Computing for Sustainable Innovation: 3rd Exascale and Scalable AI Workshop Agenda

Welcome and Introduction

Hartree Centre Overview: Addressing Grand Challenges and Delivering Scientific Innovation, Vassil Alexandrov Hartree Centre

Abstract: Hartree Centre’s overview will be focused on outlining the Centre’s challenge and mission led research and its scientific innovation approaches. In particular, a very short overview of our capabilities how we can address key such challenges by applying latest HPC and AI advances will be given. All these are impossible without wider national and international collaboration, so valuable examples and key areas of collaboration will also be presented.

Bio: Professor Vassil Alexandrov joined the Hartree Centre as Chief Science Officer in 2019, bringing with him a wealth of application-focused research expertise. His expertise and research are primarily in the area of Computational Science, Parallel and High Performance Computing, Scalable Algorithms, Monte Carlo methods and algorithms and emerging computing paradigms. He has co-authored over 140 publications.

Prior to his time at the Hartree Centre, Vassil held positions as ICREA Research Professor in Computational Science at Barcelona Supercomputing Centre (BSC-CNS) from 2010 – 2019, Distinguished Visiting Professor at Monterrey Institute of Technology and Higher Education (ITESM) from 2015 – 2018, and Professor of Computational Science and ACET Centre Director at the University of Reading from 1999 – 2010.

Sue Thorne, STFC Hartree Centre

Erik W. Draeger, Lawrence Livermore National Laboratory (LLNL)

Abstract:  As the power of HPC hardware and AI technology continues to increase, it becomes ever more important to build collaborations and partnerships to jointly tackle common challenges. Lawrence Livermore National Laboratory has long played a leading role in harnessing the world’s largest supercomputers to address important science and engineering challenges, often in close collaboration with colleagues throughout the US Department of Energy.  The recent explosion of innovation in artificial intelligence and machine learning provides a tremendous opportunity to advance our computational capabilities, boost productivity and discover new approaches and tools.  We are eager to work with the broader community to both share our experiences and learn from others.

Bio:  Erik W. Draeger is the Director of the High Performance Computing Innovation Center and RADIUSS project at Lawrence Livermore National Laboratory as well as the Scientific Computing group leader in the Center for Applied Scientific Computing.  Erik earned a Bachelor’s degree in Physics from the University of California, Berkeley in 1995 and received a PhD in theoretical condensed matter physics from the University of Illinois, Urbana-Champaign in 2001. 

The STFC AI Strategy – Adriano Agnello, STFC Hartree Centre and Valerie Farr, STFC

Abstract: The STFC vision is the creation of responsible AI knowledge and solutions at scale and pace, to maximize the value of data across the research and innovation system for growth and prosperity. This presentation will illustrate the STFC AI ecosystem, the STFC AI Strategy core strategic objectives, and key enablers.

This session will offer insights from leading UK research facilities on how national infrastructure is enabling innovation through cutting-edge compute and AI. Attendees will hear how access to these capabilities can accelerate R&D pipelines, open new markets and support strategic decision-making.

Building Digital Twins of the Universe with the DiRAC HPC Facility – Mark Wilkinson, DiRAC

Abstract: Cosmological simulations are essential tools for the interpretation of observations of the Universe and its constituents. In essence, these simulations constitute digital twins, providing the means to explore different scenarios for the evolution of the Universe which can be compared with actual data. I will present examples of state-of-the-art simulations which have been carried out by researchers using the DiRAC HPC Facility. I will also present on-going work to develop AI-based surrogate models to enhance the inclusion of “sub-grid” physics in the next generations of cosmological simulations. Finally, I will discuss the benefits of a co-design, workflow-centred approach to the deployment of large-scale computing systems for both simulations and AI.

Bio: Mark Wilkinson is the national director of the STFC DiRAC HPC Facility (www.dirac.ac.uk), which provides high performance computing resources for the theoretical astrophysics, particle physics, cosmology and nuclear physics communities in the UK. He is a Professor of Astrophysics at the University of Leicester, specialising in the study of dark matter in galaxies using a combination of observations, theoretical models and machine learning. His recent work focusses on the use of machine learning and AI to enhance and accelerate simulations in astrophysics and cosmology. He has published almost 100 peer-reviewed papers and has almost 20,000 citations.

