Accurately predicting the response of concrete and steel structures to blast loading remains challenging due to the complex interplay of numerous parameters. Conventional approaches experimental tests, analytical methods, and numerical simulations suffer from high costs, safety constraints, oversimplified assumptions, or excessive computational demands. This paper presents a critical review of recent advances in theoretical, numerical, and machine learning (ML) methods. A systematic literature search following PRISMA guidelines was conducted, covering 386 studies published. Analytical methods, particularly single-degree-of-freedom models and pressure-impulse diagrams, are evaluated for their efficiency and fundamental limitations they offer rapid predictions (seconds to minutes) but cannot capture local failures and exhibit displacement errors of 8–25
Expansive soils represent a major challenge to civil engineering infrastructure because of their volumetric instability and poor engineering characteristics. This review is based on a systematic discussion of the way fly ash improves engineering properties of expansive soil. This study focuses on the fundamental behaviour of expansive soils in terms of their mineralogy, swelling phenomenon and engineering characteristics. The review summarises the contributions of studies on the improvements of engineering properties of significance such as Atterberg limits, compaction characteristics, strength parameters, volume change behaviour and hydraulic properties. The quantitative synthesis of data published in the literature indicates that optimum fly ash contents by weight typically range from 15 to 25
Large Language Models (LLMs) have emerged as a transformative technology in artificial intelligence, significantly advancing natural language understanding, generation, and reasoning capabilities. This survey reviews the evolution of language models from early statistical approaches to modern Transformer-based architectures and summarizes key developments, including attention mechanisms, scaling laws, alignment techniques, and efficient inference methods. The paper further explores the growing impact of LLMs on everyday life and a wide range of application domains, including healthcare, finance, education, agriculture, marketing, software engineering, and scientific research. To provide a systematic perspective, LLM applications are categorized according to their maturity level and integration across major artificial intelligence subfields, such as natural language processing, multimodal learning, intelligent decision support, autonomous agents, and knowledge-based systems. The survey highlights how these models enhance automation, data-driven decision-making, personalized services, and human–AI interaction across both consumer and industrial environments. Despite their remarkable capabilities, LLMs face several critical challenges, including high computational costs, limited interpretability, hallucinations, privacy and security risks, ethical concerns, and environmental sustainability issues. Existing mitigation approaches and recent advancements are reviewed to assess their effectiveness and limitations. Finally, the paper outlines key future research directions, including trustworthy and explainable AI, efficient model architectures, domain-specific adaptation, multimodal intelligence, and human-centered alignment. This survey provides a comprehensive overview of the current landscape, challenges, and future prospects of LLMs, serving as a valuable reference for researchers and practitioners.
The rising growth of digital media forces digital images as a significant medium for information exchange, indicating the need for stronger security over open networks. However, conventional cryptographic algorithms designed to encode text data are not optimal when applied to images due to high redundancy, strong dependency of pixels, and big files. Although various algorithms are being researched in the domain of cryptography, further investigation is required to analyze the shift from deterministic ciphers to modern AI-based methods and evaluate new deep learning and biology-inspired encryption structures. This review examines current advances in encryption techniques used to secure image transfer over open networks. In particular, it presents an overview of six classes of image encoding algorithms: classical cryptography, chaos-based, biology-inspired schemes, deep learning, transformer-based and hybrid encryption solutions. The paper focuses on high-quality sources published in the last five years and assesses the strengths and limitations of their structures in terms of robustness, key security, and differential resistance. Additionally, the review discusses performance metrics, the balance between speed and complexity, and appropriate use cases for each technique, providing recommendations on possible changes to enhance encryption processes and optimize performance. Overall, the findings indicate that future development in image encryption should focus on time-efficient processing of image data and domain-specific methods. It suggests focusing on lightweight architectures (e.g., State Space Models) and implementing hybrid encryption based on classical ciphers combined with AI-driven feature extraction.
