ABSTRACT. In this talk, we examine how high-performance computing has changed over the last 10 years and look at trends in the future. These changes have had and will continue to impact our software significantly. Some of the software and algorithm challenges have already been encountered, such as management of communication and memory hierarchies through a combination of compile-time and run-time techniques, but the increased scale of computation, depth of memory hierarchies, range of latencies, and increased run–time environment variability will make these problems much harder. Mixed precision numerical methods are paramount for increasing the throughput of traditional and artificial intelligence (AI) workloads beyond riding the wave of the hardware alone. Reducing precision comes at the price of trading away some accuracy for performance (reckless behavior) but in noncritical segments of the workflow (responsible behavior) so that the accuracy requirements of the application can still be satisfied.

ABSTRACT. Variational Physics-Informed Neural Networks (VPINN) train the parameters of neural networks (NN) to solve PDEs. They perform unsupervised training based on the physical laws described by the weak-form residuals of the PDE over an underlying discretized variational setting; thus defining a loss function in the form of a weighted sum of multiple definite integrals representing a testing scheme. However, this VPINN loss function is not robust, meaning that the true error of the solution is arbitrarily far from the loss value. This means that we do not know what is the quality of the neural network approximation to the solution of PDE. It can be arbitrarily far from the exact solution. To overcome this, we employ Robust Variational Physics-Informed Neural Networks (RVPINN), which modifies the original VPINN loss into a robust counterpart that produces both lower and upper bounds of the true error. The robust loss is modifies the original loss by using the inverse of the Gram matrix computed with the inner product of the energy norm, known for a given PDE. In this case, our robust loss function is a good estimate of the true error, and we know when to stop the training to obtain a good quality result. The drawback of this robust loss is the computational cost related to the need to compute several integrals of residuals, one for each test function, multiplied by the inverse of the proper Gram matrix. We show how to perform efficient generation of the loss and training of RVPINN method on GPGPU using a sequence of einsum tensor operations. Using this method, we can solve our model 2D PDEs within 350 seconds on A100 GPGPU card from Google Colab Pro. We advocate using the RVPINN with proper tensor operations to solve PDEs efficiently and robustly. Our tensorial implementation allows for 18 times speed up in comparison to \emph{for}-loop type implementation on the A100 GPGPU card.

ABSTRACT. We present an open-source Physics Informed Neural Network environment for simulations of transient phenomena on two-dimensional rectangular domains, with the following features: (1) it is compatible with Google Colab which allows automatic execution on cloud environment; (2) it supports two dimensional time-dependent PDEs; (3) it provides simple interface for definition of the residual loss, boundary condition and initial loss, together with their weights; (4) it support Neumann and Dirichlet boundary conditions; (5) it allows for customizing the number of layers and neurons per layer, as well as for arbitrary activation function; (6) the learning rate and number of epochs are available as parameters; (7) it automatically differentiates PINN with respect to spatial and temporal variables; (8) it provides routines for plotting the convergence (with running average), initial conditions learnt, 2D and 3D snapshots from the simulation and movies (9) it includes a library of problems: (a) non-stationary heat transfer; (b) atmospheric simulations including thermal inversion; (c) tumor growth simulations; and (d) the Stokes problem.

ABSTRACT. Geomagnetic storms are large-scale disturbances of the Earth’s magnetosphere driven by solar wind interactions, posing significant risks to space-based and ground-based infrastructure. The Disturbance Storm Time (Dst) index quantifies geomagnetic storm intensity by measuring global magnetic field variations. This study applies symbolic regression to derive data-driven equations describing the temporal evolution of the Dst index using solar wind and interplanetary magnetic field (IMF) parameters. We use historical data from the NASA OMNIweb database, including solar wind density, bulk velocity, convective electric field, dynamic pressure, and magnetic pressure. The \texttt{PySR} framework, an evolutionary algorithm-based symbolic regression library, is used to identify mathematical expressions linking dDst/dt to key solar wind and IMF parameters. The resulting models include a hierarchy of complexity levels and enable a comparison with well-established empirical models such as the Burton-McPherron-Russell (BMR) and O’Brien-McPherron (OBM) models. The best-performing symbolic regression models demonstrate superior accuracy in most cases, particularly during moderate geomagnetic storms, while maintaining physical interpretability. Performance evaluation on historical storm events, including the 2003 Halloween Storm, the 2015 St. Patrick’s Day Storm, and a 2017 moderate storm, confirm that the discovered equations closely capture the geomagnetic storms. The results provide interpretable, closed-form expressions that capture nonlinear dependencies and thresholding effects in Dst evolution.

ABSTRACT. Explicitly simulating the transport of circulating tumor cells (CTCs) across anatomical scales with submicron precision—-necessary for capturing ligand-receptor interactions between CTCs and endothelial walls—-remains infeasible even on modern supercomputers. In this work, we extend the hybrid CPU-GPU adaptive physics refinement (APR) method to couple a moving finely resolved region capturing adhesive dynamics between a cancer cell and nearby endothelium to a bulk fluid domain. We present algorithmic advancements that: enable the window to traverse vessel walls, resolve adhesive interactions within the moving window, and accelerate adhesive computations with GPUs. We provide an in-depth analysis of key implementation challenges, including trade-offs in data movement, memory footprint, and algorithmic complexity. Leveraging the advanced APR techniques introduced in this work, we simulate adhesive cancer cell transport within a large microfluidic device at a fraction of the computational cost of fully explicit models. This result highlights our method's ability to significantly expand the accessible problem sizes for adhesive transport simulations, enabling more complex and computationally demanding studies.

ABSTRACT. The transactional activity of a bank’s customers contains a wealth of information regarding their behavior. By examining the trans- actions of either a specific group or all clients of the bank, it is possi- ble to gain insight into the macroeconomic environment and utilize this to anticipate various outcomes. In this paper, we proposed a dynamic behavioral model for bank clients based on econophysics principles. Ad- ditionally, we identified the non-homogeneous function in the dynamics equation using physics-informed neural network. We also interpreted this non-homogeneitythroughthelensofnewsreportsfromsocialmediaplat- forms and news agencies. The model demonstrated accurate results in a numerical simulation of the restoration of the initial dependency. We also showed the potential for creating scenarios in which news events impact the behavior of bank customers.

ABSTRACT. As climate change increases the threat of weather-related disasters, research on weather control is gaining importance. The objective of weather control is to mitigate disaster risks by administering interventions with optimal timing, location, and intensity. However, the optimization process is highly challenging due to the vast scale and complexity of weather phenomena, which introduces two major challenges. First, obtaining accurate gradient information for optimization is difficult. In addition, numerical weather prediction (NWP) models demand enormous computational resources, necessitating parameter optimization with minimal function evaluations. To address these challenges, this study proposes a method for designing weather interventions based on black-box optimization, which enables efficient exploration without requiring gradient information. The proposed method is evaluated in two distinct control scenarios: one-shot initial value intervention and sequential intervention based on model predictive control. Furthermore, a comparative analysis is conducted among four representative black-box optimization methods in terms of total rainfall reduction. Experimental results show that Bayesian optimization achieves higher control effectiveness than the others, particularly in high-dimensional search spaces. These findings suggest that Bayesian optimization is a highly effective approach for weather intervention computation.

ABSTRACT. Hailstorms are highly localized severe weather events that can cause extensive damage to agriculture, infrastructure, and property, needing accurate forecasting for effective risk mitigation. The Weather Research and Forecasting (WRF) model, a leading tool for numerical weather prediction, offers a range of physics parameterizations to simulate atmospheric processes. However, the vast number of possible configurations complicates the identification of an optimal setup for hail simulation. This study leverages a genetic algorithm (GA) to systematically optimize WRF physics parameterizations for hail prediction over Central Europe, focusing on the severe hail events of June 2022.

The GA framework encodes WRF configurations as individuals within a population, evolving through selection, crossover, and mutation across multiple generations. Fitness is evaluated using the F2 score, prioritizing recall to address the imbalance between observed hail and non-hail events. By exploring over 2.4 million potential configurations, the GA enhances the WRF model's ability to capture the spatial and temporal characteristics of hailstorms. The results show that this methodology enables the exploration of a wide range of possible configurations, demonstrating its potential to optimize parameterizations for high-impact weather events effectively.

ABSTRACT. The inverse problem of estimating the refractive index in a waveguide based on wave field measurement data is studied. A differentiable finite-difference scheme for the parabolic wave equation is constructed. The desired function of spatial coordinates, corresponding to the refractive index, is represented as a deep neural network. Optimization problem with respect to unknown refractive index function is formulated and solved. Automatic differentiation of the numerical scheme is used for efficient gradient computation. Numerical examples confirm that the proposed method outperforms the existing approaches to solving underwater and tropospheric tomography problems.

ABSTRACT. Urban mobility plays a crucial role in the functioning of cities, influencing economic activity, accessibility, and quality of life. However, the effectiveness of analytical models in understanding urban mobility patterns can be significantly affected by the spatial scales employed in the analysis. This paper explores the impact of spatial scales on the performance of the gravity model in explaining urban mobility patterns using public transport flow data in Singapore. The model is evaluated across multiple spatial scales of origin and destination locations, ranging from individual bus stops and train stations to broader regional aggregations. Results indicate the existence of an optimal intermediate spatial scale at which the gravity model performs best. At the finest scale, where individual transport nodes are considered, the model exhibits poor performance due to noisy and highly variable travel patterns. Conversely, at larger scales, model performance also suffers as over-aggregation of transport nodes results in excessive generalisation which obscures the underlying mobility dynamics. Furthermore, distance-based spatial aggregation of transport nodes proves to outperform administrative boundary-based aggregation, suggesting that actual urban organisation and movement patterns may not necessarily align with imposed administrative divisions. These insights highlight the importance of selecting appropriate spatial scales in mobility analysis and urban modelling in general, offering valuable guidance for urban and transport planning efforts aimed at enhancing mobility in complex urban environments.

ABSTRACT. The emergence of coronavirus disease (COVID-19) in late 2019 sparked a global pandemic, profoundly impacting societies and economies worldwide. To mitigate its spread, governments have implemented various preventive measures, prompting extensive research into transmission risk assessment. To evaluate the transmission risk systematically, we developed a framework integrating agent-based modeling (ABM) and computational fluid dynamics (CFD), and applied the framework to a preschool COVID-19 cluster in Singapore as a case study. Individual movement and behaviors are simulated with ABM, and CFD is employed to compute virus particle flow which is critical for transmission risk. In the case study, we categorized the infected individual's movement into three types based on the initial destinations and evaluated its impact on the transmission risk. Simulation results show that the average risk level is nearly the same for all three movement types and it changes across time depending on the degree of infected individual’s active movement.

