ABSTRACT. The direct effects of climate variability on refugee mobility during conflict-driven displacement remain poorly quantified. This study investigates how precipitation and river discharge influence journey times and route accessibility within a multiscale agent-based simulation framework. Building on our earlier implementation, we analyse the integration of 40 years of ERA5 climate reanalysis data and corresponding GloFAS river-discharge records to derive evidence-based thresholds governing movement speed adjustments and route closures. Applied to the 2016–2017 conflict in South Sudan, the analysis demonstrates that while conflict drives departures, hydrological conditions regulate when, where, and how far displaced populations can travel by introducing seasonal slowdowns and route-level bottlenecks. Although climate coupling increases computational cost by a factor of 6.4, it reveals critical accessibility constraints that are not captured in performance-focused evaluations. Overall, this study provides a more detailed characterisation of climate-induced mobility constraints and enhances the interpretability of forced migration simulations.

ABSTRACT. Operational weather prediction has long relied on physics-based numerical weather prediction (NWP), whose accuracy comes at the cost of substantial compute and complex simulation workflows. Recent transformer-based forecasters offer efficient data-driven alternatives, however transformers are physics-agnostic models. Additionally, standard transformer encoders evolve representations through discrete layer updates that may be less suited to modeling smooth latent dynamics. In this work, we propose a continuous-depth transformer encoder for weather forecasting that integrates Neural Ordinary Differential Equation (Neural ODE) dynamics within each encoder block. Specifically, we replace discrete residual updates with ODE-based updates solved using adaptive numerical integration. We also introduce a two-branch attention module that combines conventional patch-wise self-attention with an auxiliary branch that applies a derivative operator to attention logits, providing an additional change-sensitive interaction signal. To further align forecasts with governing principles, we propose a customized physics-informed training objective that enforces physical consistency as a soft constraint. We evaluate the proposed method against a standard discrete transformer baseline and an existing continuous-time Neural ODE forecasting variant, demonstrating the importance of PINN-Cast in short term weather forecasting.

ABSTRACT. This paper introduces a novel wildfire growth model designed for rapid, slope-driven wildfires in Switzerland. By combining raster- and vector-based approaches, the model leverages raster efficiency and vector prediction accuracy. As a result, the new model utilises high-resolution landscape data available in Switzerland, overcoming a key limitation of established wildfire simulation models. The proposed model builds on the raster-based Cell2Fire simulator, allowing direct comparison with the baseline in diverse landscapes. Comparisons against idealised reference fire perimeters show a significant F1 -Score improvement over the baseline, even with larger simulation time steps. These findings indicate that raster-based models can incorporate the proposed extension to improve predictions with minimal additional computational cost. Furthermore, the suitability for parallel processing makes the new model well suited for faster-than-real-time applications in high-resolution landscapes.

ABSTRACT. Nonlinear time-history evolution problems employing high-fidelity physical models are essential in numerous scientific domains. However, these problems face a critical dual bottleneck: the immense computational cost of time-stepping and the massive memory requirements for maintaining a vast array of state variables. To address these challenges, we propose a novel framework based on Heterogeneous Memory Management for massive ensemble simulations of general nonlinear time-history problems with complex constitutive laws. Taking advantage of recent advancements in CPU-GPU interconnect bandwidth, our approach actively leverages the large capacity of host CPU memory while simultaneously maximizing the throughput of the GPU. This strategy effectively overcomes the GPU memory wall, enabling memory-intensive simulations. We evaluate the performance of the proposed method through comparisons with conventional implementations, demonstrating significant improvements in time-to-solution and energy-to-solution. Furthermore, we demonstrate the practical utility of this framework by developing a Neural Network-based surrogate model using the generated massive datasets. The results highlight the effectiveness of our approach in enabling high-fidelity 3D evaluations and its potential for broader applications in data-driven scientific discovery.

ABSTRACT. In this study, we investigated the computational performance of a multigrid (MG) solver for the Poisson equation in a bulk lightning model (BLM) that explicitly calculates lightning frequency and is coupled to a meteorological model. The meteorological model employs highly anisotropic grids due to the strong disparity between typical horizontal and vertical scales of atmospheric phenomena, and it also exhibits terrain-induced horizontal grid non-uniformity. To improve the performance of the MG solver on such anisotropic grids, we implemented a vertical line Gauss-Seidel (z-line GS) smoother. Furthermore, we investigated the impact of terrain-induced grid non-uniformity on the convergence rate of the MG solver at horizontal grid spacings of O(100 m) by comparing it with the Bi-Conjugate Gradient Stabilized method preconditioned with symmetric Gauss-Seidel (SGS-BiCGSTAB). For the performance evaluations, we computed the electric potential using a snapshot of the charge distribution from a realistic simulation with problem sizes (nz, nx, ny) = (32, 128, 128), (32, 256, 256), (32, 512, 512), and (32, 1024, 1024), corresponding to horizontal grid spacings of 2000, 1000, 500, and 250 m, respectively. For the (nz, nx, ny) = (32, 1024, 1024) case, the MG solver with the z-line GS smoother reduced the elapsed time by 93% (no terrain) and 89% (with terrain) relative to SGS-BiCGSTAB. Without terrain, the iteration count for the MG solver was nearly independent of the problem size, whereas the iteration count for SGS-BiCGSTAB increased with the problem size. With terrain, the convergence rate of the MG solver deteriorated at high resolution due to terrain-induced grid non-uniformity. However, it remained superior to SGS-BiCGSTAB, whose iteration count increased significantly with the problem size. These results suggest that the MG solver offers an advantage even for simulations with the grid non-uniformity induced by small-scale topography on terrain-following coordinates at a high resolution of O(100 m).

ABSTRACT. We present a metric learning approach for mineral classifi- cation using a merging of hyperspectral and multispectral satellite data. Our method integrates spectral information from multiple providers, including EMIT, EnMAP, PRISMA, Sentinel-2 , and WorldView-3 in unified embeddings. To reflect realistic operational conditions, we intro- duce a training-time augmentation strategy in which a random subset of providers is masked, forcing the network to learn robust representations that generalize to incomplete sensor coverage. Trained on an augmented dataset derived from the USGS spectral library, the proposed network significantly outperforms classical one-class methods such as Isolation Forest and OCSVM in terms of F1-score, while maintaining competi- tive recall. Analysis of missing providers demonstrates that embeddings remain stable when only a subset of sensors is available, particularly when multispectral data are missing, whereas missing multiple hyper- spectral providers has a substantial impact. The most robust embed- dings are achieved when all hyperspectral providers are present, often in combination with WorldView-3. These results highlight the importance of hyperspectral data in discriminative mineral mapping. In general, our approach provides a flexible and effective framework for operational min- eral classification under realistic multisensor conditions.

ABSTRACT. The paper presents the interval fast parametric integral equations system (IFPIES) applied for searching domain solutions of potential 2D boundary value problems (BVPs). Previously, the IFPIES has been successfully applied in modelling uncertainly defined 2D potential BVPs and finding solutions on the boundary. The combination of the modified fast multipole technique and modified directed interval arithmetic with the PIES reduced the computational time and RAM usage in the interval PIES. To find solutions in the domain, similar techniques are applied. The method is demonstrated with the solution in the domain of uncertainly defined BVPs.

ABSTRACT. This study introduces a regularization-based strategy to improve solution accura-cy near boundaries within the integral identity associated with the parametric inte-gral equation system (PIES) for three-dimensional potential problems. Accuracy deterioration in near-boundary regions arises from singular behavior of the inte-grand functions in the underlying integral formulation. The proposed approach employs a suitably constructed regularizing function with optimized coefficients to mitigate these effects. Numerical analysis demonstrates a significant enhance-ment in solution accuracy. The regularization algorithm is independent of bounda-ry geometry, boundary representation, and imposed boundary conditions, high-lighting its generality and wide applicability.

ABSTRACT. Full GEANT4 simulations of calorimeter responses in the ALICE experiment provide high accuracy but are computationally expensive. To enable fast simulation for large-scale studies, we investigate machine-learning-based surrogate models for the FoCal hadronic calorimeter. Several generative approaches, including GANs, normalizing flows, and diffusion models, are evaluated against GEANT4 using physics-motivated metrics that compare energy profiles or the morphology of the showers. In addition, we introduce a hybrid simulation strategy in which a fast, physics-inspired approximation generates a coarse response that is subsequently refined by a compact generative model. This approach achieves a favorable balance between simulation fidelity and computational efficiency, which makes it suitable for high-rate ALICE analyzes.

ABSTRACT. Recent advances in exascale computing have increased the resolution and fidelity of large-scale simulations, while rapid progress in deep learning has accelerated efforts to couple machine learning with physics-based solvers. We present a lightweight, modular in situ coupling framework that embeds machine learning training and inference directly into simulation workflows using the ParaView and Catalyst APIs. The framework provides C++/Python interoperability via a solver-side data adaptor that packages simulation state into Conduit Nodes and a Catalyst-driven Python ``bridge script'' that converts solver fields into NumPy/PyTorch representations with minimal intrusion into the solver code. We describe the design and instrumentation required to integrate the framework and demonstrate it within a proxy (mini-app) of the HARVEY vascular flow solver. To illustrate practical usage, we implement both in situ training and in situ inference of a point-cloud autoencoder running concurrently with the solver. We report scalability and overhead characteristics and show that the approach enables distributed online ML workflows without requiring language unification or major solver refactoring.

ABSTRACT. Surrogate modelling has become increasingly important in microwave engi-neering. Fast metamodels—particularly behavioral ones—are widely employed to accelerate design tasks such as parametric optimization by replacing costly full-wave electromagnetic (EM) simulations. However, constructing reliable surrogates remains challenging due to the strong nonlinearity of circuit re-sponses and the curse of dimensionality. The difficulty is especially pro-nounced in design-oriented modeling, where validity must be ensured across wide ranges of parameters. This research introduces an innovative modeling procedure that combines dimensionality reduction with spatial domain re-striction to reduce the training data acquisition cost while enhancing predictive accuracy. Dimensionality reduction is achieved using fast global sensitivity analysis, which determines the parameter-space directions with the strongest impact on the system’s frequency characteristics. These vectors define the re-duced domain, which is additionally confined by means of principal compo-nent analysis (PCA) of pre-screened high-quality designs. As a result, the sur-rogate is concentrated on the relevant design space subsets, ensuring suitability for practical design tasks while maintaining high accuracy. The proposed meth-odology has been extensively validated against state-of-the-art benchmarks, demonstrating both competitive precision and significant efficacy gains. Its de-sign readiness has been shown through practical applications, specifically, in EM-driven circuit optimization under varying specification scenarios.

ABSTRACT. Rigorous optimization methods have become standard practice in antenna en-gineering, gradually replacing traditional interactive design approaches that re-lied on parametric studies and engineering intuition. Nevertheless, antenna op-timization remains computationally expensive due to its reliance on electro-magnetic (EM) analysis. To mitigate this, accelerated strategies have been in-troduced by limiting the occurrences of EM simulations at the algorithmic level or through surrogate modeling. Multi-fidelity approaches offer another avenue, though existing frameworks typically restrict themselves to just two levels (low and high fidelity). In this work, we propose an innovative model management scheme that adaptively adjusts EM model resolution across a continuous fideli-ty spectrum. Model selection is guided by the optimization’s convergence sta-tus and design quality indicators. The process begins with the lowest usable resolution, which is progressively refined as the optimization approaches con-vergence and the objective value improves. This strategy lowers computational costs by exploiting faster, lower-fidelity simulations when far from the opti-mum, while ensuring reliability by incorporating high-fidelity models near convergence. Extensive numerical experiments involving two microstrip an-tennas showcase the efficacy of the presented framework, showing speedups of exceeding 70% compared to the baseline approach, with only negligible per-formance degradation.

ABSTRACT. Microbial contamination of laboratory cultures is a major source of experimental failure and resource waste. This study proposes a Single-Instance Driven CNN approach for rapid genus-level recognition of common soil-dwelling contaminants (Trichoderma, Fusarium, Verticillium, Purpureocillium) from microscopy images. The pipeline retrieves standardized subimages and combines their predictions via majority voting to overcome the limitations of full-image classification. DenseNet201, ResNet50v2, and InceptionResNetV2 are evaluated across five datasets collected with manual, automated, and focus-stacked microscopy. The results show that subimage-based learning consistently outperforms full-image baselines, while cross-dataset experiments confirm robust generalization. Grad-CAM analysis indicates that subimage-trained networks focus on microorganism fragments and suppress irrelevant background, supporting the interpretability of the proposed workflow.

ABSTRACT. Hospital scheduling requires the coordinated allocation of operating rooms, inpatient beds, and healthcare staff under strict constraints and interacting performance objectives. Decisions at the surgical level propagate through downstream care units, while congestion in inpatient services may restrict surgical activity, resulting in a large-scale and tightly coupled optimization problem.

This paper proposes an integrated framework based on a unified Quadratic Unconstrained Binary Optimization (QUBO) formulation that jointly models operating room scheduling, bed management across clinical phases, and human resource allocation. Non-linear congestion effects are captured using a Choquet integral–based aggregation of soft criteria, explicitly modeling interactions between delays, bed shortages, staff overload, and cancellations. The resulting non-convex QUBO is solved using a neural-guided quantum-inspired simulated annealing algorithm.

Experiments on realistic synthetic instances demonstrate clear improvements over classical simulated annealing and unguided quantum-inspired methods in terms of feasibility, solution quality, and convergence speed.

ABSTRACT. With the widespread adoption of cloud-native technologies, Cloud-Native Applications (CNAs) have become the mainstream ap- proach for building scalable and maintainable software systems. How- ever, due to their structural diversity, dynamic runtime behaviors, and complex infrastructure dependencies, existing simulators often fall short in modeling and simulating such systems effectively. To address these limitations, this paper introduces CNASIM, a modular and extensible performance simulation library for cloud-native applications. The library supports flexible modeling of various CNA architectures and behaviors, offering customizable components for services, instances, communication mechanisms, and resource management strategies. It also provides config- urable interfaces, component-level metric collection, and extension mech- anisms for incorporating user-defined logic. Experimental results demon- strate that our approach achieves good performance in terms of modeling flexibility, simulation accuracy, and applicability to complex scenarios, making it a practical aid for system designers and researchers to evaluate and optimize cloud-native applications efficiently.