Solving partial differential equations numerically has long been the province of mesh-based methods, yet the last several years have seen Physics-Informed Neural Networks (pinns) emerge as a genuinely different kind of alternative, one that encodes the governing equations directly into the loss function of a neural network and, in doing so, sidesteps the mesh-generation bottleneck that consumes so much engineering time in classical workflows. The appeal is real, but so are the limitations, and the gap between what pinns promise and what they reliably deliver in practice remains, in our view, wider than much of the current literature acknowledges. This review attempts to close that gap, or at least to map it honestly. We cast the pinn training problem in the language of numerical analysis, formalizing the composite residual functional, tracing training dynamics through the neural tangent kernel, and deriving Sobolev-norm stability bounds that connect what the optimizer minimizes to what the engineer actually needs. Along the way, we identify five recurring failure modes: spectral bias, gradient pathology, landscape stiffness, causality violation, and breakdown near discontinuities, and for each, we provide both a quantitative diagnostic and a remediation validated in the literature. Head-to-head benchmarks against finite element, finite volume, and boundary element solvers show that high-order classical methods routinely outperform pinns on smooth problems by several orders of magnitude; yet pinns hold clear advantages in inverse identification, high-dimensional settings, and geometrically evolving domains. We survey the state of the art in fracture mechanics, multi-physics coupling, fluid-structure interaction, and uncertainty quantification, and argue that hybrid fem–pinn architectures currently offer the most credible path toward industrial use. Unresolved theoretical and computational barriers are identified throughout.
Medical image segmentation underpins diagnosis, treatment planning, and disease monitoring, yet the deep learning models that achieve the highest accuracy are typically too large and computationally demanding for the clinical workstations, mobile diagnostic devices, and point-of-care systems on which they must ultimately run. Closing this gap between research-grade accuracy and clinical hardware constraints has become a central challenge for the field. In this survey, we present a unified, efficiency-focused review of deep learning methods for medical image segmentation, organising over 150 works (2015–2026) within a single framework that spans convolutional, Transformer, and Mamba/state-space architectures together with the compression techniques that render them deployable. To the best of our knowledge, no prior survey jointly covers all three architectural families alongside model compression and controlled empirical benchmarking. We structure the field through a four-pillar taxonomy: six families of efficient architectural strategy; model compression and deployment (knowledge distillation, pruning, quantization, and hardware-aware neural architecture search, with tiered guidelines for edge, workstation, and cloud); domain challenges, clinical loss formulations, and data-efficient learning; and a standardised evaluation framework of segmentation metrics, benchmark datasets, and statistical validation. To ground the review in direct evidence, we further conduct a controlled cross-modality benchmark in which representative architectures from all six families are retrained from scratch under a single protocol across dermoscopic and endoscopic polyp datasets, with boundary-quality and deployment-cost profiling. The experiments expose two trade-offs that single-dataset, overlap-only evaluation conceals: compact encoder–decoder models remain competitive with far larger architectures on well-standardised binary tasks, while channel-attention designs that excel on large datasets can degrade sharply when training data is scarce a data-efficiency failure invisible to parameter and computation budgets. These observations motivate tiered deployment guidelines and five open problems: boundary-preserving compression, domain generalisation, clinical interpretability, data-efficient learning for rare pathologies, and real-time volumetric inference. We additionally release a GitHub project page compiling key resources for efficient deep learning in medical image segmentation: https://github.com/razanharith/efficient-medseg .
