ABSTRACT. Advanced persistent threats (APTs) evade host defenses by distributing malicious activity over long time spans and blending it into benign workloads. Recent work leverages provenance graphs to model system-wide causal dependencies for intrusion detection, but practical deployments remain hindered by extreme class imbalance, temporally dispersed evidence across entities, and the need for causal, interpretable alerts. We address these challenges with a host behavioral consistency framework on temporal provenance graphs. The core design is to model normal interaction consistency in context, then convert inconsistencies into causal evidence and aggregate it across dependent entities. This yields compact attack chains that remain interpretable while controlling false positives under extreme class imbalance. Across three DARPA Transparent Computing datasets, our method reduces false positives and improves detection at scale, with a 0.40 MCC gain on CADETS over the best baseline.

ABSTRACT. Data connectors are key components for enabling controllable data exchange in data marketplaces and constitute one of the foundational technologies for realizing data sovereignty. To support controllable data exchange in lightweight data connectors, it is crucial to design an efficient and scalable access control mechanism. However, user--space access control suffers from high decision latency, whereas kernel--space access control is limited by constrained policy capacity and inflexible policy updates. To address these challenges, we propose an efficient user--kernel access control scheme for lightweight data connectors. The scheme leverages extended Berkeley Packet Filter (eBPF) to load high-frequency access control rules into the kernel while retaining low-frequency rules in user space. To further improve access control efficiency, we design a Bloom-filter-based kernel admission filter to block invalid requests before they enter the policy decision point (PDP). To prevent long-tail requests from polluting the kernel cache, we develop a set-associative, access-frequency-aware cache replacement mechanism. To determine the optimal memory allocation between these two kernel-side components, we model access control decisions as an M/D/1 queuing system, analyze the relationship between end-to-end policy decision latency and cache miss probability, and apply golden-section search to compute the optimal memory partitioning. We implement a prototype based on eBPF. Experimental results show that, under the same policy set size, our system reduces the average service time by approximately 25\% compared with AppArmor; with $10^{6}$ policies, it reduces the average service time by up to 85.6\% compared with Casbin. Under mixed invalid and long-tail traffic with an attack ratio $\gamma > 80\%$, the proposed design reduces latency by up to 65\% compared with an unprotected LRU policy.

ABSTRACT. TEEs provide a practical foundation for secure cross-domain collaborative computation. However, establishing group-oriented secure channels among mutually distrustful enclaves remains a fundamental challenge, as it must provide mutual authentication while preserving group-level trust semantics. To address this, we present GroupTEE, a group-oriented secure communication framework for TEEs. GroupTEE unifies CA-certified identities and remote attestation into a trust-binding mechanism that cryptographically binds group membership and enclave integrity to the derived session key within a single protocol execution. We prove in ROM that GroupTEE achieves EUF-CMA authentication, session-key indistinguishability, and key consistency. Experiments show that GroupTEE reduces per-participant attestation-generation operations from O(n) in pairwise-based extensions to O(1) per session.

ABSTRACT. Modern intrusion detection systems (IDSs) in large-scale enterprise networks generate massive volumes of heterogeneous alerts, many of which are weakly correlated interaction traces that obscure the semantic progression of multi-stage attacks. While existing alert filtering techniques reduce analyst workload, they often fail to preserve behavioral continuity necessary for reconstructing coherent attack campaigns. In this paper, we propose a TTP-sequence–driven framework that integrates Cyber Threat Intelligence (CTI) with alert-side interaction evidence to reconstruct semantically coherent attack events from noisy intrusion alerts. Structured ATT&CK-aligned TTP sequences are extracted from CTI reports to form a knowledge space of canonical attack progressions. Candidate TTP paths derived from alerts are embedded into a shared semantic space, filtered using a one-class representation model, and aggregated into candidate events. A retrieval-augmented reconstruction module further refines attacker–victim structures by aligning alert-derived behaviors with CTI-derived progression templates. Experiments on three publicly available cyber range datasets (CPTC-2017, CPTC-2018, and CCDC-2018) show that the proposed framework reduces structurally inconsistent interaction paths while improving attacker–victim coherence at the event level. Reconstructed events exhibit improved semantic alignment with CTI-derived attack progression patterns compared to their originating alert paths.

ABSTRACT. Graph representation is a powerful abstraction of real-world objects and relations. Computing the Graph Edit Distance (GED) between graphs is critical in domains such as bioinformatics, machine learning, and pattern recognition. GED measures the minimum number of edit operations required to transform one graph into another. However, the high computational complexity of optimal and near-optimal methods limits their applicability to large-scale graphs, making high-performance parallel GED computation essential. To address this, we propose FAST-GED, a fast and scalable open-source framework for GED computation on GPUs. FAST-GED overcomes existing limitations by combining high accuracy with fast execution through GPU-friendly algorithmic design and efficient mapping to GPU hardware, minimizing host-device communication. The implementation is optimized and tested across multiple GPU architectures.We validate FAST-GED on real and synthetic datasets with diverse graph sizes and densities. It achieves speedups of several orders of magnitude over the Python NetworkX library while reaching optimal solutions in most cases. Moreover, it outperforms state-of-the-art approximate methods in both accuracy and scalability. We show that FAST-GED enables broader adoption of GED-based solutions in real-world applications.

ABSTRACT. Efficient execution of quantum circuits on noisy intermediate-scale quantum (NISQ) devices requires adapting logical circuits to hardware with limited qubit connectivity. SWAP networks provide a circuit-agnostic solution by prescribing SWAP schedules that guarantee adjacency between every pair of qubits, thereby decoupling routing from circuit-level structure. However, constructing short and scalable SWAP networks for arbitrary hardware topologies remains a challenging combinatorial problem.

This work introduces a fast, deterministic pipeline for generating high-quality SWAP networks on general connectivity graphs. The method is built around minimum Dominating Paths, which serve as compact structural backbones from which the full SWAP schedule is derived. The proposed four-stage procedure—path extraction, internal routing, external qubit integration, and final refinement—operates in the optimal number of SWAP gates at each step and avoids stochastic search entirely.

We evaluate the approach on IBM quantum processor topologies ranging from 5 to 131 qubits. Across all architectures beyond 8 qubits, the method consistently outperforms the state-of-the-art simulated-annealing-based baseline, achieving up to 25% reductions in network length while providing strong interpretability and significantly lower runtime. These results demonstrate that minimum-path-based constructions offer an effective and scalable foundation for circuit-agnostic routing in the NISQ era.

ABSTRACT. In the Directed Feedback Vertex Set problem (DFVS), the goal is to find the smallest possible subset of vertices of a given directed graph whose removal leaves the graph acyclic. The problem has important applications in computational science, and it was a subject of the 7th Parameterized Algorithms and Computational Experiments Challenge, an international algorithmic contest held in 2022 aiming to stimulate and enhance developments in the field. The DFVS is known to be NP-complete and constitutes a natural generalization of the Vertex Cover problem (VC). Compared to the VC, however, the DFVS seems to be vastly more difficult. One of the common approaches to tackle computationally difficult problems is to design data reduction rules. The approach is known to be highly effective (both in theory - kernelization algorithms, and practice - reduction in processed data size) for several combinatorial problems, including the VC. For the DFVS substantially fewer reduction rules are known than for the VC, and they are much less robust in practical applications. In this paper we address this issue and present several new data reduction rules for the problem. Presented algorithms are based on an extended concept of domination. We provide detailed description and analysis of designed techniques and, additionally, provide insights regarding practical applicability of discussed methods. At the end, we present a compact juxtaposition showing the impact of implemented data reduction rules for a collection of real-world graph instances.

ABSTRACT. We study how to model alignment relations between discrete tokens through a unified formalism of position spaces. From two strict partial orders---horizontal and vertical alignment---we derive complementary graph abstractions: the directed position graph and the bipartite row/column graph. We formalize alignment motifs as graph patterns and cast their detection as subgraph isomorphism. Pattern search is encoded as a CSP with standard ingredients, optionally strengthened by type, order, and cardinality constraints, and coupled with an existing filtering algorithm (ILF). Experiments on controlled families of position spaces show that on position graphs, ILF consistently reduces search effort and runtime, with gains increasing with instance size; on row/columns graphs, structural constraints are decisive, often turning otherwise intractable cases into consistently solvable ones. Our contribution is thus a graph-based modeling framework and an empirical analysis of different constraint models to search for patterns within these graphs.

ABSTRACT. Data in High Energy Physics (HEP) is primarily stored in binary columnar formats. This enables efficient reading of select columns for the physics analysis use case while employing transparent compression to reduce storage space. However, this combination of requirements makes it challenging to write HEP data in parallel. This has been partially addressed with multithreaded writing, but efficient usage in cluster environments requires extension of the existing concepts. In this paper, we discuss distributed parallel writing of columnar data into a single file. For this, we present different strategies based on a central aggregator, using MPI one-sided communication, and relying on synchronization with file locks. We evaluate the scalability of these approaches on two cluster file systems, and discuss the impact of different parameters on performance. Finally, we motivate a real-world use case for resource-intensive Monte Carlo simulation jobs.

ABSTRACT. This paper presents a novel method for efficiently sampling the Pareto fronts of three-objective optimization problems. The core principles are (1) approximating the Pareto front boundary, (2) formulating the rejection criterion, and (3) efficiently sampling for Pareto front computation. Starting from the individual minima, the boundary of the Pareto front is sampled and then projected to a hyperplane. This plane is then uniformly sampled within the bounded region. Outside samples are discarded, since they would lead to redundant Pareto optimal solutions. Finally, only the remaining samples are used to determine parameters for the Pascoletti-Serafini scalarization method by which an approximately uniform coverage of the Pareto front can be computed. Results from several test problems demonstrate that our approach frequently yields superior performance regarding efficiency and uniformity, while consistently outperforming existing methods on average.

