9:00

Matthew Szyndel | SLB

Keynote speech

Hardware agnostic reservoir simulation: enabling our clients to solve fluid flow in porous media efficiently and flexibly

The simulation of fluid flow in porous rock has been of interest to the energy sector since computers were first able to solve a useful problem within a reasonable timescale. As computing performance has increased, so too has the spatial and physical fidelity of the models used. While 40 years ago the focus of these simulations was exclusively on the efficient extraction of hydrocarbons, today we also model carbon storage, lithium extraction, geothermal energy extraction and hydrogen storage.

The Intersect™ reservoir simulator is the flagship reservoir simulator of SLB. Customer data is proprietary and general; we cannot make assumptions about the grid we are simulating at the outset as it may not be shared with us and can take any form we allow. Furthermore, to be competitive, the Intersect simulator needs to make effective use of the high-performance computing hardware that our customers have available and to offer an effective software as a service solution. Upon initial release in 2010, the Intersect simulator was designed from the ground up to be a scalable distributed parallel solution with no underlying assumptions on the grid topology and characteristics. Since that release we have extended the Intersect simulator to use hybrid shared/distributed memory parallelism (2016), and to run as a full graphics processing unit (GPU) simulator (2020), should that be the best option for the customer. Multi-GPU simulation is now our development focus. Our architecture also enables us to experiment quickly and efficiently with alternative accelerator choices.

In this talk we will outline the algorithms and architecture which underpin Intersect. We will discuss the similarities and differences between central processing unit and graphics processing unit implementations and how these implementations are kept aligned. We will present performance results for a range of hardware options and discuss scalability and future options for development.

10:00

Massimo Bernaschi | Institute for Applied Computing, National Research Council of Italy (IAC-CNR)

Performance and Energy Efficiency of BootCMatchGX

BootCMatchGX is a library designed to exploit to their best high-performance and energy-efficient Graphics Processing Unit (GPU) accelerators in solving sparse linear systems with dimensions that far exceed the memory capacity of a single node.

In this talk, we present our implementation design patterns and solutions to problems that could limit the scalability. We complement performance results with energy profiles that address the sustainability requirements of modern HPC platforms.

10:30

Gianluca Carnielli | Mathworks and Imperial College

Enabling High-Performance Scientific Computing with MATLAB and Simulink

Discover how MATLAB and Simulink enable scientists and engineers to harness the power of high-performance computing for scientific discovery and engineering innovation.

This presentation showcases scalable workflows for parallel computing, GPU acceleration, and integration with HPC clusters and cloud platforms. Learn how to accelerate simulations, process large datasets, and deploy complex models efficiently bridging the gap between desktop development and large-scale execution.

11:30

Stefano Regazzo | University of Padova

URANOS-2.0: Exascale-Ready CFD computations for Europe’s Leading GPU Architectures

URANOS-2.0 represents a significant advancement in high-performance CFD, delivering a robust, open-source, and vendor-neutral solver for compressible wall-bounded flows. Optimized for Europe’s leading GPU architectures, it ensures exceptional portability through a hybrid OpenACC/CUDA implementation and demonstrates outstanding scalability up to 1024 GPUs.

This contribution presents a comprehensive performance assessment of URANOS-2.0 and highlights its capability to handle DNS, WRLES, and WMLES with exascale efficiency. Thanks to extensive kernel refactoring, the solver achieves a twofold speedup over its previous version, establishing a new standard for high-fidelity turbulence modeling in aerospace and energy engineering.

12:00

Matthieu Robeyns | CNRS

A parallel approach for solving the Green's function based on the nested-dissection and selected-inverse methods

This talk deals with the parallelization of the solution of the retarded Green's function which is a critical and computationally expensive step in quantum transport applications based on the NEGF method.

Our approach takes advantage of the sparsity of the involved operators and reformulates this operation in the form of a selected-inverse algorithm. We evaluate the use of matrix permutation techniques based on the nested-dissection method to increase parallelism and analyze the associated added computational load resulting from fill-in in order to achieve a favorable trade-off between operational complexity and parallelism.

We finally discuss how the proposed algorithm can be parallelized using a task-based parallel programming model.

