Abstract Plant-parasitic nematodes are an important threat to turfgrass. Left unmanaged, they can cause serious reductions in the quality and playability of golf courses and sports fields. Effective nematode management depends on accurate identification of the nematode genera extracted from soil samples. However, this process requires specialized expertise in nematology, which is often limited in plant diagnostic laboratories. Recent advancements in deep learning models offer promising solutions for the future of nematode identification. In this study, we evaluated the performance of EfficientNet V2-S, MobileNetV3-L, ResNet101, and Swin Transformer V2-B convolutional neural network model architectures in the classification of seven nematode taxa associated with turfgrass. Models were trained using a dataset of 5406 plant-parasitic nematode images where the dataset was split into 70, 15, and 15% for training, testing, and validation, respectively. Data augmentation and hyperparameter optimization using a combined Bayesian optimization and Hyperband algorithm (BOHB) approach were used to improve the model performance. Balanced classification accuracy on the test set was highest for EfficientNet V2-S and Swin Transformer V2-B at 94.63% and 94.34%, respectively. MobileNetV3-L and ResNet101 had lower balanced accuracies of 90.83% and 86.33%, respectively. Testing the models on an additional dataset using a user-end platform indicated the superiority of EfficientNet V2-S to other models with 82.47% accuracy. The findings of this study indicate the potential application of deep learning tools for accurate nematode identification to aid in diagnostics.

The integration of artificial intelligence (AI) with optical fiber sensing (OFS) is transforming the capabilities of modern sensing systems, enabling smarter, more adaptive, and higher-performance solutions across diverse applications. This paper presents a comprehensive review of AI-enhanced OFS technologies, encompassing both localized sensors such as fiber Bragg gratings (FBG), Fabry–Perot (FP) interferometers, and Mach–Zehnder interferometers (MZI), and distributed sensing systems based on Rayleigh, Brillouin, and Raman scattering. A wide range of AI algorithms are discussed, including supervised learning, unsupervised learning, reinforcement learning, and deep neural architectures. The applications of AI in OFS were discussed. AI has been employed to enhance sensor design, optimize interrogation systems, and adaptively tune configurations, as well as to interpret complex sensor outputs for tasks like denoising, classification, event detection, and failure forecasting.

In recent years, automatic text summarization has become a vital area of research due to its role in improving access to and understanding of vast information across domains. The rise of social media has intensified the need for summarization tools capable of handling user-generated content such as posts, comments, and discussions. Unlike structured texts, social media content is often informal, fragmented, context-dependent, and noisy. It frequently includes slang, abbreviations, emojis, and diverse writing styles, posing unique challenges for traditional summarization methods. While conventional approaches perform well on formal text, they often struggle to capture the nuances of online discourse. This highlights the need for specialized models that can generate coherent and context-aware summaries tailored to the characteristics of social media language. Recent advances in neural architectures, particularly Transformer-based sequence-to-sequence models, have shown promise in overcoming these challenges. These models excel at capturing long-range dependencies and contextual relationships, making them well-suited for summarizing dynamic and unstructured inputs. Despite technical progress, evaluating the quality of summaries remains difficult. Standard metrics like ROUGE may not fully reflect subjective qualities such as fluency, coherence, and semantic fidelity, which are essential for human-like summarization. This paper introduces a Transformer-based summarization system designed specifically for social media comments related to topical posts. We benchmark its performance against models like ChatGPT, assessing outputs across multiple linguistic and semantic dimensions. By combining both traditional and advanced evaluation metrics, our work provides a more holistic view of summarization quality and identifies key areas for future improvement.

Purpose This paper aims to address this rail safety issue through collection and analysis of video data at grade crossings using artificial intelligence. Railroad safety is vital to transportation and the economy, particularly in the USA, where highway-rail grade crossings are hotspots of accidents, 9% of which involve vehicles stopping on tracks. Design/methodology/approach This research develops a robust artificial intelligence (AI)-assisted system to detect stopped-on-tracks incidents by overcoming unique challenges in the scene of grade crossings via neural architecture search and dynamic trajectory filtering. Findings The system achieves 94% precision in stopped-on-tracks detection and 95% recall in traffic counting. The findings suggested that road markings in the first case study facilitate incident reduction up to 50%, while a targeted intervention could potentially decrease stopped-on-tracks incidents by 80% in the second case study. Social implications This AI-based approach will potentially enhance rail safety and enable more informed decision-making in rail infrastructure management by providing insight on stopped-on-tracks behavior. Originality/value The model’s accuracy was evaluated using the test set. Its effectiveness was further validated through two case studies: one assessing the impact of road markings (e.g. dynamic envelopes) on reducing stopped-on-tracks incidents and another examining the correlation between different types of traffic congestion and these incidents.

