structured PubMed/MEDLINE search, run on 7 September 2026, and PRISMA-ScR reporting principles. Eligible studies applied AI/ML to EV-derived or closely related kidney-disease data. The review was not prospectively registered. The formal scoping evidence set was used to map kidney EV–AI studies focused on diagnosis, prognosis, and biomarker discovery; EV engineering, therapeutic design, manufacturing, and physiologically based pharmacokinetic (PBPK) modeling were synthesized separately as contextual and future-oriented literature identified through non-systematic contextual literature identification. Predictive models were critically appraised using a PROBAST-informed framework. Results: Nine reports were retained in the formal PRISMA evidence set. Direct kidney EV–AI evidence was limited and heterogeneous, with small cohorts and variable validation strategies. Several studies reported encouraging discrimination, but external validation, calibration, standardized EV workflows, and prospective clinical utility were uncommon. Contextual literature suggests potential roles for AI in EV engineering, manufacturing, and PBPK-informed therapeutic development, but these applications remain investigational in CKD. Conclusions: AI–EV integration is a promising research framework rather than a clinically validated platform. Translation will require standardized EV methods, adequately powered multicentre studies, independent validation, clinically relevant operating thresholds, prospective utility studies, and explicit manufacturing and regulatory strategies. Full article

To improve the robustness and computational efficiency of underwater direction-of-arrival (DOA) estimation in complex marine noise environments, this study proposes an improved sparse Bayesian off-grid DOA estimation method under mixed noise. Within a hierarchical Bayesian framework, temporal precision variables for non-Gaussian noise and spatial precision variables for nonuniform noise are introduced and jointly estimated to adaptively reduce the weights of impulsive snapshots and noise-contaminated array elements. A space-alternating variational inference scheme transforms the joint estimation of signal and noise parameters into low-dimensional analytical updates, thereby reducing the computational burden associated with high-dimensional posterior covariance matrix inversion. In addition, adaptive grid refinement locally densifies high-confidence spectral-peak regions, while a first-order Taylor expansion is employed for continuous off-grid compensation. Simulation results show that the proposed method achieves more stable spatial spectrum reconstruction under mixed-noise and grid-mismatch conditions, providing a favorable balance between estimation accuracy and computational complexity. Results from the SWellEx-96 sea-trial data further demonstrate its ability to stably track target bearings under practical array geometry and unknown oceanic interference. These results verify the effectiveness and applicability of the proposed method in both mixed-noise simulations and real shallow-water data. Full article

This paper proposes Sensor-Initialized Gaussian Splatting (SIGS), a method that uses panoramic depth data acquired by the LiDAR sensor in a smartphone for three-dimensional (3D) scene reconstruction and VR space generation. Traditionally, 3D scene generation has required specialized knowledge and significant time, posing challenges for its application in VR. In particular, point clouds estimated by Structure from Motion (SfM), which are used for initializing 3D Gaussian Splatting (3DGS), have had limitations in density and accuracy. SIGS addresses these challenges by utilizing high-accuracy point clouds directly acquired from an iPhone’s LiDAR sensor for 3DGS initialization. For this research, a dedicated smartphone application called Panoramic Depth Recorder (PDR) was developed to simultaneously capture RGB images, depth images, point cloud data, and camera position and rotation information while the iPhone is rotated. These point clouds are then integrated into a common world coordinate system. For outdoor scenes, geometric information over a wider range is supplemented by combining near-field LiDAR data with far-field point clouds estimated by SfM. Experiments demonstrated that SIGS improved rendering accuracy (Structural Similarity Index Measure and Learned Perceptual Image Patch Similarity) and visual quality compared to conventional methods initialized solely with SfM. The generated scenes exhibited fewer artifacts and reproduced more faithful geometric shapes, with improved floor surfaces and ceiling irregularities. The mesh data of the generated 3D scenes is designed for use in VR environments. After conversion from PLY to the FBX format using Blender, they can be imported into Unity, enabling collision detection and user movement control within the VR space. This opens up possibilities for applications such as creating digital twins of robot training environments to reproduce real-world spaces in VR applications. Full article