ABSTRACT. Achieving consensus in multi-criteria decision-making scenarios involving multiple agents with conflicting preferences is a complex challenge. This paper introduces a novel method for identifying the most critical issues that significantly impact consensus quality in preference-based conflict situations. By systematically analyzing the effect of removing individual issues on consensus outcomes -- both for the entire group of agents and within smaller coalitions -- the method highlights key factors that drive agreement. A case study illustrates the practical application of the proposed approach, showing its effectiveness in prioritizing negotiation efforts and reducing conflict intensity. The findings provide valuable insights for improving decision-making processes in real-world scenarios.

ABSTRACT. The classification of dispersed data poses challenges due to inconsistencies and conflicts arising from independently collected sources. This study introduces a coalition-based classification framework that integrates conflict analysis and rule-based learning to enhance classification performance and interpretability. The approach employs four decision rule induction methods -- exhaustive search, genetic algorithms, covering algorithms, and LEM2 -- combined with three decision-making strategies: first rule approach, all rules approach, and weighted rules approach. Experiments were conducted on datasets from the UCI Machine Learning Repository. Results indicate that the covering algorithm with weighted rules approach achieves the highest classification performance across all metrics. In contrast, LEM2 performs poorly, often failing to generate covering rules, leading to random classifications. The findings demonstrate the advantages of conflict-aware coalition formation in improving dispersed data classification.

ABSTRACT. Multi-Criteria Decision Analysis (MCDA) methods are widely applied in decision-making across various fields, including sustainability, social and environmental issues, and energy management. Despite their popularity, many commonly used MCDA methods, such as the Analytical Hierarchy Process (AHP), Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), VlseKriterijumska Optimizacija I Kompromisno Resenje (VIKOR), and Multi-Attribute Utility Theory (MAUT) show compensatory features. This means that strong performance in some criteria can compensate weaker performance in others, which contradicts the principles of strong sustainability. In contrast, non-compensatory MCDA methods, such as Preference Ranking Organization METHod for Enrichment of Evaluation (PROMETHEE) and ELimination and Choice Expressing the Reality (ELECTRE), align more closely with the principles of strong sustainability. However, these methods often lack numerical ranking capabilities and can be computationally complex, limiting their broader adoption.

This paper addresses this gap by introducing the Strong Sustainability Paradigm based VIKOR (SSP-VIKOR) method, an extension of the traditional VIKOR approach. The SSP-VIKOR method incorporates compensation reduction modeling by adjusting the sustainability coefficient, enabling a more accurate evaluation of sustainability-related decision problems. The effectiveness of SSP-VIKOR is demonstrated through an assessment of the sustainable share of renewable energy sources (RES) in selected European countries. Given the European Union’s commitment to reducing greenhouse gas emissions, improving energy security, and promoting economic development, it is crucial to accurately evaluate the sustainability contributions of RES. The findings highlight the potential of SSP-VIKOR as a practical tool for policymakers and stakeholders seeking to balance the environmental, economic, and social dimensions of sustainability.

ABSTRACT. Sustainable decision-making requires balancing environmental, social, and economic objectives while minimizing trade-offs. Traditional Multi-Criteria Decision-Making (MCDM) methods, such as the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), often adhere to weak sustainability principles, allowing excessive linear compensation among criteria. To address this limitation, a novel STOPSIS (Sustainable TOPSIS) method is introduced -- an enhanced version of TOPSIS that integrates strong sustainability paradigm by incorporating a sustainability coefficient and applying a spike suppression matrix on the input decision matrix. These modifications mitigate linear compensation, ensuring rankings that better reflect ecological and social constraints. The effectiveness of STOPSIS is demonstrated through an empirical study on offshore wind farm selection in the Baltic Sea. The results indicate significant ranking adjustments based on sustainability coefficient variations, reinforcing the method’s robustness. STOPSIS offers a structured approach to sustainable decision-making, making it a valuable tool for applications in renewable energy, urban planning, and resource management.

ABSTRACT. Chaotic systems with extreme events present significant challenges in terms of prediction and control due to their complex nonlinear dynamics and potential high dimensionality. We investigate here the use of cluster-based reduced-order modelling (ROM) and control techniques applied to such systems. These techniques aims to retain the essential dynamics of the system through the identification of cluster of similar states and by only modelling the transition between such clusters and defining a control per cluster. This effectively neglects the specific dynamics within a given cluster while retaining the main dynamics of the full-order model. The considered test case is the Moehlis-Faisst-Eckhart (MFE) system which exhibits extreme events in the form of quasi-relaminarization events. The influence of the number of clusters and the order of modelling on the accuracy of the resulting reduced-order cluster-based model is explored. A cluster-based control strategy also further proposed and applied to the MFE system to prevent extreme events. This strategy manages to achieve the objectives with a large reduction in extreme events in the MFE case, which highlights the potential of cluster-based reduced-order modelling and control.

ABSTRACT. Recent advances in artificial intelligence have changed the ability to study and model complex biological phenomena. Physics-Informed Neural Networks (PINNs) represent a novel approach that link deep learning techniques with fundamental physical principles in solving partial differential equations. This work proposes an implementation of PINNs for modeling tumor-induced angiogenesis through a system of coupled reaction-diffusion equations that track the interplay between different biological agents. We introduce a computational framework that combines neural network architectures with physics-based constraints, using an optimized loss function incorporating both empirical data and theoretical principles via strategic collocation points. Experimental results validate the reliability of our approach in predicting the intricate spatial and temporal patterns of blood vessel formation, showing the potential of PINNs as a robust computational tool for simulating complex biological processes.

ABSTRACT. Automated machine learning (ML) can streamline the characterisation and management of chronic airway conditions. With the advent of quantitative CT (qCT) imaging allowing precise extraction of structural features from scans, assessment of airway obstruction levels could be automated to compliment traditional testing. This "feature known" approach has the added potential benefit of characterising structure-function relationships through explainability measures. We therefore aimed to develop inverse models to estimate spirometry parameters from high-dimensional quantitative data using these structural metrics as constraints. With the ATLANTIS (NCT02123667) dataset, this paper experiments with a selection of ML methods, specifically k-nearest neighbours (kNN), random forest (RF) and support vector machine (SVM), to predict spirometry values (Forced Expiratory Volume (FEV1), Forced Vital Capacity (FVC) and FEV1/FVC). The dynamic ratio FEV1/FVC was predicted better by all models than FEV1 or FVC. Results show effective counteraction to high-dimensionality through iterative feature refinement guided by SHapley Additive exPlanations (SHAP), and to limited training data through dynamic Gaussian noise (DGN). Diagnostic grade prediction accuracy was achieved with DGN SHAP sequential feature selection (SFS)-kNN at 1.64\% MRE with 37/76 features. A selection of variables including expiratory tissue density and lung volume, vasculature and airway geometries were seen to be important for prediction. This novel approach is a promising method for extracting useful functional information in a dynamic airway system and linking back to potential structural abnormalities.

ABSTRACT. Tracking audience engagement in real-time offers numerous benefits. For instance, event planners can make dynamic adjustments to presentations or activities to maintain high levels of interest and participation. This enhances the overall experience for attendees by ensuring the content remains engaging and relevant. This paper proposes a model for computing the binary engagement within groups. The model does this by identifying individuals' engagement during the events' time frames, which are then combined, i.e., the engagement of the group is computed by aggregating the engagement of each individual. For each individual of the group, the engagement model incorporates the computation over time of the gaze direction, valence, and arousal, classifying the engagement into two primary levels: not-engaged and engaged. The engaged category is further divided into two sublevels: positive and negative engagement. Experimental results confirm the model's effectiveness, showcasing reliable identity tracking and accurate assessment of engagement states in dynamic scenarios.

ABSTRACT. Smart cities optimize traffic management and vehicle communication through Intelligent Transportation Systems (ITS), with Vehicular Ad-hoc Networks (VANET) serving as a core infrastructure. In these environments, security vulnerabilities can severely impact the smart city traffic system, leading to traffic congestion, blockage of emergency vehicle routes, and disruption of autonomous driving systems. Unfortunately, identifying security vulnerabilities of the system in VANET is complex due to various attack types and the dynamic nature of the network, requiring systematic verification techniques for effective analysis. Recently, Nath et al. proposed an authentication protocol for VANETs using LWE-based lattice signatures and tokens, however the protocol has not been sufficiently validated. This study utilizes AVISPA (Automated Validation of Internet Security Protocols and Applications) to analyze Nath et al.'s protocol. AVISPA, an automated security verification tool based on the Dolev-Yao(DY) attacker model, is effective in assessing various threats such as replay attacks and man-in-the-middle attacks, making it ideal for evaluating security in the VANET environment. The security analysis reveals that the protocol is vulnerable to multiple attacks due to the lack of message freshness verification and user authentication. To address these vulnerabilities, we propose countermeasures to enhance message freshness verification and user authentication mechanisms, and validate the improved protocol's security through AVISPA simulation. Finally, we verify the security of the authentication scheme which is applied the countermeasures through AVISPA. The result shows that the security of smart city vehicular networks can be strenghtened through AVISPA-based security verification and this can provide valuable insights for desining security protocols in smart cities.

ABSTRACT. The aging population has led to a growing number of elderly individuals living alone, making it crucial to address their need for quick and effective emergency assistance. Older adults, often facing physical limitations or illnesses, require reliable systems for immediate help during life-threatening situations. To meet this need, smart devices like emergency call systems are being developed, enhancing seniors' safety and improving health and social care responses. Our research explores how passive and active speech analysis on mobile devices can support automatic emergency assistance. We show that this can be achieved on Edge devices using tiny machine learning models for wake-word detection, speech-to-text conversion, and intention recognition, paving the way for safer, smarter living environments for seniors.

ABSTRACT. The integration of Artificial Intelligence (AI) with Augmented Reality (AR) has transformed human-computer interaction, offering new opportunities for immersive learning and cognitive assessment. However, the relationship between user engagement in AR environments and memory retention remains underexplored. This study proposes an AI-driven framework for predicting memory retention using behavioural interaction data captured through Microsoft HoloLens 2 sensors. The model estimates the likelihood of object recall in AR-based learning environments by analyzing key interaction metrics such as gaze duration, interaction frequency, revisit counts, and head movement stability. To validate the AI predictions, we compared model-generated retention scores with user-reported recall, demonstrating a strong alignment between predicted and actual memory performance. Our findings align with established cognitive theories, indicating that increased interaction and attentional engagement enhance memory retention. Furthermore, comparisons with prior research on perceptual judgments and spatial memory reinforce the model’s effectiveness in capturing real-world cognitive processes. This study introduces a scalable, non-invasive approach to cognitive modeling, bridging AI-driven analytics with AR-based learning. The results have broad implications for education, medical training, and workforce development, where optimizing learning efficiency is crucial. By leveraging AI for real-time memory prediction, this research paves the way for more adaptive and personalized AR learning experiences.