ABSTRACT. In multi-objective optimization, the set of optimal trade-offs---the Pareto front---often contains regions that are extremely steep or flat. The Pareto optimal points in these regions are typically of limited interest for decision-making, as the marginal rate of substitution is extreme: a marginal improvement in one objective necessitates a significant deterioration in at least one other objective. These unfavorable trade-offs frequently occur near the individual minima, where single objectives attain their minimum values without considering the remaining criteria.

To address this, we propose the concept of non-extreme individual minima that relies on the notion of $L$-practical proper efficiency. These points can serve as a less sensitive replacement for standard individual minima in subsequent related methods. Specifically, they allow for a more practical restriction of the Pareto front sampling within a refined utopia-nadir hyperbox, provide a meaningful basis for image space normalization, and can enhance decision-making techniques, such as knee-point methods, by focusing on regions with acceptable trade-offs.

We provide a computationally efficient algorithm to determine these non-extreme individual minima by solving at most $2n_J$ standard weighted-sum scalarizations, where $n_J$ is the number of objectives. To ensure robustness across varying objective scales, the method incorporates an integrated image space normalization strategy. Numerical examples, specifically a convex academic case and a non-convex real-world application, demonstrate that the method successfully excludes practically irrelevant regions in the image space.

ABSTRACT. We propose a new method for undirected graph structure learning for binary variables. Our approach is based on a modification of the FOCI selection procedure, which is applied node-wise to recover a Markov blanket for each variable. The final undirected dependency graph is obtained as the union of these Markov blankets.

To adapt FOCI to the binary setting, we introduce an empirical version of a conditional dependence coefficient and incorporate a thresholding step to control false discoveries. We establish theoretical guarantees showing that, with high probability, the proposed method correctly recovers the true graph structure in finite-sample setting.

The performance of the method is evaluated in a simulation study on artificial data generated from binary structural equation models based on logistic regression.

ABSTRACT. Island Evolutionary Algorithms (IEAs) are effective for solving hard problems such as the Quadratic Assignment Problem (QAP), but their performance depends on careful and computationally expensive parameter tuning. While ML has been used to automate solver tuning in SAT (and related constraint-solving problems), its transfer to other problem classes with a different structure is less explored. Moreover, most ML-based approaches assume a single optimal configuration, ignoring the fact that multiple parameter settings can yield comparable performance. In this work, we formulate parameter prediction as an inconsistent learning problem and investigate whether instance features can be used to predict effective IEA configurations for QAP. We compare three strategies: a regression-based performance prediction model, a multiple-choice model that treats near-optimal configurations as valid targets, and a baseline single-label model. The models are trained on effective parameters for 867 QAP instances tuned with SMAC, using instance-level features for training. The regression approach achieved low predictive accuracy and poor optimization performance, indicating that performance prediction is challenging in this setting. The multiple-choice model improved predictive generalization but did not improve downstream performance. The baseline model achieved the best results, producing configurations comparable to those obtained via expensive tuning. These findings suggest that direct configuration prediction provides a robust approach despite ambiguity in the parameter space.

ABSTRACT. Accurate forecasting of geopotential height is crucial for nu- merical weather prediction and climate modeling. While transformer- based models have shown promising results, their quadratic computa- tional complexity in sequence length limits their scalability to long tem- poral contexts. In this work, we present a comprehensive comparison of linear attention architectures—including Gated Linear Attention (GLA), DeltaNet, Gated DeltaNet, Kimi Delta Attention (KDA), and Mamba2— for geopotential height forecasting on NCEP reanalysis data. We eval- uate these architectures against full-attention transformers in terms of forecasting accuracy, computational efficiency, and memory usage. Our experiments demonstrate that linear attention mechanisms achieve com- petitive forecasting performance while offering significant computational advantages for processing long temporal sequences.

ABSTRACT. Timely wildfire response requires both rapid detection and accurate short-term forecasting of fire spread. In this work, we present an AI-driven framework for real-time wildfire nowcasting that integrates multimodal data fusion and generative modeling to accelerate decision-making across the full response pipeline. For early fire detection, we leverage social media streams as dynamic spatial cues to guide targeted satellite data analysis. Exhaustive satellite scanning over large geographic areas is computationally expensive and time-consuming. By contrast, user-generated content provides near real-time geospatial signals that significantly narrow the search space. We employ multimodal learning approaches to jointly process textual, visual, and geolocation information from social media, using these signals to trigger focused satellite verification. This hybrid strategy enables faster and more efficient fire confirmation compared to satellite-only monitoring. Once a wildfire is detected and localized, rapid forecasting of fire spread becomes critical for risk assessment and intervention planning. To address this, we develop diffusion-based generative models to predict the short-term evolution and direction of the fire front. Unlike many conventional AI approaches that rely heavily on complex preprocessing and handcrafted features, diffusion models offer strong generative capacity, robustness, and fast inference, making them well-suited for operational nowcasting. Our framework produces high-resolution spread forecasts that support decisions such as evacuation planning and the strategic placement of fire barriers. By combining socially driven detection with generative diffusion-based forecasting, our approach provides an end-to-end AI system for fast, scalable, and actionable wildfire nowcasting.

ABSTRACT. Natural disasters create complex economic dynamics driven by latent variables that are difficult to observe. We develop a framework that applies data assimilation, using ensemble Kalman filter, to improve latent state estimation in natural disaster economic agent based models. To handle nonlinear and high-dimensional dynamics, we combine these methods with reduced order machine learning methods, using autoencoders to learn compact representations of the system state while preserving its structural properties. Using a dynamic input-output model of economic production and flood impact, we show that the approach can recover key latent variables, such as inventories, from limited observations, enhancing the reliability of real-time economic assessments.

ABSTRACT. Uncertainty is an inherent and unavoidable feature of atmospheric prediction. The chaotic nature of the atmosphere, combined with model imperfections and incomplete observations, implies that no deterministic forecast can fully represent the future state of the system. For complex weather events, reliable estimates of forecast uncertainty are often as critical as the forecast itself, as they directly influence risk assessment and decision-making. Traditionally, uncertainty in numerical weather prediction has been quantified using Monte-Carlo based ensemble methods. While effective, these approaches can be computationally expensive and slow to converge, especially in highly nonlinear and convective-scale dynamical systems. This has motivated growing interest in uncertainty-aware machine learning (ML) techniques as efficient alternatives to classical ensemble methods. In this work, we investigate the potential of conformal prediction (CP) as a statistically robust, distribution-free framework for uncertainty quantification in atmospheric modelling. Our study is conducted in an idealized setting, allowing us to fully isolate and carefully evaluate the strengths and limitations of CP. The contributions of this study are threefold. First, we adapt CP to a spatio-temporal, convective-scale toy model, addressing challenges associated with spatial correlations and temporal evolution. Second, we systematically evaluate the performance of different CP variants using coverage, interval sizes, and the average interval score loss (AISL). Third, we explore how CP derived uncertainty estimates can be integrated into a data assimilation (DA) framework to improve ensemble representation. Our experiments are based on the 1D modified shallow water (MSW) model, a controlled framework for analysing convective-scale dynamics and DA algorithms. The MSW model represents horizontal wind velocity (u), fluid height (h) and rainfall (r), while conserving mass and enforcing non-negativity of rainfall. Convection is triggered through a threshold-based mechanism that introduces localized perturbations, allowing the model to reproduce regime transitions and delayed rainfall onset resembling real convective behaviour. Importantly, the MSW model strikes a balance between physical world and computational affordability, making it well suited for systematic uncertainty quantification studies. Two DA approaches are employed to generate analysis ensembles: the ensemble Kalman filter (EnKF) and a quadratic programming ensemble (QPEns) method that explicitly enforces physical constraints. We initiate a DA cycling procedure and train a convolutional neural network (CNN) that maps EnKF unconstrained analysis ensembles to QPEns constrained analysis ensembles at a lower computational cost. To ensure the CNN predictions with reliable uncertainty estimates, we apply the CP framework. CP provides distribution-free guarantees, ensuring that prediction intervals contain the true value with a user-specified confidence level. Theoretically, CP can be interpreted as a hypothesis testing procedure: for a given input, candidate outputs are evaluated for plausibility based on how conforming they are relative to previously observed data. This is quantified using a nonconformity score, which measures how unusual an input-output pair appears when compared to a calibration dataset. We focus on three widely used CP variants: standard conformal prediction (SCP), normalized conformal prediction (NCP), and conformalized quantile regression (CQR). SCP and NCP rely on point predictions and differ in how nonconformity scores are scaled, while CQR directly estimates lower and upper quantiles of the target distribution using quantile regression. Implementing CQR requires replacing the standard RMSE loss used in CNN training with the pinball (quantile) loss, enabling asymmetric uncertainty representation. This distinction is particularly relevant for variables such as rainfall, whose distributions are strongly non-Gaussian. To ensure statistically strong evaluation, CP experiments are performed using ten independent random seeds. In each experiment, five seeds are randomly assigned to calibration datasets and the remaining five to test datasets. This split is reshuffled across ten independent iterations, ensuring that calibration and test data never share initializations. Coverage probabilities and interval widths are averaged across all iterations, providing stable estimates relative to the nominal 90% target coverage. In addition, AISL matrices are computed to examine how uncertainty behaviour varies across the three prognostic variables (u, h, and r) and CP methods. Our results show clear differences between the CP methods. NCP consistently produces the narrowest and most stable prediction intervals for u and h while maintaining coverage close to the nominal level, making it the most efficient method for smoothly varying fields. SCP reliably achieves the desired coverage but at the cost of wider, non-adaptive intervals, rendering it a conservative baseline. CQR demonstrates competitive coverage and provides improved representation of rainfall uncertainty, particularly in terms of asymmetric intervals. Overall, NCP emerges as the most effective method for balancing sharpness and calibration, while CQR is advantageous when asymmetry in uncertainty representation is essential. Finally, we explore the integration of CP-derived uncertainty estimates into the DA process itself. Until this point, CP estimates are computed independently of DA and serve as supplementary uncertainty information for the CNN analysis. We extend this approach by perturbing each analysis ensemble member using a Gaussian distribution defined over the CP interval at each grid point, thereby generating CP informed analysis ensembles. Our results show that the EnKF benefits from the inclusion of CP-derived uncertainty, resulting in improved ensemble representation. By combining uncertainty-aware ML with physically consistent DA, this study contributes to the broader effort of developing trustworthy forecasting methodologies in meteorology. Our results demonstrate that ML approaches can improve point prediction accuracy and provide statistically sound uncertainty estimates. This work highlights CP as a promising pathway toward reliable uncertainty quantification in weather and climate science.

