ABSTRACT Eutrophication in inland lakes, driven by excessive total phosphorus (TP), necessitates robust monitoring methods. While hyperspectral remote sensing holds promise, accurate TP retrieval is hindered by its non-optically active nature and the complex spectral interplay in water bodies. This study aimed to develop a novel, interpretable framework to overcome these challenges by synergistically integrating diverse spectral features for improved TP estimation from field hyperspectral data. The key innovation of this study was a multi-pathway feature engineering framework that strategically merged three complementary information streams: spectral indices grounded in bio-optical theory via proxies for suspended solids (SS) and coloured dissolved organic matter (CDOM), statistically optimized features from Lasso regression, and machine learning-derived predictors from XGBoost. This fusion reconciled physical interpretability with data-driven predictive power. Applied to Dianshan Lake, the framework achieved a superior prediction accuracy (R2 = 0.81) compared to any single pathway. Furthermore, hierarchical optimization distilled the fused features into a compact core set of only 8 predictors, retaining over 90% of the model’s performance and highlighting the framework’s efficiency. The study demonstrates a scalable and physiochemically insightful approach for advancing hyperspectral retrieval of non-optically active water quality parameters.

Low-Dose Computed Tomography (LDCT) reduces radiation risk but introduces high noise levels that compromises diagnostic quality. To address this, we propose a Hybrid Generalized Efficient Layer Aggregation Network-UNet (GELAN-UNet) model, which incorporates medical priors into a progressive modular architecture. This design uses medically enhanced modules in shallower layers to capture fine details and computationally efficient blocks in deeper layers to reduce cost. Key innovations include a novel low-frequency retention path and an edge-aware attention mechanism, both crucial for preserving critical diagnostic structures. Evaluated on the public Mayo Clinic dataset, the proposed method achieves a superior peak signal-to-noise ratio (PSNR) of 45.28 dB - a 12.45% improvement over the original LDCT - while maintaining an optimal balance between denoising performance and computational efficiency. The critical importance of the low-frequency path, as revealed by ablation studies, validates the rationality of the hybrid strategy, which is further supported by comparisons with full medical and frequency-aware variants. This work delivers a high-performance denoising model alongside a practical, efficient architectural paradigm - rigorously validated through systematic exploration - for integrating domain-specific medical knowledge into deep learning frameworks.

The multi-classification of histopathological images under imbalanced sample conditions remains a long-standing unresolved challenge in computational pathology. In this paper, we propose for the first time a cross-patient pseudo-bag generation technique to address this challenge. Our key innovation lies in a cross-patient pseudo-bag generation framework that extracts complementary pathological features to construct distributionally consistent pseudo-bags. To resolve the critical challenge of distributional alignment in pseudo-bag generation, we propose an affinity-driven curriculum contrastive learning strategy, integrating sample affinity metrics with progressive training to stabilize representation learning. Unlike prior methods focused on bag-level embeddings, our framework pioneers a paradigm shift toward multi-instance feature distribution mining, explicitly modeling inter-bag heterogeneity to address class imbalance. Our method demonstrates significant performance improvements on three datasets with multiple classification difficulties, outperforming the second-best method by an average of 1.95 percentage points in F1 score and 2.07 percentage points in ACC.

Composite laminates possess complex anisotropic behavior, motivating the development of simplified yet accurate modeling approaches. In this paper, we present a study that introduces a stiffness-invariants-based constitutive model for symmetric, balanced composite laminates, highlighting a novel “quasi-Poisson’s ratio” parameter as a key innovation. The proposed method reconstructs the laminate stiffness matrices using invariant theory (trace of stiffness tensor) and a Master Ply concept, thereby reducing the number of independent material constants. The methods and assumptions (e.g., neglecting minor bending-twisting couplings) are outlined, and the model’s predictions of critical buckling loads are compared to classical laminate theory (CLT) results. Good agreement is observed in most cases, with a consistent conservative bias of CLT. The results confirm that the invariant-based model captures the dominant stiffness characteristics of the laminates and can slightly overestimate stability margins due to its idealizations. In conclusion, this work provides an efficient constitutive modeling framework that can be integrated with finite element analysis and extended to more general laminates in future studies.