ABSTRACT. Host-based Intrusion Detection Systems (HIDS) collect host system logs and generate alerts when the host is attacked. However, existing research fails to adequately capture the spatiotemporal relationships within host behaviors, limiting the accuracy of their representation and modeling. To address this, we propose a spatiotemporal graph representation learning method. This method extracts key data from system logs to construct provenance graphs. A spatiotemporal joint encoder decomposes features along spatial and temporal dimensions independently, then aggregates them to capture spatiotemporal dependencies, explicitly modeling these relationships in host behavior. Experiments on the Streamspot and DARPA-Theia datasets show that the proposed method effectively captures interaction patterns and outperforms baseline models in recall rate, false positive rate, and other evaluation metrics.

ABSTRACT. This article presents an analysis of the application of machine learning algorithms for predicting the mechanical properties of austempered ductile iron (ADI) castings. As part of the study, predictive models were developed and optimized to forecast strength parameters based on chemical composition, thickness, and heat treatment process parameters. A detailed analysis of the impact of hyperparameters on algorithm effectiveness was conducted, along with a comparison of different parameter space exploration methods. The study evaluated the performance of various machine learning algorithms, identifying Gradient Boosting as the most effective for predicting mechanical properties. An additional outcome of this research is the development of a web application integrating the predictive models, allowing users to analyze the expected properties of castings based on input data. This solution has potential applications in the foundry industry, enabling better control over production processes and reducing costs associated with experimental selection of technological parameters. The results confirm that applying machine learning algorithms can significantly improve the prediction of ADI iron's mechanical properties, paving the way for further automation and optimization of metallurgical production processes.

ABSTRACT. The accurate prediction of glass transition temperature (Tg) is crucial for materials design but often relies on melting point dependencies, limiting its applicability in inverse design problems. We present a data-driven approach leveraging machine learning (ML) and symbolic regression to predict Tg based solely on molecular structure. Using the BIMOG dataset, we extract key structural features—including molecular branching, computed from SMILES representations using RDKit and PySMILES, and atomic composition ratios (C, CH, O, etc.)—to enhance predictive accuracy. We apply multiple ML models, including Linear Regression, Random Forest, Gradient Boosting, XGBoost, and Extra Trees, achieving R2 scores comparable to traditional approaches that depend on melting point data. Finally, we employ genetic programming for symbolic regression to derive an interpretable equation for Tg. Our results demonstrate that incorporating structural descriptors allows for accurate and generalizable Tg prediction without requiring melting point information, making this method well-suited for inverse materials design. This work highlights how computational approaches can improve the tractability of complex materials problems, aligning with the broader goal of integrating physics-based and data-driven methods for materials discovery.

ABSTRACT. Bicontinuous interfacially-jammed emulsion gels(bijels) are seeing increasing interest as platforms to enhance emulsion templating and engineered soft matter for use in pharmaceuticals and tissue engineering applications. Interest in these areas arise from bijels co-continuous tortuous nature with tunable domain sizes and structure. In applications such as bio-reactors or separation systems, desired materials properties are intrinsically linked to structure through the domain size and tortuosity. In this study, we investigate the structural implications and underlying mechanisms driving the structural changes of applying magnetic fields to bijels stabilized with magnetically responsive ellipsoidal particles using a hybrid Lattice Boltzmann and Molecular Dynamics simulation model. We characterized the response of bijel templates made with no fields using oblate and prolate ellipsoids, and demonstrated that the microstructure can be modified and made anisotropic using constant magnetic fields. The degree and rate of microstructure change are magnetic field strength dependent, with the dynamics of the system controlled by the reordering of particles at the interface. For prolate particles, this is due to a local crystallization transition characterized as an increase in the six fold Steinhardt order parameter above 0.38 and for oblate particles it is due to magnetic fields adding disorder to the interface and how the particles seek to become more ordered to the field. These results identify timescales and mechanisms driving microstructural evolution, suitable to control the microstructure of bijels used in soft matter to facilitate generation of tunable and active materials.

ABSTRACT. In real-world applications, Large Language Models (LLMs) often hallucinate, even in Retrieval-Augmented Generation (RAG) set- tings, which poses a significant challenge to their deployment. In this pa- per, we introduce AggTruth, a method for online detection of contextual hallucinations by analyzing the distribution of internal attention scores in the provided context (passage). Specifically, we propose four different variants of the method, each varying in the aggregation technique used to calculate attention scores. Across all LLMs examined, AggTruth demon- strated stable performance in both same-task and cross-task setups, out- performing the current SOTA in multiple scenarios. Furthermore, we conducted an in-depth analysis of feature selection techniques and ex- amined how the number of selected attention heads impacts detection performance, demonstrating that careful selection of heads is essential to achieve optimal results.

ABSTRACT. Large language models (LLMs) have become powerful tools for understanding documents and answering questions (QA). The grounding of these answers consistently in facts in the given documents may be achieved by citing them in the generated responses. Several approaches to Retrieval Augmented Generation (RAG) have been proposed that incorporate citation to relevant documents to enhance correctness and verifiability. However, evaluation if the document is cited accurately, relies heavily on large generative models for Natural Language Inference. In this work, we evaluate various models in different evaluation schemes for the citation verification task to provide insights into how these models perform and in which evaluation schemes they excel. Our findings show that the TRUE T5 model performs well in verifying the completeness of citations, but struggles when only partial information is available. We also demonstrate that general LLMs can perform citation verification effectively, although the results in citation addition on an already generated answer as post-processing are still suboptimal. We argue that it is important to be mindful of how citation verifiers are used and understand their strengths and limitations. Furthermore, we trained a small and lightweight model, CiteVerifier, which performs exceptionally well despite being magnitudes smaller than other models, making it an ideal solution for low-resource settings.

ABSTRACT. Legal answering systems play a key role in increasing access to justice by providing citizens and lawyers with accurate legal interpretations. However, existing legal artificial intelligence models face challenges in processing layperson descriptions, adapting them to formal legal rules, and ensuring reliability in languages other than English and across jurisdictions. This study examines the use of systems that support information retrieval, used in search generation (RAG) systems to answer Polish legal questions. We evaluate different models across multiple legal datasets to identify optimal architectures and assess the impact of tuning on performance improvement. Our findings highlight that tuning in instruction-based datasets significantly improves the accuracy and contextual relevance of artificial intelligence models. Also, our point of interest is the possible improvement in the interpretation of layperson’s questions with respect to the law made by applying the results of the tested models.

ABSTRACT. University regulations are often complex and difficult to navigate. To address this, we developed PoliChat, a Retrieval-Augmented Generation (RAG)-based chatbot that provides accurate and transparent access to regulatory information. Validated at Wrocław University of Science and Technology, PoliChat integrates real-time retrieval with citation mechanisms to enhance reliability. As part of our research, we prepared and annotated a dataset of university regulations to evaluate information retrieval and answer generation performance. We examine key factors that affect RAG performance in regulatory domains, including model size, document length, summarization, retrieved context size, and prompting strategies. We introduce Analyze&Answer, a prompting method that improves response coherence and citation accuracy.

ABSTRACT. In this article, the authors present three didactic examples of using Chat GPT in mathematical programming optimization tasks: linear programming, nonlinear programming and convex programming. These examples are analysed in terms of the correctness of the methods used and the solutions obtained. The article also describes a didactic experiment with the participation of Informatics and Econometrics students of Warsaw University of Life Sciences, consisting in solving optimization tasks on their own using Chat GPT. The nal conclusions of the article also present a comparison of the approach based on the CAS methodology and the approach using Chat GPT.

ABSTRACT. This short paper describes the programming paradigm and the main constructs of the DCEx programming model designed for the implementation of data-centric large-scale parallel applications. The DCEx programming paradigm exploits private data structures and limits the amount of shared data among parallel threads in HPC applications. The key idea of DCEx is struc-turing programs into data-parallel blocks mapped on computing elements and managed in parallel by a large number of parallel tasks. Data-parallel blocks are the units of shared- and distributed-memory parallel computa-tions and communications in the memory/storage hierarchy. Tasks execute close to data using near-data synchronization according to the PGAS model. Two use cases implemented using DCEx constructs are also outlined and performance measures on different parallel machine configurations are shown.

ABSTRACT. The Multi-Objective Shortest Path (MOSP) problem generalizes the classic shortest path framework by simultaneously optimizing multiple, often conflicting, cost functions. Recent advances in MOSP have yielded algorithms that employ sophisticated heuristic-based techniques and dimensional reduction to expedite search. However, most existing methods rely on strictly sequential frameworks, leaving parallelized approaches relatively under-explored—especially for high-dimensional objectives. In this paper, we introduce SOPMOA* (Shared-Open Parallelized Multi-Objective A*), an algorithm that addresses this gap by allowing any number of concurrent sub-searchers to work cooperatively via a shared-memory priority queue. Each sub-searcher independently processes labels, performs dominance checks against locally stored partial Pareto fronts, and contributes to a global frontier of non-dominated solutions. We propose mechanisms for safe and efficient updates to shared data structures, ensuring correctness without excessive locking overhead. Empirical evaluations on benchmark multi-objective road networks demonstrate that SOPMOA* scales favorably with increasing parallelism and consistently outperforms state-of-the-art algorithms such as EMOA*, LTMOA*, and NWMOA* in both speed and robustness. The results underscore the substantial potential of shared-data parallelization in tackling challenging multi-objective pathfinding tasks.

ABSTRACT. In high-performance computing (HPC) environments, the efficient execution of Artificial Intelligence (AI) applications is critical to ensuring optimal performance and resource utilization. In this work, we extend the Parallel Application Signature for Performance Prediction (PAS2P) methodology to AI applications running on HPC Cloud systems. We have defined the AI Application Model to describe the performance behavior of AI workloads. This extension allows us to identify phases within AI applications, enabling performance analysis to focus on these phases rather than the entire application. Having concentrated on these phases, we can evaluate the efficiency of AI applications more effectively, offering valuable insights into system performance and guiding future optimizations for large-scale AI tasks on HPC infrastructure.

ABSTRACT. Fast prediction of job execution times is crucial for efficient job scheduling in the ALICE experiment at CERN. This paper proposes a latency-optimized neural network model for predicting job execution times, replacing the current static Time-To-Live (TTL) allocations with dynamic, data-driven predictions. Leveraging Nvidia A100 GPUs, we explore optimization techniques such as FP16 and INT8 quantization, 2:4 semi-structured sparsity, quantization-aware training (QAT), and graph-based model compilation to reduce inference latency. Experimental results show that FP16 precision and semi-structured sparsity provide significant latency reductions for larger batch sizes, while INT8 quantization achieves low latency for single-sample predictions, all without compromising prediction accuracy. Our results indicate that for single-sample online inference scenario, static INT8 weight quantization leads to the best results with median 0.38 ms prediction time, a 1.8x improvement from 0.71 ms baseline. We also demonstrate the need for careful application of latency reduction techniques based on expected sizes of input matrices and provide criteria to evaluate during their selection. The optimized model achieves a Root Mean Squared Error (RMSE) of 1.9 hours, a substantial improvement over the 14.23-hour RMSE of the existing manual TTL assignments. The model would integrate into the ALICE pipeline, maintaining inference latency below 40 milliseconds, and will be hosted within CERN’s infrastructure to ensure data locality and security. This work demonstrates how modern GPU hardware and advanced neural network optimization techniques can significantly enhance computational workflows in large-scale distributed systems.

