Search Results (7,009)

Russia’s invasion of Ukraine has posed serious risks to global food security, by causing substantial crop yield losses since 2022. Accurate yield estimation helps policymakers to plan compensation, yet modelling yields in conflict regions remains challenging due to significant non-meteorological disruptions. This study proposes a novel framework to quantify war-related crop yield losses by comparing estimations derived from meteorological data, representing weather-driven yield variability, with those based on Earth observation (EO) data, reflecting actual crop conditions influenced by both weather and conflict. Thus, meteorologically based yield estimates are expected to exceed those derived from EO data, with the difference indicating war-related losses. Both, meteorological- and EO-based models, are developed using transfer learning (TL) to estimate yields of maize, winter wheat, and spring barley. Models are initially trained on EU country data and subsequently finetuned with Ukrainian data. Their performance is compared to two non-TL approaches: Extreme Gradient Boosting (XGB) and Artificial Neural Network (ANN) to test their reliability. Results show crop yield losses for maize; however, since we do not detect losses in the other crops, we conclude that simply comparing meteorological- and EO-based models proves insufficient to fully isolate conflict effects due to strong interactions of EO and meteorological data. Nevertheless, TL substantially enhances prediction accuracy (R² around 0.7), exceeding alternative models by 0.05–0.2 across crops. These findings demonstrate the value of TL for yield modelling in data-scarce environments and underscore the need for improved methodologies to quantify conflict-induced agricultural losses. Full article

Germinated Andean lupin whole flour (GAL) is rich in protein, dietary fiber, essential amino acids, and bioactive compounds, representing a promising alternative for the development of sustainable plant-based foods. This study investigated the feasibility of partially replacing soy protein isolate (SPI) with GAL in high-moisture extruded meat analogues. A central composite design was applied to evaluate the effects of the GAL ratio (0:100–50:50) and feed moisture content (50–70%) on the technological properties of the extrudates. The Response Surface Methodology was used to model and optimize the process. The incorporation of GAL significantly affected the (p < 0.10) water solubility index (WSI), oil absorption capacity (OAC), cooking loss (CL), yellowness (b*), cohesiveness, and adhesiveness, generating predictive models with satisfactory goodness-of-fit (R² > 0.75). Increasing GAL levels increased the WSI from 6.21 to 18.79% and cooking loss from 1.01 to 3.84%, while reducing OAC from 239.42 to 166.57%, indicating substantial modifications in matrix organization and hydration behavior. Numerical optimization identified an optimal formulation containing 12% GAL, 88% SPI, and 65.5% feed moisture, with a desirability of 77.82%. Model validation showed relative deviations lower than 10% between predicted and experimental values. The optimized meat analogue exhibited high protein content (84.44%), favorable techno-functional properties, and improved nutritional quality, with higher levels of branched-chain amino acids (17.34 g·100 g⁻¹ protein), essential amino acids (33.32 g·100 g⁻¹ protein), and in vitro protein digestibility (90.1%) compared with the control formulation. Multivariate analyses confirmed that phenylalanine, histidine, methionine, and leucine were the main variables that discriminated between the protein sources and the extruded products. Overall, GAL demonstrated strong potential as a sustainable functional ingredient to produce nutritionally enhanced high-moisture meat analogues. Full article