Gecko-inspired soft robots offer unique advantages for confined-space operations. Here, we introduce a soft climbing robot designed for tasks such as spacecraft inspection and disaster response. Unlike conventional attachment mechanisms relying on claws, suction-based adhesion, or electromagnetic components, the proposed system integrates polyurethane-based dry adhesive footpads, which passively adapt to a wide range of surfaces and require no external power. A key innovation of this work is the incorporation of variable stiffness footpads capable of actively tuning their mechanical properties in response to surface interaction demands, thereby optimizing both attachment stability and detachment efficiency during locomotion. Experimental characterization indicates that these footpads achieve an adhesion-to-detachment force ratio of 11:77, demonstrating a favorable balance between strong attachment and reliable release. Furthermore, the robot integrates a bioinspired crawling mechanism that synchronizes limb actuation with the deformation of a flexible spine structure, effectively enhancing propulsion and maneuverability across irregular terrain. Validation experiments conducted on diverse surface types confirm the robot’s environmental adaptability and highlight its potential for deployment in constrained, unstructured, and dynamically changing operational contexts.

ABSTRACT Accurate State of Charge (SOC) estimation is critical for the safe and efficient operation of batteries in electric vehicles (EVs). While deep learning models like Transformers have shown promise, they often struggle with sensor noise and complex temporal dynamics. Similarly, hybrid approaches like VMD‐Transformer rely on fixed basis functions that lack adaptability. To address challenges such as sensor noise, nonlinear dynamics, and complex temporal dependencies, this study proposes a novel VMD‐Basisformer model that integrates Variational Mode Decomposition (VMD) with an enhanced Basisformer neural network. The key innovations of our approach include: (1) a battery‐optimized VMD process for noise reduction and multi‐scale feature extraction; (2) a dual‐path basis generation mechanism tailored to battery temporal dynamics; (3) a hierarchical attention architecture for capturing both local and global temporal dependencies. Experimental results under Dynamic Stress Test (DST) and Urban Dynamometer Driving Schedule (UDDS) conditions show that the proposed model significantly outperforms existing methods such as Transformer, Basisformer, and VMD‐Transformer. Experimental validation on LiFePO 4 battery and supercapacitor datasets under DST and UDDS conditions shows that the VMD‐Basisformer outperforms benchmark models (Transformer, Basisformer, VMD‐Transformer) in accuracy and robustness. Ablation studies confirm the critical role of InfoNCE loss in ensuring temporal consistency.

The industrial scaling of Hermetia illucens production for waste bioconversion and alternative protein provision is severely constrained by logistical challenges in breeding stock preservation and long-distance transport. To address this, we developed and optimized stage-specific cold storage protocols to significantly prolong the shelf-life of H. illucens. Through a multi-stage experimental approach, we determined that 1-d-old eggs can be successfully stored at 17 °C for 10 d with 50.0% hatchability, larvae retained a 72.8% survival rate after 30 d at 15 °C, and prepupae maintained a 75.3% eclosion rate following 60 d at 14 °C. In addition, a key innovation of this study was the strategic use of dietary cryoprotectants to markedly enhance cold tolerance. Under a severe discriminating temperature (-5 °C), supplementation with 4% proline and 0.5% trehalose elevated larval survival to over 80%. Furthermore, at the chronic stress of the developmental threshold (12 °C), 3% glycerol, nearly doubled larval survival rates compared to the control. The application of these optimized cryoprotectants with stage-specific storage temperatures effectively mitigated sublethal fitness costs, ensuring high post-storage performance in survival, pupation, eclosion, and reproductive output. Our findings provide a robust, comprehensive framework for synchronizing H. illucens supply chains, enabling viable long-distance transport, and facilitating the reliable industrial scaling of H. illucens production for the circular bioeconomy.

