The aim of this article is to present an analytical method apt to generate an optimum shaped profile for contacting parts of cantilevered pin joints that are subjected to edge contact pressure effects. The optimum inner or outer shaped profile of the outer contacting part is obtained from contact pressure optimization so as to ensure a uniform contact pressure distribution on the contact interface under applied loading. The proposed analytical approach is illustrated through the studied case of a system of cantilevered beams in frictionless contact, using direct and optimization-based methods. The obtained analytical results are compared with finite element computations at the end of the paper, showing a good agreement, especially for the optimum outer shaped profile.
This paper evaluates graph-based multivariable bias correction for ERA5 meteorological variables in the Valencian Community, a Western Mediterranean region with complex terrain and recurrent high-impact precipitation episodes. AVAMET station observations were projected onto the ERA5 grid and used to train a GraphSAGE model with Jumping Knowledge under leakage-free chronological validation. The model was compared with raw ERA5, the oracle classical reference defined in Methods, and a fixed Random Forest reference. Evaluation combined block-bootstrap uncertainty, fixed-window seasonal robustness, spatial diagnostics, multivariable correlation diagnostics, and precipitation-specific analyses. On the temporal test split, GraphSAGE+JK reduced RMSE and MAE for six non-precipitation variables: wind gust, relative humidity, surface pressure, 2-m temperature, wind direction, and wind speed. Precipitation remained unresolved: the graph model did not improve aggregate pointwise skill or high-threshold occurrence detection, although episode-level totals showed both reduced and amplified domain-integrated errors in held-out wet episodes. Overall, the results support graph-based post-processing as a local predictive correction strategy for non-precipitation variables, while identifying precipitation extremes and spatial transferability as open challenges.
More accurate simulators do not guarantee more reliable scientific inferences. In inverse problems—recovering hidden quantities from indirect data— increased model fidelity can expose poorly observed sensitivities, shift dominant error sources, and invalidate calibrated inverse settings. Simulator quality should ideally be judged through the complete inference system, not by forward accuracy alone.
In the current work, an enhanced meshfree framework has been proposed for effective modelling of thermoelastic fracture in porous graded domains. The formulation simultaneously incorporates porosity-induced material degradation, spatial material gradation and mixed-mode fracture behaviour under coupled thermomechanical loading. An adaptive nodal support domain depending on local porosity gradients and a vicinity-based selective quadrature strategy has been incorporated with the conventional element-free Galerkin method (EFGM) framework to ensure numerical stability and computational efficiency. Numerical results demonstrated good agreement with reference solutions, with average deviations below 2%, while achieving reductions in computational time of more than 70% compared to conventional EFGM formulations. Furthermore, this framework is applied to a wide range of fracture problems, including edge and centre cracks, multi-segment kinked cracks and cracks in complex geometries such as disks containing holes. The results revealed a monotonic increase in the equivalent stress intensity factor with increasing crack length and inclination, highlighting the amplified crack-driving force due to porosity-induced compliance, material gradation and strong thermo-mechanical coupling. In kinked crack configurations and disk-with-holes problems, pronounced mixed-mode effects and stress concentration interactions further intensify the equivalent stress intensity factor, demonstrating the effectiveness of proposed approach to accurately model complex fracture behaviour in porous graded domains.
Infectious myocarditis is an inflammatory cardiac disease associated with myocardial edema and severe clinical outcomes, whose global incidence increased by more than 62% between 1990 and 2019, reaching over 1.26 million cases and approximately 324,000 deaths worldwide. This growing burden motivates the development of efficient computational strategies to investigate immune-driven tissue dynamics. In this work, we present a comparative evaluation of deep learning-based surrogate models for simulating the spatio-temporal dynamics of the innate immune response in myocardial tissue. PINNs and conventional Neural Networks (NNs) are assessed against a Finite Volume Method (FVM) solver, including a GPU-parallelized implementation. The GPU-accelerated FVM achieved an average acceleration of approximately 4.8×102 compared to its CPU serial version. Both Machine Learning (ML) approaches achieved a mean acceleration of about 6.6× relative to the GPU-based FVM. Conventional NNs achieved lower error levels and better representation of localized solution features. In contrast, PINNs exhibited higher training costs and reduced local accuracy. The study highlights the role of ML surrogates as efficient tools for large-scale and multi-execution biological simulations.