With a focus on applications related to digital identities, this paper provides an extensive overview of multimodal deep hashing biometric authentication systems. We lay out precise research goals and examine the most recent approaches, such as privacy-preserving strategies and deep neural architectures. Modern multimodal hashing frameworks are identified, template security and system interoperability issues are evaluated, and future research directions are recommended. We employ a systematic literature search with clear inclusion/exclusion criteria and categorize the works by technique (e.g., CNN, RNN, Transformer), application domain, and modality (e.g., face, fingerprint, iris). We discuss recent developments, including transformer-based biometric models [2][3] and privacy techniques (secure sketches, homomorphic encryption) [4][5]. Key studies are compiled in a standardized comparative table. With an emphasis on open-source platforms (like MOSIP [6][7]), privacy-by-design, and economic effects, we cover policy frameworks (GDPR, eIDAS, and African Union privacy charters) and provide helpful suggestions for implementing digital ID systems in Africa. Future studies and the implementation of safe, privacy-conscious biometrics for identity programs are intended to be guided by our findings.

The increasing integration of the Internet of Things (IoT) in healthcare has led to massive, high-dimensional data streams, demanding advanced, adaptive learning models for timely and accurate clinical predictions, especially in sensitive domains such as palliative care. Existing models often suffer from high computational overhead, sensitivity to learning rate variations, and difficulties in handling non-stationary data, which impedes their ability to deliver accurate and prompt predictions in high-risk medical scenarios. To address these limitations, this study proposes a Hybrid Metaheuristic-Driven Deep Neural Architectures (HMDNA) that combines a Deep Neural Network (DNN) with Cuckoo Search Optimization (CSO) for sepsis detection and prognosis. The methodology follows a structured pipeline encompassing data preprocessing, model training, and optimization. Time-series ICU data were preprocessed using k-NN imputation and min-max scaling, followed by DNN training with CSO-based optimization applied at initialization, mid-training, and fine-tuning stages. Implemented using TensorFlow and trained on an NVIDIA Tesla V100 GPU, the model achieved an accuracy of 92.7%, precision of 91.8%, recall of 90.3%, and F1 score of 91.4%. These results significantly outperform baseline models including traditional DNN (85.3% accuracy), DNN + GA (88.5%), DNN + PSO (89.2%), and DNN + ACO (90.1%). The proposed model demonstrated faster convergence, better generalization, and robustness to real-time variability in healthcare data. By combining the strengths of deep learning and metaheuristic optimization, this approach ensures reliable performance in dynamic and unpredictable clinical environments. The study highlights the potential of adaptive, hybrid AI models in enhancing healthcare decision-making, particularly in critical care scenarios where prediction accuracy and model responsiveness are vital for improving patient outcomes.

The long-standing von Neumann architecture, while foundational to modern computing, intrinsically suffers from data-transfer inefficiency, which imposes severe limits on speed and energy efficiency in artificial intelligence, machine learning, and real-time data processing. Inspired by the remarkable energy efficiency of the human brain, neuromorphic computing seeks to emulate neural architectures through hardware capable of adaptive learning. Although early complementary metal-oxide-semiconductor (CMOS)-based neuromorphic systems captured basic synaptic behaviors, their narrow dynamic ranges and high operating voltages impede their application prospects. Van der Waals (vdW) materials, particularly tellurium (Te) and Selenium (Se), have recently emerged as promising platforms for next-generation neuromorphic devices due to their intriguing electronic and optoelectronic properties including considerable carrier mobilities, broadband photoresponse, and strong coupling between electrical, optical, and mechanical stimuli. These unique properties offer a direct physical analogy to biological synapses, thereby enabling neuromorphic computing. This Perspective introduces the fundamentals of synaptic behavior and neuromorphic computing and highlights the distinctive properties of Te/Se nanomaterials for synaptic devices. Then we discuss critical advances in Te/Se-based memristors, heterostructures, and electronic/optoelectronic synaptic transistors. In the end, we conclude this perspective with a discussion on the remaining challenges and future opportunities in this evolving field.