Traditional patient education often lacks personalization and engagement, potentially limiting knowledge acquisition and treatment adherence[1]. Advances in artificial intelligence (AI), including voice cloning technology and large language models such as ChatGPT, offer new opportunities to deliver personalized, scalable, and interactive health education[2-3]. However, evidence regarding the comparative effectiveness of different AI-based voice cloning strategies and the reliability of automated AI evaluation tools remains limited[4-5]. To evaluate the effectiveness of AI-assisted patient education integrating voice cloning and ChatGPT, to compare physician voice cloning with patient self-voice cloning, and to assess the reliability of ChatGPT as an automated evaluation tool for education outcomes. A prospective, three-arm, parallel-group randomized controlled trial.A total of 180 hospitalized patients requiring standardized health education were recruited from a tertiary hospital. Inclusion criteria were: age ≥18 years, clear diagnosis requiring health education, clear consciousness, and voluntary participation with informed consent. Exclusion criteria were: severe hearing impairment, severe cognitive impairment, expected hospitalization <3 days, or prior participation in similar studies.Participants were randomly assigned (1:1:1) to receive (1) traditional education (control), (2) AI-assisted education using physician voice cloning, or (3) AI-assisted education using patient self-voice cloning. All groups received identical educational content with equal duration.The primary outcome was education content compliance, evaluated using ChatGPT-4 with validated prompts and verified by expert review. Secondary outcomes included knowledge retention, education satisfaction, treatment adherence, quality of life (SF-36), and psychological status (Hospital Anxiety and Depression Scale).Participants were randomly allocated using a computer-generated random sequence. Due to the nature of the intervention, participants were not blinded; outcome assessors and data analysts were blinded to group allocation. Of 180 randomized participants, 174 (96.7%) completed the trial. Both AI-assisted groups demonstrated significantly higher education content compliance immediately after education compared with the control group (physician voice: 86.7 ± 7.3; self-voice: 92.5 ± 6.8 vs control: 73.2 ± 8.5; P < 0.001). The patient self-voice group showed superior knowledge retention before discharge, higher education satisfaction, and greater treatment adherence compared with both the physician voice and control groups (all P ≤ 0.02). At one-month follow-up, the self-voice group maintained improved adherence (Cohen's d = 0.74) and exhibited significantly lower anxiety and depression scores (all P ≤0.02), along with improved SF-36 quality-of-life domains. ChatGPT-based evaluations demonstrated high reliability compared with expert assessments (weighted κ = 0.87, 95% CI 0.82-0.91). This study introduces an innovative patient education model integrating AI voice cloning and ChatGPT, representing a novel approach distinct from previous studies that primarily relied on standard text-to-speech or professionally recorded content. The key innovation lies in utilizing patients' own cloned voices for health education delivery, leveraging the self-reference effect to enhance learning outcomes. Compared with prior research focusing on clinician-narrated content, this study provides the first empirical evidence that self-voice education produces superior outcomes across multiple domains including compliance, satisfaction, and psychological well-being. These findings contribute to the field by establishing a theoretical and practical framework for personalized AI-driven patient education. In real-world clinical settings, this approach offers a scalable, cost-effective solution to enhance patient engagement, particularly valuable in resource-limited environments where individualized education is challenging to deliver. Trial Registration: Chinese Clinical Trial Registry (ChiCTR2500101882); registration application initiated on January 15, 2025 and finalized on April 30, 2025, before participant enrollment began in May 2025.