Agriculture is undergoing rapid digital transformation to address climate change, resource scarcity, and rising food demand. Digital convergence enables real-time monitoring, predictive analytics, automation, and traceability, enhancing productivity, sustainability, and system resilience. This structured and critical literature review synthesizes recent advancements and proposes an integrative framework encompassing six core technological domains: Sensors and IoT, Geographic Information Systems (GIS), Satellite Imagery and Drones, Artificial Intelligence and Machine Learning, Robotics and Automation, and Blockchain. Beyond classification, the study analyzes cross-domain interdependencies and identifies convergence mechanisms that support precision agriculture, automation, and supply-chain transparency. A central contribution is a multi-layered conceptual model organized into five hierarchical levels: (1) infrastructure (sensing, connectivity, data acquisition), (2) geospatial contextualization, (3) intelligence (AI-driven analytics), (4) physical execution (robotic and automated systems), and (5) governance and trust (blockchain-enabled traceability). This layered architecture formalizes digital convergence as a coordinated cyber-physical ecosystem, advancing the conceptual evolution from Agriculture 4.0 toward an integrative Agriculture 5.0 paradigm capable of adaptive, real-time decision-making.
Underwater optical imaging is fundamentally limited by strong absorption and scattering, which degrade image contrast and restrict detection range. Ghost imaging (GI), which reconstructs objects through second-order intensity correlation, has emerged as a powerful alternative capable of maintaining image quality in turbid and turbulent environments where conventional imaging fails. This review systematically examines recent progress in underwater GI, focusing on two primary strategies for enhancing its performance. The first one is speckle engineering, including orthogonal patterns, laser mode modulation, and pseudo-Bessel-ring beams. The second one is propagation-model-based reconstruction, ranging from physical degradation compensation to deep learning and self-supervised frameworks. Despite significant advances, challenges remain in imaging speed, range-fidelity trade-offs, and generalization to diverse underwater conditions. Future trends point toward hardware acceleration, physics-informed learning, multi-modal integration with sonar and polarization sensing, and applications in deep-sea exploration, infrastructure inspection, and covert surveillance. As real-time capabilities and system robustness develop, underwater GI is poised to become a cornerstone technology for next-generation oceanic exploration.
Advances in upper-limb prosthetic control increasingly rely on artificial intelligence (AI) methods, particularly machine learning (ML) and deep learning (DL), for decoding motor intent from biosignals. This review presents a structured synthesis of approaches based on surface electromyography (sEMG), electroencephalography (EEG), and hybrid multimodal systems. The evolution of learning paradigms is analysed from classical ML methods to modern DL architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), Transformers, and graph-based models, with emphasis on their ability to capture spatial, temporal, and non-stationary characteristics of biosignals. Signal acquisition and pre-processing pipelines are examined in relation to electrode configurations, noise suppression, and their impact on robustness and generalization. Multimodal strategies are systematically categorized into early, late, and hybrid fusion frameworks, and evaluated in terms of accuracy, noise resilience, and computational trade-offs. A key observation is that EMG-EEG integration exploits complementary neural and muscular information, resulting in improved stability under signal variability. The review further consolidates performance of AI models across standard benchmark datasets and identifies critical limitations in current evaluation practices, including dataset heterogeneity, absence of unified cross-modal benchmarks, and inconsistent reporting protocols, which hinder reproducibility and fair comparison. System-level considerations, including edge deployment, real-time inference, and hardware integration, are analysed alongside explainability, regulatory, and ethical requirements for clinically viable systems. The paper highlights open challenges and outlines directions for developing scalable and generalizable multimodal ML/DL frameworks for biosignal-driven prosthetic control.
As Artificial Intelligence (AI) technologies rapidly advance, the application bands of microwave filters have expanded, and their deployment scenarios have become increasingly diversified. Consequently, the exponential complexity of design parameters has rendered traditional design methodologies inadequate for high-efficiency iteration and stringent electromagnetic compatibility (EMC) requirements. Thus, AI-driven Computer-Aided Design (CAD) techniques have emerged as a critical research frontier. This paper presents a comprehensive state-of-the-art review of Machine Learning (ML) and Deep Learning (DL) applications in microwave filter design. We first systematically categorize microwave filters and elaborate on their critical performance metrics across both frequency and time domains. Subsequently, we trace the evolutionary trajectories of ML and DL algorithms, establishing a foundational framework to map the development of intelligent CAD techniques in this domain. The core of this review critically categorizes and analyzes existing research based on distinct neural network architectures, alongside hybrid approaches that integrate classical optimization algorithms with AI. Finally, in the context of the emerging era of large-scale AI models, we summarize current paradigms in filter optimization, fault diagnosis, and innovative design, while outlining future perspectives and challenges to drive transformative breakthroughs in this interdisciplinary field.