ABSTRACT. Currently, more and more machine learning (ML) surrogates are being developed for computationally expensive physical models. In this work we investigate the use of a Multi-Fidelity Ensemble Kalman Filter (MF-EnKF) in which the low-fidelity model is such a machine learning surrogate model, instead of a traditional low-resolution or reduced-order model. The idea behind this is to use an ensemble of a few expensive full model runs, together with an ensemble of many cheap but less accurate ML model runs. In this way we hope to reach increased accuracy within the same computational budget. We investigate the performance by testing the approach on two common test problems, namely the Lorenz-2005 model and the Quasi-Geostrophic model. By keeping the original physical model in place, we obtain a higher accuracy than when we completely replace it by the ML model. Furthermore, the MF-EnKF reaches improved accuracy within the same computational budget. The ML surrogate has similar or improved accuracy compared to the low-resolution one, but it can provide a larger speed-up. Our method contributes to increasing the effective ensemble size in the EnKF, which improves the estimation of the initial condition and hence accuracy of the predictions in fields such as meteorology and oceanography.

ABSTRACT. Flooding results in large economic losses and loss of life, which are further aggravated by the lack of precise forecasts of flood inundation depth and extent. Recent extreme flood events have highlighted the need for reliable operational flood forecasting systems. Conventional physics-based flood models are subject to multiple sources of uncertainty and are computationally demanding, which limits their applicability for real-time operational services. Artificial intelligence (AI)-based flood models can significantly reduce computational cost and enable near real-time forecasting at high spatial resolutions. However, their operational use remains limited because evolving prediction errors cannot be systematically corrected without robust assimilation of observations. Flood processes are highly nonlinear, with errors that evolve rapidly in space and time, while Earth Observation (EO) data provide intermittent and spatially incomplete snapshots of the true system state. Deep data assimilation (DDA) offers a principled framework to address this challenge by learning state-dependent error propagation and dynamically integrating multi-source EO information into AI-based flood forecasting models. In the recently funded Indo-German project FLAIR (Flood Forecasting using AI for Regional Sustainability), we focus on developing DDA-driven observation operators that link simulated flood states to EO-derived flood extent and water surface elevation within a two-dimensional convolutional long short-term memory framework. DDA is implemented through a state–parameter augmentation strategy to update model states in real time, enabling adaptive correction under dynamically evolving flood conditions. The framework is evaluated for two human-altered test catchments with contrasting hydrological characteristics in India and Germany. Forecast performance is benchmarked against an open-loop configuration and a DDA-based CaMa-Flood model across lead times from one to seven days. A key innovation is the assimilation of reservoir water surface elevations from EO altimetry, which informs flood wave propagation and supports reservoir optimisation. The results highlight how DDA strengthens the reliability, interpretability, and operational readiness of AI-based flood forecasting systems.

ABSTRACT. Floods are among the most destructive natural hazards worldwide and frequently cause severe economic losses and significant loss of life. Therefore, reliable and timely forecasting plays a crucial role in disaster risk reduction. In recent years, approaches based on machine learning and data assimilation have attracted increasing attention for this purpose. However, a fundamental challenge remains. Predictive models generally require large volumes of high quality data, yet such data are often scarce, spatially and temporally limited, and heterogeneous in structure. This limitation substantially restricts the development of advanced flood forecasting methods. To address this problem, we elaborate three independent datasets derived from different sources, namely EFAS, EMO-1, and LamaH-CE, which include meteorological, remote sensing, hydrological, and topographic information. Based on these datasets, we design two test cases consisting of a two-dimension forecasting experiment and a data assimilation experiment. Overall, the elaborated datasets and test cases provide a solid foundation for advancing flood forecasting research using machine learning and data assimilation techniques.

ABSTRACT. Classical data-assimilation methods, like the Kalman Filter, and even most recent formulations, are model-based methods where it is assumed that the relation between the state-vector and the measure- ments can be expressed analytically up to a random error component. Here we refer to situations where the available analytical expression for themeasurementequationisasurrogateofthetrueone,inthesensethat the former approximates the latter with not negligible deterministic dis- crepancies. Moreover, the analytic expression of these discrepancies, if it exists, is presumed to be impossible to find at a reasonable cost. We assume also that an accurate measurement equation exists for measure- ments that can be done only in laboratory experiments. The aim of this paper is to show that a Deep Kalman Filter can use a surrogate mea- surement equation to form the innovations and optimally estimate the state-vector, when the supervised learning of the corrector-gain matrices hasbeendoneusinganaccuratemeasurementequation(andcorrespond- ing data). The resulting method is trajectory-dependent, in principle, and we discuss also a possible strategy for a data-driven generalization to multiple trajectories. As a general test, we show some results with an abstract dynamical system and measurement equations with additive piecewise-polynomial and/or trigonometric biases.

ABSTRACT. Operational forecasts with coupled Earth-system models (ESM) rely on different and inconsistent algorithms to initialise different ESM components. For example, at ECMWF, the atmospheric state is estimated using 4D-Var, the ocean using 3D-Var, the land surface using 2D-optimal interpolation, and soil moisture is analysed using a simplified extended Kalman filter. To consolidate the initialisation, we perform fully-coupled 3D-Var data assimilation in the latent space of a graph neural network (GNN) autoencoder (AE). We trained the GNN AE to jointly represent atmospheric fields and land-ocean surface fields in a shared latent space of reduced dimensionality. We constructed the GNN AE as an neighborhood-attention-based progressive autoencoder, defined on a hierarchical multiscale graph. This approach has advantages in computational performance and extended applicability, as it excels at non-Gaussian distributions and replaces the control-variable transform (CVT), which is problematic outside the midlatitudinal free troposphere, e.g. in the tropics, the stratosphere, and the boundary layer. By learning from simulations of classical physics-driven models, the assimilation identifies physically-explainable increments across ESM compartments. For example, observations of sea ice induce increments of atmospheric and other ocean variables and vice versa. Moreover, the increments are spatially anisotropic and respect physical boundaries, e.g., the sea-ice increments are localised predominantly along the ice edge. The spatially-decomposed approach of the hierarchical graph enables scalability to models with O(10^9) dimensions, which would be suitable for operational numerical weather prediction. Finally, the approach can be efficiently extended into 4D-Var data assimilation.

ABSTRACT. Genome-wide epistasis detection using covariate-adjusted logistic regression is computationally extremely demanding and difficult to scale to modern genome-wide association study (GWAS) datasets. We recently introduced an efficient population structure-adjusted logistic regression algorithm for genome-wide association interaction studies (GWAIS) based on proxy-covariates derived from clustering principal component analysis (PCA) covariates (Neff et al., 2025). In this approach, we substantially reduced the computational burden while retaining genetic ancestry adjustment, but finding the optimal configuration for covariate clustering remained unsolved. Here, we present a data-driven framework for automated testing of clustering configurations for covariate-adjusted epistasis screening. Starting from per-sample PCA coordinates regarded as ground truth, we systematically evaluate K-means and Gaussian mixture model clustering across candidate parameters on a small subset of the data. Graphical scatter plots help to highlight the best configuration minimizing the mean relative error (MRE) of interaction p-values compared to the ground truth. Using the same real-world GWAS datasets previously analyzed via genome-wide screening, we show that evaluating only 1 % to 5 % of the data reliably reproduces or improves the previously identified configurations. In subsequent epistasis screening, our guided proxy-covariate adjustment achieves a speedup of up to 151x while maintaining a low MRE relative to the per-sample covariate-adjusted reference model.

ABSTRACT. Accurate segmentation of histopathology images is critical for cancer diagnosis, characterization, and treatment planning. However, manual annotation at the cellular level is labor-intensive and requires expert knowledge, while variability across tissue types, staining protocols, and scanning devices limits the generalization of conventional segmentation models. In this work, we propose a few-shot learning framework for cross-domain histopathology segmentation, leveraging the selective state-space model \textit{Mamba} to efficiently capture long-range dependencies in densely packed cellular images. Our approach enables the model to generalize to new tissue types or laboratory domains using only a few annotated examples. Additionally, we incorporate explainable AI techniques to provide interpretability of segmentation predictions, supporting clinical trust. We validate our method on the PanNuke dataset, providing a few-shot segmentation split, and demonstrate that our Mamba-based model outperforms existing few-shot and baseline methods. This work advances practical and efficient automated histopathology segmentation under limited supervision and domain variability.

ABSTRACT. High-dimensional gene expression profiles are affected by technical variation such as measurement noise, missing features, and scale changes introduced by heterogeneous preprocessing. These effects can distort value-based distances, destabilize local neighborhoods, and lead to inconsistent predictions, particularly in $k$-nearest neighbors (kNN). An alternative is to compare samples through within-sample orderings between genes rather than absolute values, following the general lineage of relative expression ordering and gene-pair methods.

In this paper, we introduce a stability-aware relational kNN framework that operates in an ordering-based space using two complementary rank-derived distances: (i) an inversion-based distance that measures discordant gene-pair orderings between samples, and (ii) a displacement-based distance that aggregates absolute rank shifts.

The framework is evaluated against standard $L_p$ norms ($L_1$, $L_2$, $L_\infty$) on seven public gene expression datasets. Predictive quality is measured by macro-F1, while robustness is quantified by prediction flip rate and neighbor-set Jaccard similarity relative to an unperturbed baseline. Robustness is evaluated under five controlled perturbation scenarios with varying intensities: feature dropout, featurewise scaling, additive Gaussian noise, monotonic sample scaling (invariance check), and training-set instance dropout. The results characterize performance-stability trade-offs and show that ordering-based distances can improve neighborhood and prediction stability in several non-trivial perturbation regimes while maintaining competitive macro-F1.

ABSTRACT. The brain’s enormous complexity, together with the high global prevalence of neurologic disorders, necessitate the development of comprehensive and advanced neuromarkers to enhance both brain understanding as well as disorder prevention, prognosis, diagnosis, and treatment. I propose a combined neuroimaging and neuromodeling multi-purpose, multi-model, multi-dimensional, and multi-modal neuromarker in the form of an electronic brain atlas. This special type of brain atlas is a personalized brain atlas for everyone (pBAe), capable of systematically accommodating the condition of one’s brain over time. The pBAe is defined as a time series of navigable and quantifiable pairs of (raw brain scans; reconstructed and annotated 3D brain models), and I here address its content, data acquisition, software architecture, and construction methodology. The proposed pBAe content and functionality are illustrated through the author’s personalized brain atlas, demonstrating 3D structure, cerebrovasculature, and cranial nerves fully parcellated by color and labeled. When deployed globally, the pBAe will have a profound impact on individuals, society, science, and AI development. For individuals, it offers deeper insights into brain structure, function, and disorders; quantified brain health; lifestyle modifications; early screening with predictive capabilities; continuous brain status monitoring; and personalized medicine, including targeted therapies. Individual pBAes worldwide would collectively form a massive neurodatabase whose analyses could drive new discoveries leading to improved public health and human well-being. Ultimately, the synergy be-tween the pBAe initiative and AI can create a self-reinforcing "virtuous circle" accelerating progress in both fields.

ABSTRACT. PyRadiomics-cuda is a GPU-accelerated extension of the PyRadiomics library, designed to address the computational challenges of extracting three-dimensional shape features from medical images. By offloading key geometric computations to GPU hardware it dramatically reduces processing times for large volumetric datasets. The system maintains full compatibility with the original PyRadiomics API, enabling seamless integration into existing AI workflows without code modifications. This transparent acceleration facilitates efficient, scalable radiomics analysis, supporting rapid feature extraction essential for high-throughput AI pipeline. Tests performed on a typical computational cluster, budget and home devices prove usefulness in all scenarios.

ABSTRACT. Local government systems produce large, heterogeneous text corpora, including policy reports, internal communications, and public documentation. Extracting interpretable topics from such corpora of unstructured data requires methods that capture both global themes and fine-grained, domain-specific subtopics. Transformer-based topic modeling frameworks such as BERTopic provide effective embeddings and clustering, but their reliance on a single global clustering step imposes uniform topic granularity, often merging semantically distinct discourse regions. We introduce Per-Cluster Topic Modeling (PCTM), a scalable extension to BERTopic that performs neighbor-aware local clustering of document embeddings followed by per-cluster topic extraction. This approach generates multi-scale topic representations, enabling coarse-grained administrative themes and fine-grained subtopics to coexist within the same framework. Evaluated on a municipal office corpus spanning multiple departments, PCTM outperforms global BERTopic baselines in topic coherence, stability, and interpretability. Beyond empirical gains, PCTM provides a computational model of multi-level social systems, reflecting the heterogeneous semantic structure of municipal governance. This framework supports AI-driven analysis of complex administrative and social text corpora, offering a robust tool for structured knowledge discovery.

ABSTRACT. This paper introduces CHExNet, a temporal, two-layer net- work dataset capturing interactions centered on the Jagiellonian Univer- sity in Kraków over the period 1364–1850. Nodes represent individuals reconstructed from heterogeneous archival and bibliographic sources, and edges encode two complementary interaction layers aligned to a shared time axis: (i) educational/professional co-presence, inferred when two persons are recorded at the same institution within the same semester- length bin, and (ii) book-production collaboration, inferred from co- participation in bibliographic records. The creation of the network is based on an expert-validated authority file reconciling person identities between the Jagiellonian University Archives and the Jagiellonian Li- brary, and custom python scripts for the creation of a series of time-based adjacency matrices. We describe the aggregated and temporal structural properties of the resulting network and illustrate its analytical value with a usage scenario that models the diffusion of Polish-language book pro- duction as a discrete-time hazard process driven by multiplex exposure. CHExNet is released for reuse as a benchmark for temporal and multi- layer network analysis, historically grounded diffusion modeling, and AI methods that require long observation windows and interpretable prove- nance.

ABSTRACT. Decentralized Training and Execution (DTE) is an appealing paradigm in multi-agent reinforcement learning due to its scalability and autonomy, yet it suffers from severe non-stationarity arising from simultaneously learning agents. This work shows how a model-based approach can mitigate this issue without relying on centralized training or access to global information and achieve performance comparable to the state-of-the-art method from the currently dominant Centralized Training for Decentralized Execution (CTDE) paradigm.

For this, a model-based algorithm is proposed, where each agent independently learns an internal model of the environment dynamics from its own experience and integrates this model into the decision-making process. By acting according to its internal model rather than directly reacting to the evolving environment, an agent’s policy becomes more stable, reducing the impact of non-stationarity created by other learning agents. The proposed approach is evaluated and compared with independent Deep Q-Networks and MADDPG, a CTDE algorithm. Experimental results demonstrate that the model-based method achieves performance comparable to that of MADDPG, despite operating in a fully decentralized manner. These results indicate that model-based, decentralized approach can serve as an effective alternative to centralized training for cooperative multi-agent reinforcement learning.