ABSTRACT. Accurate wildfire prediction requires high-resolution and up to date fuel maps that account for seasonal variations in vegetation. In these fuel maps, croplands are often classified as non-burnable, however, their flammability can vary significantly depending on the season, which directly impacts wildfire spread simulations. This study presents a methodology that combines remote sensing indices and machine learning techniques to dynamically update fuel models in cropland zones, improving wildfire simulations. Using satellite data from Sentinel-2, cropland planting status is classified as "planted" or "unplanted." The machine learning model, trained on crop yield data and validated with ground truth derived from historical planting cycles and data, achieves 80 % accuracy. The updated fuel map has been applied to a wildfire that took place in Catalonia (Spain) in June 2019. The propagation results obtained using the updated fuel map for simulating fire spread closely align with the observed fire perimeter evolution, outperforming traditional approaches that assume all croplands are either fully planted or fully unplanted. The proposed methodology is efficient and fast to apply, allowing the status of the crop fields to be kept up to date with their real situation according to the season of the year.

ABSTRACT. This study examines the information flow within the housing market network, focusing on the Seoul metropolitan area from 2013 to 2022. Using a network constructed through the multivariate Granger causality test, the centrality met-rics identify key industrial cities, such as Suwon and Hwaseong, as central nodes, while the capital city, Seoul, plays a marginal role. The findings indi-cate a shift in influence within the housing market network, moving from tra-ditionally dominant cities like Seoul toward key industrial hubs.

ABSTRACT. Heart disease is the leading cause of death worldwide. To determine the factors contributing to the development of cardiac disease, computational fluid dynamics (CFD) and in vivo data, such as MRI of blood flow, are being compared and val-idated to better understand the hemodynamics of the heart in detail. The cardiac conduction system, which transmits electrical signals and controls the heart’s beating, is also being studied. However, no studies have examined the relation-ship between the cardiac conduction system and left ventricular hemodynamics. In this study, we focused on the Purkinje fibers located in the left ventricle within the cardiac conduction system. Left ventricular models with and without Purkinje fibers were compared. The results showed differences in blood flow within the left ventricle. Thus, changes in contraction caused by the Purkinje fibers affect the hemodynamics of the left ventricle and can contribute to the development of heart disease.

ABSTRACT. Dynamic Programming (DP) is fundamental to computational science education and application, traditionally taught through tabulation methods that emphasize manual loop construction. This paper introduces amodular, systematic plug-and-play framework that greatly simplifies DP algorithm design and parallelization. Our approach begins with a recursive divide-and-conquer analysis, decomposing DP into reusable components: Refactored Recursion (RR), OrderSpec, TileSpec, dp_solve, dp_tile_solve, and dag_run. These modules encapsulate recursive structures and facilitate seamless parallelization via dynamic Directed Acyclic Graph (DAG) scheduling. We demonstrate the versatility of this framework using three classical textbook problems: Longest Common Subsequence (LCS) highlights the plug-and-play simplicity, Matrix Chain Multiplication (MCM) employs transitive reduction for dependency clarity, and the CutRod problem illustrates previously obscured tiling optimizations and parallel solutions. This new modular paradigm significantly reduces DP’s learning curve, shifting educational focus from code-centric methods to intuitive, reusable patterns, bridging theoretical recursion with practical implementations and greatly enhancing computational science education.

ABSTRACT. Spitsbergen is Norway's largest island, located in the Svalbard archipelago of the Polar Circle region. The island's population is~currently about two and a~half thousand, and it works mainly in natural resource extraction, tourism, and scientific research. The subject of our interest is the town of Longyearbyen, the capital of Spitsbergen, with a population of about 1,800. It has an airport and, until recently, the only power plant in Norway that generated electricity by burning coal. The coal has been replaced recently with Diesel generators to reduce the pollution outcome, but the problem is still there. Due to thermal inversion, on regular days and nights, the pollutants generated by the power plant tend to stay near the ground level, and then as the sun begins to warm the air up, a layer of fog mixed with the pollutants rises and dissipates. The problem, however, is that on Spitsbergen, during much of the year, there is an Arctic night during which there are no sunrises, and the thermal inversion dissipation phenomenon does not occur, which results in trapping the pollutants near the ground level. We compare the pollution generated by the coal-burning power plant with the pollution generated by snowmobiles. Our computer simulations have been performed using a novel hypergraph-grammar based model of tetrahedral mesh refinements for the stabilized advection-diffusion model and the finite element method.

ABSTRACT. Automated classification of bird species is crucial for large-scale environmental monitoring, providing valuable insights into temporal and spatial changes in ecosystems. Previous studies have primarily focused on using either acoustic or visual data for bird species recognition. However, few studies have explored the simultaneous use of both acoustic and visual data to improve classification performance. In this study, we propose a dual branch network based on pre-trained models to enhance bird species classification by integrating acoustic and visual information. Specifically, ResNet50 is used for visual data, while CNN14 is employed for acoustic data. The extracted feature embeddings are then fused, and attention mechanisms are applied to further improve classification performance. Experimental results on the HPS dataset demonstrate that our proposed model achieves significantly higher accuracy compared to using audio or image data alone. The best-performing model achieved an accuracy of 96.44%, precision of 96.62%, recall of 94.30%, and F1-score of 95.01%. This study highlights the potential of combining acoustic and visual data for bird species classification and suggests that attention mechanisms can further enhance model performance.

ABSTRACT. Parallel computing on graphics cards (GPUs) has become a crucial component of modern computational programs. One of the most widely used tools for enabling scalable operations on NVIDIA GPUs is CUDA. CUDA Graphs, introduced in CUDA 10, is a functionality that enables applications to be structured as graphs, potentially re- ducing the overhead associated with repeated kernel function calls. In this paper, we contribute by providing a direct comparison with stan- dard CUDA and CUDA Graphs, for high performance computing (HPC) workloads. A set of kernels and applications from the NAS Parallel Benchmarks suite, implemented for CUDA, was selected. Subsequently, we implemented CUDA Graphs versions for these codes. The evalua- tion was conducted across four hardware platforms, using two desktop GPUs—NVIDIA GeForce RTX 2080 and NVIDIA GeForce RTX 4070 Ti—and two server-class GPUs—NVIDIA Quadro 8000 and NVIDIA A100 80GB. The primary focus of the comparison was the execution time of the benchmarks, analyzed for various problem sizes (classes S, A, B, C and D). Two applications exhibited noticeable performance gains from the implementation of CUDA Graphs. The CG code demonstrated the most consistent improvements across all cases, achieving an average relative speedup of 3.3%. The highest result of 4.13%(Class C) for this algorithm was achieved for the NVIDIA GeForce RTX 4070 Ti card. In contrast, visible negative performance was observed for MG and some instances of IS, but we attribute that to the relatively small absolute running time in which case additional overheads cannot be mitigated by the new mechanism.

ABSTRACT. Prototype selection is one of the typical goals of machine learning, which aims to reduce the number of vectors in the training set. The local set border selector (LSBo) algorithm is presented as a Pareto optimal choice between the reduction power of the training set and the classification quality. Its complexity is O(n2), which means that it is not very advantageous for larger sets. This article presents the Fast LSBo algorithm, which is based on the original idea of the LSBo algorithm. After the applied conceptual changes, the algorithm has achieved a com- plexity of O(m log m). Additionally, the analysis of Fast LSBo on several data sets shows that its classification quality and reduction power remain statistically indistinguishable from the original LSBo algorithm.

ABSTRACT. We present a method that generates simple and accurate decision rules from datasets. The method utilizes the concept of contrastive prototypes for decomposing classification problems into sub-problems, in which the extracted rules are simple and comprehensive. The key is to perform the decomposition along the decision boundary of the classification problem. This simplifies the task and assures that both classes (positive and negative) are proportionally distributed. The algorithm starts by selecting representative prototypes obtained by a prototype construction method. Then pairs of prototypes are determined. A pair must consist of prototypes from opposite classes, so that each pair defines a region containing a fragment of the decision boundary. These pairs define subspaces. Inside each subspace a small decision tree is applied to extract simple decision rules. The results indicate that a very simple set of decision rules obtained for the subspaces has prediction accuracy comparable to that of decision trees, which are much more complex and difficult to interpret.

ABSTRACT. Seismic monitoring is used to ensure the safety of workers in the rock massif. The main security threat is a rockburst, which can be predicted based on the sequence of seismic events. An important task is to develop a mining forecasting model that can take into account the structural heterogeneity of the mountain range and select the necessary forecast horizon depending on monitoring data. In the paper, we propose a flexible approach that combines multiple machine learning models designed to solve various tasks (clustering, time series forecasting) as parts of one composite model. This approach allows for adjustment of the forecast horizon of the model, which enables it to flexibly adapt to rock massifs with different geological structures and seismic monitoring stations. Also, the use of clustering models allows us to take into account the physical and mechanical features of the rockburst formation process. According to experimental results, the resulting composite model showed more accurate results for specific forecast horizons, compared with classical "hierarchical" models and machine learning models. At the same time, the obtained model allows us to interpret the results from the rock mechanics point of view.

ABSTRACT. This paper presents an innovative approach to using computational methods in designing modular housing estates with the Wave Function Collapse (WFC) algorithm. Currently, as a result of fast changing humanitarian situations in various parts of the world, like the one caused by the war in Ukraine or climate changes, many people have to leave their homes. It results in complex problem of providing large groups of people with conditions that will allow them to function with dignity for a long time. Therefore the adaptation of the WFC algorithm to design modular settlements is proposed. The applied heuristics allow the solutions to adapt to specific project requirements, generating various modular settlement designs that consider functionality and social aspects. The proposed approach is illustrated by examples of generated arrangements of housing modules for family-type floor plans.

ABSTRACT. With Human – Artificial Intelligence (AI) collaboration booming in all fields, the pace of task-based cooperation is ever-expanding. Yet, in most applications, AI induction is sidelined to test beds and is viewed as a sceptical competitor rather than a collaborator. The healthcare domain is one such field where AI support is viewed as theoretical and far from practical. While most focus is directed towards developing and training AI models, the human expert and their interaction is often overlooked. We present a simulation-based approach that optimizes AI learning policy of a domain expert’s behaviour when evaluating decision support data of a patient’s risk of acquiring type 2 diabetes Mellitus (T2DM). Using Linear and Maximum Entropy inverse reinforcement learning (IRL) algorithms, we analyse various learning strategies by including context, reward strategies and sampling rates to showcase expert characteristics through optimal policies and effective reward functions. Our results provide insights on individual individual experts’ evaluation policy and the AI model’s learning behaviour in various environmental scenarios.