High-power mainlobe jamming severely degrades the target-detection capabilities of array radars in complex electromagnetic environments. Although the difference-channel signal can be reused as an auxiliary reference to suppress the jamming component in the sum channel, the jamming component remains unmitigated in the measured difference channel of a three-channel radar without a dedicated difference–difference channel, thereby introducing additional angle-estimation errors. Addressing this, this study analyzes a subarray-level, three-channel monopulse array radar. Based on a one-dimensional uniform linear array and the sum–difference monopulse principle, we formulate a joint model for mainlobe jamming cancellation and angle estimation. We then evaluate how the jammer azimuth, jamming-to-signal ratio (JSR), number of jammers, and radar operating modes (search and tracking) affect the resulting angle error. Simulation results indicate that in search mode, cancellation restores local angle-estimation capabilities near the target direction. When the beam points near the jammer, the jammer approaches the null of the auxiliary difference beam, weakening the reference signal and causing pronounced nonmonotonic distortion in the angle-estimation curve. In tracking mode, the cancellation process yields smaller angle errors and increased stability. A two-element 4.03 GHz anechoic-chamber experiment completes the proof-of-principle validation for the core mechanism under the single-jammer search-mode scenario. These findings define the operational boundaries of using the difference-channel signal as an auxiliary reference for sum-channel jamming cancellation in three-channel radars, providing a theoretical foundation for angle-error compensation, receiving-channel design, and the refinement of interference cancellation techniques. Full article

Clustering high-dimensional data is challenging when irrelevant or weakly informative features obscure the underlying cluster structure. To address this issue, we propose two sparse k-means clustering algorithms, S-KM1 and S-KM2, that perform cluster estimation and feature selection by sparse feature weighting simultaneously. Clustering has been applied to a wide range of disciplines, including image segmentation, social network analysis, medical imaging, market segmentation, anomaly detection, etc. In clustering algorithms, the most popular and widely used method is k-means. However, the k-means algorithm always treats feature (attribute) components in datasets equally. In practice, different features may contribute unequally to the cluster structure and should therefore not necessarily be weighed equally. This has been demonstrated by Yang and Benjamin (2023) with two types of sparse possibilistic c-means (PCM) methods called SPCM1 and SPCM2, using the Lasso concept. Motivated by SPCM1 and SPCM2, we propose two sparse k-means clustering algorithms, S-KM1 and S-KM2, in this paper. The proposed methods S-KM1 and S-KM2 consider the k-means clustering subject to different feature weight constraints based on the Lasso and L1 penalty so that they can shrink the irrelevant features towards zero and achieve sparsity in features. The proposed S-KM1 and S-KM2 are compared with some of the most popular sparsity-clustering techniques on several numerical and real-life datasets using different clustering performance measures. Comparisons and experimental findings demonstrate the validity, efficacy, and effectiveness of the proposed S-KM1 and S-KM2 clustering algorithms, and they also surpass the existing advanced algorithms in terms of efficiency and usefulness. Full article

Nanomaterials improve the fresh-state behavior, mechanical performance, durability, and multifunctionality of construction materials through packing and filling, nucleation, interfacial regulation, crack bridging, transport-barrier effects, and functional responses. These benefits depend on coupled material, processing, matrix, and service variables, making conventional trial-and-error development inefficient. Machine learning (ML) offers a data-driven route for resolving nonlinear composition–processing–structure–property relationships. This review combines a CiteSpace-based bibliometric analysis of Web of Science Core Collection (WoSCC) records with a structured qualitative synthesis of ML-assisted studies on zero-dimensional nanoparticles, one-dimensional high-aspect-ratio nanomaterials, two-dimensional layered nanomaterials, functional nanomaterials, carbon dots, and emerging hybrid systems. Annual output increased from four publications in 2012 to 268 in 2025, and ML showed the strongest recent keyword burst (strength = 22.57). Across representative studies, reported test-set coefficients of determination commonly ranged from 0.858 to 0.995, although validation designs and data independence varied markedly. Tree-based ensembles were generally effective for small-to-medium tabular datasets, whereas deep learning was most defensible for images, spectra, and time-series data. Prediction estimates properties for a specified formulation; optimization uses a validated surrogate model to search the formulation or processing space and therefore requires independent experimental verification. Nano-SiO₂, carbon nanotubes, and graphene oxide dominate the evidence base, while durability, microstructure, uncertainty quantification, and multifunctional co-optimization remain underrepresented. Matrix-aware standardized databases, physically constrained models, grouped or external validation, and closed-loop experiments are required for reliable intelligent design. The bibliometric component covers 1324 Web of Science Core Collection records published during 2012–2026, whereas the qualitative component applies explicit criteria for construction relevance, definition of the ML task, and sufficiency of performance and validation reporting. Bibliometric patterns and qualitative evidence are therefore reported and interpreted separately. Full article