ABSTRACT. The rapid spread of online misinformation has led to the adoption of increasingly complex detection models, including large language models and hybrid architectures that combine deep text encoders with graph-based reasoning. Although these approaches achieve strong performance, their computational cost, limited multilingual robustness, and deployment constraints raise concerns about their practicality in real-world settings. This work revisits the role of classic graph neural networks (GNNs) for misinformation detection, focusing on performance-efficiency trade-offs under controlled and comparable experimental conditions. We conduct a large-scale benchmarking study of lightweight GNN architectures, including GCN, GraphSAGE, GAT, and ChebNet, across seven publicly available datasets spanning political fact-checking, clickbait detection, and domain-specific misinformation in English, Indonesian, and Polish. To isolate the contribution of relational inductive bias, all models operate on identical TF--IDF feature representations and are compared against strong non-graph baselines, namely Logistic Regression, Support Vector Machines, and Multilayer Perceptrons. Performance is evaluated using F1 score, with inference time reported to assess practical efficiency. The results show that classic GNNs consistently outperform non-graph baselines across all datasets, while maintaining inference times comparable to or lower than those of neural baselines. Furthermore, GNNs remain robust in low-resource settings, exhibiting only modest performance degradation when trained on limited labeled data. These findings demonstrate that classic GNNs remain effective, efficient, and practically relevant for misinformation detection, challenging the assumption that high performance necessarily requires increasingly complex architectures. Github

ABSTRACT. DNA hydrogels are emerging as high-density programmable materials with potential applications in data storage and biomedicine. However, the microscopic mechanisms underlying their assembly and mechanical response remain poorly understood. To address this, we developed a coarse-grained bead–spring model inspired by oxDNA, enabling efficient simulations of DNA self-assembly at microscopic scales. The model incorporates patchy Y-shaped motifs with tuneable valency, flexible and rigid binding sites, and Weeks–Chandler–Andersen interactions to capture both attraction and hard-core repulsion. Using LAMMPS, we investigated cluster formation, melting behaviour, radial distribution functions, and rheological properties across a range of interaction strengths and patch flexibilities. Our results reveal distinct gelation pathways, with rigid patches favouring sharp melting transitions and flexible patches introducing tuneable mechanical responses. Rheological analysis through Green-Kubo relations highlights the dependence of storage and loss moduli on bond stiffness and interaction potential, and a machine-learning surrogate accurately predicts cross-over frequencies across a vast parameter space.

ABSTRACT. Emulsions of immiscible liquids can be used as templates for fabricating microstructure materials, including membranes and porous electrodes. Control over the domain size and morphology is desirable in these applications. In this work, we investigate the formation of ternary emulsions under an applied shear flow and characterize the emerging phase morphology. Using the lattice Boltzmann method, we simulate phase separation of a ternary fluid mixture that is sheared between two moving boundaries. At low shear rates, complex structures such as double emulsions and worm-like domains emerge. At high shear rates, the domains form aligned bands whose width depends on the shear rate, fluid properties and fluid composition. These findings demonstrate that shear can be used to control domain morphology in emulsions, thereby enabling the design of porous materials and aligned scaffolds for tissue engineering applications.

ABSTRACT. Characterizing materials through manual extraction of physical properties from microstructure images is a laborious process. This work presents a workflow to extract porosity, solid fraction, grain size distribution, and pore size distribution from scanning electron microscopy (SEM) images of sintered ceramic samples using an automated pipeline. The primary challenge for extracting physical properties from SEM images is the presence of unimodal histograms in SEM images as a result of the overlapping intensity ranges for the grain and pore phases. We evaluated several different methods for noise reduction and local thresholding of SEM images. We find that topological filtering in combination with Sauvola thresholding enables segmentation and extraction of physical property data from SEM images. We validated the automated pipeline by comparing our results with the results of manual analyses performed for samples sintered at 1200$^o$C and 1400$^o$C and achieved an accuracy of 95.14\% and 99.85\% (IOU scores), respectively. The workflow provides an efficient means to automatically extract microstructure properties from SEM images as a crucial step in generating materials datasets for machine learning.

ABSTRACT. Avoiding unnecessary computations is not only an important algorithmic measure for energy savings, i.e. sustainable computing, but in many cases it shortens the time to solution. In materials modeling, the scientific workflow method has been established over years in research practice, mainly to improve reproducibility, to enable FAIR-ness of data and to tackle extreme model complexity, in particular in multiscale modeling. Scientific workflows will continue to be important also in the age of artificial intelligence, in the form of agentic workflows that are currently emerging in computational science. In this contribution, we suggest solutions for avoiding unnecessary computations tailored to the peculiarities of materials modeling workflows in high performance computing settings balancing between the costs of re-computing, and data storage and retrieval. Particularly, the caching, lazy evaluation and checkpointing techniques will be considered. The suggested approaches are implemented in a platform consisting of workflow managements system, workload management system and computing cluster) and used via the textM domain-specific language.

ABSTRACT. Multitask learning (MTL) is widely used in scientific machine learning to improve data efficiency through representation sharing. However, in many scientific domains such as polymer informatics, datasets are partially labeled and highly imbalanced, and the appropriate degree of parameter sharing remains unclear. Conventional MTL architectures assume a predefined shared trunk, making architectural coupling a heuristic design choice rather than a principled decision.

We reinterpret multitask architecture design as a structured capacity-control problem and introduce a progressive representation merging framework. Instead of imposing sharing a priori, each task is first trained independently with sufficient capacity. Parameter sharing is then introduced gradually by merging layers across tasks, transforming representation sharing into a controlled compression process. The merging depth systematically modulates structural coupling and effective model flexibility.

We evaluate the approach on a seven-task polymer property prediction benchmark with substantial label imbalance (including Tg, Tm, mechanical strength, modulus, crystallinity, and enthalpy). Results reveal a consistent non-monotonic dependence of generalization performance on sharing depth. Moderate sharing improves macro-averaged R², particularly for data-scarce tasks, while excessive compression leads to mild degradation. Task-level analysis uncovers heterogeneous regimes corresponding to saturated, gain-regime, and decoupled behaviors, reflecting different positions in a capacity–alignment landscape governed by supervision level and inter-task compatibility.

These findings suggest that multitask learning in scientific settings operates as structured capacity modulation rather than uniform regularization. Progressive merging provides a systematic mechanism to identify minimal shared representations under sparse supervision and offers a general design principle for multitask learning in data-limited scientific computing applications.

ABSTRACT. The increasing reliance of materials science research on computational methods has made adherence to the FAIR principles (Findable, Accessible, Interoperable, Reusable) an essential requirement for transparent, reproducible, and sustainable research [1]. However, due to multiple interdependent simulation and analysis steps across diverse software packages, computing environments, and scientific domains, compliance to the FAIR principles remains challenging. Moreover, additional technical considerations such as efficient use of resources, interaction with HPC systems, and database setup and management introduce steep learning curves that can discourage researchers without background in computational science or programming. While there are tools that can support FAIR data practices, such as scientific workflows [2], they often introduce additional technical barriers which can increase the disconnection between technical capabilities and usability. These issues continue to hinder the consistent adoption of FAIR practices, not only in materials science, but several other domains that rely on computational methods.

The VirtMat Tools suite is a Python-based project for virtual research environments (VREs) [3], which includes the VRE-Middleware platform, and the VRE-Language package which contains the textM and textS domain specific languages (DSLs) for computational materials science and scientific computing, respectively. These tools provide a novel approach for the creation, execution, and analysis of complex models in the field of computational materials science. Through domain specific notations and abstractions, VirtMat Tools conceals technical complexities such as the workflow and resource management systems. Additionally, it ensures data persistence through database integration, while supporting FAIR compliance through the use of unique identifiers and different layers of metadata. These features allow VirtMat Tools to support the full data lifecycle of computational models in an accessible and user-friendly manner. As a result, it significantly lowers the barrier for domain scientists to engage with high-performance computing (HPC), scientific workflows, and FAIR-compliant simulations [4].

This contribution builds upon the ongoing development of the VirtMat Tools suite to meet the growing demand for robust, flexible, and sustainable research tools. One key aspect is the decoupling of crucial objects, such as calculators, which will allow the adoption of VirtMat Tools beyond the materials science community. This is achieved by leveraging the modularity of the textM/Texts grammar, and their supporting tools. Another key aspect is to enhance VirtMat Tools’ FAIR data capabilities in two directions: first, by improving findability through the implementation of persistent digital identifiers for models within the integrated database, which can be linked to specialized web services; second, by enhancing data accessibility through more flexible research output storage, and improved management of data and metadata for faster queries.

1. Wilkinson, M. D., et al.: The FAIR Guiding Principles for Scientific Data Management and Stewardship. Sci. Data. 3(1), 160018 (2016) 2. Roozmeh, M., Kondov, I.: Workflow Generation with wfGenes. In: 2020 IEEE/ACM Workflows in Support of Large-Scale Science (WORKS), pp. 9-16. Institute of Electrical and Electronics Engineers (IEEE), GA USA (2020) 3. VirtMat Tools GitLab repository https://gitlab.kit.edu/kit/virtmat-tools 4. Kondov, I., Cortes-Mejia, R., Müller, M., Pfisterer, N., Sreenivasan, S.: Domain Specific Language for Materials Modeling. In: Paszynski, M., Barnard, A.S.,Zhang, Y.J. (eds.) Computational Science - ICCS 2025 Workshops. ICCS 2025. Lecture Notes in Computer Science, vol. 15909, pp. 221-236. Springer, Cham. https://doi.org/10.1007/978-3-031-97564-6_18

ABSTRACT. With the growing demand for secure and efficient handling of large-scale image data in computational environments, lossless data embedding methods for encrypted visual content are gaining importance. This paper introduces a Reversible Data Hiding in Encrypted Images (RDHEI) scheme that allows embedding additional data into encrypted images while ensuring perfect recovery of the carrier. In this approach, reference pixels are selected using a predefined binary mask and compressed using Delta-Huffman coding. The remaining non-reference pixels are predicted using a kernel-based approach, with the kernel weights optimized for each image using Ridge regression. The resulting prediction errors are compressed using standard Huffman coding, thereby increasing the achievable embedding rate. The proposed method is lossless and has been evaluated on standard grayscale images, including those from the BOSSbase and BOWS2 datasets and other benchmark images commonly used in RDHEI research. Experimental results show that the method achieves a higher embedding rate compared to existing RDHEI techniques.

ABSTRACT. Privacy Amplification (PA) constitutes a critical computational bottleneck in high-rate Quantum Key Distribution (QKD) systems, particularly when processing large data blocks required to mitigate finite-size security effects. In this work, we propose a high-performance GPU implementation based on the Number Theoretic Transform (NTT), enabling exact modular arithmetic under a suitable modulus and root of unity for large-scale PA. We optimize the NTT execution on GPUs by integrating a hybrid butterfly computation scheme that combines a warp-shuffle-based approach and a shared-memory-based approach, together with kernel fusion techniques and Barrett-based modular reduction to maximize memory bandwidth utilization and parallel efficiency. Experimental results on an NVIDIA L40 GPU demonstrate a throughput of 3.32 Gbps for 2^27-bit input blocks. This result indicates that software-based NTT acceleration on commodity GPUs can support hundreds-of-Mbps-class Secret-Key-Rate (SKR) QKD systems.

ABSTRACT. We present a unified integer-only kernel for hyperbolic functions tanh(x) and sigmoid(x), designed for embedded RISC-V platforms without floating-point units. The kernel combines LUT anchoring, CORDIC microrotations, and linear correction in Q20 arithmetic, producing outputs directly in Q1.16 format suitable for ML inference. Exhaustive evaluation across the entireadvwb 17-bit input space demonstrates deterministic accuracy better than 1 ULP in Q1.16, with maximum absolute errors below 1.5 * 10^(-5). Compared to soft-float implementations, our approach achieves significant speedup while reducing hardware complexity by unifying multiple activation functions in a single IP block. This makes the method highly relevant for efficient deployment of neural networks on resource-constrained RISC-V microcontrollers.

ABSTRACT. In this paper, we study cache-efficient optimization and vectorization for the dynamic programming McCaskill algorithm, which computes the RNA partition function and base-pairing probabilities under a thermodynamic model. The McCaskill algorithm operates on the full ensemble of possible RNA secondary structures without pseudoknots, weighting them by their free energies to derive both the partition function and base-pairing probabilities. This task belongs to classical bioinformatics applications commonly benchmarked on multicore HPC systems. Well-known manual and automatic polyhedral code optimizations dedicated to dynamic programming tasks often avoid vectorization and focus solely on efficient loop tiling. As a result, such codes fail to exploit the vector computational capabilities of x86 processors and those available in newer architectures such as RISC-V. In this work, we demonstrate the generation of a readable, efficient OpenMP implementation of the McCaskill algorithm that leverages vector-level and cache-aware optimizations, outperforming related approaches. We analyze vectorization limitations on platforms lacking advanced optimization tools such as oneAPI and propose an efficient implementation based on RISC-V vector intrinsics, demonstrating that hardware capabilities significantly exceed what current compilers can utilize.

ABSTRACT. Fredholm determinants play a vital role in mathematics and physics, notably in integral equations theory and random matrix theory. Recently, the Fredholm determinants were proposed as an appropriate measure (especially for infinite dimensional continuous quantum systems) of quantum entanglement, which is one of the basic quantum information resources applied to the construction of the quantum computational machines. While the most studies use sequential numerical methods to compute Fredholm determinants, parallel programming techniques can be also effective. This paper introduces a solution utilizing multi-core processors and GPUs with a suitable quadrature, implemented in Python for broad accessibility. We also show how to use proposed routines to compute von Neumann entropy for bosonic states and superposition of Fock states.

ABSTRACT. In this paper we study mixed-precision training in a hybrid neural network combining a photonic quantum processor with classical computation on an HPC cluster equipped with NVIDIA H100 GPUs. In such hybrid quantum–classical models, gradients are estimated from a finite number of quantum measurement shots, introducing stochastic sampling noise that interacts with classical floating-point precision. We experimentally compare full FP32 training with mixed-precision (BF16/AMP) execution in a hybrid neural network applied to binary MNIST classification. Using gradient variance as a diagnostic metric, we analyze the interplay between quantum sampling fluctuations and numerical rounding effects. The results show that in low-shot regimes typical of near-term photonic devices, quantum noise dominates and mixed precision does not degrade training stability, while at higher shot counts numerical precision becomes increasingly relevant. These findings provide practical guidance for precision-aware optimization in hybrid quantum–HPC workflows.

ABSTRACT. Probabilistic Neural Networks (PNNs) are memory-based networks that have been used successfully for classification and regression tasks. Training of PNNs may be performed by analytical methods, e.g., plug-in or by heuristic methods, e.g., Particle Swarm Optimization. Heuristic methods are superior to traditional PNN training techniques because of their nonparametric behavior, independence of the PNN kernel selection, and ability to optimize method parameters for a given problem. One of the recently proposed training algorithms - \textit{constrained Hybrid Metaheuristic} (cHM), overperformed other analytical and heuristic procedures on a variety of datasets. Here, we present a further evaluation of the cHM method for training PNNs for classification tasks. In particular, we study its effectiveness and stability with different hyperparameters across $10$ datasets. The results show that the cHM training procedure is stable and the parameter selection does not significantly impact the PNN training accuracy for 8 of 10 tested datasets. In addition, the best cHM hyperparameters for triggering PNNs are proposed for each dataset. This data proves that the cHM method can be successfully applied to train PNNs without bias on the inner algorithm parameters.