14:00

Mario Tacconi | ENI

Composable Hybrid Computing: Integrating Microservices with HPC

As scientific and industrial computing demands evolve, the convergence of artificial intelligence (AI), quantum computing (QC), and high-performance computing (HPC) is becoming essential for enabling next-generation workflows. This project investigates the feasibility of a composable hybrid computing architecture that could integrate AI, QC, and other emerging paradigms, such as neuromorphic processors, into traditional HPC environments. These components, designed as service-oriented architectures (SOA), aim to create a flexible, scalable, and modular ecosystem where AI-driven insights, quantum acceleration, and brain-inspired computing can be dynamically integrated to enhance simulation-based models and solve classically intractable problems.

At the core of this integration is a Kubernetes (K8s) cluster embedded within the HPC infrastructure. This cluster orchestrates microservices that expose the heterogeneous computing backends as computing services, enabling seamless communication between loosely coupled AI/QC/neuromorphic components and tightly coupled simulation workloads. By leveraging containerization, service discovery, and hardware abstraction, the system supports dynamic resource allocation, fault tolerance, and rapid deployment of diverse computational models alongside legacy HPC applications.

The hybrid architecture is designed to harness the unique strengths of each computing paradigm: HPC for large-scale numerical simulations, AI for pattern recognition and surrogate modeling, QC for solving specific classes of problems with exponential speedups, and neuromorphic systems for energy-efficient cognitive tasks. By combining these capabilities, the system enables new approaches to grand challenge problems in fields such as climate and weather modeling, nuclear fusion, materials discovery and complex systems optimization.

This project aims to demonstrate the feasibility and transformative potential of such a hybrid paradigm through a series of domain-specific use cases. It also addresses key challenges including workload orchestration, data interoperability, and infrastructure integration, laying the foundation for a new approach to intelligent, heterogeneous, and composable scientific computing.

14:30

Omar Bettinotti | Dassault Systemes

AI-Powered 3D Surrogate Physics Models: Learning from Historical Data

Estimating 3D physics fields from virtual twin simulations is critical since the very early phases in the design of industrial products. However, exploring multiple design options in such phases can be computationally expensive, especially when using high-fidelity models. Surrogate modeling provides faster alternatives for early decision-making. Machine Learning (ML) techniques have emerged as non-intrusive techniques, performing well in terms of computing times and accuracy for a wide range of design exploration. Within the ML subset, traditional statistical learning techniques interpolate scalar quantities over a design space with acceptable accuracy, whereas more advanced Deep Learning (DL) techniques based on Neural Networks are better positioned to recover the full history of transient quantities or even full 3D fields.

Starting from parametrized design and fully labeled dataset in design exploration, DL techniques have been returning astonishing results. On the other hand, the complex iterative workflow from design to modeling and simulation of complex products and assemblies often makes it impractical and challenging to consider and keep track of parameter dimensions. Therefore, a new paradigm shift is required in order to tackle unlabeled and historical datasets in input. For this accomplishment, self-supervised learning approaches can map geometry to latent representations, addressing the feasibility limits of parametrized surrogate modeling methods.

With this presentation the Authors will share their prototyping efforts and progress for a variety of test cases, ranging from quasi-statics to transient dynamics and multi-physics applications.

15:00

Simone Bna | Cineca

SPUMA: a minimally-invasive GPU porting of OpenFOAM

High Performance Computing (HPC) on hybrid clusters represents a significant opportunity for Computational Fluid Dynamics (CFD) when modern accelerators are utilized effectively. However, despite the widespread adoption of GPUs, programmability remains a challenge, especially in open-source contexts. In this talk, we present SPUMA, a full GPU porting of OpenFOAM targeting NVIDIA and AMD GPUs.

The implementation strategy, based on a portable programming model and on the adoption of a memory pool manager that leverages the unified memory feature of modern GPUs, requires minimal changes in the core of the OpenFOAM source code. Numerical tests on the two pre-exascale clusters in Europe, LUMI and LEONARDO, hosting AMD MI250X and NVIDIA A100 GPUs respectively, will be presented.

16:00

Eric Pascolo | Cineca

Funding and Infrastructure for Innovation: AI Factory, HPC, and European Opportunities for Enterprises and Academia

This talk will explore the evolving landscape of European funding programs that support innovation across both enterprises and academia, with a focus on initiatives such as AI Factory. It will highlight how these opportunities intersect with engineering and artificial intelligence, and how they can be leveraged to accelerate research and development.