ABSTRACT. The transportation sector is responsible for approximately 23% of global greenhouse gas (GHG) emissions, with road transportation contributing nearly 70% of these emissions. The widespread adoption of electric vehicles (EVs) is transforming this sector by reducing emissions and decreasing reliance on fossil fuels. However, the growing number of EVs presents significant challenges for charging infrastructure, particularly in managing long queues, extended wait times, and limited station capacity. Most existing studies on EV charging station performance assume Poisson-distributed EV arrivals and exponentially distributed charging times—assumptions that often fail to capture real-world variability. In this paper, we propose a generalized queueing model that incorporates real-world datasets of interarrival times and charging durations to provide a more accurate evaluation of charging station performance. We conduct a transient analysis of charging station operations. We assess key performance metrics such as the mean number of EVs waiting in queues, station throughput (EVs charged per hour), and the proportion of EVs unable to charge due to full station capacity. Additionally, we examine the impact of a queue management policy that incentivizes users to charge only up to a predefined state-of-charge (SoC) threshold rather than the conventional 80-100% SoC. This strategy effectively reduces charging durations, alleviates congestion, and enhances station throughput. Our findings demonstrate that optimised charging policies improve charging station efficiency, reduce service losses, and significantly enhance customer experience. By implementing effective queue management strategies, charging station operators can minimize congestion, increase revenue, and support the sustainable expansion of EV adoption.

ABSTRACT. The paper presents a multidimensional horizontal approach to solving the fuzzy Sylvester matrix equation (FSME). The use of the horizontal membership function (HMF) of the fuzzy set allows for generating a granule of information about the FSME solution. The paper presents an algorithm for solving FSME using HMF, which generates a full FSME solution. The solution obtained using the given algorithm differs from the results presented in the cited articles. The calculated granule of the FSME solution contains solutions that do not occur in the results obtained in the analyzed examples, therefore these results are underestimated.

ABSTRACT. In high-performance computing, asynchronous computation is an important optimization technique. The upcoming C++26 standard introduces a new asynchronous execution framework, stdexec, enabling the development of high-performance code using only standard C++. This paper explores the parallelization of single-GPU and multi-GPU lattice Boltzmann method (LBM) computations using stdexec and further optimizes performance through its asynchronous execution model. Additionally, stdexec is compared with other mainstream parallel computing frameworks, including CUDA, C++ standard parallelism, and OpenACC. Experimental results show that stdexec achieves approximately 61.4%-100.3% of the performance of C++ standard parallelism, and with further optimization through asynchronous execution, it can achieve comparable performance. Considering that stdexec is still a prototype, these results suggest potential for further optimizations in the future, providing additional options for high-performance computing development in C++.

ABSTRACT. The dynamics of firm populations are governed by the birth of new firms and the death of existing ones, and understanding these processes can inform urban planning and policies. Longitudinal data containing the entry and exit dates of all firms would allow direct analysis of birth and death rates and how they depend on external market factors and firm-level properties. However, real-life data sets are often incomplete, with missing records of extinct firms or the exit dates of those firms. In such scenarios, it is unclear whether and how we can extract information about the birth and death processes. Here, by modeling how these processes influence the age distributions and survival fractions of firms, we reveal how one can gain insights even from incomplete data. In particular, we show that while the age distribution alone is insufficient for inferring both the birth and death dynamics, the survival fractions can be used to infer how death probabilities depend on firm age and sector size. We illustrate our approach separately on 14 of the largest firm sectors in Singapore, and found that the observed survival fractions imply that the death probabilities decrease with firm age and increase with sector size, with multiplicative interactions between both factors. By assuming a sigmoidal dependence of death probability on both age and sector size, we infer a death model for each of the sectors that can accurately reproduce empirical data. We also demonstrate how the inferred death models can be used to reconstruct other system properties.

ABSTRACT. In this work, we explore the Transcriptomics Atlas pipeline adapted for cost-efficient and high-throughput computing in the cloud. We propose a scalable, cloud-native architecture designed for running a resource-intensive aligner -- STAR -- and processing tens or hundreds of terabytes of RNA-sequencing data. We implement multiple optimization techniques that give significant execution time and cost reduction. The impact of particular optimizations is measured in medium-scale experiments followed by a large-scale experiment that leverages all of them and validates the current design. Early stopping optimization allows a reduction in total alignment time by 23\%. We analyze the scalability and efficiency of one of the most widely used sequence aligners. For the cloud environment, we identify one of the most suitable EC2 instance types and verify the applicability of spot instances usage.

ABSTRACT. This paper presents a method for accurately estimating eigenvalues and eigenvectors, which are fundamental in many computational problems, including those in physics, engineering, and numerical simulations. By providing guaranteed error bounds and leveraging interval arithmetic, the method enhances the reliability and precision of numerical solvers.

ABSTRACT. Training and fine-tuning language models is becoming increasingly expensive. ''Model soups'' offer a promising solution by combining parameters from separately trained models to create a new one with merged capabilities. Our paper explores using heterogeneous model soups to improve LLM alignment by combining models trained with different alignment methods - a novel approach not previously explored in literature. Through empirical evaluation using an ''LLM-as-a-judge'' approach, we found that mixing different types of models can improve alignment performance, though this requires careful adaptation of interpolation techniques to account for varying alignment objectives. Our final model merges are available on HuggingFace, and we've shared our model merging source code on GitHub.

ABSTRACT. Extracting high-level conceptual models from simulation code can benefit model validation and verification, system optimisation, and cross-disciplinary communication. However, conceptual models are often embedded within implementation details, making them difficult to access and interpret. This paper explores the feasibility of using Large Language Models (LLMs) to infer conceptual models from simulation code. We conduct a preliminary investigation on an agent-based simulation (Flee), demonstrating how LLMs can extract key structural, behavioural, and temporal elements. Our results suggest that LLMs can generate meaningful conceptual representations that align with expert-created models, offering potential support for model verification. However, we also identify limitations such as omissions and misinterpretations, highlighting the need for human oversight. While our study is based on a single example, it provides initial insights into the role of LLMs in conceptual model inference and their potential integration into simulation validation workflows.

ABSTRACT. We introduce NeoN, a tool for detecting and analyzing Polish neologisms. Unlike traditional dictionary-based methods requiring extensive manual review, NeoN combines reference corpora, Polish-specific linguistic filters, an LLM-driven precision-boosting filter, and daily RSS monitoring in a multi-layered pipeline. The system uses context-aware lemmatization, frequency analysis, and orthographic normalization to extract candidate neologisms while consolidating inflectional variants. Researchers can verify candidates through an intuitive interface with visualizations and filtering controls. An integrated LLM module automatically generates definitions and categorizes neologisms by domain and sentiment. Evaluations show NeoN maintains high accuracy while significantly reducing manual effort, providing an accessible solution for tracking lexical innovation in Polish.

ABSTRACT. This paper proposes a novel hybrid approach to routing in Wireless Sensor Networks that combines Q-learning and Learning Automata models. It is designed to optimize the routing process by leveraging the strengths of both techniques: Q-learning's ability to adapt to dynamic network conditions and Learning Automata's fast adaptation and convergence in stable scenarios. Preliminary analysis proves the feasibility of the proposed approach, showing that it can improve the network lifetime and packet delivery ratio when compared with similar routing protocols.

ABSTRACT. Selecting appropriate examples for in-context learning significantly impacts the performance of Large Language Models (LLMs). In this paper, we show that leveraging LLMs’ positional biases and incorporating knowledge of class distribution can improve classification outcomes, especially for underrepresented classes. We introduce Class-few-shot, a method that balances class representation among few-shot examples. To investigate this, we conduct almost 10,000 experiments on 4 datasets and 3 models, cross-checking how different biases affect models' performance and how they interact. We show that presenting classes from the most to least numerous (Maj-Min) using an alternating pattern leads to better results than standard few-shot prompting with the same number of examples. Additionally, we investigate the impact of label correctness and compare the general few-shot and Class-few-shot results, outlining the strengths of both approaches. All our raw experiment results are publicly available on GitHub.

ABSTRACT. This paper introduces “f3cake,” an open-source numerical platform developed by the Computational Transport Phenomena Lab (CTPL) for solving PDEs using finite difference/volume methods. The name “f3cake” stands for Friendly, Fast, Conservative Algorithms, and Coding Environment, highlighting its core features: ease of use, high efficiency, and conservation of physical variables. Developed with Julia, “f3cake” supports irregular, non-uniform, and unstructured meshes and leverages Julia’s ability to make coding expressions more intuitive, closely mimicking PDE/ODE formulations. Its forward programming design simplifies implementation and debugging, accelerating development compared to traditional tools. “f3cake” can also ensures high performance of simulating for a wide range of scientific applications and physical phenomena, including fluid flow, seepage, heat transfer, and solid mechanics. Numerical examples and some guide of this platform will be given in this content.

ABSTRACT. This paper presents an algorithm designed to eliminate the direct evaluation of both strongly and weakly singular boundary integrals in the parametric integral equation system (PIES) for the analysis of three-dimensional multidomain ortho-tropic problems. Singular integrals present considerable challenges in the computational implementation of PIES, directly influencing the accuracy of the analysis results. The proposed method regularizes PIES by incorporating an auxiliary regularization function with coefficients that effectively remove singularities. As a result, the regularized PIES avoids the explicit evaluation of singular integrals, enhancing computational efficiency and precision.

ABSTRACT. Graph-based machine learning models have gained significant attention in predicting the emergence of new relationships in evolving networks. In this work, we present a study on forecasting scientific collaborations using a Graph Attention Network (GAT) with L2 regularization and dropout. We construct yearly co-authorship graphs based on historical publication data and analyze the evolution of these graphs over time, having the International Conference on Computational Science (ICCS) as an example of a living scientific community. Our approach involves training on past yearly graphs to predict the formation of new edges in future graphs. We assess the model's performance by varying the prediction window and evaluating results using link prediction metrics. The proposed method demonstrates the feasibility of utilizing deep learning techniques for predicting future collaborations based on past scientific interactions.

ABSTRACT. Numerous algorithms and parallelisations have been developed for short-range particle simulations; however, none are optimally performant for all scenarios. Such a concept led to the prior development of the particle simulation library AutoPas, which implemented many of these algorithms and parallelisations and could select and tune these over the course of the simulation as the scenario changed. Prior works have, however, used only naive approaches to the algorithm selection problem, which can lead to significant overhead from trialling poorly performing algorithmic configurations.