ABSTRACT. Development of new antennas is an inherently cognitive task that often involves re-use of structures from the literature, or their responses (e.g., for performance comparisons). The process might also be associated with reconstruction of their electromagnetic (EM) simulation models which—when performed manually—is both time-consuming and prone to errors. In this work, a proof-of-concept framework for image-based, non-parametric reconstruction of antenna EM models has been proposed. The method boils down to extraction of shape-related coordinates from photograph of a structure, followed by their processing and incorporation to a script that enables reconstruction and simulation of EM model. The approach has been demonstrated using two antennas.

ABSTRACT. Maximization of yield is essential for reducing the cost of en masse manufactured antennas. Conventional approach to the problem, based on Monte Carlo (MC) analysis, are impractical for modern antennas due to overwhelming cost associated with their simulations. In this work, a surrogate-assisted optimization of yield for multi-dimensional structures is considered. The method shifts expensive MC to an approximation surrogate that is iteratively re-set in the course of optimization. The algorithm embeds a mechanism for automatic tuning of yield for nominal designs that violate the specifications. The method has been demonstrated using two topology-agnostic antennas represented using 22 and 52 parameters, respectively. Acceptable optimization cost (time-wise) is maintained by evaluation of training designs for surrogate identification using an in-house distributed computing system. The maximized yields for considered antennas amount to 97% and 81%, respectively. Comparisons of the results against yield estimated based on direct EM simulations are also provided.

ABSTRACT. Optimizing already trained neural networks is one of the core problems in the domain of Artificial Intelligence. In this paper, we present a new approach capable of optimizing the structure of an already trained neural network. We present a new way of removing neurons from layers that is based on matrix scaling, which truly decreases the number of parameters in a model rather than zeroing weights. We present a simulation-based approach for optimizing the structure of a neural network that can select the best change to the network without causing data loss. A suite of empirical experiments demonstrated that the new approach can optimize already-trained neural networks, achieving up to 85% parameter reduction while maintaining approximately 95% training accuracy, without losing more than 5% in most cases. Optimized models use as little as 2% of the parameters of ResNet-18 with comparable performance, supported by a neuron removal method based on matrix rescaling that preserves learned information.

ABSTRACT. Research hypothesis generation from scientific literature using Large Language Models (LLMs) remains largely prompt-based, with limited work on model alignment. We present one of the few systematic reviews of existing datasets for hypothesis generation, analyzing their structure and suitability for alignment. Based on this analysis, we introduce a dataset designed for training and aligning LLMs that encodes literature-derived background knowledge as structured concept connections, along with evaluation metrics grounded in references and their content. Using this dataset, we study how alignment with structured knowledge affects the novelty and grounding of generated hypotheses.

ABSTRACT. We propose an improved Fourier Neural Operator (FNO) based approach to modeling two-dimensional Rayleigh-Bénard convection (2D RBC). By learning increments over time instead of solutions, we improve accuracy compared to the standard FNO baseline model. With 314k parameters, our improved model is extremely lean and only requires 1.26 MB when trained for 11 500 epochs on NVIDIA A100 GPUs and has an inference time of 7 ms, compared to 1.18 MB and 5 ms for the baseline model. FNO generalizes to finer meshes than were used in training, but, as can be expected, overall prediction accuracy is limited by the resolution of the training data.

ABSTRACT. Time-series forecasting and causal discovery are central in neuroscience, as predicting brain activity and identifying causal relationships between neural populations and circuits can shed light on the mechanisms underlying cognition and disease. With the rise of foundation models, an open question is how they compare to traditional methods for brain signal forecasting and causality analysis, and whether they can be applied in a zero-shot setting.

In this work, we evaluate a foundation model against classical methods for inferring directional interactions from spontaneous brain activity measured with functional magnetic resonance imaging (fMRI) in humans. Traditional approaches often rely on Wiener–Granger causality. We tested the forecasting ability of the foundation model in both zero-shot and fine-tuned settings, and assessed causality by comparing Granger-like estimates from the model with standard Granger causality. We validated the approach using synthetic time series generated from ground-truth causal models, including logistic map coupling and Ornstein–Uhlenbeck processes. The foundation model achieved competitive zero-shot forecasting fMRI time series (mean absolute percentage error of 0.55 in controls and 0.27 in patients). Although standard Granger causality did not show clear quantitative differences between models, the foundation model provided a more precise detection of causal interactions.

Overall, these findings suggest that foundation models offer versatility, strong zero-shot performance, and potential utility for forecasting and causal discovery in time-series data.

ABSTRACT. Data stream mining increasingly relies on models that remain reliable under non-stationarity. This paper targets feature drift, i.e., changes in the predictive relevance ordering of features over time. We propose FRDD (Feature Ranking-Based Drift Detector), a chunk-based detector that monitors rank-stability violations rather than classification error. FRDD builds feature-wise acceptance intervals from rank variability estimates within a reference chunk and raises an alarm when a significant fraction of monitored features violates these intervals. The approach is modular in the ranking procedure and supports both label-aware (LASSO) and label-free (Laplacian Score) operations. We evaluate FRDD under a unified adaptation policy in which every alarm triggers the same downstream model reset/retrain procedure, ensuring that performance differences reflect detector behavior. Experiments on synthetic streams with programmed abrupt, gradual, incremental, and recurring drifts and on four real-world streams show that the supervised FRDD variant is FP-controlled (few alarms) while maintaining competitive predictive quality, whereas the unsupervised variant is more reactive and may fragment slow drifts. Overall, the results indicate that monitoring feature-rank dynamics provides a practical early-warning signal for initiating adaptation in evolving streams.

ABSTRACT. In contemporary computing, we can observe simultaneous: adoption of virtualization that has become mainstream, popularity of AI workloads, as well as very considerable increase of energy costs. The latter result from high prices and huge requirements of the aforemen- tioned workloads. In this context, we provide performance and energy- aware comparison of virtual machine and container based environments against a bare metal high performance system for running three con- temporary AI workloads. We used Single Shot Detector (SSD) used in MLPerf Inference Benchmarks for Image Classification and Object De- tection, and two workloads from the standard PyTorch benchmark suite: resnet50 training and hf_Bert inference, both with 32-bit floating point precision. Compared to the state-of-the-art, we contribute by: using a multi(8)-GPU environment, considering another variable being a power cap imposed on each of the GPUs, and assessment of not only result- ing execution times and energy consumption but also the Energy Delay Product (EDP) metric. Power capping allows to minimize energy con- sumption and EDP compared to the default power limit. We have deter- mined that Docker results in virtually negligible performance overhead (0.03% on average) while KVM imposes performance penalty of 21.22% on average, compared to bare metal. For energy consumption, overheads of Docker and KVM are -0.03% and 1.63% while for EDP the overheads are 0% and 23.14% on average, respectively (39.42% in the worst case for KVM). We have also determined that power caps that minimize EDP are different for KVM and Docker/bare metal which is an important ob- servation for optimization of that essential performance-energy trade-off.

ABSTRACT. Automated image classification is a fundamental task in computer vision. One of the most powerful deep learning (DL) models employed in image classification is the Convolutional Neural Network (CNN). Its key strength is the automatic extraction of features during direct raw image processing. This work proposes an ensemble transfer learning model based on CNNs to classify wheat grains. The dataset consists of 288 images of the kernels of three wheat grain varieties. In order to build the ensemble model, VGG16, InceptionV3, Dense-Net201-V1, DenseNet201-V2, and NASNetMobile were evaluated and com-pared on the basis of the performance metrics including accuracy, Cohen’s Kappa, precision, recall and F1 score. The ensemble CNN model demonstrated a higher level of performance than its base models achieving an accuracy of 96.48% and F1 score of 97%. Despite the limited amount of data available, the study's findings indicate that the proposed approach is an effective means of enhancing wheat grain classification.

ABSTRACT. 3D scene reconstruction from 2D images is one of the most important tasks in computer graphics. Unfortunately, existing datasets and benchmarks concentrate on idealized synthetic or meticulously captured realistic data. Such benchmarks fail to convey the inherent complexities encountered in newly acquired real-world scenes. In such scenes especially those acquired outside, the background is often dynamic, and by popular usage of cell phone cameras, there might be discrepancies in, e.g., white balance. To address this gap, we present HuSc3D,a novel dataset specifically designed for rigorous benchmarking of 3D reconstruction models under realistic acquisition challenges. Our dataset uniquely features six highly detailed, fully white sculptures characterized by intricate perforations and minimal textural and color variation. Furthermore,the number of images per scene varies significantly, introducing the additional challenge of limited training data for some instances alongside scenes with a standard number of views. By evaluating popular 3D reconstruction methods on this diverse dataset, we demonstrate the distinctiveness of HuSc3D in effectively differentiating model performance, particularly highlighting the sensitivity of methods to fine geometric details, color ambiguity, and varying data availability – limitations often masked by more conventional datasets.

ABSTRACT. In CT angiography the accurate segmentation of abdominal aortic aneurysms (AAAs) is difficult due to large anatomical variability, low-contrast vessel boundaries, and the close proximity of organs whose intensities resemble vascular structures, often leading to false positives. To address these challenges, we propose an anatomy-aware segmentation framework that integrates organ exclusion masks derived from TotalSegmentator into the training process. These masks encode explicit anatomical priors by identifying non-vascular organs and penalizing aneurysm predictions within these regions, thereby guiding the U-Net to focus on the aorta and its pathological dilation while suppressing anatomically implausible predictions. Despite being trained on a relatively small dataset, the anatomy-aware model achieves high accuracy, substantially reduces false positives, and improves boundary consistency compared to a standard U-Net baseline. The results demonstrate that incorporating anatomical knowledge through exclusion masks provides an efficient mechanism to enhance robustness and generalization, enabling reliable AAA segmentation even with limited training data.

ABSTRACT. Physics Informed Neural Networks (PINNs) are an increasingly popular approach to utilizing tools and infrastructure developed for neural network training to solve PDEs. While the mainstream approach is based on strong formulations, Variational PINNs (VPINNs) have been proposed to tackle problems with lower regularity. Their issues with robustness, which manifest as a disconnect between the value of the loss function used for training and the error in the relevant Sobolev norm, can be alleviated by employing Robust Variational PINNs (RVPINNs) at the expense of efficiency due to the cost of integrating terms involving the neural network and factorizing the Gram matrix. Collocation-based Robust Variational PINNs (CRVPINN) aim to regain the efficiency of PINNs while retaining robustness by applying the RVPINN framework to variational formulations based on discrete grids and finite difference approximations. Here, we extend the CRVPINN method to three-dimensional domains and validate it on Poisson and stationary advection-diffusion problems.

ABSTRACT. Image aesthetic attribute assessment has become an explainable output of Image Aesthetic Quality Assessment (IAQA), which aims to evaluate aesthetic attributes such as rule of thirds, color harmony, and object emphasis. Image aesthetic attributes are inherently context-dependent, as their importance varies across different categories of photography scene and semantic content. Most existing IAQA methods for predicting aesthetic attributes are scene-agnostic, which becomes a fundamental obstacle to effectively modeling the dependency between scene and attributes. In this work, we propose a scene-aware model for image aesthetic attribute assessment that uses a vision transformer backbone to extract multi-level aesthetic representations. The extracted features are combined with learned scenes embedded in a two-tower module that models both general and scene-dependent patterns. An adaptive gating mechanism fuses the predictions according to scene context. Experimental results demonstrate that modeling scene dependencies improves correlation performance for both overall aesthetic score prediction and aesthetic attribute assessment, outperforming state-of-the-art attribute-based IAQA methods.

ABSTRACT. This paper reports recent benchmarks on current hardware of the simulation software Delft3D FM for test cases from real-life problems in surface water. For this typical and quite broad range of model applications there is need for efficient and scalable high performance computing (HPC). As the scope is also broad in the variety of hardware being used, optimization for HPC on specific hardware and/or real-life problems is not practical. Therefore, a general approach is followed within the field of computational science for which benchmarks play a central role. It is an iterative process over years: outcomes of benchmarks serve both for current applications to see what can be expected, on short time scales to adapt models where needed/possible given current simulations software and/or hardware, and on longer time scales to adapt simulation software for given trends in application and/or hardware. For this, benchmarks of Delft3D FM were performed before. As both hardware, simulation software, and models further developed, new benchmarks were required to get an actual status. The paper starts with a small historical overview of HPC for Delft3D FM and the numerical background. Also background is given of the real-life applications from which test cases are selected. This underlying information of simulations software and test cases, is necessary to evaluate outcomes of the new benchmarks. Benchmark results are reported for the separate test cases and followed by a global analysis. This will be a basis for a next iteration, to further improve Delft3D FM.

ABSTRACT. Gaussian Splatting (GS) is a recent and pivotal technique in 3D computer graphics. GS-based algorithms almost always bypass classical methods such as ray tracing, which offer numerous inherent advantages for rendering. For example, ray tracing can handle incoherent rays for advanced lighting effects, including shadows and reflections. To address this limitation, we introduce MeshSplats, a method which converts GS to a mesh-like format. Following the completion of training, MeshSplats transforms Gaussian elements into mesh faces, enabling rendering using ray tracing methods with all their associated benefits. Our model can be utilized immediately following transformation, yielding a mesh of slightly reduced reconstruction quality without additional training. Furthermore, we can enhance the quality by applying a dedicated optimization algorithm that operates on mesh faces rather than Gaussian components. Importantly, MeshSplats acts as a wrapper, converting pre-trained GS models into a ray-traceable format. The efficacy of our method is substantiated by experimental results, underscoring its extensive applications in computer graphics and image processing.

ABSTRACT. Efficient management of hospital Emergency Departments (EDs) is essential to address problems such as service saturation or bottlenecks caused by an unexpected influx of patients and/or resource limitations. Adjusting healthcare personnel to the needs of the system at any time is a difficult challenge to address, as it requires anticipating future situations. Our research explores the use of a Digital Twin (DT) that simulates the operation of the ED to address this challenge. The DT, which operates at a faster speed than the real system, allows us to anticipate future scenarios and prevent potential problems to anticipate their behavior. Furthermore, a heuristic optimization method designed to efficiently explore multidimensional and combinatorial search spaces allows us to find the configuration of healthcare personnel that manages to maintain normal operating of the ED, assuming limitations in the personnel resources in the service. This article describes in detail the entire process proposed to determine the healthcare staff configuration that best fits the requirements of the real service.