The session will also present upcoming infrastructure developments at Cineca, including advancements in high-performance computing (HPC), the integration of AI Factory resources, and the emergence of quantum computing capabilities—positioning Italy at the forefront of European digital innovation.

16:30

Ruggero Poletto | CFD FEA Service

Scalability comparison between Calculix and CodeAster

This talk aims at investigating the Code_Aster and Calculix scalability. In particular, the following aspects will be considered: (i) mesh size; (ii) hardware type; (iii) parallelization methods; (iv) algebraic solvers.

The benchmark case refers to a linear analysis using 3D elements in a mechanical engineering analysis and the objective is providing guidelines to achieve the best scalability configurations.

17:00

Giacomo Baruzzo | Department of Information Engineering, University of Padova

High-Performance Scientific Computing in Veneto: The CONVECS Iniziative

The CONVECS project (COmuNità VEneta per il Calcolo Scientifico) starts in 2025 as an Operation of Strategic Importance of the Veneto Regional Government, which has allocated significant funding under the POR-FESR framework. The project aims to enhance, consolidate, and interconnect the high-performance computing (HPC) infrastructures currently available in the universities and research labs of the national physics institute in the Veneto region to provide a scalable and unified computing environment accessible to both researchers and technology-driven enterprises, with tangible impact on science, industry, and society.

This talk will present the mission, structure, and expected outcomes of CONVECS, highlighting its role in positioning Veneto as a European hub of excellence in scientific computing and innovation. Beyond serving the academic community, CONVECS will empower local companies to collaborate with universities on advanced research projects, fostering the development of next-generation applications in artificial intelligence, digital twins, smart manufacturing, agrifood, health, and energy. By enabling interdisciplinary collaboration and transnational partnerships, CONVECS represents a critical step forward in strengthening the regional innovation ecosystem and supporting the digital transformation of society.

9:00

Stefano Zampini | KAUST

Keynote speech

Device-accelerated solvers with PETSc

Despite the revolutionary impact of GPUs on various computational tasks, challenges still emerge when trying to efficiently utilize these architectures for the solution of large nonlinear systems of partial differential equations, highlighting the need for tailored algorithms.

In this talk, I will discuss recent developments towards exploiting GPUs with the Portable and Extensible Toolkit for Scientific Computing (PETSc), an award-winning software for the numerical solution of large systems of nonlinear and time-dependent partial differential equations.

The focus will center on diverse components, encompassing Krylov methods, direct solvers, preconditioners, multi-GPU communications, distributed sparse matrix-matrix multiplications, and quasi-Newton methods. The objective is to furnish the audience with valuable insights into attainable and realistic performance expectations.

10:00

Gianluigi Rozza | Sissa

Surrogate Modelling by Reduced Order Methods and Machine Learning to enhance Scientific Computing: focus on CFD

Partial differential equations (PDEs) are invaluable tools for modeling complex physical phenomena. However, only a limited number of PDEs can be solved analytically, leaving the majority of them requiring computationally expensive numerical approximations. To address this challenge, reduced order models (ROMs) have emerged as a promising field in computational sciences, offering efficient computational tools for real-time simulations. In recent years, deep learning techniques have played a pivotal role in advancing efficient ROM methods with exceptional generalization capabilities and reduced computational costs [1, 2, 3], especially for parametric settings and turbulent flows.

In this talk we explore how classical ROM techniques can be elevated through the integration of some deep learning models. We will introduce hybrid approaches, which consider both physics-based and purely data-driven techniques [4, 5, 6, 7], as well as aggregated ones, where the model is built as a combination of different pre-trained models [8]. Examples will deal with parametric flows in presence of compressibility and turbulence, as well as industrial applications.