ABSTRACT. Peptide-based data storage is a promising alternative to conventional storage technologies, offering high density and longevity. However, the storage channel introduces insertion, deletion, and substitution errors, and must avoid problematic secondary structures during synthesis. This paper explores error-correcting coding techniques for peptidebased storage through simulation. We developed and tested two models: an LDPC-based model using simple and composite mappings, and a HEDGES-based model concatenated with Reed-Solomon codes. Our simulations demonstrate that the HEDGES-based model effectively corrects all types of indel errors at rates up to 2%, achieving code rates between 0.27 and 0.53. The LDPC-based model with a code rate of 0.5 reduces homopolymers but only corrects substitution errors. Our findings suggest that HEDGES codes are applicable to peptide-based data storage systems and outperform LDPC-based approaches for this channel.

ABSTRACT. In this paper, we propose a modular system architecture that leverages a Natural Language Processing approach to process diverse medical data streams. The framework integrates outputs through an expert system layer (based on fuzzy logic) to improve explainability and deliver reliable, explainable, and clinically actionable recommendations. Preliminary text-based analyses of cancer patient interviews illustrate the potential of this approach to differentiate cancer types and optimize diagnostic workflows. By addressing the challenges of integrating various technologies for text-based analysis, this work contributes to the development of AI systems capable of accelerating cancer diagnosis, enhancing clinical decision-making, and improving patient outcomes.

ABSTRACT. Biological function emerges from the organization of complex interaction networks spanning multiple scales, from molecular regulation to large-scale brain connectivity. Traditional graph-theoretic descriptors have been widely used to characterize such systems, yet they predominantly encode topological features and often overlook the geometric structure that governs network robustness, redundancy, and resilience to perturbations—properties that are critically altered in disease. Motivated by the need for quantitative descriptors that capture these latent organizational principles, we investigate network curvature as a geometric biomarker for discriminating healthy and pathological biological states. We consider two heterogeneous and conceptually distinct case studies: gene co-expression networks derived from normal and cancer tissues, and whole-brain connectome networks from healthy individuals and patients with major depressive disorder. By leveraging network curvature, we quantify disease-associated alterations in both local cohesiveness and global network organization. Across molecular and neurobiological domains, we observe consistent and systematic curvature shifts distinguishing healthy from diseased networks. These findings support network curvature as a general, domain-independent indicator of biological network dysfunction, highlighting its potential to reveal fundamental geometric signatures of disease that are not captured by conventional connectivity measures.

ABSTRACT. Network-based EEG analysis has recently attracted attention as a tool for studying alterations in the functional organisation of the brain associated with neurological conditions such as epilepsy and psychogenic non-epileptic seizures (PNES), which represent a diagnostic challenge due to their clinical similarity to epileptic seizures and the absence of epileptiform activity in EEG recordings. In this work, we propose a systematic framework for the construction and analysis of EEG-based functional brain networks, with the aim of evaluating the impact of phase-based connectivity estimators and graph construction strategies on network topology. Functional connectivity matrices were calculated using the phase locking value (PLV), phase locking index (PLI), and weighted phase locking index (WPLI) across multiple frequency bands. Brain networks were constructed using four binarisation and filtering approaches: median threshold, minimum connected component, minimum spanning tree, and orthogonal minimum spanning tree. Global and local graph-theoretic measures were analysed to characterise network organisation in healthy controls, epileptic patients, and PNES patients. The results demonstrate that methodological choices have a marked impact on network topology and critically affect the sensitivity of graph-theoretical measures in distinguishing between clinical groups.

ABSTRACT. Classical network sampling and graph exploration methods are typically designed to maximize coverage, reduce bias, or estimate local statistics. In many scientific and societal applications, however, decisions are governed not by individual edges but by global structural properties such as connectivity, cycle structure, and percolation regimes. This paper introduces a decision-centric, topology-aware framework for adaptive graph exploration in which sensing actions are selected to minimize uncertainty in task-relevant topological descriptors rather than generic reconstruction error. Persistent-homology-based summaries of a partially observed graph are treated as the state of a closed-loop decision process. Structural uncertainty is quantified directly in topological space using Wasserstein dispersion of persistence diagrams and entropy-based measures of Betti summaries, and exploration is formulated as an optimal experimental design problem over graph queries. The resulting control policy selects nodes, edges, or subgraph expansions that are expected to maximally reduce decision risk over competing topological hypotheses, including the existence of a giant component, the percolation state, or loop-rich versus tree-like regimes. Experiments on synthetic and real-world networks show that topology-guided exploration reaches decision-stable conclusions with substantially fewer queries than uniform and non-topological uncertainty-driven baselines, while maintaining comparable reconstruction fidelity. These results establish a computational framework for network sensing in which topology is elevated from a post hoc analysis output to a control state, enabling resource-efficient, decision-driven exploration of large and partially observed graphs.

ABSTRACT. The problem of wildfires is very common in many countries around the world, such as Canada and Chile. One example is the wildfire that occurred in Grasslands National Park in Canada in April 2013. Another example is a series of wildfires between February 1 and 5 in the Valparaiso region of Chile. Computer simulations of fires allow us to predict their future evolution, which, in turn, can enable better control of firefighters’ work. For this to be possible, the simulators developed must be efficient and effective. In our article, we analyze the parallel scalability of the WILDFIRES-IGA-ADS simulator on a concurrent server with shared memory. Our simulator employs the isogeometric finite element method with tensor product B-spline basis functions, explicit time integration schemes, and a linear computational cost solver resulting from the Kronecker product structure of the mass matrix.

ABSTRACT. We introduce the Automatically Variationally Stable FE (AVS-FE) method [1] for the numerical analysis of the response of soft nearly incompressible solids embedded with (small-scale) hard magnetic constituents [2], that are subject to steady state electromagnetic fields. To model the response of such materials, we use the time harmonic Maxwell equations in conjunction with the nonlinear nearly incompressible hyperelasticity model.

The AVS-FE method is similar in methodology to the Discontinuous Petrov-Galerkin (DPG) method [3], in the sense that it employs piecewise discontinuous test functions that are established such that the resulting FE discretizations are unconditionally stable. However, it distinguishes itself by using FE trial functions that have a global order of continuity as used by classical FE or mixed FE methods. As a result, there is no need to solve for flux variables over the mesh skeleton and the field variables, such as stress and displacement, are continuous throughout the computational domain. The AVS-FE method has been implemented by using a time-harmonic eddy current model to govern the electromagnetic field variables and a nearly incompressible neo-Hookean model to represent the response of the solid. Numerical results are presented of a simulation of a hard magnetic soft beam subject to a time-harmonic electro-magnetic field induced by a Helmholtz coil.

[1] V. M. Calo, A. Romkes, and E. Valseth, “Automatic Variationally Stable Analysis for FE Computations: An Introduction”. In: G. R. Barrenechea, J. Mackenzie (Eds.), Boundary and Interior Layers, Computational and Asymptotic Methods BAIL 2018, Springer International Publishing, pp. 19–43, (2020).

[2] R. Zhao, Y. Kim, S.A. Chester, P. Sharma, X. Zhao, “Mechanics of hard-magnetic soft materials”. In: Journal of the Mechanics and Physics of Solids 124 (2019), pp. 244–263. issn: 0022-5096. doi: https://doi.org/10.1016/j.jmps.2018.10.008. url:https://www.sciencedirect.com/science/article/pii/S0022509618307646.

[3] L. Demkowicz and J. Gopalakrishnan, “Analysis of the DPG Method for the Poisson Equation”, SIAM Journal on Numerical Analysis, Vol. 49 (5), pp. 1788–1809, (2011).

ABSTRACT. This paper presents the implementation of CO2 sequestration simulation using the IGA-ADS (IsoGeometric Analysis using Alternating Directions Solver) framework. We formulate the problem based on Darcy’s Law for two-phase flow in porous media, solving coupled equations for pressure and gas saturation. The spatial discretization employs B-spline basis functions with explicit time stepping for saturation and implicit solving for pressure using the MUMPS direct solver. We demonstrate the solver’s capabilities on three reservoir configurations with heterogeneous porosity and permeability fields. Performance analysis on mesh sizes up to 256x256 points shows that pressure computation dominates the computational cost. The solver successfully handles physically realistic parameters and shows promising scalability on HPC resources. Future work includes inverse problem formulation for optimal injection well placement and hybrid solvers combining IGA-ADS with physics-informed neural networks.