Background: Postoperative hemoglobin (Hb) reduction reflects the physiologic extent of perioperative blood loss after total knee arthroplasty (TKA). Whereas previous studies have relied on transfusion as a binary endpoint, transfusion decisions are highly variable across institutions, obscuring the underlying hematologic trajectory. This study aimed to develop and interpret machine learning (ML) models to characterize and quantify the determinants of postoperative Hb reduction in a standardized cohort of non-transfused TKA patients. Methods: A retrospective cohort of 866 patients who underwent primary TKA under a standardized operative protocol—with identical cemented posterior-stabilized implants and uniform cementing technique—was analyzed (1 January 2014–31 March 2024). During the study period, a consistent 1 g intra-articular tranexamic acid (TXA) regimen administered through the drain was introduced and applied to a subset of patients, allowing TXA use to be modeled as a binary predictor. Four ML algorithms (Linear Regression, Random Forest, XGBoost, and Stacking Regressor) were trained using preoperative, demographic, and perioperative variables. Fivefold cross-validation assessed model performance, and SHapley Additive exPlanations (SHAP) values were used to identify influential predictors and enhance interpretability. Results: Across all ML models, preoperative Hb emerged as the strongest determinant of postoperative Hb reduction, followed by TXA use, body mass index (BMI), and platelet count. Ensemble models captured non-linear and interacting effects more effectively than linear regression. Test-set performance was modest (best R² = 0.330), consistent with the influence of unmeasured physiologic factors such as hidden blood loss, fluid dynamics, and inflammatory responses. Accordingly, the primary value of the framework lies in the exploratory and transparent assessment of determinant importance rather than in individual-level prediction. Conclusions: This study provides an interpretable, exploratory ML framework for identifying factors associated with percentage Hb reduction after TKA. Preoperative Hb was the dominant determinant, while TXA use and higher BMI were recurrently associated with smaller predicted percentage reductions. Given the modest test-set performance and the absence of external validation and clinical utility assessment, the models should not be interpreted as tools for individual-level prediction or clinical decision making. Full article

Expanded Polystyrene (EPS) offers a viable pathway for balancing the fresh and hardened properties of infrastructure with environmental sustainability. This study evaluated six concrete mixes with varying EPS Styrofoam aggregate ratios and three water-to-binder (w/b) ratios, all of which incorporated silica fume and a superplasticizer. Eight machine learning (ML) algorithms (SVMs, GPR, ANNs, etc.) and a deep-learning LSTM model were utilized to preliminarily predict trends in EPS concrete properties. The experimental results indicated that the fine aggregate replacement outperformed the coarse aggregate replacement, retaining approximately 76% of the density of the control mixture, along with other property reductions. The fine aggregate replacement resulted in a compressive strength reduction of up to 46.4%, with losses in tensile strength of 20.9% and an improvement in workability of 3.4%. Finally, various Artificial intelligence (AI) models identified trends in the predictions of EPS properties based on the mixing ratio within a limited experimental dataset. In addition, explainable AI with SHAP analysis, quantifying feature contributions, ensured that the coarse aggregate replacement exerted a more significant negative impact on the mechanical properties and density than the fine aggregate replacement. Although these mixtures offer significant weight reduction, their use in structural applications requires further verification, as the reduction of nearly half of the compressive strength is significant. These findings provide strategies and a framework to facilitate the precise practical application of EPS concrete in nonstructural or lightly loaded applications. Full article

Credit risk assessment is a core component of financial decision-making. This study develops an explainable machine learning framework for modeling loan approval decisions on heterogeneous tabular data, centered on a Cross-Attentional Tabular Transformer that applies bidirectional cross-attention between numerical and categorical feature groups. The prediction target is historical loan-approval status, treated as a proxy for, not a direct measure of, borrower default risk; a supplementary validation on a dataset with an authentic default label is also reported. Class imbalance is addressed through focal loss, and post hoc interpretability is provided through SHAP analysis. Three classifiers, Random Forest, Gradient Boosting, and the proposed transformer, are evaluated on a 5000-sample credit dataset using accuracy, precision, recall, F1-score, ROC-AUC, and average precision. Gradient Boosting achieves the best performance (accuracy 0.9640, F1-score 0.9189), with Random Forest comparable; the proposed transformer reaches 0.9530 accuracy and 0.8949 F1, without surpassing the ensembles and at substantially higher computational cost. A five-split robustness comparison additionally evaluates XGBoost, LightGBM, CatBoost, and calibrated logistic regression: all three Gradient-Boosting variants and both classical ensembles exceed the transformer’s performance on every metric, while calibrated logistic regression does not. The evaluated baseline set excludes deep tabular architectures such as TabNet, FT-Transformer, SAINT, and TabPFN-style methods. Across the three primary classifiers, SHAP identifies credit score, employment status, and income as the dominant features, consistent with domain expectations. The results characterize the observed performance–efficiency trade-off between ensemble methods and attention-based tabular learning under the evaluated data conditions. Full article