ABSTRACT. Equivalence graphs (e-graphs) are data structures used in rewriting systems for representing expressions from a language, e.g. a programming language. An algorithm called equality saturation expands an e-graph representing a single expression to an e-graph representing all expressions equivalent to the starting one in accordance with a set of rewriting rules. Although effective, this process is computationally expensive, and parallelization remains largely unexplored. We evaluate parallel strategies for saturation and extraction, demonstrating significant speedups in matching and extraction phases, while highlighting fundamental limitations in parallelizing rule application within current architectures. Additionally, we introduce a general-purpose parallel Union-Find data structure featuring efficient lock-free path compression.

ABSTRACT. The study presents a reconstruction approach tailored to the demands of low‑dose, single‑slice computed tomography, emphasizing both radiation reduction and high computational efficiency. By focusing on scenarios in which only one diagnostically relevant slice is needed, the method avoids the overhead of volumetric processing and delivers images in a fraction of the time required by conventional algorithms. Although the highest spatial frequencies appear slightly softened, the reconstructed slices maintain sufficient diagnostic quality for rapid clinical assessment. This trade‑off is particularly advantageous in emergency settings, where immediate visualization is more important than maximal sharpness. The approach therefore offers a practical solution for trauma care, accident evaluation, and other time‑critical situations in which fast, low‑dose imaging supports urgent medical decisions. Future extensions may incorporate artificial intelligence to enhance fine‑detail recovery and adapt reconstruction quality to varying dose levels, creating a hybrid framework that combines the reliability of physics‑based modeling with the flexibility of data‑driven enhancement.

ABSTRACT. We implement the Lights Out problem on a 2D grid and on Möbius ladder graphs and evaluate the performance of Grover's search on real quantum hardware. We use two instances using 9 and 16 qubits, and implement them on publicly available quantum hardware by IBM and IQM. Our experiments show improvements in IBM hardware between the Heron r1 and Heron r2 generations, highlighting progress in IBM hardware during the 2023-2024 period. The Lights Out circuits produced output distributions close to uniform on IQM devices. To diagnose device limitations, we additionally ran a small Grover SAT baseline, finding that IQM Garnet performs more reliably than other tested IQM devices. We also observed that QPUs of the same manufacturing revision can differ significantly in performance (a newer device is not guaranteed to be better), and that calibration has a significant impact on the performance of quantum devices, so the choice of device strongly depends on calibration quality.

ABSTRACT. Agricultural sustainability assessment requires methods that can compare environmental performance across diverse regions using consistent, well-accepted indicators. We present a reproducible computational workflow that transforms official European statistics (Eurostat), CORINE Land Cover and Copernicus High Resolution Layers into spatially explicit sustainability indicators---applicable to any European NUTS2 region. The workflow computes established metrics including Energy Return on Investment (EROI), Energy-Landscape Integration Assessment (ELIA), greenhouse gas emissions, and nutrient balances, enabling transparent cross-regional comparison. Applied to three contrasting regions (NL23 Flevoland, ES53 Illes Balears, ITF2 Molise) over 2010--2020, we demonstrate that shifts in agricultural composition---not land area or intensity changes---drive 56--92\% of observed sustainability transitions. This finding reveals that \emph{what} is produced matters as much as \emph{how} or \emph{where}. Since composition changes dominate, land-management instruments alone (e.g., buffer strips, cover crops) may be insufficient to address the primary driver of change. Our approach integrates (i) mass-balance-preserving downscaling enabling re-aggregation to any spatial boundary, (ii) standardised socioecological indicators with explicit, configurable coefficients, and (iii) Shapley decomposition providing axiomatic, unbiased attribution of observed changes to land, intensity, and composition factors.

ABSTRACT. Generating graphs subject to strict structural constraints is a fundamental computational challenge in network science. Simultaneously preserving interacting properties—such as the diameter and the clustering coefficient— is particularly demanding. Simple constructive algorithms often fail to locate vanishingly small sets of feasible graphs, while traditional Markov-chain Monte Carlo (MCMC) samplers suffer from severe ergodicity breaking. In this paper, we propose a two-step hybrid framework combining Ant Colony Optimization (ACO) and MCMC sampling. First, we design a layered ACO heuristic to perform a guided global search, effectively locating valid graphs with prescribed diameter and clustering coefficient. Second, we use these ACO-designed graphs as structurally distinct seed states for an MCMC rewiring algorithm. We evaluate this framework across a wide range of graph edge densities and varying diameter-clustering-coefficient constraint regimes. Using the spectral distance of the normalized Laplacian to quantify structural diversity of the resulting graphs, our experiments reveal a sharp contrast between the methods. Standard MCMC samplers remain rigidly trapped in an isolated subset of feasible graphs around their initial seeds. Conversely, our hybrid ACO-MCMC approach successfully bridges disconnected configuration landscapes, generating a vastly richer and structurally diverse set of valid graphs.

ABSTRACT. Reaching for grasping, and manipulating objects are essential motor functions in everyday life. Decoding human motor intentions is a central challenge for rehabilitation and assistive technologies. This study focuses on predicting intentions by inferring movement direction and target location from multichannel electromyography (EMG) signals, and investigating how spatially and temporally accurate such information can be detected relative to movement onset. We present a computational pipeline that combines data-driven temporal segmentation with classical and deep learning classifiers in order to analyse EMG data recorded during the planning, early execution, and target contact phases of a delayed reaching task.

Early intention prediction enables devices to anticipate user actions, improving responsiveness and supporting active motor recovery in adaptive rehabilitation systems. Random Forest achieves 80% accuracy and Convolutional Neural Network 75% accuracy across 25 spatial targets, each separated by 14° azimuth/altitude. Furthermore, a systematic evaluation of EMG channels, feature sets, and temporal windows demonstrates that motor intention can be efficiently decoded even with drastically reduced data. This work sheds light on the temporal and spatial evolution of motor intention, paving the way for anticipatory control in adaptive rehabilitation systems and driving advancements in computational approaches to motor neuroscience.

ABSTRACT. This paper presents a semi-supervised patch-memory k-nearest neighbors approach for hazard detection and localization in indoor mobile robots. A convolutional neural network (CNN) backbone is used to extract multi-scale representations, which enable anomaly assessment at both the global image and patch levels via a memory bank of normal patches. Evaluation is performed on a generative benchmark extension of the public MDDRobots dataset, comprising 2,669 test images from independent sequences under varying environmental conditions and covering eight hazard categories. The experimental results show robust performance, with ResNet-18 achieving the highest image-level detection accuracy, measured by the area under the receiver operating characteristic curve (ROC AUC = 0.7282) and VGG-16 obtaining patch-level localization (ROC AUC = 0.7918). The proposed approach is interpretable and suitable for safety-critical robotic perception.

ABSTRACT. Human–Artificial Intelligence (AI) collaboration is advancing with the commercialization of AI-based decision support systems. There is a rigorous effort to improve the AI model's ability to align with human experts' solutions. Yet, deriving the human expert’s decision process to select a solution is a complex task. We present an experiment to model the human optimal seeking behaviour as part of the human complexity state when perceiving information of process models during evaluation. Data from an independent targeted survey acquires human experts’ subjective understandability, correctness and usability of medical procedures represented via a free-to-explore process model state space. Descriptive process modelling metrics, replay fitness, precision, generalization and simplicity relative to each medical procedure is introduced as external context and the user’s exploration behaviour as internal context. Using Markov Decision Processes and reinforcement learning, we model and analyse the human exploration-exploitation behaviour during perceiving to derive the underlying policy-reward function, therefore revealing subjective decision-making criteria of human experts when selecting optimal process models. Our results provide insights into modelling human agents’ decision behaviour by exploration and exploitation states as part of the state space environment.

ABSTRACT. Scientific simulations grow increasingly complex, and require massive compute resources (e.g., supercomputers) and energy to execute within reasonable time-bounds.To address these requirements, new HPC processors emerge regularly. Quantifying the performance and energy-efficiency impact of these HPC processors on existing HPC applications is non-trivial, yet critical for building the next generation of (super)computers. In this work, we demonstrate how comprehensive benchmarking contributes to quantifying the (potential) performance and energy efficiency impact of six different processors on HemoCell, a complex coupled scientific simulation. To this end, we constructed representative benchmarks and ran hundreds of HemoCell simulations to provide both coarse-grained data, focusing on metrics such as utilization and speed-up, and fine-grained data, based on hardware event counters. Our detailed data analysis provides further insights into the impact of higher memory bandwidth, the need for higher cache capacity, the relevance of high core-density, and the effect of low-power designs. Our results show that increased cache capacity is the highest-impact feature for both the performance increase and energy consumption reduction of HemoCell.

ABSTRACT. The article suggests a model describing the dynamics of offenses in terms of not wearing a seat belt and exceeding the established speed limit, taking into account the set regulatory norms and parameters of national specifics. The study examines the issue of drivers' sensitivity to a sudden change in government policy, the impact of penalty policies on driver behavior within the framework of the model.

ABSTRACT. Fuzzy Petri nets serve as a sophisticated modeling paradigm, merging classical Petri net theory with fuzzy logic. This powerful combination facilitates the representation of and reasoning within systems characterized by imprecise, ambiguous, or uncertain information, a common challenge in many real-world applications, particularly decision support systems. This study sought to evaluate the decision-making efficacy of an example decision support system model operating under uncertainty. The evaluation was contingent upon the specific fuzzy Petri net class selected, the type and size of the dataset employed, and the predefined assessment criteria. The research incorporated three distinct classes of fuzzy Petri nets, varying in modeling power, along with four numerical datasets possessing diverse statistical characteristics, and several metrics for gauging model effectiveness and efficiency. The experimental phase involved net models of a simple train traffic control system, utilizing bespoke specialized software developed for the automatic simulation of various net models. The authors propose that this work addresses challenges outlined in the comprehensive review article on the state of research in fuzzy Petri net theory and applications by K. Zhou and A. Zain, published in Artif. Intell. Rev., 45, 405-446 (2016).

ABSTRACT. Learning from distributed data, maintained independently and analyzed without full central integration, poses significant challenges for building coherent and reliable classification models. In such environments, local datasets may differ not only in content but also in class distributions, which can affect the quality of the resulting global model. This paper extends the authors’ previously proposed distributed classification framework that integrates conflict analysis, coalition formation, and decision rule induction. The main novelty lies in incorporating a class balancing stage applied independently to each local dataset prior to the construction of the system. Six techniques representing undersampling, oversampling, and hybrid strategies are examined: Random Undersampling, NearMiss, Tomek Links, Random Oversampling, SMOTE, and SMOTE-Tomek. Decision rules are induced using four rough set-based algorithms: the exhaustive search algorithm, the genetic algorithm, the covering algorithm, and LEM2. Final decisions are determined through three strategies: first rule approach, all rules approach, and weighted rule approach. The experiments were conducted on two datasets from the UCI Machine Learning Repository: Car Evaluation and Balance Scale. The proposed approach was compared with a baseline without class balancing and with a balancing-only variant omitting coalition formation. The results indicate that class distribution adjustment improves imbalance-sensitive metrics under severe class imbalance, with a moderate reduction in overall accuracy. Under such conditions, coalition formation appears to contribute to improvements in both predictive performance and class-balanced outcomes.

ABSTRACT. Counterfactual explanations provide interpretable insights into classifier decisions by identifying minimal input modifications that alter predictions. While extensively studied for tabular data, visual counterfactuals present unique challenges requiring semantically meaningful changes rather than imperceptible perturbations. Current approaches predominantly employ diffusion models, GANs, and VAEs, while normalizing flows remain underexplored despite offering tractable likelihood computation. We introduce PLACE (Plausible LAtent Counterfactual Explanations), a method that leverages conditional normalizing flows for explicit density estimation in counterfactual generation. Operating in the latent space of a pre-trained autoencoder, PLACE optimizes a novel composite loss function balancing validity, proximity, and plausibility. The plausibility term directly maximizes log-likelihood under the target class distribution, enabled by the flow's tractable density computation. Experiments on CelebA and MNIST demonstrate that PLACE achieves competitive performance across multiple metrics while uniquely satisfying explicit plausibility constraints through substantially improved log density scores. Our method balances computational efficiency with multi-objective optimization, validating normalizing flows as an effective approach for probabilistically constrained visual counterfactual explanations.

ABSTRACT. We present a preliminary study of a system for face spoofing detection based on a multi-stream CNN architecture that operates on a single RGB image taken with a simple camera, such as the front-facing camera of a smartphone. The network extracts and processes features through four parallel streams: a local texture convolution with Convolutional Block Attention Module, a color analysis with large-kernel convolutions, a context analysis with dilated convolutions, and frequency analysis with the Fast Fourier Transform to identify high-frequency artifacts. The proposed architecture achieves an accuracy of 99.28%, and an AUC value of 99.94%, on the CelebA-Spoof dataset, and values of 97.13% and 99.18% for the CASIA-FAS dataset, respectively.

ABSTRACT. In the Maximum Induced Acyclic Subgraph problem, the objective is to determine a largest subset of vertices that induces an acyclic subgraph. This problem is equivalent to the Directed Feedback Vertex Set problem (DFVS), in which one seeks a minimum-size set of vertices whose removal makes a given directed graph acyclic. Both problems are known to be NP-complete. Consequently, it is natural to address them using heuristic methods and preprocessing techniques. In contrast to many other combinatorial optimization problems, however, a limited number of data reduction rules are known for the DFVS. In this paper, we introduce the new concept of partial domination and develop data reduction rules derived from this notion. Moreover, we present a linear-time algorithm for identifying cycle-dominators, a problem of considerable importance in various areas of computational science, most notably in control-flow graph analysis. We provide a thorough description and analysis of the proposed algorithms, along with results from computational experiments that demonstrate their practical effectiveness.

ABSTRACT. A comprehensive understanding of ground-level air pollutants is essential for developing policies and strategies that effectively reduce associated human health, environmental, and economic risks. Nitrogen dioxide is a key air pollutant that poses its own risks to human health and is also essential to the formation of secondary air pollutants, including ozone and particulate matter. Air quality monitoring technologies enable concentration measurements at high temporal frequencies but are spatially constrained by site locations. In contrast, remote sensing satellites expand spatial coverage, yet accurately capturing short-term exposure variability remains challenging. In 2023, NASA successfully launched the TEMPO (Tropospheric Emissions: Monitoring of Pollution) geostationary satellite, which enables high spatial resolution retrievals at hourly intervals. Advancements in deep learning and computer vision effectively leverage spatial context to more accurately estimate ground-level pollution. Our research integrates these technological advancements within a deep learning framework capable of estimating hourly ground-level NO_2 concentrations at spatial scales as fine as 10 meters across the continental United States. This methodology can expand knowledge of sources driving diurnal variation of ground-level NO_2, inform emissions control policy and technologies, and deepen understanding of drivers of atmospheric chemical processes. Our thorough experimental evaluation offers several useful insights.