ABSTRACT. One of the commonly used methods for performing simulations is the high-order polygonal discontinuous Petrov-Galerkin method. In the paper, the graph grammar approach for modeling mesh refinements for the polygonal discontinuous Petrov-Galerkin method is introduced. The computational mesh is represented by hypergraphs, and mesh adaptations are represented by graph grammar productions

ABSTRACT. Background:The long-term memory of the disease dynamics and seasonal forcing due to climatic impacts, along with the delayed responses, have made it difficult to accurately predict dengue outbreaks. Traditional compartmental models are based on the Markovian assumption, and the data-driven models may not be easily interpretable from an epidemiological perspective. Fractional calculus provides a sound framework to endow models with memory; nevertheless, combining fractional methods with machine learning techniques requires special attention. Methods:A fractional physics-informed neural network (FPINN) based on a SEIR-SI dengue compartment model, using Caputo fractional derivatives and L1 time discretization, is proposed. An ensemble averaging of five pre-defined fractional orders is simulated to model the memory effects. The model includes multi-scale temporal encoding, residual learning, and time-dependent transmission rates adjusted by environmental factors. The adaptive physics loss weighting balances the physics and data loss functions during training. Results:The effectiveness of the proposed approach is demonstrated using the weekly dengue incidence data from San Juan, Puerto Rico. The fixed-alpha FPINN ensemble achieves strong predictive performance, with a test-set MAE of 6.25 and an R^2 of 0.896. In comparison to data-driven time series baselines, the ensemble reduces the MAE by 29.1% relative to a Long Short-Term Memory (LSTM) model and by 36.2% as compared to a Temporal Convolutional Network (TCN). It also improves MAE by 6.7% over the best individual fixed-alpha model (alpha=0.95). In addition to enhanced accuracy, the ensemble provides well-calibrated predictive uncertainty as well. Conclusion:The findings suggest that the combination of fractional-order memory with physics-informed learning constitutes an effective and interpretable model for dengue forecasting. The fixed-alpha ensemble FPINN method improves the stability, accuracy, and uncertainty estimation, providing a sound model for epidemic forecasting.

ABSTRACT. Predicting the clinical outcomes in patients with antineutrophil cytoplasmic antibody (ANCA)-associated vasculitis (AAV) remains a challenge and early identification of patients who are at risk of severe disease course is crucial. To address this, we applied machine learning (ML) and federated learning (FL) techniques to the POLVAS dataset – the largest multicenter clinical database of vasculitis cases in Poland and one of the largest AAV datasets in Europe. Our goal was to predict the key outcomes: increased risk of death and the need for renal replacement therapy (RRT), an independent risk factor of death. We also analysed the significance of individual input features for the predictive capabilities of our models. Furthermore, we compared the performance of centralized model with FL models, which allow for data privacy preservation – a key factor given the highly sensitive medical data involved. We achieved a prediction performance of 0.86 AUC for RRT prediction, and 0.81 AUC for death prediction using a centralized approach and 0.86 weighted mean AUC for RRT prediction and 0.80 for death prediction with the FL approach. The presented results show that FL can effectively predict the risk of RRT and mortality in AAV patients, addressing privacy concerns without compromising the accuracy.

ABSTRACT. This paper presents a data-driven framework for optimizing descriptive norms in medical contexts, bridging micro-level agent behaviors and macro-level dynamics using partial differential equations (PDEs) and reinforcement learning. Descriptive norms are modeled as opinion distributions in a continuous space, evolving through diffusion and migration terms influenced by perceptual kernels and external potentials. The approach extends prior models by integrating Gaussian Mixture Models (GMMs) from real-world COVID-19 data. Top-down optimization employs Q-learning to align opinion distributions with target GMMs, reshaping clinical propensities (e.g., azithromycin usage) post-guidelines, achieving 82\% similarity and faster convergence than random parameters. Bottom-up optimization uses causal effects from double machine learning on indicators like mortality, lymphocyte count, and consciousness decline, guiding PDE parameters to emergent norms that outperform baselines (e.g., 36.8\% reduced consciousness decline)

ABSTRACT. Type 2 Diabetes (T2D) is a heterogeneous metabolic disorder driven by the interplay between endocrine stress and glucose regulation. We adapt an 18-variable ordinary differential equation (ODE) model of the hypothalamic-pituitary-adrenal (HPA) and metabolic systems to a UK Biobank pilot cohort (N=100) representing metabolic extremes (50 healthy, 50 T2D). Scaling mechanistic models is challenged by inherent population metabolic diversity. To address this, we implement an individualized scaling framework anchoring glucose uptake and hepatic production to baseline fasting glucose. Results demonstrate that this approach reproduces clinical bimodal distributions and phenotypic attractors without manual refitting. By achieving statistical parity using only baseline glucose and mental health data, this framework enables scaling neuroendocrine-metabolic models to the full UK Biobank for virtual trials and personalized interventions.

ABSTRACT. This papers develops quantum-classical hybrid algorithms for Darcy flow and the Poisson equation. We pay attention to physical-property preserving. Weak Galerkin finite element methods are developed and implemented as Python code utilizing Qiskit libraries. Numerical experiments for 1-dim problems are conducted to examine fidelity, efficiency of iterations, and errors.

ABSTRACT. Red blood cell distribution width (RDW), a routinely measured biomarker of red blood cell (RBC) size heterogeneity, is strongly associated with cardiovascular events, cancer progression, and mortality. However, its mechanistic role remains poorly understood due to limited experimental control and the computational difficulty in modeling heterogeneous cell populations. Therefore, we develop and verify a high-resolution, massively parallel computational framework that explicitly models physiologically relevant RDW levels and quantifies their impact on blood flow and cell transport. Using large-scale parallel simulations across representative vascular geometries, including straight vessels, bifurcations, stenoses, and expansions, we show that bulk hemodynamics are largely insensitive to RDW variation, whereas RBC transport and circulating tumor cell (CTC) margination are strongly affected. These results highlight the importance of RBC heterogeneity in cell-level transport modeling and provide mechanistic insight into the links between RDW and disease risk. Our framework enables the scalable incorporation of patient-specific RDW into predictive simulations, supporting precision modeling in cardiovascular and oncological applications.

ABSTRACT. Solving the p-Laplace equation numerically is challenging due to inherent degeneracies and singularities. We present a new hybrid strategy combining Continuous Data Assimilation, Picard iteration with Anderson Acceleration, and Newton’s Method. By incorporating coarse-mesh observations via a nudging term, the scheme provides essential spatial regularization. Our results show that this integration stabilizes nonlinear iterations and significantly improves convergence speed. Notably, the CDA-enhanced solver captures localized singular behaviors and suppresses oscillations in highly nonlinear regimes where standard methods typically fail.

ABSTRACT. This paper presents a novel framework for pore-scale simulation of Stokes flow in complex porous media, with particular emphasis on the systematic construction of the global stiffness matrix (A matrix). The methodology integrates three key innovations: (i) an embedded boundary formulation on structured Cartesian grids that represents irregular solid grains through a cell-type classification system, eliminating the need for conforming mesh generation; (ii) a unified data structure (AVariable) that simultaneously stores field values and their Jacobians, enabling forward-mode automatic differentiation through operator overloading; and (iii) a systematic restriction procedure that reduces the global system from the full Cartesian grid to only active fluid cells. The framework architecture follows object-oriented design principles with clear component responsibilities, enabling extensibility to additional physics while maintaining computational efficiency through sparse matrix operations.

ABSTRACT. Abstract. Pore-network models (PNMs) are powerful tools for simulating fluid flow and transport in porous media, with applications spanning petroleum engineering, hydrology, and materials science. Classical PNMs simplify porous media into interconnected pore-throat networks and model flow using empirical resistance relationships (e.g., Hagen- Poiseuille flow). However, these models often overlook the anisotropic nature of pore-scale resistance and struggle with large-scale simulations due to computational limitations. This paper first provides a quick review of classical pore-network models, including their formulation, assumptions, and limitations. Then, we introduce a novel tensor resistance-based pore- network model that incorporates anisotropic flow resistance via tensor algebra and leverages domain decomposition to enhance computational efficiency for large-scale porous media systems. The new model decomposes the porous domain into subdomains, solves for tensor resistance within each subdomain, and couples subdomain solutions via interface conditions. Validation against experimental data and classical PNM results demonstrates that the tensor resistance PNM captures anisotropic flow behavior with 15-20% higher accuracy while reducing computational time by 40% for large networks.

ABSTRACT. This paper presents a fully automated machine learning framework for generating high-complexity, free-form planar antenna designs conditioned on a specified operational frequency range. The proposed system eliminates human engineer bias by combining generative and predictive neural networks into an end-to-end antenna design pipeline. Its core component is a conditional variational autoencoder (CVAE) that produces pseudorandom yet physically feasible antenna geometries guided by frequency-band labels. Three latent-space sampling strategies—posterior, linear interpolation, and convex interpolation—are explored to balance design diversity and conditional consistency. Auxiliary models implemented as multilayer perceptrons (MLPs) and a convolutional neural network (CNN) perform iterative verification of generated structures, including viability detection, label classification, and reflection response estimation, without the need for costly full-wave electromagnetic simulations. The framework is trained and validated using two large-scale datasets of pseudo-random antennas—1DDesign-to-Label-1.8.3 and 1DDesign-to-Response-2.8—comprising a total of over 320,000 geometries with corresponding performance labels and reflection responses. Experimental validation using independent CST Microwave Studio simulations demonstrates that the best model achieves an exact-match accuracy of 0.634, a relaxed-match accuracy of 0.769, a macro-average accuracy of 0.947, and a macro-average F1 score of 0.756. The average generation and verification time per design is 0.703 s. The proposed approach enables rapid, bias-free antenna design exploration and candidate preselection, significantly reducing design time as well as dependence on expert knowledge and EM computation. The framework’s modular structure allows further enhancement, making it a versatile foundation for next-generation computer-aided antenna engineering.

ABSTRACT. This research proposes a novel methodology for rapid multi-objective (MO) electromagnetic (EM)-driven design of antenna systems. Our approach inte-grates machine learning (ML) with neural network metamodels. At each itera-tion, the surrogate, optimized using an MO-oriented evolutionary algorithm, generates multiple candidate Pareto-optimal designs. These solutions are further refined locally using approximate response Jacobians estimated from the surro-gate model. The metamodel itself is continuously updated with EM data col-lected during the optimization process. By combining surrogate-driven explora-tion with gradient-based refinement, the method accelerates convergence and enables more accurate identification of Pareto fronts compared to conventional techniques. The framework has been validated on two planar antennas, show-casing excellent cost efficiency: the mean MO expenses are under 300 EM sim-ulations, corresponding to roughly 80% savings relative to offline surrogate-assisted optimization, and substantial speedup over baseline ML methods lack-ing gradient enhancements. Moreover, the proposed technique produces supe-rior Pareto fronts in terms of both distribution uniformity and solution quality, as confirmed by dominance-based performance metrics.