In this work, we investigate this problem in the case of Molecular Dynamics simulations. We present three algorithm selection strategies: an approach which makes performance predictions from past data, an expert-knowledge fuzzy logic-based approach, and a data-driven random forest-based approach. We demonstrate that these approaches can achieve speedups of up to 4.05 compared to prior approaches and 1.25 compared to a perfect configuration selection without dynamic algorithm selection. In addition, we discuss the practicality of the strategies in comparison to their performance, to highlight the tractability of such solutions.

ABSTRACT. Surrogate models can play a pivotal role in enhancing performance in contemporary High-Performance Computing applications. Cache-based surrogates use already calculated simulation results to interpolate or extrapolate further simulation output values. But this approach only pays off if the access time to retrieve the needed values is much faster than the actual simulation. While the most existing key-value stores use a Client-Server architecture with dedicated storage nodes, this is not the most suitable architecture for HPC applications. Instead, we propose a distributed architecture where the parallel processes offer a part of their available memory to build a shared distributed hash table based on MPI. This paper presents three DHT approaches with the special requirements of HPC applications in mind. The presented lock-free design outperforms both DHT versions which use explicit synchronization by coarse-grained resp. fine-grained locking. The lock-free DHT shows very good scaling regarding read and write performance. The runtime of a coupled geochemical simulation was improved between 14% and 42% using the lock-free DHT as a surrogate model.

ABSTRACT. Within this study, we propose and evaluate an approach to structure and improve context provided in RAG-based solutions for code generation. The approach is based on combination of semantically relevant API and code selection and filtering for better context representation in following LLM prompt. The experimental evaluation performed with CodeGen-350M-mono and several popular benchmarks such as RepoCoder, CoderEval, CoIR show good overall performance (even in comparison to bigger LLMs). Also, the experimental evaluation shows improvement with narrower and more focused context representation (project-scope API instead of popular public API).

ABSTRACT. The increasing frequency and severity of wildfires in the Western United States demand improved fire response tools. Initial Attack Fire Response within the first few hours of ignition is critical in preventing fires from escalating. WIFIRE Firemap has been instrumental in supporting early fire suppression efforts through real-time fire behavior modeling. However, wildfires often burn for days or weeks, necessitating longer-term predictive capabilities. To address this challenge, we extended Firemap to forecast fire spread from the first few hours to five days. This advancement integrates two long-term fire behavior models, ELMFIRE and GridFire, enabling real-time, data-driven decision support. The enhanced Firemap platform improves strategic wildfire response planning, allowing firefighters and emergency managers to anticipate fire spread on extended timelines. We present how these extensions were used during the Los Angeles firestorms of 2025, demonstrating their potential to mitigate wildfire risks, protect communities, and improve firefighting strategies, and make recommendations for effective use of extended attack tools for decision support.

ABSTRACT. The increasing occurrence and complexity of forced displacement require robust predictive models to aid humanitarian responses. However, existing predictive models for forced displacement lack accurate and timely data, have gaps in existing datasets, struggle with the unpredictability of human behaviour, do not account for rapidly evolving political and environmental factors and introduce methodological uncertainties. Hence, this paper proposes an artificial intelligence (AI)-enhanced agent-based modelling (ABM) approach to assist effective humanitarian planning and efficient resource allocation. This novel approach aims to predict the movements of internally displaced people and the arrival of forcibly displaced individuals in neighbouring countries. Importantly, it introduces a combination of quantitative models and qualitative insights from expert knowledge, along with humanitarian reports. Our AI-enhanced ABM approach (i) uses an agent-based simulation tool, Flee, incorporating behavioural assumptions and customisable rulesets for scenario modelling, (ii) analyses near real-time signal data from geospatial data and social media activity to satellite imagery with AI techniques, and (iii) refines the ABM model with AI-generated inputs to enhance the granularity, accuracy, and reliability of predictions.

ABSTRACT. In this talk, we present our recent studies on modelling and simulation aimed at unravelling the plasticity deformation and failure behaviours of multi-principal element alloys (MPEAs). First, we explore shock-induced deformation and spallation in CoCrFeMnNi HEAs, uncovering anisotropic deformation mechanisms and highlighting the role of chemical short-range ordering (SRO) in reducing ductility while enhancing spall strength. Next, we examine the Hall-Petch to inverse Hall-Petch transition under shock loading, providing a pressure-dependent model that captures the interplay between hardening and softening mechanisms in nanocrystalline HEAs. Subsequently, we present a temperature-sensitive crystal plasticity model to characterize the deformation and failure resistance of Cantor alloy-like MPEAs under tensile and cyclic loading, linking SRO effects to improved mechanical responses and enhanced fatigue resistance. Finally, we present a discrete dislocation plasticity framework to reveal the serrated plastic flow of Cantor alloys, emphasizing the microstructural factors that govern temperature-dependent stability. These studies collectively demonstrate the power of multi-scale simulations in elucidating the complex deformation and failure behaviours of MPEAs, offering valuable insights for advancing alloy design and engineering to achieve superior mechanical performance under dynamic and thermal conditions.

ABSTRACT. The objective of the paper is to describe computational methods of control synthesis for a certain type of nonlinear driftless control systems. Such systems are previously found to be treated as simplifications (called homogeneous approximations) of more complicated nonlinear systems that still preserve most crucial properties of the original ones like controllability. The class of systems in question have a special feedforward form that is much easier to integrate and allows to solve concrete problems in control theory. Here we continue our research with describing the computational procedure for control synthesis as the extension of existing software libraries in Python language. We show that our approach leads to faster computation times compared to standard methods. The results are illustrated with some numerical experiments and simulations.

ABSTRACT. Digital Twins (DTs) have driven significant innovation across industries, creating virtual replicas of physical assets that enable continuous learning, optimization, and informed decision-making. Digital Twins Systems of Systems (SoS) pose open challenges that relate to representation, orchestration, and management at scale and call for innovative approaches for collaborative modelling of their ecosystem. Federated Digital Twins (FDTs) have emerged as a solution, enabling integration and resource sharing between independent DTs, fostering collaboration and unlocking the full potential of interconnected systems. This work proposes a framework inspired by connectionism theory to model FDTs as a system of systems, drawing on federated systems and cognitive neuroscience to facilitate collaboration and emergent communication patterns. A Smart Connected Farming case study is used as a proof of concept for the proposed framework.

ABSTRACT. In this paper, we investigate the application of a reinforcement learning algorithm known as the Dueling Double Deep Q-Network to discover bus scheduling strategies and compare them against conventional approaches. In particular, we look into real-time control strategies where buses may choose to stay or leave at bus stops. We explore both waiting time and travel time as the optimization objectives. The results for uniform bus frequency show that average waiting time can be reduced by allowing buses to stay longer at stops with higher passengers' arrival rate but at the cost of increased average travel time. This is also supported by our analytical calculation on a theoretical bus loop model. We then apply our method to a model based on a real world bus loop in Nanyang Technological University. The results highlight the potential benefit of reinforcement learning methods to find novel strategies that can be better than conventional approaches.

ABSTRACT. Floating Offshore Wind Turbines (FOWT) provided new potential in harvesting wind energy in far offshore deep-sea regions and contributed to the world decarbonization Net-Zero target. Providing structural health monitoring (SHM) is crucial for ensuring the structural integrity of FOWT in lifecycle. How-ever, the SHM is technically challenging with high Operational and Maintenance Expenditure (OPEX). Recently, Digital Twin (DT) and advanced sensor technologies offer alternative solutions to provide effective strategy in SHM re-motely. Data-driven DT with deep learning models can formulate highly nonlinear dynamics systems. Yet, these existing models only perform the “black box” prediction without explicitly modeling the spatial-temporal relationship and con-sider only homogenous loading exerted in contrast to the complicated loading combination of FOWT with wind, wave and sea current. To address the existing modelling limitations, a new Graph Neural Network (GNN)-Encoder-Decoder-Long Short-Term Memory (LSTM) surrogate model of FOWT is presented in this work, which can perform 50 times faster than the real-time of simulation data set with accurate prediction of wind turbine tower bottom forces in the dominant dynamic modes force-aft and side-side directions. The training data is based on the software QBlade simulation and focuses on the OC4 5MW DeepCwind FOWT structure. A holistic quantitative analysis is carried out to validate the tractable latent space vectors for this complex FOWT system.

ABSTRACT. We propose an iterative quantum-assisted least squares (i-QLS) optimization method that leverages quantum annealing to overcome the scalability and precision limitations of prior quantum least squares approaches. Unlike traditional QUBO-based formulations, which suffer from an exponential qubit overhead due to fixed discretization, our approach refines the solution space iteratively, enabling exponential convergence while maintaining a constant qubit requirement per iteration. This iterative refinement transforms the problem into an anytime algorithm, allowing for flexible computational trade-offs. Furthermore, we extend our framework beyond linear regression to non-linear function approximation via spline-based modeling, demonstrating its adaptability to complex regression tasks. We empirically validate i-QLS on the D-Wave quantum annealer, showing that our method efficiently scales to high-dimensional problems, achieving competitive accuracy with classical solvers while outperforming prior quantum approaches. Experiments confirm that i-QLS enables near-term quantum hardware to perform regression tasks with improved precision and scalability, paving the way for practical quantum-assisted machine learning applications.

ABSTRACT. Emergency Departments (EDs) face increasing complexity due to rising patient demand, resource constraints, and the need for efficient service coordination. Traditional simulation models, while useful, cannot be easily adapted to a different hospital environments, making it difficult to transfer and scale solutions. Based on previous work, with a simulator working in a hospital, this work describes a modular Agent-Based Modeling and Simulation (ABMS) approach for increasing flexibility adaptation and reuse in ED simulations. The proposed technique, which deconstructs existing models into individual components, will allow hospitals with different workflows and operational constraints to construct customized simulations. To validate this methodology, we develop a structured, modular framework using NetLogo and Python. The suggested metasystem enables adaptive simulation-based decision assistance for emergency departments, which improves resource allocation and operational planning.

ABSTRACT. In recent years, the interplay between machine learning (ML) and cheminformatics has driven advancements in bioactivity prediction. However, the challenge of model explainability remains a significant barrier to adopting these approaches in drug discovery. This study addresses critical shortcomings in existing modeling techniques by examining the assumptions of feature independence and contribution additivity that are the foundation of traditional explainability methods. We investigate fingerprint-based and molecular graph models within quantitative structure-activity relationship modeling. While these models demonstrate impressive predictive performance, they offer limited actionable insights for medicinal chemists. To assist researchers in developing useful and interpretable activity prediction models, we propose a new benchmark based on the pharmacophore concept, commonly used in preliminary compound filtering. Furthermore, we introduce PharmacoScore, a novel evaluation metric designed to assess whether ML-based explanations prioritize essential pharmacophore components over non-critical features. Our findings highlight a crucial misalignment between ML model explanations and established pharmacophore principles, revealing a pressing need for innovative interpretability strategies in cheminformatics. This work not only offers a valuable resource but also sets the stage for future research, enhancing the transparency of ML in drug discovery.