This study presents a Design for Additive Manufacturing (DfAM)–driven redesign of an industrial robot vacuum gripper for Fused Deposition Modeling (FDM), focusing on the systematic transformation of a multi-part, machined aluminum assembly into a lightweight, support-minimized polymer component suitable for continuous industrial operation. Beyond a practical redesign, the work contributes a geometry-centered DfAM methodology that links internal channel topology, overhang control, and functional interfaces to manufacturability, vacuum performance, and cost efficiency. The development follows three iterative design revisions, progressing from a geometry-adapted baseline toward a fully DfAM-optimized solution. A key innovation is the introduction of support-free internal vacuum channels with triangular cross-sections, enabling complete elimination of soluble support material within enclosed cavities. This redesign reduces the internal vacuum volume by 44%, leading to faster vacuum response while maintaining functional suction performance. The optimized overhang angles, filleted load paths, and DfAM-compliant suction cup seats significantly reduce post-processing requirements and improve structural robustness. Experimental validation under industrial operating conditions confirms that the final design achieves reliable vacuum performance and mechanical durability. Compared to the original configuration, the optimized gripper demonstrates a substantial reduction in manufacturing complexity, with printing time reduced by approximately 50% and total part cost decreased by 26%, primarily due to eliminated tooling, reduced support material, and simplified post-processing. The presented results demonstrate that DfAM principles, when applied systematically at both global and internal geometry levels, can yield quantifiable functional and economic benefits. The findings provide transferable design guidelines for support-free internal channels and functional interfaces in FDM-manufactured vacuum components, offering practical reference points for researchers and practitioners developing end-use additive manufacturing solutions in industrial automation.

Abstract This paper introduces a quantum-enhanced finite element method (FEM) designed for noisy intermediate-scale quantum (NISQ) devices, leveraging variational quantum algorithms (VQAs) to solve engineering partial differential equations (PDEs). We demonstrate the framework by solving the Euler-Bernoulli beam and the NAFEMS T4 heat transfer problems, which involve Dirichlet, Neumann, and Robin boundary conditions. A key innovation is a ``set-to-zero" strategy that incorporates boundary conditions through a correction matrix, $K_{bc}$, allowing for flexible imposition at any node without domain decomposition. The global stiffness matrix is decomposed into a constant number of Pauli terms, $O(1)$, using the method by Sato et al., while boundary terms are handled with a sublinearly scaling Partial Pauli Measurement (PPM) technique. The algorithm achieves logarithmic qubit scaling ($n = \lceil \log_2 N \rceil $ qubits for N degrees of freedom) and employs shallow, hardware-efficient circuits with empirically trainable depth for small-scale systems. Validation on the Qiskit statevector simulator shows high accuracy. For the Euler-Bernoulli beam problem with 4 to 64 degrees of freedom, the algorithm achieves relative errors of 0.5–1.5\% and fidelities of 0.998–0.999. For the NAFEMS T4 heat transfer benchmark, a 5.4\% relative error is observed. The VQA converges robustly within 77–350 iterations, though barren plateaus are a known challenge for scaling to larger systems. This work establishes a scalable framework for quantum FEM, offering a significant memory advantage over classical methods and advancing the potential for quantum-enhanced engineering simulations.&amp;#xD;

Graduate education in biomedical science faces persistent challenges in rural and under-resourced regions, including limited access to research training infrastructure and experiential learning opportunities. The University of South Dakota's Graduate Research Initiative for Scientific Enhancement (G-RISE) program addressed these barriers by embedding structured training within a Carnegie-classified high research activity institution (R2) in a state designated by the NIH Institutional Development Award (IDeA) program as historically underfunded. From 2020 to 2025, G-RISE supported 11 Ph.D. students, most of whom were first-generation college graduates or from rural backgrounds, through a curriculum emphasizing rigorous research, mentor development, microcredential coursework, science communication, and career exploration. Trainees achieved 100% Ph.D. retention and graduated one year faster than their peers (4.08 vs. 5.07 years), with comparable publication rates (1.04 vs. 1.16 publications/year). Department-wide outcomes also improved during the funding period: the median time-to-degree decreased to 5.0 years, attrition dropped, and graduates averaged 5.7 peer-reviewed publications, more than twice the pre-G-RISE average. Additionally, there were increases in graduates earning nationally competitive fellowships. Key training innovations, including interdisciplinary microcredential electives and formal mentor training, were adopted across the broader graduate program, strengthening institutional capacity. These findings illustrate that targeted, scalable interventions can improve educational outcomes and research productivity in institutions with limited NIH training infrastructure. The USD G-RISE model offers a replicable framework for programs, especially in rural or less resourced settings, seeking to enhance biomedical training. Further, it underscores the importance of aligning training strategies with local strengths and workforce needs.