Electromagnetic (EM)-based optimization is imperative for most microwave circuits. This is because designs produced by circuit theory methods require further adjustments to ensure satisfactory performance. Unfortunately, repetitive EM simulations incurred by optimization procedures are associated with considerable expenses. Expediting the process is essential to reduce the time span of design cycles and time-to-market in the case of industrial products. One possibility is utilization of variable-resolution EM simulations; however, available frameworks typically employ two discrete levels of fidelity (coarse/fine) combined with often sophisticated model correction methods. This paper proposes an alternative approach to accelerated microwave component tuning. Our methodology involves a continuous spectrum of EM simulation resolutions, which are controlled using the convergence and design quality indicators of the underlying optimization routine. A management scheme adjusting the model resolution is developed to utilize the lowest usable resolution at the search process onset, which is continuously increased as the algorithm approaches convergence and the objective function reaches satisfactory levels. The final stages are carried out with high-fidelity resolution to ensure accuracy. Verification experiments conducted using three microstrip circuits demonstrate over sixty percent of average savings over the reference (single-resolution) gradient-based routine, and noticeable and consistent acceleration regarding several state-of-the-art expedited versions. Selected designs are experimentally validated.
Stochastic matrix games are an important class of stochastic games that enhance traditional matrix games by using state-dependent transition between game states. However, analytical and numerical solution methods for stochastic matrix games are often impractical due to the need to solve high degree equations or perform computationally expensive iterative procedures. To overcome these challenges, we present a novel fully connected neural network architecture established for efficiently estimating the game values of stochastic matrix games. In this context, we randomly generate 5×105 games for each type of game of size 3 × 3, 4 × 4 and 5 × 5 and solve them using Shapley iteration with tolerance level of 10−5. Then, we decompose these stochastic matrix games into deterministic, stochastic and indicator matrices to construct proper datasets. Next, we use a recent vectorization process to transform these matrices into a vector form to feed the neural network. The results demonstrate that the proposed neural network efficiently and accurately estimates the game value of 3 × 3 , 4 × 4 and 5 × 5 stochastic matrix games with mean errors of -0.0070, -0.0154, and 0.1060, respectively. We also evaluate the performance of the model by presenting absolute mean errors, absolute median errors, and root mean square errors for each size of games. Once trained, the model provides the game values in a constant time which offers a significant reduction in computation time compared to traditional iterative approaches. Consequently, the proposed neural network provides an efficient alternative approach for solving stochastic matrix games.
Artificial Intelligence (AI) has become a central methodology in health science, enabling data-driven modelling and inference in several biomedical domains. In recent years, AI-based approaches have been increasingly employed in pharmacology and pharmacovigilance to detect, characterize, and predict Drug Side-Effects (DSEs). Concurrently, network science has provided a principled framework for representing the highly interconnected nature of biological systems, including drug–target, protein–protein, and disease–gene interaction networks. In this context, Graph Neural Networks (GNNs) have emerged as a powerful tool for operating on graph-structured data, by propagating and aggregating information across nodes and edges. In this paper, we present a method for inferring candidate DSEs by using a Heterogeneous GNN and a multilayer network model. Specifically, we treated the issue as a link prediction problem by applying a GNN to a heterogeneous topology, consisting of drug–drug, DSE and chemical drug–gene associations, modelled on a multilayer network. It jointly exploits drug–drug interactions, drug–gene relationships, and known DSE associations, to effectively integrate information across layers. Our method was based on a relation-weighted heterogeneous GNN model that learns the relative importance of different association types while enabling multi-hop message passing. The experimentation demonstrated that the proposed method consistently outperforms baseline GNN architectures, achieving improved discrimination and robustness on large-scale data. Furthermore, results highlight the effectiveness of relation-aware multilayer modelling for candidate DSE inference, e.g., in supporting pharmacovigilance and drug development.
A key challenge in expediting drug discovery is accurately predicting molecular properties. Optimizing drug-like qualities remains a complicated and resource-heavy process. While methods utilizing molecular fingerprints and descriptors are informative, they frequently have difficulty in fully representing the intricate structural relationships in molecules. Graph neural networks (GNNs), on the other hand, have emerged as a powerful approach, demonstrating their ability to effectively depict molecules as graphs and address this shortcoming.This research introduces an innovative hybrid model that integrates conventional chemical descriptors with GNN techniques for predicting molecular properties. Our model integrates molecular fingerprints and descriptors with features extracted by GNNs, enhanced through a graph attention mechanism that effectively captures both local and global structural information. By leveraging these complementary data sources, the model delivers significantly improved prediction accuracy compared to several existing methods, achieving superior average performance across various tasks.Moreover, this approach achieves competitive performance without relying on resource-intensive pre-training phases. This highlights the computational efficiency and robustness of the proposed method, demonstrating that favorable results can be obtained with lower computational costs, making it a compelling choice for addressing challenges in molecular property prediction.