This paper investigates the existence and global exponential stability of almost periodic solutions in a class of impulsive Cohen–Grossberg-type bidirectional associative memory (BAM) neural networks with time-varying delays. Real neural systems often experience sudden perturbations and nonuniform temporal interactions, leading to complex oscillatory behaviors. To capture these effects, a new impulsive Cohen–Grossberg BAM model is developed that integrates both delays and impulsive influences within a unified framework. Using the theory of almost periodic functions, fixed point methods, and impulsive differential inequalities, new sufficient conditions are derived for the existence and stability of almost periodic solutions. A Lyapunov functional combined with a generalized Gronwall-type inequality provides rigorous global exponential stability criteria. Numerical simulations confirm the theoretical analysis. The results extend existing studies and offer new insights into how delay and impulsive factors jointly shape the stability and dynamics of hybrid neural systems, contributing to the design of robust and delay-tolerant neural architectures.

Many network architectures exist for learning on meshes, yet their constructions entail delicate trade-offs between difficulty learning high-frequency features, insufficient receptive field, sensitivity to discretization, and inefficient computational overhead. Drawing from classic local-global approaches in mesh processing, we introduce PoissonNet, a novel neural architecture that overcomes all of these deficiencies by formulating a local-global learning scheme, which uses Poisson's equation as the primary mechanism for feature propagation. Our core network block is simple; we apply learned local feature transformations in the gradient domain of the mesh, then solve a Poisson system to propagate scalar feature updates across the surface globally. Our local-global learning framework preserves the features's full frequency spectrum and provides a truly global receptive field, while remaining agnostic to mesh triangulation. Our construction is efficient, requiring far less compute overhead than comparable methods, which enables scalability—both in the size of our datasets, and the size of individual training samples. These qualities are validated on various experiments where, compared to previous intrinsic architectures, we attain state-of-the-art performance on semantic segmentation and parameterizing highly-detailed animated surfaces. Finally, as a central application of PoissonNet, we show its ability to learn deformations, significantly outperforming state-of-the-art architectures that learn on surfaces. https://github.com/ArmanMaesumi/poissonnet

The Rubik's cube, a widely recognized combinatorial puzzle with an astronomically vast state space, has been the subject of various research experiments with neural networks used as heuristic estimators to navigate the state-space exploration. However, prior efforts have overlooked the intriguing symmetries inherent to this domain. Drawing on geometric deep learning principles, this article introduces a novel neural architecture that explicitly leverages these symmetries, grounded in a rigorous group-theoretical analysis. The design of the proposed symmetry-invariant model is then validated empirically through an innovative universal procedure for detecting model symmetry invariance. Finally, experimental results demonstrate that the symmetry-aware neural architecture exhibits enhanced generalization and problem-solving efficacy compared with the state of the art.

We propose a transformer-based neural operator as a novel parametric arbitrary-scale super-resolution technique for reconstructing three-dimensional (3D) turbulence fields with direct numerical simulation (DNS) resolution from spatially sparse data. This technique consists of three components: a convolutional neural network-based encoder extracting low-resolution flow features as regular-grid contexts, an interpolation-based conditioning module embedding the contexts onto a set of sparsely distributed points, and a Galerkin-type attention decoder mapping the selected point coordinates to their corresponding flow quantity values. Thanks to the neural operator architecture, the super-resolution technique learns a parameter-dependent mesh-invariant mapping from arbitrary points over continuous domains to their corresponding flow values. The model performance was assessed on DNS data of a 3D turbulent channel flow at two Reynolds numbers (Reτ=130 and 350). At the two upsampling factors (×4 and ×8), the proposed models, trained with a few sparse samples, consistently outperform their convolutional neural network-based counterparts, trained using full high-resolution fields. For variable upsampling factors, our model naturally generalizes across scales, reproducing 3D high-resolution flow fields from a range of coarseness levels. Conversely, its convolutional neural network-based counterpart fails this task as it uses fixed-grid image-to-image mappings. In terms of computational efficiency, utilizing sparse data significantly reduces training time while achieving better accuracy when compared to training on full-field high-resolution data. Notably, our model allows for querying arbitrary points or regions, avoiding the need to generate and store entire high-resolution fields, making it feasible for super-resolution tasks where large-scale, full-resolution turbulence fields cannot fit into graphics processing unit memory.