ABSTRACT. Continuous lesson planning requires an incessant search of large volumes of pedagogical materials, an exhausting task that frequently results in frustration for teachers and other educational professionals. Although digital platforms are allies in the search for teaching materials, they commonly face the cold-start problem, the initial difficulty in recommendation due to the lack of prior data for collaborative filtering. To mitigate this challenge, this study proposes and evaluates a recommendation system with content-based filtering, extracting information from the documents themselves to suggest educational activities, named as Learning Sequences. We conducted a comprehensive analysis comparing traditional information retrieval techniques (Bag of Words, TF-IDF) and dense word embedding models, including static (Word2Vec, GloVe, FastText) and contextualized (BERT) representations. Additionally, the PageRank algorithm was adapted to operate on textual similarity graphs, based on the relevance of the documents, named as Global and Local PageRank. The methods were evaluated through multiple ranking and classification metrics, such as Hit Rate, MRR, NDCG, and F1-Score. The results show an efficient method to recommend learning sequences based on textual similarity using TF-IDF and cosine distance-based method.

ABSTRACT. Machine learning models can aid content moderation and make big data analysis possible. However, the quality of their performance depends on the datasets they are trained on, which very often are prepared with the help of human annotators. In the context of prejudice detection, this poses ethical concerns related to exposing humans to large volumes of hate speech, which comes with psychological burden and mental health risks. To address this issue, the machine learning community has come up with two possible solutions. One approach is the manual labelling of the dataset by the researchers themselves and therefore mitigating the risk for unprepared annotators [e.g., 2, 4]. The other solution focuses on practices grounded in care for the mental wellbeing of annotators at every stage of the construction of a dataset [3]. One limitation of those approaches is that they are not easily scalable. They work well for small datasets, but the creation of large volumes of annotated data consumes a lot of resources. This is especially problematic in the context of hate speech classification, where the quality of the dataset is crucial for the accuracy, reliability and validity of the classifier. To address this issue, we propose the Neo-Psychometric Machine Learning approach. Broadly speaking, this approach uses psychometric measures and domain knowledge for the broader purpose of addressing Machine Learning tasks. In this study, we propose a theory-driven classifier architecture. We utilize Contextualized Construct Representations [1] – similarity metrics of the clas- sified text and relevant psychometric scale item embeddings extracted from a sentence embedding model – as features for traditional classification models. We demonstrate this approach on the task of classifying anti-trans social media posts from TIDEs dataset (Transgender and Nonbinary Community-Labeled Dataset for Transphobia Identification in Digital Environments), containing 3,509 an- notated posts (42.7% Transphobic). Our Anti-Trans Hate Speech classifier as- sumes trans-related posts as inputs. It utilizes the pre-trained all-MiniLM-L12-v2 model, without fine-tuning, and its associated tokenizer to produce the text em- beddings. Then, we calculate the distance metrics to questionnaire items selected to operationalize the Transmisogyny theory [5]. Finally, we use one of the three classifiers tested: Logistic Regression, XGBoost and Support Vector Machine. Our results show that the Neo-Psychometric NLP approach can be success- fully applied to the task of classifying anti-trans social media posts (see Table 1, which presents the comparison of the Neo-Psychometric Model to other classifiers trained by Lameiro et al. [3]). The classifiers based on the Neo-Psychometric NLP approach achieved comparable performance to the DeBERTa black box classifier, while reducing the risk of overfitting on construct-unrelated features and providing potential for theory-driven explanations. While the Neo-Psychometric approach would, in our view, not be best suited in applications where the aim is to maximize accuracy, profit, or other per- formance metrics, we believe it can be applied in situations where theoretical backing, explainability and reducing construct-unrelated model bias is crucial – with good results.

Bibliography [1] Atari, M., Omrani, A., Dehghani, M.: Contextualized construct representa- tion: Leveraging psychometric scales to advance theory-driven text analysis (2023). https://doi.org/10.31234/osf.io/m93pd [2] Channon, L., Mathieson, N.: Automated detection of mainstreamed trans- phobic content on youtube. Bulletin of Applied Transgender Studies 4(1-3), 41–75 (2025). https://doi.org/10.57814/49jz-0663 [3] Lameiro, F., Dunagan, L., Card, D., Gilbert, E., Haimson, O.: Tides: A trans- gender and nonbinary community-labeled dataset and model for transphobia identification in digital environments. In: Proceedings of the 2025 ACM Con- ference on Fairness, Accountability, and Transparency. p. 1411–1423. FAccT ’25, Association for Computing Machinery, New York, NY, USA (2025). https://doi.org/10.1145/3715275.3732095 [4] Leitner, M., Dorn, R., Morstatter, F., Lerman, K.: Character- izing network structure of anti-trans actors on tiktok (2025), https://arxiv.org/abs/2501.16507 [5] Serano, J.: Whipping girl: a transsexual woman on sexism and the scape- goating of femininity (2007)

ABSTRACT. In the paper, we show how to enhance classification results of already trained models using the authors’ Invariant Dataset Augmentation (IDA) method. The IDA method allows to increase the classification rates considering as an example the classification of the skin lesions using a small image set i.e. PH2 with 200 images, one big dermoscopic dataset i.e. Derm7pt with 1011 images, and one big COVID-19 dataset with 3175 images. To solve the problem of the lack of rotation invariance, IDA method was used and the dataset was increased in an eightfold or twofold and invariant way. In the research, we applied the IDA methods and compared the results of VGG19, XN and Inception-ResNet-v2 CNN networks in three skin lesions features classification defined by well-known dermoscopic criterions e.g. the Three-Point Checklist of Dermoscopy or the Seven-Point Checklist for PH2 and Derm7pt. For COVID-19, VGG19, XN and Inceptionv3 CNN net-works were used. Due to Invariant Dataset Augmentation, the classification rate parameters like true positive rate by almost 20%, false positive rate as well as the F1 score and Matthews correlation coefficient have been significantly increased opposite to type II error that has significantly decreased. In the paper, the confusion matrix parameters result in 98-100% accuracy, 98-100% true positive rate, 0.0-2.3% false positive rate, tests F1=0.95 and MCC=0.95. In the COVID-19 case and viral pneumonia the approach to classify COVID-19 in chest X-ray images is shown, we use 3175 X-ray images: 540 cases of COVID-19 extracted from available datasets, 1294 viral pneumonia cases and 1339 cases taken from healthy patients. 25 % of the images have been used as a training set. The classification results of COVID-19 using VGG19 and Xception networks resulted in >97.5% of the accuracy of COVID-19, >98.5%, of the balanced accuracy, around 97%, of the true-positive rate (sensitivity); ~1.0 F1 test and >0.97 Matthews correlation coefficient. That general approach can provide higher results while using CNN networks in other applications.

ABSTRACT. In this paper, we consider a temporally high-order, decoupled, linear, and fully discrete finite element method for the Biot model. The generalized BDF method and the simplified auxiliary variable method with correction are adopted to discretize the proposed Biot model in time. Compared with the existing works for classical BDF for the Biot model, which is addressed by several novel techniques. For an arbitrary time step, we analytically prove the time-discretized energy stability. Through a series of 2D and 3D numerical experiments, we verify the stability and accuracy of the proposed scheme.

ABSTRACT. Simulations to predict two-phase flows in porous media require the solution of large systems of nonlinear equations, typically addressed using a Newton-type method. A major computational cost in this process is often the assembly, storage, and evaluation of the underlying sparse Jacobian matrices. It is well-known that automatic differentiation is capable of exploiting this sparsity structure by first representing the Jacobian by a matrix compression from which the Jacobian is then reconstructed. This work introduces a novel matrix compression strategy that enables the efficient construction of a sequence of Jacobian matrices with varying sparsity patterns. In contrast to a standard approach which, in a static fashion, computes a matrix compression from scratch at every Newton iteration, the new dynamic strategy reduces the computational effort by updating all sparsity-exploiting data structures incrementally. For an immiscible flow of two fluids moving through the pore space without mass exchange, the resulting dynamic approach is shown to speed up the construction of matrix compressions by a factor of up to 400.

ABSTRACT. Parallel computational fluid dynamics (CFD) simulations are commonly executed on modern high-performance computing systems with hardware topologies, where communication costs vary significantly across memory hierarchies and network links. OpenFOAM, a widely used open-source CFD framework, provides a scalable MPI-based parallel infrastructure. However, its default process placement and communication strategies are largely topology-agnostic, which can lead to suboptimal performance on NUMA-based cluster architectures.

In this work, we present a topology-aware communication optimization for OpenFOAM that targets both local and global communication in iterative linear solvers. Local communication is optimized by reassigning MPI ranks to physical cores based on the communication structure induced by the domain decomposition and the underlying hardware topology. In addition, a custom global reduction schedule is introduced that aligns collective communication with the hierarchical organization of compute nodes, NUMA domains, and NUMA nodes.

Experimental results show a substantial redistribution of communication toward faster hardware paths, leading to better runtime with the largest improvements observed for communication-intensive solver configurations. These findings demonstrate that topology-aware communication strategies can significantly enhance the performance of parallel CFD simulations on modern HPC systems.

ABSTRACT. In recent years, emergency recovery parachutes have been considered as a passive safety measure for flying cars to mitigate damage to onboard passengers and people on the ground in the event of catastrophic failures such as electrical power loss. However, the design and verification of parachute systems for flying cars remain challenging due to issues such as low-altitude deployment and possible interference between suspension lines and propellers. Therefore, a numerical framework capable of reproducing the coupled behavior of a vehicle–parachute system in a virtual environment is essential. In this study, a fluid–structure interaction (FSI) simulation framework is developed by weakly coupling a moving-boundary fluid solver based on an unstructured moving-grid finite-volume method solving the Euler equations with a structural solver using Project Chrono. Numerical wind-tunnel FSI simulations of a parachute and free-fall FSI simulations of a flying-car–parachute system are conducted. The results show that reinforcement cables suppress canopy inflation, reducing the projected diameter, with a difference of approximately 19% in projected area. Comparison between the rigid and elastic models confirmed that drag fluctuations occur in the elastic model due to periodic canopy deformation. In the free-fall simulation, the descent velocity decreases from 9.81 m/s to 8.53 m/s in about 1.43 s, approaching steady descent where drag balances the system weight. Oscillatory attitude motion suggests static instability of the parachute–payload system. These results demonstrate the importance of considering elastic deformation in the design of emergency recovery parachutes and provide fundamental insights for the safety design of flying cars.

ABSTRACT. This study presents a computationally efficient framework for early- stage design evaluation of next-generation air mobility systems. A coupled fluid– rigid body motion analysis combining the Moving Computational Domain (MCD) method with a multi-axis sliding mesh technique was quantitatively val- idated against wind tunnel experiments of a quadcopter subjected to frontal dis- turbance winds. To reduce computational cost, an inviscid fluid formulation was employed while preserving dynamic motion coupling. The results demonstrate that the predicted pitch angle under disturbance winds (0-5 m/s) agreed with ex- perimental measurements within a maximum error of 1.56°, accurately reproduc- ing the aircraft’s dynamic equilibrium response. Rotor rotational speeds were predicted with an average relative error of approximately 10 %, and their varia- tion with wind velocity exhibited strong correlation with experimental trends. These findings confirm that the proposed framework provides practical predic- tive capability with significantly reduced computational cost, making it suitable for rapid behavior assessment during early design stages. The present study rep- resents a step toward the realization of a scalable physics-based digital twin for multirotor aircraft. Future work will focus on improving rotor thrust prediction accuracy through refined geometric modeling while maintaining the computa- tional efficiency of the inviscid approach.

ABSTRACT. As electric vertical take-off and landing (eVTOL) aircraft diversify in configuration, quantitative comparison of dynamic motion performance at the conceptual design stage has become a critical challenge.This study proposes a computational framework that integrates route-tracking control into a fluid--rigid body coupled simulation based on the Moving Computational Domain (MCD) method and multi-axis sliding mesh technique.A PID controller combined with a look-ahead point algorithm enables autonomous simulation of a common four-phase flight route (ascent, forward acceleration, deceleration, and descent).The framework is applied to two eVTOL configurations: a compact coaxial Octorotor (400 kg) and a large Dodecarotor (1400 kg). Four motion performance metrics---maneuverability, responsiveness, energy efficiency, and ride comfort---are quantitatively evaluated. Fluid--rigid body coupling captures unsteady aerodynamic effects overlooked by conventional rigid-body models: the Octorotor experiences 24\% greater aerodynamic drag during forward acceleration, and its fluid-coupled energy consumption (1.87 kWh) is 60\% higher than estimates based solely on rotor RPM. Furthermore, while the Dodecarotor offers 1.4$\times$ better per-passenger energy efficiency due to its larger rotor disk area, its ride comfort is significantly compromised (RMS acceleration: 3.75 vs. 0.238 m/s$^2$), identifying a fundamental trade-off between efficiency and comfort. This framework provides high-fidelity quantitative guidance for vehicle concept selection and control optimization in model-based eVTOL development.

ABSTRACT. The prediction of masses of atomic nuclei using machine learning can complement theoretical models and advance the exploration of poorly known domains of the nuclear chart. We propose a machine learning technique based on gated recurrent units (GRU), which have demonstrated competitive performance in nuclear-mass prediction by exploiting long-term dependencies. By integrating multiplicative interactions and product-unit transformations within recurrent units, we report significant improvements in nuclear-mass prediction. Computations are performed in the complex domain to jointly capture amplitude and phase dynamics. For interpolation and temporal-extrapolation tasks based on the atomic mass evaluation (AME2016 and AME2020), the complex additive-multiplicative product-unit gated recurrent unit (AM-PU-GRU) consistently achieves the lowest prediction errors, with an interpolation RMSE of 0.227 ± 0.004 MeV and an extrapolation RMSE of 0.179 ± 0.015 MeV. These results surpass other state-of-the-art machine learning models and also outperform the real-valued GRU baseline and product-unit ablation variants, while remaining robust to different theoretical priors, including WS4 and SEMF. Our findings establish complex-valued product-unit recurrent networks as a new benchmark for sequence-based nuclear-mass prediction.

ABSTRACT. Understanding nonlinear feature interactions is crucial in science and engineering, yet standard multilayer perceptrons (MLPs) often capture such interactions only implicitly, leading to entangled representations that can impair robustness and interpretability. We investigate product-unit residual networks (PURe) that integrate multiplicative product units with residual connections to explicitly model cross-feature couplings while stabilizing optimization. We conduct a systematic evaluation on an interaction-driven synthetic benchmark and two real-world datasets, assessing predictive accuracy, robustness to Gaussian feature noise, and performance under limited training data, and we compare real- and complex-valued variants under a matched parameter budget. Beyond accuracy, SHAP interaction analyses show that PURe learns more concentrated and structurally coherent interaction patterns than MLP baselines. Overall, PURe achieves competitive or improved performance, better robustness and sample efficiency in low-data regimes, and enhanced interaction-level interpretability.