Accurate regional demand forecasting supports reliable operation of urban electric vehicle (EV) charging networks. However, public charging demand exhibits spatial heterogeneity, multi-scale periodicity, and uncertainty. This paper proposes Urban-CSTPNet, a multi-relational spatio-temporal probabilistic forecasting framework. Five semantically explicit graphs represent geographical adjacency, spatial distance, historical demand correlation, pricing-pattern similarity, and static regional attributes. Sample-level time-conditioned graph gating fuses these relations using historical demand states and calendar context. Independent recent, daily, and weekly branches capture short-term variation, daily repetition, and weekly regularity, and are combined through temporal gating. The model produces multiple conditional quantiles and applies horizon-specific conformalized quantile regression using an independent calibration set. Experiments on the Shenzhen UrbanEV dataset at 1, 3, 6, 12, and 24 h horizons achieve a mean absolute error (MAE) of 60.44, root mean squared error (RMSE) of 192.44, and Pinball Loss of 16.96. These values are 11.51%, 6.04%, and 9.50% lower than those of ST-MGF-Q. For calibrated 90% intervals, the prediction interval coverage probability (PICP), prediction interval normalized average width (PINAW), and Interval Score are 0.9024, 0.0175, and 349.60. Results confirm improved forecasting accuracy, probabilistic quality, and empirical interval reliability. Full article

Accurate estimation of downhole weight on bit (DWOB) is important for drilling parameter optimization and safe drilling in extended-reach horizontal wells. However, DWOB is difficult to obtain continuously because surface weight on bit (SWOB) is attenuated by drill-string friction, wellbore trajectory, and borehole-wall contact. To address this problem, a physics-informed convolutional neural network combined with bidirectional gated recurrent unit (CNN-BiGRU) prediction model optimized by the whale optimization algorithm (WOA) is proposed using surface sensor data and downhole measurement-while-drilling (MWD) measurements. The model first uses a one-dimensional CNN to extract local features from drilling parameters along the measured-depth direction and then employs a BiGRU to capture depth-series dependencies. Meanwhile, a drill-string frictional attenuation relationship is embedded into the loss function as a physical prior, and WOA is used to optimize network hyperparameters and the physics-constrained weight. Field data from a horizontal well were used for validation. The proposed model achieved a coefficient of determination (R²)of 0.9245 on the test set, with root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE) values of 0.1856, 0.1135, and 0.0124, respectively. The results demonstrate that the proposed data–physics hybrid framework improves the accuracy and physical consistency of DWOB prediction. Full article

Deep learning surrogate models trained with mean-squared-error loss produce statistically accurate but physically unconstrained flood predictions: water may flow uphill, appear spontaneously, or smooth over street-level corridors. In this work, a physics-informed training framework is developed for CNN-LSTM models that predict urban flood depths at 15 min intervals over a $128\times 128$ spatial grid. Three differentiable penalty terms are embedded directly into the loss function: (i) a gravity loss that penalizes depth increases against the water-surface-elevation gradient, (ii) a continuity loss enforcing local mass conservation with rainfall-adaptive thresholds, and (iii) a topography-aware false-alarm penalty modulated by the topographic wetness index (TWI). The framework is evaluated on the Norfolk, Virginia, flood dataset spanning two major storm events (August 2017 and September 2022) comprising 300 samples, with all variants trained on identical splits and robustness assessed over repeated random splits and leave-one-storm-out tests. A road-proximal evaluation restricted to a TWI-derived street mask quantifies street-level skill. The physics-constrained model achieves near-zero gravity violations (∼10⁻⁶) and the highest street-channel recall (0.77 ± 0.09 versus 0.44 ± 0.10 for the unconstrained baseline), the capability most relevant to downstream traffic routing, and its recall advantage more than doubles on a held-out storm, while a uniform false-alarm variant attains 16% lower mean absolute error but suppresses street recall to 0.25. The proposed TWI-modulated penalty reconciles this trade-off: it improves upon the uniform variant on every metric measured, recovering 60% higher street recall at the lowest MAE among all constrained variants and the best street-level F1 score. These results expose a fundamental tension between aggregate pixel-level error metrics and application-specific physical plausibility, and demonstrate that terrain-aware loss modulation offers a principled resolution. Full article