ABSTRACT. Measuring multiple reaction monitoring (MRM) transitions for peptides is a critical task in targeted proteomics, enabling the precise identification and quantification of peptides in complex biological samples. This study presents a novel method for MRM transition prediction, leveraging a hash-based representation inspired by convolutional neural networks. Peptide fragments are encoded as sparse count vectors, incorporating local sequence context through efficient hashing. We employ gradient-boosted decision trees (GBDTs) for prediction, capitalizing on their robustness to high-dimensional, sparse input data. Compared with the baseline models, the proposed method significantly improves prediction accuracy, achieving a mean Hits@5 score of 3.4318 for the hash-based model and 3.5405 for a hybrid model that integrates hash-based encoding with target frequency features. These results are statistically significant, demonstrating the efficacy of our approach. Furthermore, the method is computationally efficient, with inference optimized by transpiling trained models into Zig, producing lightweight executables capable of high-speed processing. Our results highlight the potential of this approach for scalable peptide MRM transition prediction, offering a practical solution for high-throughput proteomics applications.

ABSTRACT. Advancing single-cell analysis requires tools that not only enable precise experimental measurements but also offer predictive capabilities to guide device optimization and expand experimental possibilities. This study addresses this need by developing a digital twin framework for mechano-NPS, a high-throughput microfluidic platform for single-cell analysis. By creating a virtual replica that integrates models of fluid dynamics and cellular behavior, the digital twin serves as a critical tool for both device development and hypothesis exploration. The foundation of the digital twin was established by accurately modeling the fluid dynamics within the mechano-NPS device, with simulations at various inlet pressures verified against analytical solutions. To ensure biological relevance, cellular models were rigorously tested to replicate key behaviors within the platform. The digital twin’s performance was validated against experimental data, focusing on cell velocity and whole cell deformation index (wCDI). While variances in cell velocity highlighted systematic biases, the strong agreement of simulated wCDI with experimental results underscores the digital twin’s reliability. This framework not only demonstrates the potential to enhance the mechano-NPS platform but also exemplifies how digital twins can transform experimental approaches in cellular biology.

ABSTRACT. Analyzing multi-featured time series data is critical for space missions: efficient event detection is essential for automatic analysis, potentially onboard. Limited onboard computational resources and data downlink constraints necessitate robust methods for identifying regions of interest in real time. This work presents an adaptive outlier detection algorithm based on the reconstruction error of Principal Component Analysis (PCA) for feature reduction, designed explicitly for space mission applications. The algorithm adapts dynamically to evolving data distributions by using Incremental PCA, enabling deployment without a predefined model of all possible conditions. A pre-scaling mechanism normalizes feature magnitudes while preserving relative variance within feature types. We demonstrate the algorithm’s effectiveness in detecting space plasma events, such as distinct space environments, dayside and nightside transients, and transition layers through NASA’s MMS mission observations. Additionally, we apply the method to NASA’s THEMIS data, successfully identifying a dayside transient using onboard-available measurements.

ABSTRACT. Magnetic Resonance Imaging (MRI) reconstruction remains a vital research domain, with ongoing efforts directed at enhancing image quality while minimizing acquisition time. Conventional k-space filling techniques have traditionally relied on Fourier transform properties and interpolation methods. However, recent innovations have introduced deep learning and compressed sensing as optimization strategies for this process. This study introduces a novel super-resolution framework that incorporates Toeplitz matrices for structured k-space completion, integrated with deep learning-based models and compressed sensing methodologies. Given their intrinsic connection to convolutional operations, Toeplitz matrices provide a mathematically sound foundation for defining k-space structures while ensuring data consistency.

Within this framework, deep neural networks are employed to infer the underlying k-space distribution in PROPELLER sequences from sparsely sampled data, while Toeplitz matrix constraints are utilized to maintain coherence. Additionally, the application of compressed sensing principles—incorporating sparsity priors and regularization techniques—improves both robustness and image quality, facilitating high-fidelity reconstructions from substantially undersampled acquisitions. The proposed approach is validated using both simulated and real MRI datasets, demonstrating that it effectively reduces reconstruction error and enhances image quality in comparison to traditional interpolation methods and standalone deep learning models. The findings indicate that combining sequentially rotating blade raw data acquisition with structured priors based on Toeplitz matrices, deep learning-driven inference, and compressed sensing optimization can yield more precise and computationally efficient MRI reconstructions, ultimately contributing to faster scan times in clinical applications.

ABSTRACT. In today's intricate decision-making field, characterized by diverse variables and conflicting objectives, the demand for reliable, knowledge-driven approaches is crucial. Multi-Criteria Decision Analysis (MCDA) methods serve as relevant tools in this area, offering structured evaluation frameworks to guide comprehensive decision support processes. However, the variety of MCDA methods introduces a significant challenge. These methods often provide divergent rankings, necessitating compromise solution methods to include those differences and propose cohesive recommendations. While established techniques address this need, they often overlook the potential synergy between positional rankings and preference scores, limiting the depth of insights provided. To fill this gap, this study introduces a novel Compromise Fuzzy Ranking (CFR) method, which integrates positional rankings and preference scores to offer nuanced and comprehensive consensus in multi-criteria evaluations. CFR's efficacy is demonstrated through simulation and theoretical examples, highlighting its potential to enhance decision support systems and navigate the complexities of contemporary decision-making challenges.

ABSTRACT. This paper introduces a novel approach to large-scale group decision-making, addressing the challenge of aggregating diverse expert opinions effectively. Integrating the Ranking Comparison (RANCOM) method into aggregation procedures minimizes the impact of individual inaccuracies on results. The study compares six aggregation techniques through simulation experiments and a practical case study on solar panel evaluation. Furthermore, it proposes a new fuzzy ranking-based approach to aggregate group results. The research underscores the significance of addressing inaccuracies in individual assessments and subtle differences in expert opinions. Key contributions include an enhanced method for aggregating expert opinions and a robust framework for large-scale group decision-making, improving the reliability of outcomes in complex multi-criteria scenarios.

ABSTRACT. Most Multi-Criteria Decision Analysis (MCDA) methods usually rely on importance weights for criteria to represent the knowledge and preferences of the decision-maker or expert. While this approach effectively prioritizes criteria, it may not suffice in cases where the resulting ranking contains ties or very close solutions. To address this limitation, we introduce the Frobenius SPOTIS (Fro-SPOTIS) method, which incorporates information about performance ordering in alternatives into the decision-making process. To assess the difference between ordering, Frobenius distance between rankings is used. The method can be fine-tuned using a tolerance parameter $\tau$, allowing for flexibility in its application. To demonstrate the application of the method, we present a simple experiment in which we resolve ties in ranking. Additionally, we conduct two sensitivity analysis experiments to explore the features and limitations of the Fro-SPOTIS method, providing a comprehensive understanding of its capabilities and potential areas for improvement.

ABSTRACT. Fast and accurate flood forecasting models are fundamental for managing flood risk and mitigating the negative impacts that floods can have on the society and the environment. For the pan-European area, currently, the European Flood Awareness System (EFAS) is the official flood forecasting and early warning system. It's forecasts derive from a process-based rainfall-runoff model (LISFLOOD), which requires large amounts of high quality hydro-meteorological data, that usually are not uniformly available, affecting the quality of the outputs. Running process-based models at high spatial resolution for large spatial scales requires high-performance computer clusters to deliver timely forecasts. Recently, the use of machine learning-based forecasting models like long short-term memory (LSTM) as surrogate models for traditional process-based models has gained popularity. Compared to traditional process-based models, machine learning-based forecasting models offer the advantage of lower computational resource requirements and greater tolerance for variations in input data quality. Additionally, machine learning-based compression models, such as convolutional autoencoder (CAE), can further reduce computational costs by compressing the data. Moreover, large-scale models, frequently exhibit lower accuracy in forecasting river discharge in smaller watersheds, in which data may be less available and rivers are smaller, but still significant, and cannot be overlooked. These watersheds exhibit a fast response to rainfall events, thus requiring fast river discharge forecasts to guarantee enough lead time to take action in case of an imminent flood. To enhance forecasting accuracy, data assimilation (DA)—a technique that integrates data from multiple sources to optimize forecasting outcomes—can be effectively employed to improve the precision of river discharge forecasting. In this work, we propose a latent three-dimensional variational data assimilation (3D-Var) method combined with machine learning models to deliver fast and accurate river discharge forecasting. We tested our method on the real-world datasets (EFAS and Lamah-CE) and achieved an average 53.6% improvement in forecasting accuracy measured by Mean Squared Error (MSE) compared to LSTM forecasting, while delivering one-day lead-time river discharge forecasting in approximately one minute for an area of around 30,000 km^2.

ABSTRACT. The goal of many applications in energy and transport sectors is to control turbulent flows. However, because of chaotic dynamics and high dimensionality, the control of turbulent flows is exceedingly difficult. Model-free reinforcement learning (RL) methods can discover optimal control policies by interacting with the environment, but they require full state information, which is often unavailable in experimental settings. We propose a data-assimilated model-based RL (DA-MBRL) framework for systems with partial observability and noisy measurements. Our framework employs a control-aware Echo State Network for data-driven prediction of the dynamics, and integrates data assimilation with an Ensemble Kalman Filter for real-time state estimation. An off-policy actor-critic algorithm is employed to learn optimal control strategies from state estimates. The framework is tested on the Kuramoto-Sivashinsky equation, demonstrating its effectiveness in stabilising a spatiotemporally chaotic flow from noisy and partial measurements.

ABSTRACT. When computations of the dynamic behavior of a digital twin includes the recursion of an internal state, data assimilation can be used to adjust the numerical values of the state. The optimal linear adjustment of this state on the basis of observations and simulations is known as a Kalman filter, in which an optimal linear gain is computed based on covariance information to minimize the variance on the state error. This paper illustrates that such covariance information can be learned and used to find an optimal trade-off between the observations and simulations for state adjustment. Although the concept of learning covariance information is well understood by the Ensemble Kalman Filter (EnKF), this paper emphasizes the underlying approach how to learn covariance information with the purpose of convergence and minimal variance of the state error. The concept is illustrated for a dynamic digital twins of an linear oscillatory mechanical system and a non-linear dynamic wildfire progression. The examples illustrate that the results on data assimilation heavily depends on the quality of the covariance information.