Molecular dynamics simulations provide a microscopic description of atomic motion and have become a central tool for studying chemical and physical processes. Their broader application, however, remains limited by three fundamental challenges. These include the need for accurate interaction potentials, the difficulty of sampling rare events efficiently, and the challenge of extracting mechanistic information from large trajectory data sets. Recent advances in machine learning offer new strategies to address these limitations by learning computationally expensive operators directly from data while incorporating physical constraints such as conservation laws and symmetry. This review examines how machine learning assisted molecular dynamics is advancing two research areas that share a common mathematical framework but operate at very different physical scales namely reactive molecular chemistry and astrophysical systems. In reactive chemical environments, machine learning models learn surrogate force fields capable of describing bond breaking and bond formation, enabling efficient prediction of reaction pathways, species evolution, and free energy landscapes during processes including aggregation, oxidation, dissociation, and transport. Similar concepts are increasingly applied in astrophysical modeling where machine learning approximates complex physical operators that are otherwise computationally prohibitive, including mappings in nonadiabatic dynamics, turbulence closures in simulations of core collapse supernovae, and efficient representations of cosmological and spectroscopic fields. Across these domains the central principle is the approximation of physical operators while preserving the structure of the governing equations. By integrating quantum trained force fields, enhanced sampling guided by learned coordinates, and data driven analysis of trajectories, machine learning augmented molecular dynamics provides a unified computational framework capable of extending simulations across wide spatial and temporal scales. This perspective highlights how the integration of machine learning with molecular dynamics bridges precision and scalability, linking molecular processes such as polycyclic aromatic hydrocarbon chemistry and soot formation with large scale astrophysical phenomena governing the evolution of stars and galaxies.
Air pollution has become one of the most significant environmental and public health challenges worldwide, necessitating accurate prediction systems to support effective air quality management and informed decision-making. Deep learning has emerged as a powerful approach for modelling the complex nonlinear and spatiotemporal relationships that characterise air quality data. This review provides a comprehensive synthesis of recent advances in deep learning techniques for air pollution prediction. Following the PRISMA 2020 guidelines, the literature was systematically reviewed to evaluate deep learning architectures, environmental data sources, model performance, and emerging research trends. The review traces the evolution of deep learning models from Artificial Neural Networks (ANNs) to Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), Transformer-based models, transfer learning, and hybrid architectures. It further examines the integration of ground monitoring stations, satellite remote sensing, meteorological observations, IoT-based sensors, and multi-source data fusion for improving prediction accuracy and model robustness. Comparative analysis indicates that hybrid and Transformer-based models consistently achieve superior predictive performance, while RMSE, MAE, and R² remain the most widely adopted evaluation metrics. The review also identifies key research challenges, including data scarcity, model interpretability, computational complexity, and benchmark standardisation, and highlights emerging directions such as Explainable Artificial Intelligence, Graph Neural Networks, Physics-Informed Neural Networks, Federated Learning, and Edge AI for developing accurate, scalable, and intelligent air quality prediction systems.