Cancer stem cell (CSC) persistence drives recurrence and therapy resistance in oral squamous cell carcinoma (OSCC), but what keeps cells locked in this stem-like state is poorly understood. In this narrative review, we propose that CSC state is sustained not by any single pathway but by joint dysregulation of three interacting cell-biological systems: membrane potential (V_mem), communication between neighbouring cells through gap junctional intercellular communication (GJIC), and the Hippo–YAP pathway. We argue that these systems act together on one common point—the YAP protein, retained in the nucleus—which switches on a SOX2-centred stemness gene programme and stabilises a self-reinforcing CSC state. Drawing on evidence from cancer genomics, developmental bioelectricity, connexin biology, and OSCC-specific studies, we reconstruct how membrane depolarisation, loss of gap junction coupling, FAT1 mutation, and Hippo pathway inactivation could converge on persistent nuclear YAP, and how betel quid—the principal risk factor across South and Southeast Asia—may engage all three systems at once. Because the model holds that each input reinforces the others, it predicts that targeting several together should displace CSC state more durably than targeting any one alone. We set out the testable predictions this framework generates. Full article

Liquid crystal (LC) is a key enabling technology for continuously phase-reconfigurable microwave devices, offering analogue-tuning capabilities distinct from discrete alternatives such as MEMS and p-i-n diodes. However, the insertion loss of LC-based phase shifters is inevitably influenced by conductor surface roughness—a factor often neglected in idealised simulations. This paper presents, for the first time, a rigorous numerical quantification of how metal surface roughness affects the insertion loss and phase shift of a 60 GHz LC-filled coaxial delay line (0–180° phase shifter) with radial conductor surfaces instead of conventional planar ones. Using full-wave finite-element simulations incorporating Groisse’s phenomenological model and Huray’s snowball model, four surface configurations are analysed at 54–66 GHz: perfectly smooth conductors, roughness on both inner and outer conductors simultaneously, and roughness applied to each conductor individually. Results show that roughness induces a measurable increase in insertion loss—worst when both conductors are rough—but its impact on differential phase shift remains minimal (<0.32°). Huray’s model predicts conductor losses 1.77 times higher than Groisse’s model, yielding more conservative metrics. For the insertion loss evaluation in Case 2 at 60 GHz under the reference isotropic LC state, Groisse’s model predicts 1.90921 dB, while Huray’s model predicts 2.16319 dB, a 0.25 dB discrepancy (12% uncertainty relative to the mean). The inner conductor dominates roughness-induced losses due to concentrated current density, suggesting prioritised surface finishing of the core line. This study isolates loss mechanisms in a coaxial LC structure, providing insights into low-loss reconfigurable devices. Practical PCB copper foil fabrication methods are also evaluated with quantitative analysis of non-ideal cylindrical geometries. Full article