Adaptive fuzzy fractional-order terminal sliding mode control using adaptive neural network observation architecture with application to ultra-precision stages

Accurate surface geometry representation is crucial in 3D visual computing. Explicit representations, such as polygonal meshes, and implicit representations, like signed distance functions, each have distinct advantages, making efficient conversions between them increasingly important. Conventional surface extraction methods for implicit representations, such as the widely used Marching Cubes algorithm, rely on spatial decomposition and sampling, leading to inaccuracies due to fixed and limited resolution. We introduce a novel approach for analytically extracting surfaces from neural implicit functions. Our method operates natively in parallel and can navigate large neural architectures. By leveraging the fact that each neuron partitions the domain, we develop a depth-first traversal strategy to efficiently track the encoded surface. The resulting meshes faithfully capture the full geometric information from the network without ad-hoc spatial discretization, achieving unprecedented accuracy across diverse shapes and network architectures while maintaining competitive speed.

Dose equivalent rate forecasting: A comparison of time series methods and machine learning approaches.

Beyond the brain: a computational MRI-derived neurophysiological framework for robotic conscious capacity.

An automated method for neural architecture design in lithium-ion battery state of health estimation

A continuous encoding-based representation for efficient multi-fidelity multi-objective neural architecture search

ZEP-NAS: Enabling green-aware model design via zero-cost emission proxy in neural architecture search

Advancements in deep learning have revolutionized materials discovery by enabling predictive modeling of complex material properties. However, single-modal approaches often fail to capture the intricate interplay of compositional, structural, and morphological characteristics. This study introduces a novel multimodal deep learning framework for enhanced material property prediction, integrating textual (chemical compositions), tabular (structural descriptors), and image-based (2D crystal structure visualizations) modalities. Utilizing the Alexandriadatabase, we construct a comprehensive multimodal dataset of 10,000 materials with symmetry-resolved crystallographic data. Specialized neural architectures, such as FT-Transformer for tabular data, Hugging Face Electra-based model for text, and TIMM-based MetaFormer for images, generate modality-specific embeddings, fused through a hybrid strategy into a unified latent space. The framework predicts seven critical material properties, including electronic (band gap, density of states), thermodynamic (formation energy, energy above hull, total energy), magnetic (magnetic moment per volume), and volumetric (volume per atom) features, many governed by crystallographic symmetry. Experimental results demonstrated that multimodal fusion significantly outperforms unimodal baselines. Notably, the bimodal integration of image and text data showed significant gains, reducing the Mean Absolute Error for band gap by approximately 22.7% and for volume per atom by 22.4% compared to the average unimodal models. This combination also achieved a 28.4% reduction in Root Mean Squared Error for formation energy. The full trimodal model (tabular + images + text) yielded competitive, and in several cases the lowest, error metrics, particularly for band gap, magnetic moment per volume and density of states per atom, confirming the value of integrating all three modalities. This scalable, modular framework advances materials informatics, offering a powerful tool for data-driven materials discovery and design.

Chemical Exchange Saturation Transfer (CEST) MRI has demonstrated its remarkable ability to enhance the detection of macromolecules and metabolites with low concentrations. While CEST mapping is essential for quantifying molecular information, conventional methods face critical limitations: model-based approaches are constrained by limited sensitivity and robustness depending heavily on parameter setups, while data-driven deep learning methods lack generalizability across heterogeneous datasets and acquisition protocols. To overcome these challenges, we propose a Lorentzian-model Informed Neural Representation (LINR) framework for high-quality CEST mapping. LINR employs a self-supervised neural architecture embedding the Lorentzian equation - the fundamental biophysical model of CEST signal evolution - to directly reconstruct high-sensitivity parameter maps from raw z-spectra, eliminating dependency on labeled training data. Convergence of the self-supervised training strategy is guaranteed theoretically, ensuring LINR's mathematical validity. The superior performance of LINR in capturing CEST contrasts is revealed through comprehensive evaluations based on synthetic phantoms and in-vivo experiments (including tumor and Alzheimer's disease models). The intuitive parameter-free design enables adaptive integration into diverse CEST imaging workflows, positioning LINR as a versatile tool for non-invasive molecular diagnostics and pathophysiological discovery.