ABSTRACT. Validating training data for reasoning models typically requires expensive trial-and-error fine-tuning cycles. In this work, we investigate whether the utility of a reasoning dataset can be reliably predicted prior to training using intrinsic data metrics. We propose a suite of quantitative measures and evaluate their predictive power by fine-tuning 8B and 11B models on semantically distinct variants of a Polish reasoning dataset. Our analysis reveals that these intrinsic metrics demonstrate strong and significant correlations with downstream model performance. Crucially, we find that the predictors of utility are scale-dependent: smaller models rely on alignment-focused metrics to ensure precision, whereas larger models benefit from high redundancy, utilizing verbose traces to solve complex tasks. These findings establish a scale-aware framework for validating reasoning data, enabling practitioners to select effective training sets without the need for exhaustive empirical testing.

ABSTRACT. In recent years, convolutional neural networks have become the dominant tool in image processing and data classification tasks. Despite their high effectiveness, classical CNN architectures are characterized by a large number of parameters and limited capacity for adaptive control of information flow between layers. In response to these limitations, this paper proposes a biologically inspired control and pruning mechanism based on glial cell modeling. The presented architecture extends a classical convolutional network with a glial controller consisting of modules functionally corresponding to astrocytes, oligodendrocytes, and microglia. This controller generates dynamic channel gates that modulate the activations of convolutional feature maps in a manner dependent on the current state of the network. The learning process is carried out alternately for the glial and convolutional parts, enabling stable adaptation of both model components. The effectiveness of the proposed approach was experimentally verified on the publicly available CIFAR-10 datasets and on an original, deterministic sequential dataset built from network traffic logs containing user URLs. In the case of image data, classical classification scenarios were applied, while for sequential data, training sequences representing computer network user behavior were used. The obtained results indicate an improvement in classification quality compared to the base model lacking glial mechanisms, while maintaining the interpretability of the controller's operation. Additionally, the analysis of masks generated by the microglia module allowed for estimating the potential reduction in the number of convolutional parameters by identifying redundant channels. The results confirm the validity of using glial cell mechanisms as an adaptive control layer in deep neural networks, both in image processing and network data analysis tasks.

ABSTRACT. The solution of large, sparse linear systems often dominates the computational effort of scientific applications and is a frequent optimization target. Modern libraries provide numerous solver and preconditioner configurations, but their performance varies significantly across problem instances. Previous works have addressed the selection of an optimal solver, but are typically limited by the problem set addressed (e.g., only symmetric positive definite matrices), the use of expensive matrix features, or the complexity of the approach.

This work proposes a modular, low-cost embedding-based framework for solver selection that decouples performance modeling from feature representation and downstream prediction. Solver-problem relationships are learned directly from observed performance data, while inexpensive numerical features are used to project unseen problems into the learned embedding space. The framework focuses on multilabel prediction and evaluation using user-centric metrics, such as MAPE and nDCG, which better reflect relative performance.

Experiments on 621 matrices from the SuiteSparse matrix collection across 101 PETSc solver configurations demonstrate a 17% increase in top-prediction accuracy over classical feature-based models when expensive numerical features are included, along with reductions of 37% in mean average percentage error (MAPE) and 46% in top-prediction error (1-error). When restricted to a reduced feature set, the embedding approach remains competitive, while still consistently achieving ca. 24% lower MAPE and 1-error across a broad range of problems.

ABSTRACT. Parameter settings in evolutionary algorithms and metaheuristics are important because such parameter values can influence the performance of algorithms under evaluation. For a given algorithm, there are many different numerical experiments to show that the algorithm can work well in practice; however, theoretical analysis tends to lack behind and in most cases there is no theoretical analysis at all. In this work, we show that theoretical analysis using the theory of dynamical systems and evolution of population variance can give some good results in terms of parameter ranges for the bat algorithm. Results from numerical experiments are also consistent with theoretical bounds. Such analyses can provide good insights from different perspectives about the algorithmic characteristics such as variance evolution, transition between exploration and exploitation and convergence behaviour.

ABSTRACT. The problem of fitting reduced data representing the sequence of interpolation points of curve ɣ in arbitrary Euclidean space is examined here. In our setting, the corresponding interpolation knots are assumed to be unknown. In this work the so-called piecewise Bézier cubic fitting scheme \hat{ɣ} is applied to partially interpolate reduced data based on local replacement of the unknown knots. In exchange of partial interpolation, the fitting scheme in question preserves all geometrical properties ingrained by Bézier cubic curves (designed mainly for modelling purposes). The main theoretical contribution of this work establishes linear, quadratic or in-between sharp asymptotics in estimating ɣ by \hat{ɣ} subject to the character of various admissible samplings. The convergence orders proved in this work are ultimately verified numerically in affirmative as sharp. The tests are conduced on 2D and 3D curves with the aid of Mathematica software package. Finally, possible extensions and applications of this work are briefly outlined.

ABSTRACT. This paper explores the architecture of a hybrid video encoder implemented in Java, addressing the challenges of high-performance multimedia processing within a managed environment. The study focuses on balancing algorithmic complexity with execution efficiency, specifically managing temporal dependencies and memory overhead inherent in multi-threaded video compression.

The proposed architecture incorporates four key optimization mechanisms: a complete inter-frame processing pipeline, motion estimation with Half-Pel precision, a P-SKIP mode, and entropy coding. To maximize throughput, the encoder utilizes a Group of Pictures (GOP)-parallel model, where multiple threads independently process frame sequences of up to 30 frames.

Experimental results demonstrate bitrate reductions reaching up to 48\% compared to the baseline intra-only version while achieving a structural similarity index (SSIM) exceeding 0.95. The implementation demonstrates that GOP-level parallelism in Java effectively maintains high encoding throughput and stability across diverse many-core architectures. The study confirms that combining sub-pixel estimation with adaptive complexity management optimizes Rate-Distortion performance while effectively minimizing computational overhead.

ABSTRACT. Technical indicators are often treated as interchangeable features, yet in practice they behave like small programs whose outputs depend on conventions: cleaning, warm-up handling, thresholds, and execution lag. We evaluate 22 classical indicators as deterministic algorithms on daily OHLCV (open, high, low, close, volume) data for 415 Warsaw Stock Exchange equities (2015-06-28 -- 2023-12-29). Every instrument is processed under one fixed protocol: uniform cleaning rules, fixed signal mappings, a one-day execution lag, and a proportional cost model (10 bps per entry or exit). Decision stability is tested by rescaling the close series by ±1%, recomputing indicators and their discrete actions, and measuring day-level agreement. Slow trend families (moving averages, Moving Average Convergence Divergence (MACD), and Directional Movement Index / Average Directional Index (DMI/ADX)) barely react in this test (median stability near 100%), while stochastic oscillators ip often enough that their median stability falls to about 66-69%. The perturbation is intentionally local: it targets small close-price discrepancies across data vendors and is not a general stress test. Each signal then drives a minimal long-or-cash simulation executed on the next day with fixed costs. Redundancy diagnostics (correlations, Principal Component Analysis (PCA), and mutual information) show strong within-family overlap and exact identities (for example, the Bollinger middle band equals SMA(20), i.e., the 20-day simple moving average). The protocol yields a short list of stable, non-duplicate indicators and documents where common defaults are fragile; any apparent profitability is treated as provisional until it survives out-of-sample checks and a stricter execution model.

ABSTRACT. This study investigates the feasibility of ultra‑large ensembles by MLWP models. By integrating eight MLWP models—five deterministic and three probabilistic—we constructed a 408‑member “Grand Ensemble” using JMA initial conditions. Verification using Z500 during winter 2024 shows that the Grand Ensemble improves ensemble‑mean RMSE and CRPS compared with individual MLWP models, while achieving a spread comparable to GEPS. These results suggest that combining various MLWP models enhances prob-abilistic forecast accuracy.

ABSTRACT. We apply a convolutional neural network (CNN) to ensemble predictions and obtain improvements of more reliable probabilistic predictions. The CNN post-processor improved probabilistic skill and rank histogram, yet decreased the ensemble spread. Since this behavior looks counter-intuitive, we demonstrate that the primary driver for this improvement was the variance of the ensemble mean, not the ensemble spread itself. To analyze this, we decomposed the total forecast variance into two components: the temporal variance of the ensemble mean and the mean of the ensemble variance (squared ensemble spread). The analysis revealed that the core mechanism was the CNN's correction of large-scale biases (e.g., seasonal biases). This correction allowed the ensemble mean to track the observed climate more accurately, substantially increasing its temporal variance. This enhanced mean variance was the dominant factor that caused the overall forecast distribution to align more closely with that of the observations. Consequently, this study demonstrates that a reduction in ensemble spread does not necessarily equate to a loss of reliability. The key to improving the probabilistic skill and flattened the rank histogram was not ensemble spread inflation but bias correction, which enhanced the temporal variance of the ensemble mean. This finding highlights that for biased dynamical systems, the variance of the ensemble mean can be more critical than the ensemble spread itself for achieving a realistic overall forecast distribution.

ABSTRACT. Artificial Intelligence (AI) weather models are improving rapidly, and their forecasts are already competitive with long-established traditional Numerical Weather Prediction (NWP). In order to have confidence in this new methodology, it is critical that we understand how the predictions are being made. This is a huge challenge as these AI weather models largely remain black boxes. In other areas of Machine Learning (ML), mechanistic interpretability has emerged as a framework for understanding ML predictions by analysing the building blocks responsible for them. Here we present an open-source, highly-adaptable tool which uses aspects from mechanistic interpretability. It provides organisation to AI weather model data from the processor and allows for initial analyses, including cosine similarity and Principal Component Analysis (PCA), enabling the user to identify latent feature directions with meteorological features. Applying our tool to the graph neural network GraphCast, and presenting preliminary case studies for mid-latitude synoptic waves and specific humidity, we demonstrate its ability to find linear combinations of latent channels corresponding to interpretable features.

ABSTRACT. Geosteering is the real-time adjustment of a well trajectory during drilling to keep the borehole within a geological target. It is a sequential state-estimation and decision-making problem under incomplete information; noisy downhole measurements provide indirect evidence about subsurface boundaries, and each trajectory adjustment affects both future observations and feasible drilling paths. The core difficulty is epistemic. The geological configuration is fixed, but the decision-maker must act under incomplete information and update its state of knowledge as new measurements arrive. In this work, we treat geosteering as an offline sequence decision problem and explore machine-learning strategies trained from previously generated trajectories, avoiding costly online exploration during training. Specifically, we propose an offline workflow that combines particle-filter belief updates with Transformer sequence modeling to produce steering policies, and we evaluate decisions using reservoir-contact and feasibility metrics. The goal is to assess whether sequence models can produce stable long-horizon steering behavior and whether the supervised training objective---the per-step action imitation loss defined as mean squared error (MSE) between the predicted steering action and the recorded action at each decision step---correlates with domain performance measured by reservoir contact.

The offline dataset is generated from simulations, i.e., drilling episodes under diverse boundary configurations and measurement noise are recorded. To represent incomplete information explicitly, we use particle filtering to maintain a distribution over plausible boundary locations conditioned on the observed log response. The particle filter is used to expose the learner to a distribution of plausible boundary interpretations along the trajectory and, when desired, to provide compact summary features that can be included in the policy input. Each episode is represented as a sequence of interleaved tokens comprising return-to-go, state features, and actions, following the Decision Transformer formulation built on the Transformer architecture. Return-to-go encodes the desired cumulative objective, the state aggregates trajectory context together with log-derived features, and the action specifies the next steering command. Supervised training aligns predicted actions with historical actions, avoiding computationally expensive and potentially unstable value bootstrapping.

We evaluate learned policies in a synthetic geosteering environment. Measurements are locally ambiguous, and steering smoothness constraints limit feasible actions. Performance is quantified using reservoir contact ratio (RCR) as the primary domain metric, complemented by feasibility indicators linked to steering aggressiveness. Across controlled ablations, we observe a mismatch between the per-step action imitation loss (the MSE used for training) and long-horizon decision quality measured by RCR, consistent with the results reported. Short context windows can achieve lower one-step MSE yet exhibit myopic behavior and boundary drift, whereas longer contexts better exploit accumulated evidence and improve boundary tracking and RCR. These results emphasize that offline sequence policies should be assessed using domain-aligned objectives rather than training loss alone, motivating dataset and model design choices that preserve long-range dependencies relevant to subsurface control.

A preliminary version of this work was presented at a NeurIPS 2025 Workshop. Building on that version, this presentation adds two novelties. First, we train on an expert-generated dataset to complement simulator-generated trajectories and to reduce reliance on a single data-generation mechanism. Second, we investigate a trajectory-level Transformer formulation (Trajectory Transformer). This model is motivated by the need to condition steering behavior on long measurement histories and to capture the potentially multi-modal nature of expert decisions under incomplete information.

ABSTRACT. Deep learning models for generative imaging have made impressive strides, yet their application to medical imaging remains limited by data scarcity. We investigate the potential of diffusion models for generating abdominal CT slice images conditioned on textual prompts and structural guidance. We finetuned ControlNet on the RAOS abdominal CT dataset, aiming to generate CT slices in three anatomical planes (axial, coronal, and sagittal) while accurately reflecting the organs specified in the prompt. During training, we optimized only the ControlNet parameters responsible for encoding the conditional inputs, keeping the pre-trained Stable Diffusion backbone (including the denoising U-Net) frozen to preserve its generative prior. We compared two forms of structural conditioning: Canny edge maps and organ-wise semantic segmentation masks. Quantitative evaluations using FID, FD-DINOv2, and CLIP scores, combined with qualitative visualizations, demonstrate that the fine-tuned ControlNet models produce significantly more realistic and controllable outputs than fine-tuned Stable Diffusion alone. Notably, segmentation-conditioned ControlNet combined with text prompts generated the sharpest, most anatomically coherent slices and achieved superior evaluation scores compared to prompt-free generation. These results highlight the importance of synergizing textual and structural conditioning for reliable medical image synthesis.

ABSTRACT. Most state-of-the-art vein recognition systems are constrained by closed-set assumptions, limiting their scalability in realistic, high-security medical and access control environments. In this paper, we rigorously evaluate the computational boundaries of Deep Metric Learning (DML) under strict open-set, subject-disjoint protocols. By constraining the feature manifold to a unit hypersphere and optimizing it via batch-hard triplet loss, our ResNet50-CBAM framework achieves near-perfect open-set performance (99.45% OSCR, 1.57% EER) on the large-scale MMCBNU_6000 benchmark. However, our comprehensive cross-dataset analysis reveals a critical computational trade-off: metric learning exhibits significant sensitivity to the density of the training manifold in low-data regimes (e.g., the UTFVP dataset). This study not only demonstrates a highly robust, real-time approach for open-set biometrics but also exposes the vulnerability of deep embeddings to data scarcity and domain shifts, highlighting a fundamental challenge for deployable computer vision systems in medical and security applications.