Accurate correction of gridded numerical weather prediction precipitation forecasts remains challenging because many end-to-end deep learning correction models treat meteorological variables as undifferentiated data channels and therefore provide limited physical interpretability. Inspired by general principles of biological environmental sensing, selective information processing, and regulatory constraint learning, this study proposes PCPNet, a bio-inspired and physically constrained precipitation correction framework. The framework does not imitate a specific biological organ or species; instead, it abstracts three information-processing principles into a meteorological correction task. First, key weather-system cues, including low-level shear lines, trough-ridge effects, upper-level jet-stream forcing, vorticity-divergence-related vertical motion, and water-vapor flux convergence, are quantified as structured diagnostic fields. This transforms the subjective synoptic diagnosis of forecasters into automated grid-based sensing features. Second, these diagnostic cues are fused with numerical weather prediction variables and terrain descriptors in an encoder–attention–decoder network, allowing the model to emphasize dynamically important precipitation-triggering regions. Third, water-vapor conservation and terrain-forcing relationships are embedded as differentiable constraint losses, providing training-time constraint-based regulation that guides the corrected precipitation field toward physically consistent solutions. The method is evaluated from 2021 to 2023 in Hunan Province, China, using hourly numerical weather prediction model outputs as input features, China Meteorological Administration Land Data Assimilation System gridded analysis data as the training target, and independent meteorological station observations for strict cross-validation. PCPNet reduces the mean absolute error by 22.1% compared with the uncorrected China Meteorological Administration Land Data Assimilation System gridded precipitation products and outperforms Linear Regression, Bagging, Boosting, Multi-Layer Perceptron, TabNet, and Tree-based Progressive Regression Models by 12.9%, 13.5%, 16.9%, 10.8%, 14.9%, and 15.9%, respectively. The single-day event analysis provides an initial demonstration of heavy precipitation recovery capability, while comprehensive validation across long-term continuous weather events is planned for future operational deployment to further verify model stability. These results indicate that bio-inspired sensing and regulatory constraint learning can improve both the accuracy and interpretability of precipitation nowcasting correction. Full article

Accurate network traffic prediction is important for the autonomy, resilience and resource orchestration of 6G and AI-native communication infrastructures, while also supporting green networking and digital twin network applications. However, existing data-driven prediction models face several limitations: over-reliance on massive labeled data, physically implausible predictions, black-box non-interpretability and over-parameterization that impairs edge deployment. To address these issues, this paper proposes a Physics-Informed Network Traffic Prediction (PINTP) framework for graph topology network traffic prediction, which formalizes network traffic evolution as Graph-based Advection–Diffusion–Reaction (ADR) equations and embeds physical regularization into the neural architecture. The framework adopts a hybrid differentiation paradigm unifying automatic differentiation for temporal dynamics and spectral graph theory-derived operators for discrete spatial topologies, and designs a physics-constrained composite loss function with data-driven collocation to balance data fidelity and physical consistency. Experiments are conducted in two complementary settings: a 100-node synthetic random-graph benchmark that evaluates the full graph-topological formulation, and a topology-unavailable real-world telemetry proxy based on Alibaba Cluster Trace v2018 for evaluating sparse-label physics-informed temporal regularization. Comparative analysis with mainstream baselines, including Multilayer Perceptron (MLP), Spatio-Temporal Graph Convolutional Network (STGCN), Graph WaveNet, Transformer, Temporal Convolutional Network (TCN), and XGBoost, shows that the proposed PINTP/PINN implementation achieves a test $R^2$ of 0.898 and MSE of 0.000723 on the 100-node synthetic graph benchmark, close to the strongest Transformer result ($R^2=0.900$, MSE = 0.000710), while using substantially fewer trainable parameters. PINTP/PINN also outperforms Graph WaveNet, STGCN and TCN in this setting, indicating that physics-informed regularization can remain competitive as graph size increases. On the Alibaba proxy task, PINTP/PINN achieves the strongest result among the evaluated models with a test $R^2$ of 0.963. In an independent Alibaba ablation protocol, physical regularization (e.g., $\lambda =10.0$) reduces the mean squared error by 89.15% compared with pure data-driven models and helps mitigate overfitting. This work presents a systematic PINTP framework for graph topology network traffic prediction, achieving competitive prediction accuracy with high parameter efficiency and a degree of physical interpretability. It helps address several limitations of traditional data-driven models, indicates potential for future deployment-oriented studies on real-time network management and resource-constrained edge analytics, and provides an interpretable modeling route for physics-informed network analytics in next-generation communication systems. Full article

