Mesoscopic simulations are useful for predicting properties of supercapacitors thanks to their ability to bridge the gap between atomistic detail and device-scale behavior, enabling the study of ion transport, charge storage, and pore-scale phenomena. In particular, mesoscopic lattice-gas simulations can access electrode length scales that remain outside the reach of fully atomistic molecular dynamics (MD) simulations due to the high computational costs of the latter. However, mesoscopic simulations require pore-resolved microscopic data that typically depend on pore size and electrolyte composition. This work presents a practical coupling between LAMMPS, a classical MD code with a focus on materials modeling, and the lattice-gas code LPC3D, in which LPC3D acts as the driver of an ensemble of independent LAMMPS simulations, one for each pore size. The coupling is implemented in Python by combining mpi4py based communicator splitting with the LAMMPS Python interface, enabling concurrent MD executions within a single MPI allocation. In order to parameterize pore-dependent fields for large-scale simulations of heterogeneous porous electrodes, post-processing procedures extract pore-resolved equilibrium number densities from MD trajectories and transfer these quantities to LPC3D. We describe deployment on high performance computing (HPC) systems in CPU-only mode (IntelMPI) and in GPU-accelerated mode (IntelMPI with Kokkos and NVIDIA backends), and we report the strong and weak scaling behaviors of the coupled workflow on multi node architectures.
The long-term safety assessment of ONKALO®, the world’s first deep geological repository for nuclear waste at Olkiluoto (Finland), requires predicting the hydro-geochemical evolution of a fractured crystalline bedrock over thousands of years. This is done by using a comprehensive/holistic hydro-geochemical reactive transport modelling methodology to add new concepts (e.g., cation exchange, cement leachates) and improve other processes in existing models (e.g., anion exclusion, matrix diffusion, surface hydrology). A significant bottleneck is the Nernst-Planck (NP) formulation used to represent multi-component diffusion processes like anion exclusion. In particular, the NP approach is numerically stiff in large-scale models with approximately 50 million degrees of freedom. This paper presents a cation exchange (CE) formulation equivalent to the NP equations that enables anion exclusion description and reduces significantly the computational cost of large-scale 3D conservative and reactive transport simulations. It also includes an analysis of the scalability of PFLOTRAN on the supercomputer LUMI for production simulations, and the study of the palaeo-hydro-geochemical evolution of the Olkiluoto site using both NP and CE approaches. The results demonstrate that a CE model can be tuned to match the NP formulation, providing a practical tool for further modelling the hydro-geochemical evolution of Olkiluoto.
Long-range electrostatic interactions constitute a major computational bottleneck in particle-based simulations, particularly when periodic boundary conditions are employed. Mesh-based Ewald methods, such as the particle-particle particle-mesh (P3M) approach, reduce the computational complexity to O(N log N) but introduce challenges related to parallel scalability, communication overhead, and efficient utilization of modern heterogeneous architectures. While highly optimized implementations exist in established molecular dynamics packages, these are typically tightly coupled to full simulation frameworks, limiting their flexibility and reuse in emerging HPC applications.In this work, we present a performance-portable library for electrostatic solvers, providing implementations of both classical Ewald and P3M methods targeting CPU and GPU architectures. The library is designed to enable efficient execution on multi-GPU systems while maintaining portability across heterogeneous platforms. We investigate scaling and performance characteristics of the P3M method, with particular focus on the FFT-based long-range component, which is known to limit scalability. Strong-scaling experiments demonstrate that small problem sizes are dominated by communication overhead, leading to reduced parallel efficiency, whereas larger systems achieve near-ideal scaling over a wide range of GPU counts.Our results highlight the importance of balancing computation and communication in mesh-based electrostatics and demonstrate that portable, decoupled implementations can achieve competitive performance on modern HPC systems. The presented approach facilitates integration into diverse simulation workflows and provides a foundation for further optimization of long-range solvers in exascale computing environments.