In the global health context of aging population and increasing burden of chronic diseases, disease prediction is a core means to optimize healthcare resource allocation and precision prevention. The fusion of artificial intelligence technologies and high throughput sequencing provides a new paradigm for resolving microbe-disease associations. Aiming at the general failure of existing models to effectively capture potential correlations among samples, as well as the challenges of high dimensionality, sparsity, and small samples in rare disease prediction scenarios, this paper proposes an enhanced multilayer perceptron (MLP) model guided by multiple sample-feature graphs and sample graphs. The model constructs multiple sample-feature graphs for all features based on Manhattan distance, and utilizes graph convolution to learn the embedding representations of sample-feature graphs, and thereby parameterizes the first layer of the MLP. Simultaneously, to incorporate inter-sample associations, the model constructs sample graphs using cosine similarity and a network-enhanced denoising method. It then employs graph convolution to learn the node embedding representations of the sample graph, which are used as complementary input features for the MLP. The experimental results show that parameterizing the first layer of the MLP helps to alleviate the overfitting problem; the graph structure can improve the prediction performance; and incorporating the sample graph positively contributes to enhancing the model’s capabilities. Compared with other methods on the IBD, Colorectal, Obesity, Cirrhosis, C-T2D, and EW-T2D datasets, the model’s AUC values are 4.3%, 3.3%, 9.6%, 2.3%, 4.2%, and 10.6% higher than those of the best baseline models, respectively, demonstrating significantly superior prediction performance.
The present research work analyzes the performance of a finite Markovian model for the fault-tolerant redundant system by integrating several realistic operational features, including retrial, feedback, reboot mechanisms, and admission control F−policy. A recursive methodology is employed to compute the queue length distributions, which are further utilized to evaluate key performance metrics that describe the behavior of the proposed model. To address uncertainty in the system parameters, the crisp model is transformed to a fuzzy queueing framework and analyzed using parametric non-linear programming and Zadeh’s extension principle. A comprehensive cost function is developed to determine the optimal system parameters that minimize overall operational costs. Numerical simulations are used to obtain the analytical findings for both the crisp and fuzzy models. The genetic algorithm and particle swarm optimization metaheuristic techniques as well as numerical optimization viz. quasi-Newton method are applied to evaluate the optimal parameters. The sensitivity analysis and relative sensitivity analysis are conducted to examine the influence and relative influence of system parameters on the cost function. From an application perspective, the proposed study highlights the applicability of the proposed model to fault-tolerant routers in telecommunications networks.
Adaptive collocation sampling is critical for the accuracy of Physics-Informed Neural Networks (PINNs), yet its systematic evaluation across classical and hybrid quantum architectures remains lacking. We propose a Kernel Density Estimation (KDE)-based adaptive sampling strategy that dynamically concentrates collocation points in regions of high PDE residual, and rigorously evaluate it across four PINN architectures—Plain PINN, ClassicalAug-4, ClassicalAug-8, and Hybrid Quantum PINN (HQPINN)—on the 1D viscous Burgers’ and Allen–Cahn equations. KDE-based sampling produces large, consistent improvements across all architectures: Cohen’s d effect sizes exceed 1.0 for every architecture–PDE combination, with win rates of 87.5%–100% across eight-seed paired comparisons and mean relative L2 error reductions of 12%–17% across eight-seed validated configurations. Critically, quantum and bottleneck-matched classical augmentation achieve statistically comparable accuracy on both PDEs: under uniform sampling, HQPINN is marginally better on both PDEs (|Δ|≤0.009), while under KDE, ClassicalAug-4,slightly outperforms HQPINN on both PDEs (|Δ|≤0.008); none of these differences are statistically significant. This rigorous null result — enabled by our parallel-gate architecture with controlled baselines — provides a principled benchmark for the quantum-PINN community and establishes a clear foundation for identifying the problem regimes where quantum advantage may emerge. We further provide the first eight-seed comparison of KDE against three established adaptive methods (RIS, RAD, RAR) on both classical and quantum architectures, finding that while RAD and RAR achieve lower mean L2 errors on specific architecture–PDE combinations (e.g., RAD on Burgers’: 0.225 vs. KDE’s 0.269; RAR on Allen–Cahn: 0.741 vs. KDE’s 0.872 on Plain PINN), KDE offers the lowest cross-run variability (CV of 17.5%–28.9% vs. 22.8%–38.7% for baselines) and the most consistent cross-architecture improvement, making it a reliable general-purpose strategy. Hard boundary condition enforcement yields substantial accuracy gains for stiff PDEs (ϵL2 from 0.83 to 0.13 on Allen–Cahn). All results are validated across eight random seeds with Holm–Bonferroni-corrected hypothesis tests.