Posture transition behavior in sows approaching parturition provides an important physiological cue for farrowing prediction. However, manual monitoring is time-consuming, labor-intensive and difficult to sustain under nighttime production conditions, while existing machine vision approaches remain limited in their ability to represent continuous posture transitions in complex farm environments. Here, we propose an event-level posture transition recognition framework that integrates YOLOv11n with an RGB–Flow cross-modal temporal network. YOLOv11n is first used to detect basic sow postures at the frame level, after which candidate transition events are automatically generated and refined according to temporal state changes. For each event segment, RGB appearance features and optical-flow motion features are extracted to construct dual-branch spatio-temporal representations. We further develop a multi-scale cross-modal attention temporal network (MS-CMATNet) for event-level behavior classification. The network captures local temporal dynamics through a multi-scale module, enhances interactions between RGB and Flow representations through cross-modal attention, and improves feature discriminability and stability by incorporating temporal–channel attention blocks (TCBAM) and an auxiliary cross-modal consistency loss (AuxCross). Experiments show that MS-CMATNet achieves an Accuracy of 88.14%, a Macro-Recall of 84.04%, and a Weighted-F1 score of 87.66% under the fixed training/validation split, outperforming the compared machine learning models, deep temporal models, and representative temporal and cross-modal baselines. Repeated stratified cross-validation and paired t-tests further confirm that MS-CMATNet achieves statistically reliable improvements over most compared baselines, particularly in Macro-F1 and Weighted-F1. These findings demonstrate the potential of the proposed framework for automated farrowing prediction in smart livestock farming. Full article

With the growing global emphasis on data privacy protection, particularly the enforcement of the “right to be forgotten” under the GDPR, effectively deleting private information from models has become an urgent challenge. Graph-structured data presents a particularly challenging unlearning scenario due to its non-Euclidean nature and strong relational dependencies, which are prevalent in real-world applications such as social and recommendation systems. To address this issue, graph unlearning has been introduced to eliminate the influence of deleted data on models while preserving their overall performance. The effectiveness of graph unlearning is typically evaluated by three key metrics: model performance, unlearning efficiency, and robustness against membership inference attacks, which together determine the overall quality of an unlearning method. Existing graph unlearning methods fall into exact and approximate regimes. Most studies focus on edge/node-level unlearning, and existing attempts at feature-level unlearning remain limited. Exact unlearning methods that adopt the SISA partition and retraining paradigm may inadvertently reintroduce the features intended to be unlearned during the aggregation phase, thereby leading to incomplete unlearning. Approximate methods, on the other hand, often incur excessive information loss in feature-level removal, which degrades predictive accuracy. Accordingly, we propose a graph unlearning framework specifically designed for feature-level unlearning, consisting of two main stages. In the first stage, we zero out the features of the unlearned nodes at each layer to block their propagation through the GNN, thereby reducing their influence on neighboring node representations. In the second stage, we induce misclassification of the unlearned nodes to progressively degrade model representations and learned knowledge associated with them, enabling more thorough feature-level unlearning. Experiments on multiple graph datasets and models demonstrate that our method achieves favorable overall unlearning performance in most settings, offering a balanced trade-off between accuracy, unlearning efficiency, and unlearning effectiveness. Full article

The full-scope simulator plays a critical role in nuclear power plant emergency drills, personnel training, and accident analysis. Traditional system programs lack sufficient computational performance to meet real-time requirements when simulating complex accident scenarios in reactor systems. This study focuses on the small-break loss-of-coolant accident (SBLOCA) in nuclear power plants, generating large-scale datasets through digital simulations. After data preprocessing and normalization, a light gradient boosting decision tree (LightGBM) regression model was developed using machine learning algorithms. SHAP (SHapley Additive exPlanations) analysis identified the contributing factors, enabling the model to predict key parameters such as peak fuel cladding temperature, primary reactor coolant pressure, and pressurizer water level. The model achieved a mean square error (MSE) below 0.002 and a coefficient of determination (R²) exceeding 0.98, with a prediction speed approximately 32,500 times faster than traditional system programs, requiring less than $4\times {10}^{-4}$ seconds per data point. This study provides a novel solution for complex condition simulations in nuclear power plant full-scope simulators. Full article