This paper presents the current state of the European High Performance Computing Joint Undertaking (EuroHPC JU) infrastructure and its evolving access ecosystem supporting High-Performance Computing (HPC), Quantum Computing and Artificial Intelligence (AI) across Europe. It provides an overview of the EuroHPC supercomputing ecosystem, highlighting recent deployments, upgrades and the introduction of Europe's first operational exascale and hybrid quantum-HPC systems. The paper further reviews the EuroHPC JU access framework, including traditional HPC Access Calls, AI Factory Access Calls and the newly launched Quantum Access Pilot Call, outlining their objectives, evaluation procedures and target user communities. Based on proposal and allocation data collected between 2021 and May 2026, the paper analyses proposal submissions, success rates, awarded computational resources, research domains distribution, Principal Investigator demographics and geographical participation across the different access calls. The results illustrate the sustained growth in demand for EuroHPC resources, particularly for AI-driven applications, and demonstrate how the introduction of AI Factory access mechanisms and the Quantum Access Pilot Call are expanding support for scientific research and industrial innovation, supporting academia, startups, SMEs, commercial companies and the public sector. The presented analysis provides a comprehensive overview of the evolution of the EuroHPC JU access framework and its role in enabling European scientific excellence, technological innovation and digital sovereignty through a world-class computing infrastructure.
Many problems in soft matter research involve particles suspended in a solvent, where hydrodynamic interactions play a crucial role. A well-established simulation approach for such systems combines molecular dynamics for the particles with the lattice-Boltzmann method for the solvent. We present the coupling of the coarse-grained molecular dynamics code ESPResSo to the waLBerla library, a high-performance framework for lattice-Boltzmann and other stencil-based methods. This coupling code, developed in the context of the EuroHPC Centre of Excellence MultiXscale, enables large-scale, multi-GPU simulations of coupled particle–fluid systems within ESPResSo. Beyond hydrodynamics, the integration also provides a diffusion–advection–reaction solver coupled to electrostatics and lattice-Boltzmann for electrokinetic simulations. A key advantage of using waLBerla is its code generation infrastructure, which facilitates both the adaptation of algorithms and their optimisation for different hardware architectures. We demonstrate the scalability of coupled simulations with multi-GPU benchmarks on MareNostrum 5.
Radiative transfer is a fundamental component of weather and climate modeling, often accounting for a significant portion of total simulation time in Earth System Models (ESMs). RTE+RRTMGP-CPP is a C++ frontend as well as a CUDA-based implementation of the Radiative Transfer for Energetics (RTE) solver and the Rapid Radiative Transfer Model (RRTMGP), both widely used to compute longwave and shortwave radiative transfer in vertical columns of atmospheric models. However, despite existing GPU acceleration, high computational intensity remains a bottleneck for scalability on modern supercomputers.In this work, we present a scalable, mixed-precision implementation of RTE-RRTMGP-CPP. Driven by recent advances in machine learning, GPU manufacturers have integrated specialized low-precision hardware units offering significantly higher throughput and energy efficiency than traditional 64-bit floating-point numbers. By selectively integrating these low-precision data types and arithmetic in performance-critical GPU kernels of RTE-RRTMGP-CPP, we improve computational efficiency while maintaining scientific accuracy within an acceptable tolerance. Using automatic precision tuning, we are able to find multiple configurations of the application that provide Pareto-optimal trade-offs between numerical accuracy and execution speed.Our results show that mixed-precision tuning accelerates the radiative transfer solver, achieving speedups of up to 2.1 × on Leonardo (NVIDIA A100) and 2.2 × on LUMI (AMD MI250X) compared to double precision and carefully selected configurations offer attractive performance-accuracy trade-offs compared to uniform single precision. Overall, this work demonstrates how to exploit the low-precision capabilities of modern GPUs and prepare weather and climate models for upcoming exascale systems.