Mark was the editor of the 2019 community-led white paper “UKRI National Supercomputing Roadmap 2019-30” and  chaired the editorial board for the peer-reviewed “UKRI Science case for UK Supercomputing” which was published in 2020. He recently chaired the STFC AI Strategy Development Working Group. He currently co-Chairs the STFC Exascale Working Group and is a member of the UKRI Advisory Group for Digital Research Infrastructure (AGD).

Isambard-AI: a template for sustainable, leadership-class AI supercomputers – Simon McIntosh-Smith, University of Bristol

Abstract: In this talk we’ll introduce the new Isambard-AI supercomputer, recently listed at #11 in the Top500. Isambard-AI delivers nearly 5,500 NVIDIA Grace-Hopper GPUs for under 5MW in a new modular datacentre. The system employs next-generation direct liquid cooling to achieve high energy efficiency, for a typical PUE of under 1.1. The Isambard-AI capability has been delivered in record speed, and we’ll describe how this has been achieved.

Bio: Professor Simon McIntosh-Smith is the founder and Director of the Bristol Centre for Supercomputing, which runs the UK’s Isambard-AI service. He began his career in industry as a microprocessor architect, first at Inmos and STMicro in the 1990s, before co-designing the world’s first fully programmable GPU at Pixelfusion in 1999. In 2002 he co-founded ClearSpeed Technology where, as Director of Architecture and Applications, he co-developed the first modern many-core HPC accelerators. He previously founded the HPC Research Group in Bristol, where his research interests include advanced computer architectures and performance portability.

Federating National Compute and Data – Jon Hays, UKRI

Abstract: There are many benefits to be realised from federation, but it is also not without it challenges and pitfalls. Not least that the word federation means different things to different people and their expectations on why, what, and how it can be delivered are highly varied.

In this presentation we will introduce the National Federated Compute Service Network Plus (NFCS+) project. The primary objective of this project is to build community around large-scale federated computing and produce a roadmap for UKRI to advise and recommend a future direction to achieving a national federated compute (and data) service that will help us realise the benefits while avoid the pitfalls.

Bio: Jonathan Hays is a Professor of Physics at Queen Mary University of London, where he is the head of the Particle Physics Research Centre. He has had a strong involvement in large-scale computing for over two decades as a user, developer, provider and now as Science Director of STFCs IRIS Federation – that coordinates and provides access to large-scale compute and data services across the science remit of STFC. He is Project lead for the National Federated Compute Serivce NetworkPlus Project.

Keynote on USA’s Department of Energy AI programme – Rick Stevens, Argonne

Abstract: In this talk, I outline the concept and strategic importance of an AI Factory for Science. This is a specialized platform designed to industrialize and scale AI model development, deployment, and maintenance to accelerate scientific discovery. I highlight the need for large-scale, optimized AI infrastructure tailored to advanced scientific workloads, emphasizing requirements such as massive GPU resources, integration with high-performance computing, and domain-specific models. I also discuss key applications across scientific fields, including energy, physics, medicine, and space, supported by partnerships and advanced hardware. Finally, I address the transition plan, cost considerations between on-premises and cloud deployments, and the evolving role of AI-enhanced supercomputing in driving future breakthroughs.

Bio: Part quantum theorist, part AI strategist, and part national laboratory futurist, Rick Stevens is a Professor of Computer Science at the University of Chicago and the Associate Laboratory Director of the Computing, Environment and Life Sciences (CELS) Directorate and Argonne Distinguished Fellow at Argonne National Laboratory. His research spans the computational and computer sciences from high-performance computing architecture to the development of tools and methods for bioinformatics, cancer, infectious disease, and other challenges in science and engineering — all in the service of unlocking the next leap in scientific discovery. Moving effortlessly between biology, materials science, supercomputing, and macroeconomics, he spends his weekends camping under the stars… or calculating optimal shielding for a fusion starship.

Putting UK Supercomputing on a long-term footing – Farhat Raza, Department for Science Innovation and Technology

Explore the tangible impact of large-scale computing and AI across sectors. This session will highlight real-world case studies where scalable digital technologies have improved productivity, operational efficiency and innovation for both businesses and public service delivery.