ABSTRACT. This paper introduces Bayesian optimization into the training process of a neural network classifier model that allows for determining the operating range of freeform microstrip antennas for randomly generated designs. The proposed approach addresses the challenge of manual hyperparameter tuning and leads to improved classification accuracy compared to existing implementations.. Moreover, the optimization process includes selected architectural parameters, resulting in a more generalizable solution that does not require expert knowledge in electromagnetic simulation procedures. As a result of the conducted experiments, a multilayer perceptron network architecture was developed, achieving high classification accuracy for frequency band intervals of the designed microstrip antennas. The application of Bayesian optimization simul-taneously allowed for the effective minimization of the number of model archi-tecture parameters.

ABSTRACT. This work introduces a convolutional neural network capable of estimating the reflection response of 50-vertex free-form planar antennas across the 3–8 GHz band. By directly inferring electromagnetic behavior from geometry, the proposed approach eliminates reliance on expert-driven design rules and electromagnetic simulations. This is the first study employing machine learning to predict the complete frequency-dependent reflection response of complex free-form antenna structures. The contributions include: (i) a publicly available dataset comprising 108,710 pseudo-random antenna geometries and their corresponding reflection responses; (ii) a convolutional neural network that maps complex antenna shapes to wideband reflection characteristics; and (iii) an experimental analysis covering performance metrics and comparison with alternatives. Antenna geometries are encoded as a 2×51 tensor of vertex coordinates, while the network outputs 251 reflection coefficient values sampled at 20 MHz intervals over the 3–8 GHz range. Model hyperparameters were determined using Bayesian optimization, and robustness was assessed through 10-fold cross-validation and evaluation on an independent test set. The proposed model attains a mean squared error of 3.789, a mean absolute error of 0.902, a root mean squared error of 1.947, and a coefficient of determination of 0.437 over the considered frequency band. With an average inference time of 6.06 ms, the framework supports rapid screening of large design populations, effectively removing electromagnetic simulation as a bottleneck during early-stage exploration.

ABSTRACT. Modern oncology remains constrained by a persistent methodological divide: biological and clinical knowledge is predominantly articulated in narrative form, while predictive treatment planning requires formal, executable models capable of simulating tumor evolution under therapeutic pressure. Here, we argue that the emergence of Large Language Models (LLMs) enables a fundamental reconfiguration of this paradigm. Their capability motivates the concept of conversational modeling, in which modeling pipelines are initiated, shaped, and iteratively refined through natural language interaction rather than explicit mathematical programming. We introduce the Language-Driven Therapy Design (LDTD) framework and assess whether contemporary LLM-based tools can autonomously construct predictive models of cancer dynamics under therapy starting solely from narrative descriptions. By comparing LLM-generated models with professional, expert-curated predictive repositories for anticancer treatment, we demonstrate that narrative-driven models already achieve comparable predictive performance. These findings suggest that conversational modeling can substantially lower the barrier to advanced computational oncology, pointing toward a broader democratization of scientific modeling in data- and knowledge-constrained domains.

ABSTRACT. High-fidelity 3D human scans are essential in various engineering applications, including entertainment and medicine. Recent advancements in neural scene representation techniques, such as Neural Radiance Fields (NeRF), offer promising alternatives to traditional 3D reconstruction methods like Structure from Motion (SfM). This paper contributes to Computational Science by presenting a comparative study of two computational paradigms for 3D human reconstruction—NeRF and SfM—with particular emphasis on their temporal stability. The analysis focuses on the reconstruction of a detailed human body by evaluating reconstructions carried out across three consecutive time frames for each test scene. The evaluation employs extensive point cloud analysis based on four quantitative criteria: reconstruction completeness, resolution, surface smoothness, and point cloud consistency. The results demonstrate that NeRF surpasses photogrammetry in capturing fine volumetric structures and reconstructing regions that are only visible in a limited subset of input views, yielding a reconstructed area approximately 9% larger in the analysed full-body scans. Conversely, SfM demonstrates superior performance in terms of surface smoothness and temporal stability, producing reconstructions with surface smoothness approximately 45% higher than those obtained with NeRF. The complementary strengths of these methods highlight opportunities for future research to leverage both artificial intelligence and traditional techniques for enhanced 3D reconstruction.

ABSTRACT. Clay‑based 3D printing has emerged as a low‑carbon alternative to cementitious additive manufacturing in construction. However, pore space, inhomogeneities, shrinkage, and crack formation in clay-sand-water mixtures are still poorly understood, particularly with respect to the influence on the structural integrity of clay-printed elements. Recent studies on characterizing the internal structure of clay‑printed elements predominantly rely on surface inspections and destructive testing. Quantitative, non-destructive methods focusing on detecting and localizing defects within the internal structure of clay-printed elements have so far largely been unexplored. This paper investigates the use of mobile X‑ray computed tomography (CT) to analyze the internal structure of clay-printed elements made of clay-sand-water mixtures used in additive manufacturing. In particular, a clay-printed sample is scanned, together with manually extracted wet and dried samples (representing the initial and final material states in clay-based 3D printing), using the mobile CT system ENCI (extracting non-destructively cuneiform inscriptions). The reconstructed volumes of the CT scans are used to investigate the pore space of each sample. Renderings are used to compute the spatial porosity distribution and to identify potentially weak zones affecting the structural integrity of each sample. The results show that CT is capable of providing valuable information on the porosity of clay-printed elements.

ABSTRACT. This paper investigates the adaptation of classical rule-based classification algorithms (LEM2, PRISM, and RIPPER) to the federated learning paradigm. We propose a federation discretization framework that establishes global bin boundaries using aggregated feature statistics, ensuring consistent rule vocabulary across clients without sharing raw data. We conducted experiments on five medical datasets, which had varying numbers of clients (2–5) and discretization granularities (2–7 bins). The results show that the global model consistently achieves high levels of predictive accuracy and interpretability, outperforming averaged results of local models by 7.5–9.1% in balanced accuracy across algorithms. Notably, federated LEM2 significantly outperformed even its centralized counterpart (p < 0.001), while federated RIPPER and PRISM achieved statistical parity with centralized baselines. The resulting rulesets are compact (8.8–54.8 rules) and interpretable (1.44–8.12 conditions per rule), with coverage ranging from 19% to 26% depending on the algorithm. These results demonstrate that federated rule learning can provide both transparency and competitive performance for clinical applications where data is limited and privacy is essential.

ABSTRACT. Circulating tumor cells (CTCs) interact with the vascular endothelium under hemodynamic forces that influence where metastatic seeding occurs. Wall shear stress (WSS) has been shown to regulate adhesive signaling, yet its mechanistic role in shaping spatial patterns of CTC arrest within complex vascular networks remains poorly understood. In this work, we present a three dimensional multiscale simulation framework that couples tissue scale blood flow with receptor ligand adhesive dynamics whose kinetics depend on local WSS. We evaluate three modes of WSS regulation, including modulation by WSS magnitude, modulation by the local spatial gradient of WSS, and modulation by peak upstream WSS gradient within a region of influence. Using physiologically realistic vessel geometries and flow conditions, we track thousands of CTCs across tissue scale vascular networks using a high throughput adaptive refinement strategy executed on the Aurora supercomputer. Across a curved microvessel and a reconstructed bioprinted double bifurcation network, WSS magnitude dependent adhesion produces strong spatial biases in CTC arrest. In the bioprinted network, inclusion of WSS regulated adhesion leads to a thirty one fold increase in active receptor engagement in the lower entry fork compared to a non WSS modulated baseline, consistent with experimentally observed regional enrichment. These results demonstrate that local WSS profiles act as a physical driver of preferential CTC attachment and motivate inclusion of shear regulated adhesion in predictive models of metastatic transport.

ABSTRACT. This work proposes the use of Adaptive Resolution Schemes (AdResS) in molecular dynamics simulations that include three-body interactions described by the Axilrod-Teller-Muto (ATM) potential.It is known that the computation of three-body forces greatly increases the number of calculations, which is the motivation to investigate appropriate speedup techniques. In this work, AdResS is used to allow for the computation of three-body interactions only within a portion of the simulation domain. This is implemented using force resolution regions, such that across the domain three-body forces are introduced smoothly from a pure two-body force region. The proposed scheme is tested using a relevant number of single-site, Lennard-Jones molecules in homogeneous case studies. In order to measure performance, the achieved speedup is expressed in terms of the geometric configuration of the force resolution regions where the three-body forces are calculated. These tests were carried out at node-level to measure the parallel and AdResS speedups using several force resolution region configurations with increasing portions of three-body presence. The effects of AdResS on meaningful ensemble averages and correlation functions is then investigated. It is shown that AdResS, for three-body interactions, introduces inconsistencies in simple ensemble averages, like the temperature and density. For this reason, it is required to use correction schemes to compensate for the unwanted effects. In this work we make use of the thermodynamic force for the correction of artificial density gradients in our simulations. These gradients are a well-known source of error in density measurements in AdResS, which originate from a pressure difference introduced by the coarse-grained interactions.

ABSTRACT. Cardiovascular diseases remain the predominant cause of death worldwide. Although computational fluid dynamics is an effective tool for predicting hemodynamics and drug particle transport in arteries, its clinical adoption remains limited due to the complexity of the workflow and computational limitations. This paper presents a scalable approach for predicting hemodynamic parameters and particle exit fractions in large networks from results obtained for ideal bifurcations. A total of 828 3D CFD simulations, combined with particle tracking, were performed on a high-performance computing environment with variations of inlet velocity, particle density, inlet and outlet diameters, and bifurcation angles on Y and T-shaped bifurcations. The simulation dataset was used to train three ML algorithms: Linear Regression, k-Nearest Neighbour, and Random Forest Regression for the prediction of velocity and particle exit fraction at the bifurcation outlets. Larger vascular networks were built by successively combining predictions from single bifurcation variants, thereby reducing computational costs. Our model successfully predicted the distributions of particle exit fraction and velocity in larger networks with better accuracy and reduced runtime than a 1D model of a realistic hepatic artery geometry and a reconstructed model.