Third-party large language models often lack the domain expertise, language coverage, and regulatory compliance that EU institutions require. We performed continual pretraining of Mixtral Mixture-of-Experts 8 × 7B model on over 100 billion tokens across all 24 official EU languages from EURAMIS, the European Commission’s translation support system. To mitigate catastrophic forgetting, we applied hierarchical data packaging with token density normalization, integrated replay data using language-specific ratios, and employed warm-up cycles with adaptive learning rate scheduling. We evaluated on standard benchmarks, a custom EU Formal Language sentence completion task, and toxicity probes. The results revealed trade-offs between multilingual adaptation and high-resource language preservation. Replay mechanisms partially mitigate degradation. This paper provides empirically grounded guidance for practitioners developing multilingual LLMs for public sector applications using European HPC infrastructure.
Uncertainty in the terrestrial carbon cycle remains a major constraint in climate projections, partly driven by the uncertainties affecting the land surface representation and variability in Earth system models. To address this limitation, we present a data-driven framework —AI4Land— for generating high-resolution historical reconstructions and future projections of key land surface variables. The framework follows a two-phase approach using a U-Net architecture. In the first phase, which is the focus of this work, it reconstructs annual land use and land cover by integrating coarse-resolution scenario data with static geophysical features. In a planned second phase, the resulting high-resolution maps will be used to predict dynamic biophysical variables, particularly leaf area index, at finer temporal scales. Trained on Earth observation data, the models learn to reproduce spatially explicit and physically consistent land surface patterns, extending temporal coverage to periods lacking direct observations. AI4Land was developed and trained on MareNostrum5, demonstrating how GPU-accelerated HPC infrastructure enables global-scale climate AI pipelines. The final product is a suite of open-source emulators designed for real-time coupling with digital twin platforms, such as those developed under the Destination Earth initiative. By delivering realistic and evolving land surface conditions on demand, this work aims to reduce critical uncertainties and improve the predictive power of next-generation climate simulations.
Density functional theory (DFT), combined with the plane-waves (PW) representation, forms the basis of a large class of first-principles materials simulation codes. In these applications, Fast Fourier Transforms (FFTs) constitute the core computational kernel that strongly impacts overall performance. In the Quantum ESPRESSO suite (QE), an open-source code widely used for material design, FFTs are tightly integrated with the application’s parallelization strategy and data distribution. This coupling enables efficient application-aware optimization, although it makes direct benchmarking against standard distributed FFT libraries nontrivial. However, the emergence of highly optimized distributed FFT APIs such as cuFFTMp, together with topology-aware communication libraries such as NCCL, motivates a reassessment of this application-specific implementation to evaluate its performance against current HPC standards and identify optimization opportunities on modern architectures.In this work, we analyze and benchmark FFTXlib, the distributed FFT library used in QE, against NVIDIA’s cuFFTMp on the pre-exascale systems Leonardo Booster and MareNostrum5 ACC. From this study, we derive insights that can inform future optimizations and guide new developments of PW-DFT codebases. We further introduce an optimization based on an adaptive batching strategy that relies on the specific data distribution of FFTXlib and significantly improves the overall strong-scaling behaviour. We demonstrate its impact through production-scale benchmarks of representative workloads (a 1413-atom water system and a 1532-atom carbon nanotube system) on three top-tier EuroHPC platforms, including NVIDIA and AMD accelerators (Leonardo Booster, MareNostrum5 ACC, LUMI-G).
Differential expression analysis traditionally relies on normalization assumptions that are difficult to verify in practice. The propr R package addresses this through differential proportionality, that is, the analysis of ratios of gene expression rather than normalized counts, which reflects mRNA stoichiometry directly and without normalization assumptions. However, the approach carries a significant computational cost that has limited its applicability to large datasets. Here we present a substantially redesigned version of propr, featuring a modernized software architecture, package-wide performance improvements, and GPU acceleration of core computational components. These advances reduce runtime approximately 20-fold for large bulk RNA-seq data and 400-fold for single-cell data, enabling rapid, iterative data exploration and bringing large-scale single-cell atlases within practical reach of ratio-based differential expression analysis.