ABSTRACT Due to the inherent bandgap limitations of traditional semiconductors, achieving a broadband response typically requires complex heterostructures, which suffer from lattice mismatch and compromised performance. Herein, we propose a novel oxygen plasma treatment WSe 2 homojunction strategy to overcome this trade‐off. A simple mask‐assisted magnetron sputtering technique was employed to fabricate an interdigital p‐n homojunction photodetector. The key innovation lies in creating a staggered type‐II band alignment via localized oxygen doping on pure WSe 2 , which effectively promotes charge separation across the broadband. The device exhibits remarkable detectivity values of 4.58×10 9 Jones (365 nm), 3.57×10 10 Jones (470 nm), 1.29×10 11 Jones (550 nm), 1.13×10 12 Jones (850 nm), and 1.52×10 12 (1064 nm) at a bias of 0 V. Clear “HIT” pixel images with distinct edges were obtained at all these wavelengths. At 1064 nm, the device achieved a high on/off ratio of 2,208 and fast rise/fall times of 59 µs and 18 µs, respectively. Through wide‐spectrum imaging and high‐frequency response testing, the device demonstrates excellent potential for next‐generation, high‐performance broadband imaging systems.

This systematic review outlines pivotal advancements in deep learning-driven protein structure prediction and design, focusing on four core models—AlphaFold, RoseTTAFold, RFDiffusion, and ProteinMPNN—developed by 2024 Nobel Laureates in Chemistry: David Baker, Demis Hassabis, and John Jumper. We analyze their technological iterations and collaborative design paradigms, emphasizing breakthroughs in atomic-level structural accuracy, functional protein engineering, and modeling multi-component biomolecular interactions. Key innovations include AlphaFold3's diffusion-based framework for unified biomolecular prediction, RoseTTAFold's three-track architecture integrating sequence and spatial constraints, RFDiffusion's denoising diffusion for de novo protein generation, and ProteinMPNN's inverse folding for sequence–structure co-optimization. Despite major progress in applications such as binder design, nanomaterials, and enzyme engineering, challenges persist in dynamic conformational sampling, multimodal data integration, and generalization to non-canonical targets. We propose future directions, including hybrid physics-AI frameworks and multimodal learning, to bridge gaps between computational design and functional validation in cellular environments.

Abstract Landslide Early Warning Systems (LEWS) aim to anticipate slope failures by identifying critical hydrometeorological conditions, but their accuracy is often limited by simplified approaches that do not fully capture subsurface hydrological processes. To strengthen planning and management strategies for landslide hazard-prone areas in alpine regions, we developed the SCLAM model, which integrates the SNOW-17 snowmelt model, the coupled routing and excess storage (CREST) hydrological model, and a landslide model that combines the infinite slope method with a random forest approach. The SCLAM model was validated in the Upper Garonne River Basin (Pyrenees, Spain), a region that experienced multiple-occurrence landslide events on 18 June 2013 due to intense rainfall and snowmelt. The SCLAM model showed fast performance, with an average accuracy of 75% in identifying landslide initiation areas at a 30-m pixel scale. Furthermore, a key innovation of this study is the use of the baseflow excess variable from the CREST model as an indicator of subsurface hydrological conditions. As a result, we proposed a prototype LEWS capable of representing landslide-prone areas both spatially and temporally. The baseflow excess variable proved to be a reliable temporal indicator of slope instability, enabling the system to identify potential landslide triggers and reduce the number of false alerts. Spatial warnings were generated by aggregating landslide probabilities of failure within subbasins, using the percentage of unstable area to classify landslide warnings into four levels. Together, these components provide a practical foundation for an operational early warning system capable of supporting real-time decision-making.