ABSTRACT. The vascular endothelium forms a selective barrier lining all blood vessels, tightly regulating the exchange of nutrients, oxygen, and waste between blood and tissues. The integrity of this barrier depends on the dynamic equilibrium between endothelial cell-cell contacts, cytoskeletal organization, and intracellular signalling. Loss of endothelial barrier integrity and reduced blood perfusion are hallmarks of inflammatory diseases such as Chronic Kidney Disease and Long COVID. Intracellular signalling pathways regulating vascular endothelium integrity rely on RhoGTPases, which act as molecular switches, controlling cytoskeletal reorganization and cell shape. Two closely related members of RhoGTPases, RhoA and RhoB, act as negative regulators of endothelial integrity, yet operate through fundamentally different regulatory mechanism and on different time scales. RhoA is regulated by rapid, signal-dependent GEF/GAP-mediated cycling and GDI-controlled sequestration, producing a fast and transient loss of barrier integrity in response to acute stimuli. RhoB, in contrast, lacks GDI regulation and maintains constitutive low-level activity through continuous protein synthesis and ubiquitin-dependent degradation cycles, linking its excess accumulation under inflammatory conditions to slow, persistent barrier disruption. Remarkably, RhoGTPase networks can display bistability and hysteresis, properties that have been proposed to underlie transitions from acute to chronic inflammatory states. How these two signalling modes interact, and how bistability and history-dependent behaviour connect to barrier function, remain poorly understood. To address these open questions, we propose a multiscale modelling framework for endothelial cell monolayers, covering three different scales: (i) the molecular and subcellular scale, capturing the reaction networks governing RhoA and RhoB activity within single cells; (ii) the cellular scale, describing the spatio-temporal distribution of protein activity across individual cells, incorporating diffusion and spatial heterogeneity across neighbouring cells; and (iii) the monolayer scale, integrating multicellular dynamics to describe collective barrier behaviour. Here we focus on the first scale, which constitutes the biological and mathematical foundation for the entire multiscale pipeline. At the molecular scale, a minimal ODE-based model has been developed to capture the core biological distinctions between RhoA and RhoB regulation. The model integrates GEF/GAP/GDI-mediated signal-dependent kinetics for RhoA alongside synthesis and ubiquitin-dependent degradation kinetics for RhoB. Analytical and numerical stability analysis have been used to inspect the dynamical behaviour of the system, identifying distinct dynamical regimes, including bistability and oscillatory behaviour. Yet, numerical explorations are hindered by a limited availability of robust experimental data, a challenge that is particularly relevant in the biomedical sciences. We developed a framework that addresses this challenge by providing dynamical insights in the absence of extensive empirical data. Our approach applies stability analysis as a tool to explore the landscape of possible dynamical behaviours. Unlike existing methods that prioritize broad, non-targeted classifications of behaviour in the whole parameter space, our approach uses stability analysis to efficiently identify parameter regions consistent with specific dynamical objective, such as bistability or sustained oscillations, relevant to the biological hypotheses under investigation. This target-oriented strategy sacrifices exhaustive coverage but enables more efficient numerical exploration, reducing both computational demands and time for problem bounded by the curse of dimensionality. The proposed algorithm has been effectively tested on analytically-tractable systems of ODE and applied to real case scenarios. By generating plausible hypothesis or excluding unpromising parameter regions, this method offers a complementary tool for biomedical modelling, bridging the gap between theoretical analysis and experimental decision-making in data-scarce contexts. Ultimately, the integration of these numerical strategies into a coherent multiscale pipeline constitute a significant step in linking molecular signalling to the endothelial barrier dysfunctions characteristic of inflammatory diseases.

ABSTRACT. Stimuli responsive gels offer an attractive route to design actuators [1], and smart reactors [2]. However, for cross linked gel systems the link between molecular level polymer-solvent interactions and macroscopic swelling and shrinkage due to solvent responsiveness remains under explored. In this work, we implement quantum chemistry based thermodynamic modeling via Conductor-like Screening Model for Realistic Solvents (COSMO-RS), and mesoscale dissipative particle dynamics (DPD) simulations to understand the polymer-solvent interactions and establish a multiscale framework for designing responsive gel networks.

To understand the solvent responsiveness of gels during solvent exchange process, alginic acid hydrogels are used as a model system. Initially, hydrogels (i.e. with water in the gel matrix) are produced. Solvent exchange is performed using organic solvents of different natures to form lyogels (i.e. with organic solvent in the gel matrix) that probes solvent responsive behavior. For such lyogels, volume of the gels is determined before and after solvent exchange. This enables to understand volumetric swelling or shrinkage ratios of lyogels relative to the hydrogels. The data reveal a rich, non monotonic response: some alcohols cause moderate shrinkage, while highly interacting aprotic solvents such as dimethylsulfoxide and dimethylformamide induce pronounced swelling. This highlights the importance of gel-solvent interactions in different solvents.

To qualitatively predict these trends, COSMO-RS calculations are performed. Polymer chain of alginate comprises of alpha-L-guluronic acid (G) and beta-D-mannuronic acid (M) monomeric units of varying ratios. For the COSMO-RS calculations, monomers, dimers, trimers and hexameric sequences of G and M are chosen. For each monomer-solvent and oligomer–solvent pairs, the infinite dilution activity coefficients (IDACs) are computed [3]. From these calculations it is found that solvents that experimentally promote gel swelling systematically exhibit lower IDACs. This provides a qualitative understanding of polymer-solvent interactions through the predictive thermodynamic model that requires only a single polymer chain to be executed.

While the thermodynamic analysis provides insights on equilibrium volume changes, many applications rely on the rate at which gels respond to a change in solvent quality. To address this, a DPD model of gels, represented as randomly cross linked bead–spring networks in explicit solvent is developed. Gels are first equilibrated in a good solvent, which corresponds to strong polymer–solvent affinity, and then instantaneously exposed to a bad solvent with poor polymer-solvent affinity. This mimics the experimental solvent exchange process. This model helps with the exploring the effect of porosity on the shrinkage kinetics and the resultant gel structure. It is observed that different porosities result in different pore growth rates and shrinkage kinetics. In particular, shrinkage kinetics of low porosity gels corresponds well with Tanaka’s theory [4]. However, increasing porosity causes delay in shrinkage because of larger pores and slower pore growth rate.

To connect thermodynamic solvent quality to polymer chain length, we complement the DPD simulations with COSMO‑RS data for alginate oligomers of varying chain lengths. By computing radii of gyration for chains with different degrees of polymerization and comparing corresponding oligomer IDACs, a thermodynamic bridge between COSMO-RS derived measures of solvent quality and the interaction parameters from DPD that control the swelling–shrinkage kinetics of cross linked gels could be established. This yields a unified multiscale description of solvent responsive gel behavior.

Bibliography [1] Cheng, F., Chen, H., & Li, H. (2021). Recent progress on hydrogel actuators. Journal of Materials Chemistry B, 9(7), 1762–1780. https://doi.org/10.1039/D0TB02524K [2] Hu, X., Karnetzke, J., Fassbender, M., Drücker, S., Bettermann, S., Schroeter, B., Pauer, W., Moritz, H.-U., Fiedler, B., Luinstra, G., & Smirnova, I. (2020). Smart reactors – Combining stimuli-responsive hydrogels and 3D printing. Chemical Engineering Journal, 387, 123413. https://doi.org/10.1016/j.cej.2019.123413 [3] Eckert, K. M., Müller, S., Luinstra, G. A., & Smirnova, I. (2024). Exploring pNIPAM lyogels: Experimental study on swelling equilibria in various organic solvents and mixtures, supported by COSMO-RS analysis. Fluid Phase Equilibria, 586, 114182. https://doi.org/10.1016/j.fluid.2024.114182 [4] Tanaka, T., & Fillmore, D. J. (1979). Kinetics of swelling of gels. The Journal of Chemical Physics, 70(3), 1214–1218. https://doi.org/10.1063/1.437602

ABSTRACT. Accurately modeling multiscale spatiotemporal flows remains a central challenge in fluid mechanics, particularly when long-horizon temporal stability and fine-scale spatial fidelity must be achieved simultaneously. Existing operator-learning models often suffer from autoregressive instability, while diffusion-based approaches lack explicit mechanisms for causal time evolution. We introduce Uni-Flow, a unified autoregressive-diffusion framework that explicitly decouples temporal evolution from spatial refinement. A physics-informed autoregressive module evolves low-resolution latent dynamics to preserve large-scale structure and ensure stable long-horizon rollouts. A conditional diffusion refinement module then reconstructs high-resolution physical fields, recovering fine-scale structures through a small number of deterministic denoising steps. This separation resolves the long-standing tension between temporal stability and spatial accuracy in data-driven flow modeling. We validate Uni-Flow across progressively complex systems: (i) two-dimensional Kolmogorov turbulence at Re = 1000, (ii) three-dimensional turbulent channel inflow generation at Re_\tau= 180, and (iii) patient-specific pulsatile aortic flow simulated using a high-fidelity lattice Boltzmann solver. Across all cases, Uni-Flow preserves invariant statistics, including energy spectra, probability density functions, and autocorrelation structures, while significantly outperforming purely autoregressive baselines in long-term stability. In the cardiovascular case, Uni-Flow reduces a multi-hour, multi-GPU hemodynamic simulation to second-scale inference on a single GPU, achieving task-level over five orders of magnitude wall-clock acceleration while maintaining physiologically consistent pressure dynamics. These results establish Uni-Flow as a general and architecture-agnostic framework for generative modeling of complex multiscale flows, enabling stable, high-resolution, and computationally efficient surrogate simulations across turbulence and physiological fluid dynamics.