Large-scale traffic simulation is increasingly relevant for urban and regional mobility analysis, particularly in scenarios that require extensive routing computations, repeated executions, and substantial data generation. In such settings, distributed and high-performance computing resources are important not only for reducing execution time, but also for enabling simulation campaigns of practical scale. This work builds on our previous efforts in distributed route computation and focuses on broader large-scale simulation workloads with Ruth, a mesoscopic traffic simulator designed for large-scale scenario execution and mobility data generation. The methodology relies on distributed simulations involving over 900,000 vehicles on regional road networks exceeding 190,000 nodes, together with runtime observations from different communication backends and execution modes. We present execution results obtained on Karolina and discuss simulator-side and runtime-level optimisations supported through EuroHPC access to Karolina and LUMI, which enabled larger workloads and more efficient simulations. The work also considers a dedicated front-end connected to the LEXIS Platform to simplify access for non-expert users, as well as exploratory integration with external decision-support environments such as FLOREON+, including support for scenario visualisation and analysis. This work extends the use of Ruth from kernel-level distributed execution towards broader urban and regional simulation. The findings highlight practical aspects of execution, access, and data generation in distributed environments, while supporting ongoing optimisation efforts and larger future scenarios, including nation-wide traffic simulation studies.
We present a GPU-accelerated parallel domain decomposition framework for the two-dimensional extension of the angiogenesis PDE system, which governs the spatiotemporal evolution of five interacting biological fields: endothelial cell density, protease and inhibitor concentrations, extracellular matrix density, and oxygen concentration. The model is discretized via the method of lines with second-order central differences in space and integrated in time with a fourth-order Runge-Kutta scheme. The computational domain is partitioned into P = P1 x P2 non-overlapping subdomains, each assigned to one GPU, with inter-GPU communication handled by an asynchronous halo-exchange protocol that overlaps InfiniBand data transfer with local computation. Results confirm a parallel stability constraint ∆t ≤ h2/(4dmax), uniform in P. Numerical experiments confirm second-order spatial accuracy, strong-scaling efficiency and weak-scaling efficiency.
The PRECISION (PRedictive Exploration of Correlation In Specific Interactions and Optimization Networks) project of Dompe’ Farmaceutici (an Italian pharmaceutical company) aims to support the large-scale analysis of protein–ligand binding free energies through molecular dynamics (MD) simulations and high-performance computing (HPC) workflows. Building on the experience gained within the LIGATE project, in which more than 4000 protein–ligand complexes were simulated, this work focuses on overcoming the computational bottlenecks associated with the analysis of large MD datasets. Traditional methods like Molecular Mechanics/Poisson-Boltzmann Surface Area (MM-PBSA) and Molecular Mechanics/Generalized Born Surface Area (MM-GBSA) are often computationally demanding and difficult to scale efficiently for large screening campaigns. To address these limitations, here we present an optimized Binding Free Energy Exscalate (BFEx) workflow for HPC infrastructures by introducing an improved MPI-based parallelization strategy and refactoring the original Bash-based implementation into a modular Python framework. The optimized workflow exploits the computational resources of the Leonardo supercomputer to significantly reduce execution times, decreasing the analysis time from more than 24 hours to approximately 10 minutes for the tested systems. The proposed implementation improves the scalability, maintainability, and portability of the workflow, while also simplifying future developments such as GPU acceleration and the efficient processing of large MD simulation datasets for future applications, including machine learning-based analyses. In addition, the modular and portable design of the optimized workflow facilitates its deployment on different HPC infrastructures beyond the Leonardo supercomputer.