In data science, Missing Data (MD) are handled by approaches from basic imputation (e.g., mean) to more complex (e.g., Machine Learning (ML)). Assessing these methods consists in amputating data then imputating back amputated values to compare with ground truth. The existing amputation methods are most often Random Amputation (RA) or parameterized approaches using Missing Data Mechanisms (MDM) involved. MD from real-world datasets can follow complex patterns and the specific MDM involved can be impossible to determine with certainty. Thus, existing data amputation methods are often difficult to properly apply on real-world datasets.In this article, we propose Missing Mechanisms Respectful Data Amputation (MMRDA) as a new method to generate synthetic MD and we study the impact of data amputation methods on the results of imputation techniques assessment.MMRDA has been compared to the RA method and assessed on open datasets. MMRDA significantly outperforms RA. Four well-known imputation techniques were used to compare imputation performance post-MMRDA with imputation performance post-RA. Differences were significant and the best imputation performance is not always achieved by the same imputer depending on the amputation technique used for a given dataset. This supports that MDM involved affect imputation and that the attribution of complex mechanisms is still to be explored. Therefore, we recommend all researchers to pay close attention to the amputation method used during imputation techniques assessment or comparison, especially on real-world datasets.
Transformer models have been widely applied across various domains, often treating spatio-temporal data as video-like sequences due to the success of generative video prediction. However, this paper argues that transformers are not always optimal for spatio-temporal data with long forecast horizons and strong periodicity. Focusing on metocean forecasting (specifically sea ice, ocean, and atmospheric data) the study evaluates transformer-based models against convolutional neural networks (CNNs). For long-term sea ice forecasting in the Arctic, transformers such as TimeSformer and SwinLSTM failed to capture annual dynamics, including summer melt. In contrast, a lightweight CNN baseline outperformed existing state-of-the-art numerical and data-driven forecasts, improving error metrics by up to 30%. Similarly, in atmospheric bias correction, CNNs proved superior, reducing errors in Global Forecast System fields by 20% relative to transformers. The narrative shifts with ocean forecasting, where transformer models enhanced by a contrastive pre-training achieved comprehensive superiority. They significantly reduced errors across all ocean variables, including a 40% reduction for mixed layer depth. These three case studies demonstrate that transformer limitations exist but are conditional rather than absolute, but CNNs remain the appropriate choice when data is limited or fine spatial structure is critical. This paper is an extended version of the ICCS-2025 conference paper [1].
Role-based collaboration is pivotal for solving collaborative problems, with group role assignment as its core. Traditional group role assignment models only consider oppositional conflict, that is, full conflict across all roles, and complete cooperation, that is, full collaboration across all roles, failing to capture real-world nuanced interactions where agents conflict or cooperate only within specific roles. Therefore, we propose general conflict and general cooperation, with a critical constraint: cooperating agents for a role must be simultaneously assigned or excluded to ensure collaborative integrity. This paper proposes a new problem in collaboration based on the environments-classes, agents, roles, groups, and objects (E-CARGO) model, which is called the group role assignment with general conflict and cooperation (GRAGCC) problem, establishing key constraints: general conflict agents cannot share roles, and general cooperation agents must be jointly assigned or excluded. Moreover, we derive the feasibility and necessary conditions and propose practical algorithms to maximize group performance. Simulation experiments validate the conditions and algorithms, demonstrating that GRAGCC outperforms other representative models by more accurately capturing realistic agent interactions, reducing inefficiencies arising from unaddressed partial conflicts, and effectively harnessing cooperative gains.
Understanding the evolutionary dynamics and future trajectory of agricultural net carbon sink (ANCS) was critical for aligning agricultural development with China’s dual-carbon goals. This study systematically analyzed the dynamics of ANCS across 31 provincial-level regions in China from 2010 to 2024. Agricultural carbon emissions (ACE) and carbon sinks (ACS) were estimated using the emission factor method based on panel data. The relationship between ANCS and both agricultural economy and total power of agricultural machinery were examined using the Tapio decoupling model. The LMDI method was applied to decompose the key drivers of ANCS dynamics, and the GM (1,1) model was employed to forecast ANCS trends up to 2030. The results revealed that China’s ANCS increased markedly from 350 Mt in 2010 to 570 Mt in 2024. The relationship between ANCS and both agricultural economy and total power of agricultural machinery gradually transitioned from coupling to decoupling. Decomposition analysis identified the level of agricultural mechanization (+377.79 Mt) and agricultural machinery efficiency (+216.67 Mt) as the dominant positive drivers of ANCS, whereas agricultural labor scale (-151.10 Mt) and ANCS intensity (-118.74 Mt) emerged as the primary inhibitory factors. Projections indicated that China’s ANCS would maintain a steady growth trajectory through 2030, consistent with China’s carbon peaking target. These findings provided scientific evidence and practical support for policy formulation aimed at low-carbon agricultural transformation and the realization of China’s carbon peak and carbon neutrality goals.