Purpose Web citations are essential for scholarly integrity, but their reliability is threatened by link rot and content drift. Fields like Library and Information Science (LIS), which depend heavily on web references, face significant challenges in preserving digital citations. This longitudinal study (2005–2025) aims to investigate the decay and recovery of web citations in LIS journals, offering actionable solutions to preserve digital scholarship. Design/methodology/approach This 20-year longitudinal study employed a quantitative approach to examine web citation decay in LIS literature. We analyzed 2,886 citations from 608 articles published in four leading journals. Findings Three key findings emerged from the analysis. First, web citations are now decaying exponentially, with accessibility dropping from 87% for citations 0–5 years old to 38% for those over 10 years old. Furthermore, permanent link rot has tripled from 5% in 2012 to 15% in 2025. Second, preservation outcomes vary dramatically by domain (e.g. .edu domains show 93% accessibility versus 42% for .com domains) and content format (e.g. PDFs maintain 92% accessibility compared to 41% for database-driven content). Third, although recovery tools have improved (the Wayback Machine’s success rate increased by 171%), their benefits are offset by new challenges such as failures caused by dynamic content, which now account for 19% of all failures. The study advocates for mandatory archiving protocols and persistent identifiers to safeguard scholarly records, highlighting the urgent need for systemic reforms. Originality/value This study offers three key innovations: (1) It provides the first longitudinal evidence that link rot is accelerating despite improvements in archiving tools, thereby revealing a paradox in preservation efforts; (2) it identifies dynamic content as a significant new error category, accounting for 19% of failures; (3) it demonstrates a phenomenon of “preservation resistance” in 15% of citations that are unrecoverable by any method. These findings redefine the challenges of modern link rot.

Bioluminescence, visible light produced by a living organism, is a key innovation in the diversification of deep-sea fishes. It is useful for a myriad of behaviors and interactions in deep-sea organisms, including communication, predation, camouflage via counterillumination, and predator avoidance. In this study, we investigate the deep-sea tubeshoulders (Platytroctidae), fishes that possess a unique postcleithral tube organ associated with their shoulder girdle that excretes bioluminescent fluid, a feature that unites all members of this poorly studied family. Many tubeshoulders also possess additional bioluminescent structures and luminescent tissues, including a series of tube organs on the caudal peduncle unique to Platytroctes apus that are hypothesized to be similar in structure and function to the postcleithral tube organ. Herein, we present the first histological analysis of the caudal tube organs in P. apus and use histological methods to investigate the morphological diversity in postcleithral tube organ structure across 14 platytroctid species, representing 10 of 13 valid genera. We show that the postcleithral tube organ generally exhibits a conserved morphology across genera and species. However, several species-specific anatomical differences are noted. In some individuals, we observe the presence of luminescent-fluid cells within the tube organ in various stages of development, which may provide evidence for inferring the type of secretory gland found in this novel bioluminescent light organ. We also show that the structure of the caudal tube organs in P. apus are similar to the postcleithral tube organ present in all members of Platytroctidae, likely indicating a similar luminescent fluid emission function.