References
[1] Benner, P., Grivet-Talocia, S., Quarteroni, A., Rozza, G., Schilders, W., & Silveira, L. M. (2021). System-and data-driven methods and algorithms (p. 378). De Gruyter.
[2] Benner, P., Schilders, W., Grivet-Talocia, S., Quarteroni, A., Rozza, G., & Miguel Silveira, L. (2020). Model Order Reduction: Volume 2: Snapshot-Based Methods and Algorithms (p. 348). De Gruyter.
[3] Benner, P., Schilders, W., Grivet-Talocia, S., Quarteroni, A., Rozza, G., & Miguel Silveira, L. (2020). Model order reduction: Volume 3: Applications (p. 466). De Gruyter.
[4] Ivagnes, A., Stabile, G., Mola, A., Iliescu, T., & Rozza, G. (2023). Hybrid data-driven closure strategies for reduced order modeling. Applied Mathematics and Computation, 448, 127920
[5] Ivagnes, A., Stabile, G., & Rozza, G. (2024). Parametric Intrusive Reduced Order Models enhanced with Machine Learning Correction Terms. arXiv preprint arXiv:2406.04169.
[6] Zancanaro, M., Mrosek, M., Stabile, G., Othmer, C., & Rozza, G. (2021). Hybrid neural network reduced order modelling for turbulent flows with geometric parameters. Fluids, 6(8), 296.
[7] Stabile, G., Zancanaro, M., & Rozza, G. (2020). Efficient geometrical parametrization for finite‐volume‐based reduced order methods. International journal for numerical methods in engineering, 121(12), 2655-2682.
[8] Ivagnes, A., Tonicello, N., Cinnella, P., & Rozza, G. (2024). Enhancing non-intrusive reduced-order models with space-dependent aggregation methods. Acta Mechanica, 1-30.

10:30

Xavier Trias | Polytechnic University of Catalonia

On the evolution of Poisson solvers for extreme-scale CFD simulations

This work addresses a fundamental question: is the numerical complexity of solving Poisson’s equation increasing or decreasing for very large DNS and LES simulations of incompressible flows? As Reynolds numbers grow, simulations require both finer meshes and smaller time steps. There are two competing effects influencing the convergence of Poisson solvers: namely, the former increases the condition number of the discrete Poisson's equation whereas the latter leads to better initial guesses. Physical and numerical arguments are combined to derive power-law scalings at very high Reynolds numbers.

A theoretical convergence analysis of both Jacobi and multigrid solvers defines a two-dimensional phase space, partitioned into two regions depending on whether the number of solver iterations decreases or increases with the Reynolds number.

These insights can help to define a road-map for the development of Poisson solvers, providing guidance on how CFD codes may need to evolve for future HPC architectures approaching the zetta (and eventually yotta) scale.

11:30

Simone Montangero | University of Padova

Platforms and methods for quantum computing

We provide an overview of the current quantum computing landscape, highlighting some of the most advanced hardware platforms along with a selection of recent results and methods that are driving the technological progress of this potentially disruptive field.

12:00

Carlos Roig | CIMNE

SYCL based acceleration of a Structured CFD Solver

We present a novel structured mesh solver tailored for large scale incompressible external flow simulations, specifically designed for extreme-scale computing.

The method leverages a minimal data structure that eliminates the need to store the mesh explicitly in memory and uses SYCL to bring the possibility of targeting different accelerators. These features are combined with an implicit representation of the analysed geometries via distance functions (i.e., level set) and a boundary condition imposition done with the Shifted Boundary Method (SBM). Thanks to these, the solver is able to handle complex domains while keeping the benefits of a structured mesh representation. Besides, a fractional step numerical scheme is employed aiming to maximize computational efficiency. Specifically, velocity prediction is performed explicitly using a 4th order Runge-Kutta method and pressure correction is solved by solving a Laplacian system using Conjugate Gradients, both highly parallelizable.

14:00

Alfonso Amendola | ENI

High Performance Computing for the Energy Transition: challenges and opportunities

Energy sector has traditionally used high performance computing modelling techniques in key areas such as seismic imaging, geological modelling, and reservoir flow simulation.

In this talk, we highlight new areas of application of high-performance computing in Eni, with practical examples drawn from advanced material simulation and wind resource modelling and optimization, showing that these applications might enhance competitiveness and sustainability of the Energy sector.