Accurate characterization of particle dynamics in complex swirling flows is essential for understanding flow behavior and transport mechanisms. However, extracting three-dimensional (3D) rotational motion parameters of particles from High-Speed Motion Analyzer (HSMA) imaging experiments remains challenging due to the limitations of conventional manual frame-by-frame analysis, which suffers from low efficiency, subjective errors, and poor scalability. In this study, an intelligent imaging-based sensing system is proposed for automated 3D measurement of coupled particle self-rotation and revolution motion in hydrocyclone swirling fields. The proposed system integrates dual orthogonal high-speed cameras with a machine learning-driven processing framework, enabling an end-to-end workflow consisting of particle detection, multi-object tracking, motion-state recognition, kinematic parameter calculation. To achieve robust particle detection under different imaging conditions, You Only Look Once version 8 (YOLOv8) detectors are trained following a mixed-view strategy: the Y-view detector is trained on the Y-camera data, whereas the X-view detector is trained on the merged dual-camera data. ByteTrack is subsequently employed to associate particle detections across consecutive frames and generate continuous trajectories. For motion-state recognition, EfficientNet-B1 is adopted to classify the relative positional states of internal particle markers. Furthermore, a Self-Attention Generative Adversarial Network (SAGAN) is introduced for minority-class sample generation to alleviate class imbalance and improve classification robustness.Based on the extracted trajectories and state information, a kinematic model is integrated to automatically estimate particle revolution and self-rotation velocities. Experimental results demonstrate that the proposed YOLOv8 detectors achieve reliable detection performance, with mean average precision at an Intersection over Union (IoU) threshold of 0.5 (mAP@0.5) values of 79.84% and 71.62% for the two imaging views, respectively. The combined YOLOv8-ByteTrack framework achieves Higher Order Tracking Accuracy (HOTA) and Multiple Object Tracking Accuracy (MOTA) values of 63.791 and 90.49, respectively. After SAGAN-based data augmentation, the classification accuracy of EfficientNet-B1 increases from 98.62% to 99.54%. System-level validation using more than 60 randomly selected particle trajectories shows average measurement accuracies of 97.59% for revolution velocity and 91.8% for self-rotation velocity, while reducing the processing time from more than 600 s manually to 11.1 s per trajectory. The proposed machine learning-assisted particle 3D motion sensing system provides an efficient and reliable solution for automated extraction of particle kinematic information under complex flow imaging conditions, offering new opportunities for intelligent monitoring and mechanism analysis of multiphase flow systems. Full article

This work presents a data-driven approach for state-of-charge estimation of rechargeable zinc-air batteries based on electrochemical impedance spectroscopy. Due to the nonlinear electrochemical behavior and flat discharge voltage profile of zinc-air batteries, accurate state-of-charge estimation remains challenging. The investigated cells employ a three-electrode configuration with a dedicated gas diffusion electrode for discharge and a separate electrode for charging. This work focuses exclusively on discharge operation, as the two current paths involve physically distinct electrodes with fundamentally different impedance characteristics. High-dimensional impedance spectra are combined with physically interpretable features derived from a simplified equivalent circuit model and compressed via principal component analysis. A long short-term memory network models the relationship between the resulting feature representation and state-of-charge, with Bayesian hyperparameter tuning applied to optimize architecture and training configuration. Performance is compared against baseline models including multilayer perceptrons. The model is trained on multiple battery cells and evaluated on a completely held-out cell to assess cross-cell generalization. The results show that principal component analysis compression of the combined impedance spectrum and equivalent circuit feature vector is the decisive optimization step, achieving a mean absolute error of 1.04% on an unseen test cell. In contrast, the choice of model architecture has a smaller impact on performance. Full article

Surface electromyography (sEMG) reflects neuromuscular control, but multichannel recordings are high-dimensional and difficult to interpret. This study evaluated whether muscle-synergy activations provide a compact input representation for lower-limb kinematic reconstruction. Twelve-channel sEMG and hip, knee, and ankle angles were obtained from 120 healthy male participants performing seven tasks in the Gait120 dataset. A four-synergy representation was derived using nonnegative matrix factorization, and within-participant cross-task similarity was assessed. Under participant-wise five-fold cross-validation, fold-specific shared dictionaries were estimated exclusively from training participants, and separate task-specific models reconstructed joint trajectories across five locomotor tasks. ExtraTrees achieved the lowest RMSE in most task–joint combinations and was used for exploratory detailed comparisons. Shared-synergy activations reduced the regressor-input representation from 12 variables to four activation coefficients per time point. Task-level mean RMSEs were 5.60–6.29° for synergy activations and 5.52–6.28° for raw sEMG. After Holm correction, no statistically significant difference was detected in 14 of the 15 task–joint comparisons, while one favored synergy activations. These findings support shared-synergy activation as a compact and physiologically interpretable input representation and provide a basis for neuromuscularly informed modeling of lower-limb movement across locomotor conditions. Full article