ABSTRACT. Alpine skiing is a complex sport where technique is the key. The ability to detect turns, along with their intricate patterns, can provide valuable insights into performance analysis and injury prevention for skiers. Modern turn detection systems are often costly or cumbersome to use, limiting accessibility for recreational skiers, professionals, and coaches alike. In this paper, we present our AI-based solution focused on IMU sensors embedded in smartphones, which are widely available to the general public. Our approach addresses the challenges posed by noisy IMU data from mobile devices, varying skiing techniques, and diverse environmental conditions. We collected skiing data from 11 skiers of varying skill levels, who skied freely (without designated tracks) across three different ski resorts. The dataset consists of measurements captured by smartphones, including IMU signals from accelerometers, gyroscopes, and orientation sensors. To process the data and extract individual turns, we developed a gradient-based algorithm paired with optimization techniques specifically tailored to the constraints of smartphone sensors. Our proposed algorithm achieves robust turn detection while maintaining computational efficiency, enabling analysis on mobile devices. Experiments demonstrate that the model achieves an F1 score of 0.943 on test data. This highlights the potential of using smartphone-embedded sensors for sports analytics, making advanced motion detection more accessible to a broader audience. Our findings open pathways for personalized feedback systems and scalable solutions in ski analytics.

ABSTRACT. We present a mixed finite element projection method for the unsteady Stokes equations. Each time step involves the solution of a predictor and a projection problem. The predictor problem accounts for the viscous effects and uses a stress-velocity mixed formulation. The projection problem is based on a velocity-pressure mixed formulation and enforces the incompressibility. Unconditional stability and first order in time accuracy is established. A second order multipoint flux mixed finite element method on simplicial grids is utilized for the space discretization. This allows for local elimination of the stress in the predictor problem, resulting in a positive definite velocity system. Similarly, the velocity in the projection problem is locally eliminated, resulting in a positive definite pressure system. At the end of each time a second order accurate H(div)-conforming piecewise linear velocity is computed, which is pointwise divergence free. Numerical experiments are presented to illustrate the performance of the method for several benchmark problems.

ABSTRACT. This paper investigates the fluid-solid coupling problem in fractured porous elastic media. The geometry of the fractures is explicitly considered as a potentially non-planar interface. The model equations are of mixed-dimensional type, where the flow equations on the d −1 dimensional fracture surfaces are coupled with the d dimensional porous matrix. This paper considers a strongly compressible fluid flow model, where the density is chosen as the primary variable, in contrast to the slightly compressible model discussed by Girault et al. (Girault et al. Mathematical Models and Methods in Applied Sciences. Vol. 25, No. 4 (2015) 587645), which takes pressure as the primary variable. We derive a thermodynamically consistent mathematical model and present its weak formulation. Energy stability is established for both the continuous and semi-discrete (in time) cases. The proposed model and numerical framework provide a solid foundation for simulating strongly compressible flows while maintaining thermodynamic consistency and stability.

ABSTRACT. Recent advancements in reinforcement learning (RL) have highlighted its potential for active control tasks in domains like fluid dynamics and energy systems. This study investigates the application of model-based reinforcement learning (MBRL) to active flow control (AFC), focusing on two complex problems: drag reduction around a cylinder and subsurface reservoir production optimization. Two MBRL algorithms, probabilistic ensembles with trajectory sampling (PETS) and model-based policy optimization (MBPO), are adapted and compared. PETS employs a model predictive control framework using the cross-entropy method for action optimization, while MBPO utilizes a soft actor-critic framework for policy optimization based on model-generated rollouts. Our experiments demonstrate that both MBRL algorithms achieve performance comparable to the state-of-the-art model-free proximal policy optimization (PPO) algorithm while significantly improving data efficiency—requiring only 10% and 30% of the simulation data for the drag reduction and reservoir production tasks, respectively. Ablation studies reveal distinct algorithmic behaviors: PETS exhibits robust hyperparameter insensitivity but is prone to local optima in complex environments, whereas MBPO consistently discovers global optima but requires more meticulous hyperparameter tuning. These insights underscore the promise of MBRL for AFC applications and offer practical guidelines for deploying these techniques in data-intensive physical systems.

ABSTRACT. This paper presents a novel finite element algorithm for solving quasi-static Biot poroelasticity model. By introducing a total pressure variable, we reformulate the Biot system into a coupled Stokes-parabolic system. To efficiently solve this system, we propose a split parallel approach. The system is decomposed into a Stokes subproblem and a parabolic subproblem. These subproblems are then solved in parallel using a stabilization technique. This parallel approach eliminates the need for sequential or iterative decoupling, significantly reducing computational time. The algorithm is proven to be unconditionally stable, and these theoretical results are validated through numerical experiments.

ABSTRACT. This study employs a combined approach of molecular dynamics (MD) simulation and classical density functional theory (cDFT) to investigate hydrogen adsorption on graphite surfaces. By leveraging MD parameters in cDFT calculations, we derive an effective wall-potential for hydrogen-graphite interactions. Our methodology demonstrates excellent agreement between the two techniques, accurately capturing enthalpy and entropy of adsorption. Furthermore, by analyzing MD simulation trajectories using the small system method, we directly calculate the thermodynamic correction factor of the surfaces. The results provide valuable insights into the development of efficient graphite-based materials for hydrogen storage and transport applications, enabling the design of optimized systems for sustainable energy solutions.

ABSTRACT. Solving recursive equations through iteration can be a computationally expensive endeavour, and the time required to reach an optimal solution delays the progress of any dependent processes. To address this issue for a specific use case of decision-making in an agent-based model, a method of replacing the iterative function used in said model, a Bellman equation, with a feedforward multilayer perceptron was developed. A hyperparameter grid search was performed to determine the combination of architecture, learning rate, and batch size which produced results deviating the least from those of the original iterative method. With the resulting neural network, accepting four inputs and yielding two outputs, the time required to compute outputs scales sublinearly with the number of agents. Excluding training time, for a set of 1,000 agents, the selected neural network produces output at over 66,000 times the speed of the original function. It achieves this acceleration while maintaining a 99.3% accuracy in adaptation strategy selection and 0.10 mean absolute error in consumption, leading to its ready adoption as an acceptable replacement for the original method.

ABSTRACT. In agent-based models (ABMs), traditional statistical inference faces challenges due to intractable likelihoods and computational costs. This study evaluates neural posterior estimation (NPE) and neural ratio estimation (NRE) for parameter inference in ABMs and compares them with approximate Bayesian computation (ABC). NPE and NRE are argued to be more efficient than traditional methods such as ABC and circumvent some of their limitations. The assessment of the methods focuses on the satisfaction threshold in Schelling's model of residential segregation, including regions of high variance and non-equilibrium dynamics. As these simulation-based methods still require summary statistics as high-level descriptions of the ABM, we propose a general approach to construct them based on spatial and/or temporal information and evaluate how the different summary statistics affect performance. Both NPE and NRE generally outperform ABC regardless of summary statistics. Most notably, NRE excels when employing the most detailed spatio-temporal information, but adding spatial or temporal information alone is not always beneficial for NPE, NRE and ABC. This holds true for different training budgets and when estimating multiple parameters. Hence, the study underscores the importance of spatio-temporal information for accurate parameter inference in this ABM, but information redundancy can degrade performance as well. Therefore, finding optimal high-level descriptions to capture fundamental emergent patterns in the model through summary statistics might prove crucial in cases where the systems are governed by more complex behaviour.

ABSTRACT. While emergent communication in artificial agents has been widely studied, interactions between previously separated populations remain underexplored, despite their real-world relevance. Our aim is to build a model of two pre-learned populations that meet and attempt to communicate. We develop an agent-based language evolution model, where agents are designed to resemble human internal development as closely as possible. These agents participate in ’language games’—atomic, scripted communication scenarios. When merging two pre-learned populations, we observe a significantly higher rate of successful communication compared to training all agents together from the beginning. This effect persists even after extended simulation of the merged population. Our findings suggest that merging pre-learned populations can enhance communication efficiency, offering practical insights for designing collaborative AI systems.

ABSTRACT. The emergence of collective cooperation within competitive environments is well-documented in biology, economics, and social systems. Traditional evolutionary game models primarily investigate the evolution of strategies within fixed games, neglecting the simultaneous coevolution of strategies and the environment. Here, we introduce a game selection model where both the strategies employed by agents and the games themselves evolve dynamically through evolutionary processes. Our results demonstrate that these coevolutionary dynamics foster novel collective phenomena, including changed cooperative interactions. When applied to structured populations, the network's architecture, and agent properties such as risk-aversion and bounded rationality significantly influences outcomes. By exploring the interplay between these factors, our model provides novel insights into the persistent social dilemmas observable in real-world systems.

ABSTRACT. The Coordinate Rotation Digital Computer (CORDIC) algorithm is widely recognized for its fast real-time processing capabilities, making it highly suitable for hardware implementations in diverse applications such as signal processing, high-performance computing, and edge computing devices. Despite its advantages, the traditional CORDIC algorithm’s iterative computational method introduces significant challenges, including a complex structure and high hardware resource consumption, which can limit its efficiency and scalability in certain applications. In this article, we introduce an innovative and efficient variation of the CORDIC algorithm designed to address these challenges. Our proposed algorithm significantly reduces the number of required operations while maintaining computational accuracy, thereby optimizing performance. Furthermore, we demonstrate that this streamlined algorithm can be effectively implemented on Field-Programmable Gate Arrays (FPGAs), leveraging their reconfigurable hardware to achieve enhanced processing speeds and reduced resource utilization. This advancement not only improves the feasibility of using CORDIC in resource-constrained environments but also expands its applicability in modern computing contexts.

ABSTRACT. This paper investigates the time and energy consumption of LU (with MKL library) and WZ multithreaded matrix factorization algorithms on Intel and AMD processors, utilizing OneAPI and Clang compilers. The study evaluates how processor architecture and compiler optimizations impact time and energy use during matrix factorization.

We describe the experimental setup, including hardware specifications, software configurations, and methods for collecting time and energy metrics. Both algorithms are tested under various conditions to assess their suitability for energy-efficient high-performance computing.

The results show variations in execution times and energy consumption based on the processor and compiler used. For LU factorization on Intel Xeon processors, Intel OneAPI optimizations prove most effective, while for WZ factorization on AMD EPYC processors, the Clang compiler demonstrates better performance. Choosing the right compiler options can reduce time and energy consumption by up to 6.5 %.

ABSTRACT. The digital signal processing (DSP) of Internet of Things (IoT) devices using edge-based machine learning (ML) requires fast arithmetic operations and high energy efficiency. Hardware implementations of transcendental functions in ML and deep learning chips are impractical, necessitating effective software algorithms for functions like 1/3 √x across precision levels. This paper analyzes 1/3 √x approximation, introducing algorithms that use magic numbers and piecewise linear approximation to optimize relative error, precision, and computation speed. A mathematical analysis determines key algorithm parameters, with single-precision implementations in C. These algorithms were tested on various hardware platforms and compared for speed, accuracy, and relative errors.

ABSTRACT. In this article, the modified wavenumbers for several numerical approximations of the Caputo fractional derivatives based on the Lagrange interpolation method over uniform mesh are derived. With regard to the modified wavenumbers, the Fourier analysis of differencing errors is performed to quantify the resolution characteristics of these approximations. Numerical experiments are performed, and the plots of original wave numbers against modified wave numbers are presented for each approximation method.