We present LBFAST, a GPU-oriented lattice Boltzmann solver based on a lightweight moment-represented formulation, in which post-collision populations are reconstructed on the fy from a reduced set of moments rather than stored explicitly. This approach significantly lowers the memory footprint, enabling large three-dimensional simulations within the constraints of modern accelerator architectures, where VRAM capacity and bandwidth are critical resources. The method is assessed through standard single-and two-component benchmarks demonstrating good accuracy and stability. Extensive scaling experiments on multi-GPU systems show near-ideal weak scaling up to 512 GPUs and sustained performance across different velocity sets. The combination of reduced memory usage, competitive throughput, and stable energy efficiency makes the proposed formulation a practical route for large-scale lattice Boltzmann simulations on current and emerging HPC platforms.
Automatic text summarization (ATS) plays a vital role in streamlining information retrieval, yet existing multi-document sum-marization (MDS) systems—particularly for low-resource languages like Hindi—often face limitations in sentence selection opti-mization, redundancy control, and coherence maintenance. To bridge these gaps, this study introduces a Binary Grasshopper Opti-mization Algorithm (BGOA) for MDS. The proposed BGOA models the summarization process into binary optimization problem which maximizes a fitness function incorporating coverage, cohesion, and non-redundancy. Experimental evaluations on the DUC-2006 (D1) and DUC-2007 (D2) benchmark datasets demonstrate that BGOA consistently outperforms classical metaheuristics such as BPSO and CSO, as well as recent baselines like MDS, CDI, and SKT5. Statistically, BGOA achieved improvements of up to 10.3% in ROUGE-1, 21.9% in ROUGE-2, and 3.6% in F-score over the nearest baseline on D1, and up to 30.4% in ROUGE-1, 24.3% in ROUGE-2, and 32.3% in F-score on D2. These consistent gains confirm that the BGOA framework provides a robust, scalable, and efficient optimization-based solution for high-quality summarization in both resource-rich and resource-constrained environments.
In contemporary years, high-utility itemset mining (HUIM) has become an important topic of research. Every organization focuses on increasing its profits by maximizing the sales of its products. In top-K HUIM, researchers try to reduce running time and space usage to find important itemsets efficiently. However, some existing algorithms require a high running time to discover the important items because of inefficient pruning techniques and data structures. Therefore, traditional top-k HUIM algorithms struggle with efficiency and scalability. Inside this paper, we have introduced a novel algorithm, FTKEH (Fast Top-K Efficient High utility itemset mining), which utilizes the Red-Black tree data structure. The softmax function is used to raise internal thresholds that normalize utility values and handle them smoothly; therefore, our algorithm is scalable for large-scale datasets. An efficient pruning strategy that focuses on the first quadrant of the used matrix is also presented, which is better than the many pruning strategies used in previous years in terms of computational time complexity and memory consumption.We have taken six real-world benchmark datasets (Retail, supermarket analysis, superstore, Amazon sales report, bike sales, and accident). Experimental results demonstrate that FTKEH shows a marked improvement over the state-of-the-art top-K HUIM algorithms with respect to runtime, scalability, and memory efficiency. Our proposed FTKEH algorithm works much better on sparse datasets and comparatively better on dense datasets as compared to state-of-the-art algorithms.
ASD is a neuro-developmental condition usually found in the early days of childhood. The main characteristics are the challenges in social interactions, emotional regulation, and communication abilities. Early identification is crucial, as prompt intervention can greatly enhance developmental outcomes and overall quality of life. However, traditional diagnostic methods are often subjective and require significant time. In order to aid the detection of autism in the early stage, we have develeped the system, which uses the DL algorithm, VGG16 convolutional neural network (CNN). This will be analyze the facial images, which in return will use the publically available dataset from the source Kaggle. This dataset has images of kids with & without autism. The mouth, nose, and eyes were among the facial features from which important classification features were taken. Our VGG16-based model outperformed other approaches with a classification accuracy of 98.1%. These findings illustrate the value of DL models as precise, practical, and non-invasive instruments for the early detection of ASD. This has significant implications for clinical practice and educational support.