ABSTRACT. Parallel programming remains notoriously difficult, primarily due to the cognitive load of managing synchronization and non-deterministic state across distributed nodes. While Behaviour-Oriented Concurrency (BoC) simplifies this on single-system architectures by unifying parallelism and coordination, its reliance on shared memory limits its horizontal scalability. This paper presents ARBOC (Arbitrated Behaviour Oriented Concurrency), a distributed runtime that extends BoC across multiple systems while preserving its core semantic guarantees. ARBOC separates coordination from execution through centralized arbitration, scheduling behaviours via a dependency-aware directed acyclic graph and executing them on distributed workers without exposing programmers to message passing, distributed locks, or explicit synchronization. We describe the execution semantics, architecture, and implementation of ARBOC, and evaluate it on a range of parallel and coordination-intensive benchmarks. Results show that ARBOC scales effectively for parallelisable compute-heavy workloads while retaining predictable Behaviour under contention, demonstrating that BoC’s abstraction naturally generalizes to distributed environments.

ABSTRACT. This article presents the design and implementation of a low-cost modular autonomous navigation system for conventional wheelchairs, focusing on a distributed architecture. Unlike traditional centralised systems, the proposed solution leverages the ROS2 Data Distribution Service (DDS) protocol to establish dynamic node discovery between an external processing unit and a local unit. This approach allows SLAM and complex trajectory planning tasks to be delegated to an external computing station, while the local unit manages the acquisition of data from the LD-19 LiDAR and the control of the actuators in real time using an STM32 board. The navigation logic employs a sequential waypoint approach in which a boundary system is implemented to prevent skipping points and maintain a straight trajectory. The results validate that this distributed configuration ensures low latency and high scalability, offering a robust and economical technical alternative for transforming standard mobility devices into intelligent assistance platforms.

ABSTRACT. We propose a lightweight hybrid approach to clickbait detection that combines OpenAI semantic embeddings with six compact heuristic features capturing stylistic and informational cues. To improve efficiency, embeddings are reduced using PCA and evaluated with XGBoost, GraphSAGE, and GCN classifiers. While the simplified feature design yields slightly lower F1-scores, graph-based models achieve competitive performance with substantially reduced inference time. High ROC--AUC values further indicate strong discrimination capability, supporting reliable detection of clickbait headlines under varying decision thresholds.

ABSTRACT. Quantum optimization has grown rapidly in the past decade, supported by advances in both hardware and software frameworks such as Quadratic Unconstrained Binary Optimization (QUBO). In contrast, Constraint Programming (CP) provides a highly expressive declarative approach for modeling combinatorial problems—an expressiveness that QUBO lacks. Bridging these two worlds requires systematic methods to translate CP models into quantum‑compatible formulations. While some automated conversions exist for linear arithmetic and certain global constraints, extensional constraints—capable of expressing any finite‑domain relation—have not yet been addressed. This paper introduces a method for automatically translating table constraints, a widely used form of extensional constraint, into QUBO. The conversion is efficient: binary tables require no ancillary variables, and tables with $l$ $n$-ary tuples require only $l$ additional Boolean variables per element constraint.

ABSTRACT. With fault-tolerant quantum computing on the horizon, there is growing interest in applying quantum computational methods to data-intensive scientific fields like remote sensing. Quantum machine learning (QML) has already demonstrated potential for such demanding tasks. One area of particular focus is quantum data fusion—a complex data analysis problem that has attracted significant recent attention. In this work, we introduce an automated QML (AQML) approach for addressing data fusion challenges. We evaluate how AQML-generated quantum circuits perform compared to classical multilayer perceptrons (MLPs) and manually designed QML models when processing multisource inputs. Furthermore, we apply our method to change detection using the multispectral ONERA dataset, achieving improved accuracy over previously reported QML-based change detection results.

ABSTRACT. The potential capabilities of quantum computers motivated the development of cryptographic protocols suitable for securing communication against the adversaries with access to large fault-tolerant quantum computers. However, even though current quantum computers are limited in terms of size and precision, they can still be useful for finding loopholes and weaknesses in the post-quantum cryptographic protocols. In this work, we present an attempt to utilized the capabilities of Quantum Generative Adversarial Networks (QGANs), one of the promising architectures used in quantum machine learning, for this purpose. We describe an example application of QGAN architecture for the purpose of loading the probability distribution of the hash-based digital signatures into the memory of a quantum computer. Our results confirm that near-term hybrid quantum-classical methods posses capabilities required for this purpose. The presented approach can be used as a first step in the workflow enabling the utilization of quantum computing for attacking post-quantum cryptographic primitives.

ABSTRACT. We introduce a cavity-enhanced optical architecture for collective quantum pro-cessing in which logical qubits are encoded in the polarization subspace of recir-culating intracavity modes. The physical carrier and computational degree of free-dom are explicitly separated: harmonic cavity bundles provide a stable resonant substrate, while programmable polarization transformations implement single-qubit operations. A polarization-selective nonlinear interaction in the entangle-ment region generates tunable controlled-phase gates, enabling a universal gate set. A parameter-scaling analysis shows that order-unity conditional phases are attainable in centimeter-scale cavities using experimentally accessible solid-state nonlinear media, without requiring extreme nonlinear coefficients, millisecond photon lifetimes, or sub-hertz laser stabilization. The results indicate that resonant recirculation provides a physically plausible platform for cavity-based collective quantum architectures.

ABSTRACT. Neural operators offer promise for efficient solutions to time-dependent partial differential equations, but face challenges in long-term prediction due to complex dynamics, gradient accumulation, and error propagation. To address these limitations, we propose a novel curriculum learning strategy, temporal weighted loss. This method mitigates overfitting to early dynamics by dynamically adjusting the weights applied to the loss across the temporal sequence, prioritising initial time steps during early training. This approach enhances model generalisation and prediction accuracy for extended time horizons, demonstrating improved performance compared to baseline curriculum learning techniques.

ABSTRACT. During the day, the temperature in urban areas rises with height, leading to a negative vertical temperature gradient that causes the warm air near the ground to rise. This phenomenon causes the pollution generated near the ground to rise and disperse, which is generally a desirable outcome. It can be modeled as vertical convection. However, during the night, the ground loses heat rapidly through radi ation, and the air close to the ground cools while the warmer air remains above it, blocking the pollution from rising. This is a thermal inversion phenomenon where the temperature increases with height instead of decreasing. This phenomenon causes the pollution generated by vehicles, houses, and industry to remain trapped near the ground, resulting in a high concentration of urban pollution. In this paper, we model this thermal inversion phenomenon using time-dependent three dimensional advection-diffusion equations and solve it numerically using Physics Informed Neural Networks. The training is performed over a four-dimensional space-time domain. We divide the temporal axis into a selected number of time segments and train the neural network segment by segment, introducing the initial conditions taken from the final time moment of the previous slice.

ABSTRACT. Physics-Informed Neural Networks (PINNs) solve physical systems by incorporating governing partial differential equations directly into neural network training. In electromagnetism, where well-established methodologies such as FDTD and FEM already exist, new methodologies are expected to provide clear advantages to be accepted. Despite their mesh-free nature and applicability to inverse problems, PINNs can exhibit deficiencies in accuracy and energy metrics compared to FDTD. This study demonstrates that hybrid training strategies can bring PINNs closer to FDTD-level accuracy and energy consistency.

A hybrid methodology addressing common challenges in wave propagation is presented. Causality collapse in time-dependent PINN training is addressed via time marching and causality-aware weighting. To mitigate discontinuities introduced by time marching, a two stage interface continuity loss is applied. To suppress cumulative energy drift in electromagnetic waves, a local Poynting-based regularizer is developed.

In the developed PINN model, high field accuracy is achieved with an average 0.09% NRMSE and 1.01% L2 error over time. Energy conservation is achieved with only a 0.024% relative energy mismatch in the 2D PEC cavity scenario. Training is performed without labeled field data, using only physics-based residual losses; FDTD is used solely for post-training evaluation. The results demonstrate that PINNs can achieve competitive results with FDTD in canonical electromagnetic examples and are a viable alternative.

ABSTRACT. The aerodynamic optimization of wind turbine blades is traditionally bottlenecked by computationally expensive, non-differentiable high-fidelity simulations that preclude efficient gradient-based design. This work presents a novel differentiable computational framework that integrates deep learning surrogates with physical solvers to enable end-to-end gradient-based optimization of rotor performance. We propose a hybrid architecture where NeuralFoil (a neural network predicting airfoil coefficients) is embedded within a GPU-accelerated, differentiable Blade Element Momentum Theory (BEMT) solver with guaranteed convergence.

Unlike standard iterative solvers, our implementation leverages the Implicit Function Theorem (IFT) to compute exact gradients of the power coefficient ($C_P$) with respect to high-dimensional geometric parameters, achieving numerical stability even in low-speed regimes where traditional tools like QBlade often diverge. We demonstrate the efficacy of this framework on the NREL 5MW reference turbine, using an Augmented Lagrangian method to satisfy geometric and aerodynamic constraints rigorously. The proposed pipeline runs fully in parallel on GPUs, offering a scalable, physics-informed pathway for next-generation renewable energy engineering.

ABSTRACT. Magnetic resonance imaging (MRI) acceleration has traditionally relied on compressed sensing and learned reconstruction from undersampled k-space. While modern deep networks substantially improve reconstruction quality, acquisition itself is still largely treated as fixed or only weakly adaptive. We propose generative control of k-space: a foundation-model-driven, closed-loop framework in which sampling, reconstruction, uncertainty estimation, and physics consistency are unified into a single decision-making system. Our architecture combines physics-informed reconstruction, transformer-based global context modeling, and generative diffusion/GAN priors that assess plausibility, uncertainty, and hallucination risk. The system maintains a probabilistic belief over both image and k-space and selects new measurements by minimizing a composite risk that integrates uncertainty, generative inconsistency, and physics residuals. Representative experiments show that the proposed approach reaches decision-stable, physically consistent reconstructions with substantially fewer samples than uniform, compressed-sensing, and image-centric adaptive baselines, while reducing hallucination risk and preserving image quality. This reframes MRI from a passive inverse problem into an active, self-optimizing measurement process and suggests a new role for foundation models in scientific imaging.