A Trinity of Scale: Exascale Simulation, AI, and the Fusion Moonshot – Rob Akers, UKAEA

At UKAEA, we are tackling a grand challenge: to deliver net power safely and affordably to the grid from magnetically confined fusion — a tightly coupled system-of-systems problem where traditional “build, test, learn” approaches are too slow, costly, and risky. There is neither time, money, nor political appetite to learn from the “rapid unscheduled disassembly” of a tokamak. Instead, we must expose emergent phenomena and pre-empt Black Swan failure modes through the transformative confluence of high-performance computing and AI at the exascale. Our approach is to design in silico, intelligently combining high-fidelity simulations (predominantly open tools) with AI methods, surrogate models, and lower-order solvers. This talk will explore how scalable compute and AI are helping reduce cost and tame complexity in fusion powerplant design — and preview how UKAEA’s fusion “Moonshot” is positioned at the heart of a second giant leap: the UK’s first AI Growth Zone, designed to ensure the UK is an AI maker, not an AI taker.

Discovery of high entropy ceramics for extreme environments – Richard White, Lucideon

Lucideon is developing novel ceramic based materials for extreme environments, in particular extreme temperature, high entropy ceramics, where the number of potential compositions far exceeds an experimental programme of fabrication and evaluation.

The Need of Compute and AI to tackle the NHS’s Increased Demand – Janet King, NHS England NW Digital Transformations Director

Focusing on the future of mobility, this session will demonstrate how advanced simulation and AI tools are being used to optimise battery performance, accelerate design cycles and maintain a competitive advantage in the fast-evolving electric vehicle landscape.

Accelerating Battery Development with high-fidelity microstructure and particle scale models: through manufacturing to performance – Francois Usseglio-Viretta, NREL 

Linking atomistic simulations on supercomputers to larger scales for designing better batteries – Chris Skylaris, University of Southampton and Faraday Institution

Software tools for the next-generation digital twin of batteries for exascale machines – Karthik Chockalingam, STFC Hartree Centre

Improving Battery Safety with AI-Powered Battery Management – Brain Smith, Eatron Technologies

The government recently shared that ‘successful deployment of fusion energy would be globally transformative and allow the UK to export the technology to a global fusion market expected to be worth trillions of pounds in the future.’ In this session, we will discuss how cutting-edge technologies are essential to building comprehensive alternative energy sources.

Generative AI for Inertial Fusion Energy Science – Vadim Elisseev, IBM

Fusion energy research has long captured the public imagination for its applications to fundamental physics, material sciences, and as a clean, low-carbon source of electricity. Recent breakthroughs at the National Ignition Facility (NIF), where lasers were used to initiate nuclear fusion in hydrogen isotopes, mark significant progress toward realizing inertial fusion energy (IFE). However, achieving sustainable IFE remains a formidable challenge due to the vast and complex parameter space involved in optimizing conditions for thermonuclear ignition. To address this complexity, we investigate the application of generative artificial intelligence (Generative AI) methodologies, focusing on how the integration of large language models (LLMs), and deep reinforcement learning (DRL) can accelerate the autonomous discovery of optimal IFE configurations. We present our ongoing research into how these AI techniques can enable more efficient exploration, control, and optimization of fusion experiments, potentially transforming the path toward viable fusion energy.

Fusion Innovation at Exascale: The need for high-fidelity modelling for real-world prototypes – Gurdeep Singh Kamal, Tokamak Energy

Tokamak Energy is pioneering the development of fusion power through advanced spherical tokamaks and high-temperature superconducting (HTS) magnet technologies. As we push the boundaries of plasma performance and magnet design, the demand for high-resolution, multi-physics modelling and simulation grows exponentially. This talk will explore how Machine Learning and Exascale Computing can unlock new frontiers in fusion innovation. We will present our current modelling capabilities, focusing on integrated high-fidelity simulations. model driven experimental design and whole system modelling needs. By utilising exascale computing, we aim to take advantage of the enhanced predictive power and speed to allow for faster iteration, deeper insight, and ultimately, a shorter path to commercial fusion.