Hyperspectral imaging (HSI) is a powerful imaging technique that captures information over a wide wavelength range. Remote sensing through HSI is the most advanced form of Earth observation, providing rich spectral data. Every material has unique spectral information, called a spectral signature. These spectral data can be used for a detailed analysis of any material on Earth’s surface. The spectral signature contains the reflectance values for every wavelength. Hyperspectral data are highly dimensional; therefore, machine learning (ML) algorithms are used for dimensionality reduction, feature extraction, and classification. This study presents a comprehensive approach to soybean classification using spectral data across 131 wavelengths from 2086 samples. The samples fall into one of six categories: emerge-early, early mid, late, critical, mature, and harvest. Advanced feature extraction techniques such as Gray-Level Co-occurrence Matrix (GLCM), Wavelet Transform, Local Binary Pattern (LBP), and Local Ternary Pattern (LTP) were applied individually and in various combinations to improve the ability of a model to distinguish between different classes. Each feature set was evaluated using six classification algorithms: Decision Tree, K-Nearest Neighbors (KNN), Naive Bayes, neural networks, support vector machines (SVM), and ensemble methods. The models were then assessed based on confusion matrices by calculating the overall classification accuracy, kappa coefficient, and per-class accuracy to determine the optimal feature-classifier pairing. Classification was then performed using the modified Convolutional Neural Network (CNN) deep learning (DL) technique, and the accuracy was evaluated. The results using ML and DL techniques help us understand how to build strong and reliable classification systems for agricultural data.
Hearing loss, commonly referred to as hearing impairment, involves a partial or complete inability to hear. Deaf people find it very difficult to converse with those who are not aware of sign language. Communication between deaf and non-deaf individuals remains a significant challenge. To address this challenge, we utilized two approaches: the first one integrates MediaPipe with an Artificial Neural Network (ANN), and the second one employs a Convolutional Neural Network (CNN). The second approach yields better results. However, in live application, the first approach (MediaPipe and ANN) performs well in a noisy background. With the use of artificial intelligence, we convert American Sign Language into standard English for non-deaf people and vice-versa so that the deaf can also comprehend it. Experimental observations show that he achieved accuracy is 90.5 percentage and 92 percentage by using MediaPipe + ANN and CNN respectively. We also developed a real-time chat application that integrates these models, making everyday communication easier and more inclusive for people with hearing impairments.
The rapid growth of smart city infrastructures has led to an increased demand for secure, efficient, and scalable artificial intelligence (AI) frameworks capable of processing large volumes of heterogeneous and privacy-sensitive data at the network edge. Conventional federated learning (FL) frameworks, although promising, face significant limitations due to high communication overhead, limited computational resources at the edge, and susceptibility to adversarial attacks. To address these challenges, this paper presents AMQFMD:A Multi-Agent Model for Quantum Federated Machine Learning in Distributed Smart City Environments, a novel multi-agent model that combines quantum computing and federated learning to using quantum parallelism and quantum key distribution (QKD) for accelerated and secure training. The proposed framework utilizes a multi-agent architecture to enable adaptive collaboration, resource-aware task allocation, and secure model aggregation. By integrating quantum-enhanced optimization, AMQFMD achieves faster convergence, improved energy efficiency, and enhanced privacy preservation. Experimental evaluations on IoT-enabled smart city datasets, including The MultiIoT Benchmark dataset, demonstrate that AMQFMD consistently outperforms conventional FL and hybrid edge-cloud approaches, achieving an accuracy of up to 98.27%, throughput of up to 1,626 txn/s, and the highest Robust Performance Index (RPI) of 9.15 in healthcare monitoring. These results validate AMQFMD as a scalable and secure solution for next-generation smart city applications, establishing a foundation for integrating quantum computing with edge intelligence in real-world deployments.