Abstract As the More-than-Moore (MtM) paradigm reshapes semiconductor manufacturing toward functional integration, material diversity, and high-precision three-dimensional (3D) architectures, chemical mechanical polishing (CMP) has evolved from a planarization step into a critical enabler of heterogeneous integration. This review comprehensively examines the key challenges and technological innovations driving CMP in the MtM era. In front-end-of-line (FEOL) processes, high selectivity among diverse materials should be ensured while suppressing defects. In back-end-of-line (BEOL) and advanced packaging, CMP is required to deliver sub-nanometer flatness across complex multilayer and hybrid bonding structures. The emergence of wide-bandgap materials such as SiC and GaN introduces additional demands for chemically enhanced and hybrid CMP techniques to overcome their extreme hardness and chemical inertness. Furthermore, recent advancements in artificial intelligence (AI)-driven process prediction, in-situ sensing, and eco-friendly consumables are accelerating CMP’s transformation into a data-informed, sustainable manufacturing platform. This review redefines CMP as a foundational technology for next-generation integration and outlines its future trajectory in advanced semiconductor fabrication.

We present the Imran-Class Gravastar (ICG), a novel analytical model for compact stellar objects that provides an exact, isotropic, and horizonless solution to Einstein’s field equations. The model’s key innovation lies in its curvature-dependent geometric formulation, which generates a self-consistent quadratic equation of state [Formula: see text] directly from spacetime geometry, eliminating the dependence on empirical nuclear equations of state. This geometrically derived approach ensures automatic satisfaction of all energy conditions and causality requirements while maintaining dynamical stability ([Formula: see text]) throughout the stellar interior. The analytical framework produces closed-form expressions for mass, radius, and compactness, yielding maximum masses of [Formula: see text]–[Formula: see text] with radii [Formula: see text]–[Formula: see text], in agreement with modern observational constraints from NICER and gravitational-wave observations. Unlike traditional neutron star models, the ICG configuration exhibits systematically softer pressure profiles (5–15% reduction), lower compactness ([Formula: see text]), and reduced surface redshifts ([Formula: see text]), while preserving full dynamical stability against radial perturbations. The model predicts fundamental oscillation modes at [Formula: see text] kHz with well-defined overtone structures, placing these signatures within detectable ranges for current and next-generation gravitational-wave observatories and providing testable discriminants from conventional neutron stars. The ICG framework establishes a new theoretical bridge between conventional neutron stars and horizonless compact objects, offering a geometrically derived, fully analytical alternative to numerical equation-of-state approaches for strong-field gravity research.

To address challenges in small target detection, complex backgrounds, and lightweight deployment for high-voltage line insulators, this paper proposes a Mobile Lightweight Insulator Defect Detection (MLIDD) system with a novel SPM-YOLO algorithm. Key innovations include: 1) an SPD-Conv module replacing traditional downsampling to preserve fine-grained details of small targets; 2) a PPA attention module enhancing spatial perception via multi-branch feature extraction and adaptive fusion; 3) integration of the MobileNet V4 backbone to reduce computational complexity. Experiments show SOM-YOLO improves detection accuracy and robustness while enabling efficient model compression. The system was successfully deployed on the Jetson Orin NX platform, achieving real-time, highprecision defect detection in UAV inspections.

ABSTRACT Holographic near‐eye 3D display can achieve revolutionary applications in various fields such as communication, medical treatment, and industrial operation, and it is the frontier hotspot of international research. To realize the immersive viewing experience, a large‐sized and wide‐viewing‐angle holographic near‐eye 3D display performance is indispensable. However, the existing equipment does not have this characteristic. Here, a holographic near‐eye 3D display system is proposed, which uses the unique combination of liquid crystal (LC) holographic lens and holographic tilt mapping algorithm to overcome the above challenges. The key innovation of the system is a customized polarized LC holographic lens to support large‐size and wide‐viewing‐angle holographic reproduction by performing corresponding real‐time reflective diffraction modulation on different polarized light. This is designed together with the holographic tilt mapping algorithm, which combines two display modes: large‐size and wide‐viewing‐angle. The co‐design of our unique system and algorithm has doubled the size and viewing angle, which represents an important progress in holographic near‐eye display. The proposed system is very simple in structure and easy to integrate, which is expected to promote the development of holographic near‐eye display.