ABSTRACT. The primary objective of this study was to determine and compare the values of Multidimensional Recurrence Quantification Analysis measures for gait data representing sagittal plane joint angles of the lower limb and pelvis, recorded in the CAREN Extended environment under two conditions: with mechanical perturbations and without disturbances. The research group consisted of 14 young women. A statistical analysis of the computed results was performed, including an assessment of the effect size of the observed differences. The findings demonstrated that Multidimensional Recurrence Quantification Analysis measures clearly distinguish natural gait from gait subjected to externally induced mechanical perturbations.

ABSTRACT. Large Language Models (LLMs) offer unprecedented opportunities for molecular property prediction, yet exhibit concerning capabilities for generating plausible-sounding but factually incorrect chemical descriptions. Whilst traditional evaluation metrics measure linguistic similarity, they cannot detect chemically impossible claims. Domain benchmarks focus on entity-level verification rather than systematically validating whether claimed structures are computationally possible. We develop a multi-dimensional validation framework combining reference-based metrics with three independent layers: grammatical fluency assessment, lexical validity checking, and RDKit-based computational chemistry validation. Evaluating three architecturally distinct LLMs (MolT5, LLaMA-3.1, TxGemma) across multiple decoding strategies on 451 antibiotic compounds, we demonstrate that multi-dimensional validation identifies distinct error types invisible to individual metrics. Our interactive dashboard enables inspection at multiple granularities, transforming validation from post-hoc comparison into evidence-based analysis. Our findings reveal systematic blind spots: perfect grammar accompanies fundamental chemical errors, aggregate metrics mask substantial performance differences, and domain-specific pre-training fails to transfer across chemistry tasks. This work sets foundations for trustworthy AI in chemistry, providing quality assurance infrastructure necessary for deploying LLMs in high-stakes applications that carry material consequences.

ABSTRACT. Metal artefacts remain a major challenge in computed tomography (CT) imaging, particularly in the presence of metallic implants, where they degrade image quality and obscure clinically relevant anatomy. Although numerous deep learning and model-based approaches for metal artefact reduction (MAR) have been proposed, their performance and generalisation are often limited by the scarcity of realistic training data. In clinical practice, paired CT images with and without metal artefacts are rarely available, and existing public datasets are therefore insufficient for supervised learning. As a result, many studies rely on synthetic artefacts, leading to domain gaps, especially in three-dimensional CT data.

We propose a data generation framework for realistic CT metal artefact synthesis that enables volumetrically consistent artefact generation. Building on \textit{StyleGAN}, the method introduces a variational content encoder and procedurally generated metal guidance masks to condition artefact synthesis on plausible streak propagation patterns, while adversarial training ensures visual realism and diversity.

The proposed framework enables the rapid generation of high-quality CT images with controllable and anatomically consistent metal artefacts that are coherent across adjacent slices. Volumetric consistency is quantitatively evaluated using a gradient-based cosine similarity metric, demonstrating improved inter-slice coherence compared to random artefact generation. The generated data provide an effective basis for training and evaluating MAR methods, providing a controllable and computationally efficient alternative for data generation in MAR research.

ABSTRACT. Human annotation is fundamental to building automatic systems, providing the labeled data that underpins data-driven modeling across diverse domains. Traditional approaches seek consensus among raters, treating disagreement as error to be eliminated and failing to capture the complex reality of human interpretation. However, in many contexts, diversity of annotator perspectives is informative rather than problematic. Rather than averaging away these differences, there is growing recognition of the importance of explicitly modeling annotator variability, giving rise to the perspectivist approach, which embraces multiple viewpoints. This is especially critical in tasks involving inherently subjective judgments that vary systematically across social groups, risking systems that reflect dominant group biases. In this study, we apply a Bayesian Nonparametric Multidimensional Item Response Theory model to multi-rater annotation, adopting a formulation where annotated texts are treated as persons carrying latent traits and annotators function as items. This allows us to automatically identify groups of annotators and assign to each text a set of scores over latent dimensions whose number is inferred directly from the data. We demonstrate the approach through a case study involving social media comments on immigration, annotated independently by multiple raters for the presence of racist content. The model uncovers the structure of annotator heterogeneity, offering a model-based alternative to consensus-based labeling.

ABSTRACT. Moving complex computational models into the world of humanitarian policy requires a structured way to manage knowledge and the stakeholder co-design process. This paper details the methodological architecture of the 'Homecoming' project, a high-fidelity agent-based model developed in partnership with the United Nations High Commissioner for Refugees to forecast displacement and return movements in Ukraine. We describe the project's evolution from its academic research roots through three distinct phases: initial prototyping, rapid adaptation following power grid attacks, and final maturation into a policy instrument. Our main contribution is our Expert-Informed Knowledge Repository methodology, which aligned academic research with the strict requirements of an international humanitarian policy model. We detail how weekly sessions, a structured GitHub Wiki, and a seven-category taxonomy enabled the translation of qualitative expertise into quantitative simulation parameters through continuous human-in-the-loop refinement. The effectiveness of this approach is evidenced by the model’s adoption across multiple high-level humanitarian and governmental forums. By documenting these experiences, we provide an evidence-informed roadmap for future high-fidelity computational science projects aimed at supporting evidence-based policy in volatile humanitarian contexts.

ABSTRACT. The sustainability of shared resources is increasingly challenged by the rapid expansion of AI infrastructure, whose electricity and water demands intensify pressure on already stressed regions. While coevolutionary game theory shows how environmental feedback shapes cooperation, little is known about human–AI alignment in shared resource systems. We develop a spatial agent-based model in which humans and AI compete over a water commons with environment-dependent payoffs linking local consumption to ecological regeneration. Our findings demonstrate that sustainability in spatial human–AI commons is an emergent property of feedback-mediated incentive alignment rather than resource abundance. Increasing capacity alone can weaken cooperation by allowing extraction pressure to diffuse across the grid, generating spatial traps in which cooperative clusters sustain a stressed environment. Only when environmental responsiveness tightly couples extraction to local consequence can stable interior equilibria or damped oscillations arise. These findings recast alignment in socio-technical systems as a problem of spatially structured feedback design. Alignment is then less a question of intent than of how tightly systems can bind action to consequence across the decision-making space.

ABSTRACT. Understanding how individual decisions shape environmental outcomes is critical as climate targets grow more ambitious. Traditional mean-field and two-strategy frameworks overlook individual-level cognition, dynamic risk perception, and evolving social norms. We develop a spatially explicit agent-based model (ABM) in which heterogeneous agents choose between climate-friendly and degradative actions via adaptive utility functions integrating environmental preference, social pressure, and ecosystem feedback, augmented with sub-models for dynamic risk-perception, shifting-baseline effects, and memory-based spillovers. Complemented by a mean-field approximation, phase plots and clustering analysis reveal critical tipping points, multi-stability, oscillatory behavior, and scale-free cluster formation across rationality and adaptation-speed regimes — dynamics absent from the mean-field model. These results identify the cognitive and social thresholds that determine whether socio-ecological systems collapse or recover, offering a basis for designing targeted, climate-positive interventions.

ABSTRACT. Expanding networks of interconnected devices pose novel security issues, particularly in commonplace IoT settings. This study presents a biometric authentication system that generates encryption keys on demand from facial characteristics, thereby obviating the need for permanent storage or external key servers. This technique facilitates rapid user authentication and secure messaging by generating session keys locally, making the system suitable for devices with constrained resources and fluctuating connectivity. The system demonstrated reliability and efficiency in actual tests, providing robust security against key duplication and brute-force attacks. Performance simulations demonstrate that the system operates reliably, even under substantial user and session loads. The findings indicate that biometric key generation may serve as a beneficial alternative to conventional security approaches in smart homes, healthcare, and industrial IoT applications.

ABSTRACT. Ensuring visibility at pedestrian crossings is crucial for urban safety, yet traditional audits remain costly, infrequent, and localized. This paper proposes repurposing standard dashcams as nodes in a scalable Mobile Sensor Network for proactive infrastructure monitoring. The introduced concept employs YOLOv11-based tracking, semantic segmentation, and monocular depth estimation to quantify visual occlusions autonomously. To validate the feasibility of this approach for city-scale analytics, a pilot deployment was conducted on a custom dataset covering over 70 hours of driving. The system successfully executed over 7,700 automated audits, demonstrating the capability to generate comprehensive reports on pedestrian crossing safety without human intervention. These results confirm that fusing low-cost visual sensors with deep learning effectively democratizes access to safety data, offering a foundation for a continuously updated, self-monitoring, innovative city system.

ABSTRACT. Hyperparameter optimization (HPO) is usually framed as a fully automated search that returns a single ``best'' configuration, leaving human operators with little insight into alternatives or trade-offs. However, in many applied settings, practitioners arrive with concrete wishes such as higher precision, less labeling effort, or faster models that standard HPO tools cannot address directly. In this work, we propose a surrogate-based explainable AI (XAI) framework that turns informal requirements into counterfactual queries over the configuration space. A CatBoost model is trained on the resulting configuration-performance log and analyzed with XAI tools, then queried by the DiCE and CFNOW explainers to generate counterfactual configurations under fixed soft constraints. Experiments on four standard binary classification benchmarks show that the framework can provide locally reliable guidance. In particular, DiCE often suggests alternative settings that result in consistent improvements when the main model is retrained.

ABSTRACT. The Internet of Things (IoT) has become a key enabling technology for modern smart systems, where reliable and efficient data transmission is essential for both operational safety and user satisfaction. Wireless communication is widely adopted in IoT deployments due to its ease of installation, device mobility, and flexibility; however, battery-powered operation imposes strict constraints on energy consumption, processing capability, and network parameters such as transmission range and packet size. While short data frames reduce radio activity and power usage, they also introduce challenges related to the interoperability of low-power wireless protocols with IP-based networks. To address these limitations, the Internet Engineering Task Force introduced the 6LoWPAN adaptation layer, enabling IPv6 communication over IEEE 802.15.4 links. This paper presents an experimental evaluation of the 6LoWPAN protocol implemented on the nRF52840 system-on-chip (SoC), focusing on its impact on throughput and communication reliability. In addition, the communication range of the IEEE 802.15.4 radio link was assessed using the nRF52840 USB dongle development platform. The results demonstrate that protocol overhead introduced by 6LoWPAN reduces achievable network throughput compared to native IP communication, while header compression mechanisms partially mitigate this effect by improving application-layer efficiency.

ABSTRACT. Standard collaborative filtering is efficient for large retail datasets but overlooks the temporal dynamics of customer behavior, while neural and complex models capture these dynamics at the cost of heavy computational demands. We propose a hybrid recommendation method that retains the efficiency of classical approaches while incorporating lightweight temporal modeling to capture implicit feedback and personalized cyclic purchasing patterns. The experiments verify that the proposed method achieves comparable performance to state-of-the-art methods, maintaining a linear computational overhead, and provides interpretable features. In particular, compared to the baseline, our method improves ranking quality, most notably increasing NDCG by 6.4% at @5, while providing uniform gains across all metrics (≈ 2%) at larger cutoffs. Applied to a real-world retail use case with large-scale transactional data, the method demonstrates its practicality and effectiveness for personalized product recommendations in physical retail stores.

ABSTRACT. Dataset distillation, a technique for generating compact synthetic datasets to enable efficient model training and knowledge transfer, relies on two critical procedures in distribution matching frameworks: feature extraction and distribution alignment. Despite the significance, existing studies have not systematically investigated how different feature extractors influence the performance of distilled datasets. To address this gap, we comprehensively compared four feature extractors: convolutional neural network (CNN), ResNet18, multilayer perceptron (MLP) and lightweight Vision Transformer (ViT), and further analyzed the impact of dynamic and fixed feature extractors. This study provides empirical guidance for selecting feature extractors in distribution matching frameworks of dataset distillation.

ABSTRACT. Large language models (LLMs) are evolving from passive text generators into agentic systems that can plan, retrieve evidence, call tools, and collaborate with humans in iterative decision workflows. In financial investment decision-making, this shift enables new capabilities— automated research synthesis, multimodal signal ingestion, scenario generation, portfolio proposal, and constrained execution—yet it also amplifies failure modes such as hallucination, tool misuse, prompt injection, distribution drift, and compliance risk. This paper integrates four research themes into a single conceptual framework, Perception → Cognition → Collaboration → Governance. We review representative methods and systems, propose a finance-oriented reference architecture centered on provenance-rich evidence packets and permissioned tool access, and outline an evaluation checklist that extends beyond profit-and-loss to include reliability, controllability, security, and audit readiness. We conclude with open research directions toward robust, reproducible, and economically meaningful LLM investment agents suitable for real-world deployment.

ABSTRACT. We propose a novel method for evaluating query reformulation in Retrieval-Augmented Generation (RAG) without relying on costly manual annotations of relevant documents. The approach measures both semantic similarity and downstream retrieval performance. We benchmark multiple reformulation techniques and large language models on a Polish university-regulations dialog dataset with an official document corpus as the retrieval base. The resulting evaluation pipeline enables multi-perspective assessment and can be readily adapted to new domains, document collections, reformulation methods, or settings with annotated relevance data.

ABSTRACT. Regulated enterprises cannot adopt generative AI when the system offers no proof of how an answer was produced or who was authorized to see the evidence behind it. The most capable LLMs are cloud- hosted and off-limits under data sovereignty mandates. Organizations fail to provide sanctioned alternatives, so employees use consumer AI tools and leak exactly the very data that restrictions were meant to protect. Without traceability, auditability, verifiability, accountability (and other "-lities") built into the architecture by design, enterprise stakeholders will not trust AI-generated knowledge. Without trust, adoption does not happen.

This paper presents TrustChain-RAG, an architecture born from five years of operating a multilingual knowledge pipeline (100M+ words, six European languages) for a Life Sciences organization and prior patents. The framework combines multilingual Context Graphs with lattice-aligned RBAC, locally-hosted RAG, and a private blockchain audit trail. Cross-lingual bridge nodes in the context graph outperform the best current multilingual embeddings (BGE-M3) by +15pp in precision on special- ized domains. Evaluation across three regulated domains, six European languages, and five LLMs (7B-72B parameters) shows 50-62% hallucination reduction over standard RAG. Controlled ablations isolate this to graph topology rather than retrieval volume or fine-tuning. The accuracy margin holds at +12 FA points regardless of model scale, from 7B to 70B. RBAC enforcement holds at 1.3% leakage versus 6% for metadata pre-filtering. Failure analysis at the 70B scale shows the error bottleneck shifting from model reasoning to knowledge structure: ontology gaps, bridge calibration, confirming that the framework and the model address different limitations.