14:30

Giulio Monacchi | Saipem

The transformative potential of Machine Learning applied to offshore pipelay

The construction of an offshore pipeline is a complex process that involves specialized vessels operating in challenging marine environments. Thousands of steel pipes are welded onboard and laid from the pipelay vessel down to the seabed, which can be located at depths of 2000 meters or more. The pipeline catenary, hanging from the vessel stinger to the seabed touch-down point, is subjected to significant loading conditions. Ensuring the integrity of the pipeline infrastructure from installation onwards requires sophisticated engineering tools to simulate the pipeline's mechanical behavior during laying and to calculate the stresses involved. Tens or even hundreds of different loading scenarios must be simulated to verify the pipeline's integrity in all possible laying configurations along the route, considering variations in pipe data (diameter, steel and concrete thickness), changes in water depth, all applicable sea states, and imposed loads, which demands significant engineering man-hours.

This presentation provides an overview of the research activities conducted to evaluate how Artificial Intelligence (AI), particularly Machine Learning (ML) techniques, can expedite such calculations in a specific case study, either as an alternative or in support of the well-established Finite Element (FE) physics-driven approach typically adopted. The case study focuses on calculating the local plastic deformation of the pipeline under unexpected loading conditions during offshore laying. A summary of the steps undertaken will be presented. Initially, data exploration was conducted, involving the analysis of available input data and results from previous Finite Element Analysis (FEA) on specific pipelay scenarios. This process included identifying potential correlations between the input data and FEA results, examining data distribution and determining the most effective preprocessing methods for subsequent Machine Learning (ML) applications. Subsequently, the most suitable ML model was selected, designed, and implemented to predict pipe plastic deformation. The model was trained and tested using a synthetic dataset derived from prior FEA simulations. Finally, its performance was assessed by comparing predictions for new, unseen scenarios with results from new FEA simulations.

15:00

Fabio Villa | ESSS

Reducing time to solution in ANSYS: from GPU acceleration to AI solution

Reducing time to solution in ANSYS is critical for accelerating engineering workflows and enhancing productivity, especially in computational fluid dynamic. In this presentation we will explores the advancements solutions proposed by ANSYS.

From the traditional HPC solution to GPU-accelerated. GPU acceleration enhances simulation performance, reduces time to solution, and enables more complex or higher-fidelity analyses. Also the new AI solution introduces a paradigm shift by enabling predictive modeling and surrogate simulations. By integrating machine learning techniques, ANSYS workflows can bypass iterative solvers for certain tasks, achieving near-instantaneous results with acceptable accuracy. This evolution not only reduces computational load but also opens new possibilities for real-time design iteration and optimization. The transition from GPU to AI signifies a transformative step in simulation efficiency.

16:00

Sebastian Gries | Frauhnofer SCAI

Grey-Box Approach to Linear Solvers: Tailored Strategies in Different Practical Applications

While it appears to be a simple linear algebra operation of inverting a matrix, the linear solver part typically accounts for the largest portion of runtime in a numerical simulation. Which methods to choose to design a robust and efficient solution strategy depends on the type of application that you are considering. With further emerging physical complexity in the simulation models, it is hardly possible to simply utilize a standard text-book method if optimal performance and robustness were expected.
With a Grey-Box Approach, we refer to identifying a suitable linear solution method for a particular type of application.
While this requires some initial design-work, it can be used robustly and efficiently in a black-box manner with different simulations of that kind afterwards.

We will discuss this approach with practical examples from different industrial simulation applications. The range of solver strategies will consider application-specific algebraic-multigrid approaches. But we will also demonstrate how adjusted incomplete-factorization-type methods can be successfully applied in applications that do not (yet) require the strength of multigrid.
In all cases, after finding the basic strategy, a Machine-Learning-based parameter optimization can be applied to realize a case-based performance optimization.

16:30

Joel Guerrero | Leonardo

How HPC and physical modeling enables industrial innovation

In today’s digital era, industries such as aerospace, automotive, and energy are transforming how they design and build products, thanks to the power of High-Performance Computing (HPC) and physics-based modeling. This presentation explores how these technologies help companies innovate faster, smarter, and more sustainably.

With HPC, engineers and scientists can run highly complex simulations that once took weeks or were simply intractable due to limited computing power and time, in just a few hours or days. These simulations allow teams to test ideas virtually, explore design alternatives, and solve problems before anything is physically built.
When combined with multi-physics simulations and physics-based modeling, companies can accurately predict how products will perform under various conditions, improving safety, efficiency, and reliability. We’ll also explore how these tools support virtual prototyping and digital twins, and how artificial intelligence (AI) is transforming the way data is consumed and used to guide design decisions.