Physics-informed digital twins in computational engineering require more than a simulation model or a machine-learning surrogate. They connect a physical asset to a computational representation, learn from data, quantify uncertainty, update under sensor information, and support engineering decisions under explicit latency and credibility constraints. This scoping review synthesizes six research questions on physics-informed learning, neural operators, uncertainty quantification, data assimilation, application-domain maturity, and evidence-informed method selection. Seventy-nine database-identified scholarly reports inform the synthesis; the RQ-specific counts are 15, 14, 13, 26, 27, and 17, with overlap permitted across questions. Each headline claim is appraised through test-set construction, baseline adequacy, in- versus out-of-distribution evaluation, uncertainty and failure analysis, validation provenance, and complete sensing-to-decision latency. Numerical errors, speedups, and update times are interpreted within their study-specific evaluation settings rather than as universal selection thresholds. The evidence indicates that physics-informed neural networks can support bounded inverse, surrogate, and sparse-data tasks, but their maturity is constrained by optimization pathologies, boundary enforcement, numerical baselines, and distribution shift. Neural operators provide moderate evidence of acceleration for repeated partial differential equation prediction, but only limited evidence as complete digital-twin backbones because offline cost, geometry and regime transfer, uncertainty calibration, and operational synchronization are inconsistently reported. Uncertainty quantification and data assimilation are essential credibility and synchronization layers, yet calibration non-identifiability, misuse of information-based covariance bounds, ensemble miscalibration, filter divergence, observability limits, and incomplete end-to-end validation remain common. A non-compensatory evidence-gate procedure converts the critical synthesis into method-selection guidance while preserving negative, conflicting, and unreported evidence.
Future prediction in visual systems is shifting from pixel-level frame generation toward latent-space forecasting, agent-centric reasoning, and planning-oriented decision support. This review summarizes this transition and analyzes how pretrained visual representations, temporal modules, graph-based interaction models, and language-conditioned planning reshape both modeling assumptions and evaluation protocols. The discussion first covers the move from direct video prediction to forecasting over compact semantic representations produced by frozen or adapted vision backbones. It then reviews recurrent, convolutional, transformer, state-space, diffusion, and world-model predictors for latent dynamics, followed by common evaluation pitfalls, including persistence shortcuts, static-scene dominance, metric mismatch, and insufficient horizon testing. The review further connects latent forecasting with multi-agent trajectory prediction, future-aware interaction reasoning, and instruction-conditioned inspection planning. The literature and case studies indicate that high feature similarity or low displacement error does not necessarily imply useful future understanding. Robust evaluation should therefore include simple baselines, horizon-wise metrics, dynamic-region or agent-level tests, multimodal uncertainty, and downstream decision performance.
Uncertainty is an intrinsic feature of real-world decision environments, manifesting as vagueness, imprecision, and ambiguity that classical probability theory cannot adequately model. Fuzzy set theory provides a rigorous mathematical framework for representing graded membership and linguistic uncertainty and has subsequently evolved into a rich family of generalizations including intuitionistic, type-2, neutrosophic, hesitant, Pythagorean, Fermetean, T-spherical, picture and q-rung orthopair fuzzy sets. This paper presents a comprehensive bibliometric and systematic review, derived from fully screened Web of Science corpus of more than 3200 publications following a PRISMA 2020-compliant search, screening and extraction protocol, of fuzzy set theory and its real-life applications from 2000 to 2024. Through analysis of publication trends, collaborative networks, productive authors, institutions and countries and keyword clustering, we map the evolution of the field across three phases: foundational consolidation (2000–2009), rapid theoretical expansion (2010–2017) and interdisciplinary application explosion (2018–2024). The review identifies primary application domains- control systems, multi-criteria decision-making, medical diagnosis, supply chain management, environmental assessment and finance, robotics and autonomous systems, smart cities and industry 4.0 and examines the most impactful methods and representative studies within each. Three cutting-edge application frontiers are highlighted: the integration of fuzzy logic with deep learning, smart city infrastructure management and quantum-inspired fuzzy computing. The paper also provides quantitative comparison of fuzzy set theory with related uncertainty frameworks, synthesizing key findings and identifies promising future research directions. This review serves as a comprehensive reference for researchers and practitioners seeking to harness fuzzy methodologies for complex real-world uncertainty problems.