ABSTRACT. Deep learning models for 12-lead electrocardiogram (ECG) analysis achieve high diagnostic performance but lack the intuitive interpretability required for clinical integration. Standard feature attribution methods are limited by the inherent difficulty in mapping abstract waveform fluctuations to physical anatomical pathologies. To resolve this, we propose a cross-modal method that projects feature attributions from high-performance 12-lead ECG models onto the CineECG 3D anatomical space. Our study reveals that while models trained directly on CineECG signals suffer from reduced accuracy and incoherent attributions, the proposed mapping mechanism effectively recovers clinically relevant feature rankings. Validated against a ground-truth dataset of 20 cases annotated by domain experts, the mapped explanations yield a Dice score of 0.56, significantly outperforming the 0.47 baseline of standard 12-lead attributions. These findings indicate that cross-modal mapping effectively filters attribution instability and improves the localization of pathological features, combining the diagnostic expressiveness of standard protocols with the intuitive clarity of anatomical visualization.

ABSTRACT. Accurate detection and classification of cervical lesions in Pap smear images is essential for effective cervical cancer screening. Conventional deep learn-ing methods often treat segmentation and classification as separate tasks and tend to produce overconfident predictions. We propose a unified uncertainty-aware multi-task deep learning framework that performs simultaneous cervi-cal lesion segmentation and classification. The architecture combines a pre-trained ResNet50 encoder with a U-Net-style decoder for pixel-level segmen-tation and an attention-guided evidential classification head for uncertainty-aware diagnosis. Dirichlet-based evidential learning provides predictive uncer-tainty, while a homoscedastic uncertainty-based loss dynamically balances the two tasks. Evaluation on Pomeranian, Herlev, and SIPaKMeD datasets demonstrates strong performance, with segmentation accuracy of 99.19, 87.84%, and 93.85%, respectively, and classification accuracy of 99.14% in binary and 97.04% in multiclass for SIPaKMeD. Uncertainty correlates nega-tively with segmentation quality, and low Expected Calibration Error (ECE) indicates well-calibrated predictions. These results suggest that the proposed framework provides both accurate and reliable predictions, making it suitable for clinical decision support.

ABSTRACT. Accurate 3D coronary artery segmentation in CT angiography (CTA) is a key prerequisite for quantitative vessel analysis and downstream decision support. In this paper, we study three modern 3D segmentation backbones---a 3D U-Net baseline, a residual CNN (SegResNet), and a transformer-based model (SwinUNETR)---on the ImageCAS coronary artery dataset. We employ a unified preprocessing and patch-based training protocol implemented with the MONAI framework, including isotropic resampling, orientation normalization, intensity standardization, and ROI-based sampling. In our experiments, SegResNet achieved the best Dice score of 0.780, outperforming SwinUNETR (Dice 0.722) and the 3D U-Net baseline (Dice 0.616). Beyond segmentation accuracy, we further demonstrate two practical downstream modules enabled by the predicted masks: (i) centerline-based diameter estimation via 3D skeletonization and Euclidean distance transform (EDT), and a distribution-based comparison between ground truth and predictions using quantile-profile MSE; and (ii) a proof-of-concept projection-based candidate visualization module that applies a YOLOv8 detector to three orthogonal 2D projections, reported qualitatively (no standardized stenosis annotations are available for quantitative evaluation in ImageCAS). These additions support interpretable vessel geometry analysis and provide practical intermediate visual artifacts for downstream work.

ABSTRACT. This paper presents a Petrov-Galerkin immersed finite element method for coupled Darcy and Darcy-Forchheimer flows in fractured porous media. The matrix follows Darcy's law, while fractures are governed by the nonlinear Darcy-Forchheimer law to capture inertial effects. Fractures are treated as lower-dimensional interfaces embedded in the matrix, leading to an elliptic interface problem with multiple coupling conditions. The method does not require the mesh to align with fractures. Instead, interface conditions are built into local basis functions on elements cut by fractures. This allows the immersed finite element space to naturally represent both continuous and discontinuous pressure across fractures, making it suitable for both conductive and blocking fractures. Local function constructions are given for rectangular and triangular meshes. Numerical examples and benchmarks demonstrate the effectiveness of the proposed method on nonconforming meshes.

ABSTRACT. Vortical flow is common in severe arterial stenoses and blood-contacting medical devices, yet its independent contribution to platelet aggregation remains difficult to quantify because vorticity coexists with elevated shear, strong shear-rate gradients, and recirculation-driven residence effects. Here we present a three-dimensional, continuum-scale computational framework that links the local fluid-mechanical environment to early platelet aggregation. We examine a disease-motivated, high expansion-ratio double-stenosis microfluidic geometry operated at arterial wall shear rates of 1,000–5,000 s−1. Hemodynamics and evolving aggregate-induced flow disturbances are resolved with computational fluid dynamics (CFD), and predicted aggregate growth dynamics and morphology are validated against in vitro microfluidic measurements. Platelet aggregation is modeled by coupling advection–diffusion transport to shear-regulated, mechanosensitive adhesion kinetics. The simulations reproduce a vorticity-mediated redistribution of translocating platelets initiated near the stenosis apex and focused into the post-stenotic expansion, which in turn governs where aggregates nucleate and how they spread. Under high shear, post-stenotic recirculation advects platelet trajectories farther downstream, yielding skewed aggregation patterns consistent with experiments. Increasing wall shear to 5,000 s−1 amplifies vortex-driven transport and platelet influx, elevating the supply of surface-bound rolling platelets at the distal aggregate edge and promoting faster growth with greater longitudinal expansion than at 1,000 s−1. Collectively, these results indicate that vorticity modulates early aggregate morphology and growth rate primarily by shaping local shear exposure and rolling-platelet transport. This framework provides a mechanistic basis for predicting vorticity-facilitated early arterial thrombosis in complex vortical flows relevant to advanced stenosis and device interfaces.

ABSTRACT. Hydrogen is an ideal clean energy carrier, and its transportation via hydrogen-blended natural gas (HBNG) through existing natural gas infrastructure represents an economical and efficient pathway for the global energy transition. However, achieving high-purity hydrogen recovery from HBNG requires highly selective separation technologies, with palladium-silver (Pd-Ag) alloy membranes being a primary candidate due to their unique solution-diffusion transport mechanism. In this study, a comprehensive multi-scale computational framework is established with the Materials Studio to systematically investigate the transport and permeation mechanisms of hydrogen through Pd-Ag alloy membranes. At the electronic scale, the CASTEP module is utilized to perform first-principles calculations on periodic bulk phase models, elucidating the thermodynamic solubility and kinetic diffusion energy barriers of hydro-gen atoms within the metallic lattice. This stage focuses on the "solution-diffusion" behavior by evaluating energy as hydrogen traverses octahedral and tetrahedral interstitial sites. Subsequently, surface models are employed to explore the competitive adsorption energies of hydrogen molecules versus various impurity gases in natural gas, providing a theoretical basis for understanding surface poisoning. To bridge the gap between atomic interactions and macroscopic performances, the Forcite module is implemented to conduct molecular dynamics (MD) simulations on a composite system incorporating the metallic membrane and multi-component gas mixtures. These simulations analyze the dynamic permeation and selectivity under varied conditions, revealing how silver alloying effectively modulates lattice strain and electronic density to enhance hydrogen permeability. The results demonstrate that separation performance is influenced by a synergistic relationship between bulk diffusion kinetics and surface site occupancy. This research offers theoretical foundation for optimizing alloy membrane designs and supports the strategic advancement of hydrogen purification technologies within the hydrogen energy industry.

ABSTRACT. Geological Carbon Storage (GCS) is a critical technology for mitigating anthropogenic CO2 emissions, but accurate reactive transport modeling (RTM) of subsurface ionic species remains challenging due to neglected electromigration effects. This paper focuses on the numerical modeling of Electrical Double Layer (EDL) effects a key electromigration phenomenon and its implications for GCS processes. Building on the theoretical framework of Poisson-Nernst-Planck (PNP) equations, we develop a reactive transport modeling workflow that integrates EDL effects with chemical kinetics and diffusive transport. The characteristic time scales of carbonate reactions and electromigration are analyzed to justify the selection of kinetic models for reaction terms. Two core GCS scenarios are investigated: CO2 dissolution in nanoscale water films attached to solid surfaces and CO2 diffusive leakage through caprock/well cement nanopores. Numerical results demonstrate that EDL effects are length-scale dependent: negligible at the micrometer scale but dominant at the nanometer scale, significantly altering ionic concentration distributions, pH conditions, and reaction kinetics. For CO2 leakage through nanopores, EDL-induced electrostatic forces modify ionic fluxes facilitating positive ion transport and repelling negative ions leading to lower pH and increased corrosion risk. High salinity conditions mitigate but do not eliminate EDL impacts, highlighting the necessity of incorporating electromigration effects in GCS RTM. The proposed modeling framework advances the accuracy of subsurface transport simulations, providing critical insights for GCS project design and risk assessment.

ABSTRACT. Solving multicomponent phase equilibrium PT flash calculation efficiently hinges on the rapid solution of the Rachford-Rice equation. To improve the convergence speed of the classical Newton method for solving this equation, this paper proposes an momentumed Newton method that embeds the momentum technique into the classical Newton method as its core. By integrating the momentumed Newton method with successive substitution iteration (SSI), a novel fast combined momentumed Newton-SSI solver is designed for multicomponent PT flash calculation. Numerical experiments demonstrate that the proposed combined momentumed Newton-SSI method exhibits superior convergence efficiency compared with the combined Newton-SSI method, achieving faster convergence in practical computations.