Digital twins are increasingly used as tools for managing engineering systems through continuous synchronization between physical assets and their virtual counterparts, enabling improved monitoring, prediction and control throughout the system lifecycle. Maintaining accurate alignment requires reliable state estimation from sparse sensor measurements along with adaptive calibration of model parameters to account for degradation, environmental variability and operational uncertainties. Physics informed neural networks address these requirements by embedding governing equations with observational data within a unified optimization framework that supports simultaneous state reconstruction and parameter identification while preserving physical consistency. This review examines the current state of the art in PINN based digital twin synchronization and model updating, presents the mathematical foundations that connect PINNs to classical collocation methods and describes how automatic differentiation enables efficient enforcement of physical constraints. Applications spanning manufacturing, aerospace, energy and infrastructure demonstrate that PINN based digital twins achieve high diagnostic accuracy, low inference latency and significant reductions in operational cost. Key challenges that remain include parameter identifiability, scalability limitations arising from the curse of dimensionality, real time deployment constraints, gaps in validation methodology and robustness to distribution shift.
Oil and gas pipelines are the most economical and reliable means of long-distance fluid transport worldwide. However, these systems are subject to various forms of degradation that compromise their structural integrity and reduce service life. To address these challenges, composite repair technologies have been extensively developed and applied for pipeline rehabilitation. While the effectiveness of these composite repair systems has been demonstrated through extensive hydrostatic testing and has gained industry acceptance, further improvements in design methodology are essential. This paper comprehensively reviews the analytical approaches used to evaluate composite repaired pipelines, including experimental testing, numerical simulations, and theoretical modeling. The advantages and limitations of each method are systematically summarized, compared, and critically reviewed. Furthermore, this review highlights potential research methods and identifies key challenges and emerging directions in composite pipeline rehabilitation. This review highlights specific data that is needed in the computation of important information for composite-repaired pipeline, such as minimum composite repair thickness, failure pressure etc. The goal is to offer valuable insights into the current state and future developments of composite repair strategies, paving the way for future research.
The Gazelle Optimization Algorithm (GOA) is a swarm-based metaheuristic inspired by the agile and adaptive movement of gazelles, designed to solve complex optimization problems. GOA has attracted increasing attention in science and engineering owing to its reported simplicity and promising performance across various optimization applications. Its promising performance has led to the development of numerous variants and enhancement strategies, including improved, hybrid, and multi-objective models. This paper presents a comprehensive survey of GOA, focusing on its variants, enhancement mechanisms, and applications across engineering and scientific domains. A Systematic Literature Review (SLR) methodology is employed to ensure a structured and transparent analysis. A total of 73 primary studies are systematically analyzed and classified based on algorithmic modifications, hybridization techniques, and application areas. The results show that GOA has been widely applied in seven major domains, particularly in energy systems, machine learning, and engineering optimization. Among the existing studies, enhancement-based variants dominate the literature, followed by hybrid approaches and multi-objective extensions. However, GOA still faces several challenges, including premature convergence, limited theoretical analysis, and scalability issues. This survey provides a structured overview of GOA developments and highlights future research directions for advancing robust and scalable optimization methods.
Artificial intelligence (AI) technology is rapidly evolving in healthcare by optimizing the analysis of medical images and sensor data. Transfer Learning (TL) is now playing a vital role in this evolving field of healthcare through its applications in artificial intelligence systems, ensuring high performance even in situations where there is limited, heterogeneous, and variable data sets collected in different environments. In this review, we discuss the basic principles and architectures of artificial intelligence-driven transfer learning and its applications in various artificial intelligence systems, including recent developments in various forms of deep learning techniques such as convolutional learning, recurrent learning, transformer learning, multimodal learning, and foundation learning, and how these techniques are benefiting from transfer learning in optimizing images and sensor data in healthcare. Despite its advantages, transfer learning also presents some challenges in this field of healthcare. In conclusion, transfer learning combined with recent advancements in images and sensor technology presents a promising approach in artificial intelligence systems in optimizing images and sensor data in healthcare.