Digital Engineering for Fusion Energy – Andrew Davis, UKAEA 

Fusion energy systems present some of the most demanding engineering challenges of our time, requiring robust performance under extreme thermal, nuclear, mechanical, and electromagnetic loading. As the design of pilot-scale fusion devices accelerates, multiphysics driven simulation is playing a central role in enabling predictive, integrated, and agile engineering workflows. This talk explores how digital engineering—anchored by high-fidelity simulations, advanced models, and multi-scale, multi-physics coupling—is transforming the design and qualification of fusion structural components. Drawing on examples from national and international fusion programs, we demonstrate how digital workflows are helping to reduce design iteration cycles, improve system reliability, and bridge gaps between simulation and experiment. The presentation will also highlight open research questions and the role of cross-disciplinary collaboration in advancing the digital infrastructure needed to support timely delivery of pilot-scale fusion reactors.

Artificial Intelligence to accelerate development of sustainable energy solutions – Jonathan Booth, STFC Hartree Centre

In the development of solutions for sustainable energy, AI can accelerate or bypass expensive simulations, and aid the training of advanced control and optimisation agents. Here, we show a selection of examples mostly focused on fusion energy, and perspectives and requirements for scalable computing paradigms.

 AI for Large-Scale Experimental Facilities – Jeyan Thiyagalingam, STFC

Sustainability Challenges for Future Computing – Jim Sexton, IBM

Abstract: The Semiconductor Sector is facing an unprecedented set of challenges in the next few years as new computing capabilities, particularly in AI, are driving demand for computing capabilities beyond anything previously imagined. The impact of power demands to drive those capabilities, and the sustainability of the overall production and deployment of computing is enormously challenging for the sector. In this presentation, we discuss the complex interlocking challenges which are arising and suggest opportunities to address these challenges through innovation in the underlying technologies that underpin future computing approaches.

Bio: Dr. James Sexton is an IBM Fellow, Future Computing Systems, at IBM Research Europe, Dublin Laboratory.  Dr. Sexton received his Ph.D. in Theoretical Physics from Columbia University, NY. His areas of interest lie in High Performance Computing, Computational Science, Applied Mathematics and Analytics.  Prior to joining IBM, Dr. Sexton held appointments as Professor at Trinity College Dublin, as postdoctoral fellow at IBM T. J. Watson Research Center, at the Institute for Advanced Study at Princeton and at Fermi National Accelerator Laboratory.

As computing plays an increasingly significant role in an organisation’s environmental footprint, adopting green computing practices can meaningfully reduce IT-related emissions. This session will explore how sustainable computing strategies can support organisations in reaching their wider carbon reduction targets, highlighting practical steps and technologies that can drive greener, more energy-efficient digital operations.

Leveraging custom hardware with LBANN – Brian van Essen, LLNL 

Abstract:  In this talk, we will describe the unique architecture of the El Capitan HPC supercomputer as well as some of the innovative optimizations that we have developed for AI workloads targeting this system.  This includes a discussion of the new version of the LBANN 2.0 HPC-centric deep learning toolkit that is optimized for El Capitan and is integrated into PyTorch 2.x so that the latest state of the art community algorithms can be combined with custom algorithms that are designed for large-scale scientific AI workloads.

Bio:  Brian Van Essen is a Senior Computer Scientist at the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory (LLNL) and a member of the Livermore Computing Advanced Technologies Office (LC ATO). Dr. Van Essen joined LLNL in October of 2010 after earning his Ph.D. in Computer Science and Engineering from the University of Washington in Seattle. He also holds a M.S in Computer Science and Engineering from the University of Washington, a M.S in Electrical and Computer Engineering from Carnegie Mellon University, and a B.S. in Electrical and Computer Engineering from Carnegie Mellon University.

Automating Research Software Engineering with AI – David Beckingsale, LLNL                      

Abstract: Large language model (LLM) coding assistants—such as Anthropic’s Claude Code and GitHub Copilot—are emerging as valuable partners in research software engineering tasks. This includes things like improving build systems, adding continuous integration (CI) pipelines, and generating documentation: often repetitive and formulaic, yet traditionally tricky to automate with standard scripts. This talk will explore how LLM-based coding agents can bridge that gap. We will share real examples of using Claude Code and Copilot to automate build configuration, CI testing setups, and documentation generation. Early results show that with expert oversight, these tools can handle routine development chores, allowing RSEs to focus on more complex, high-value problems. The presentation will also include qualitative feedback from RSEs who have used these tools, highlighting benefits and lessons learned in practice.

Bio: David Beckingsale is a computer scientist at Lawrence Livermore National Laboratory (LLNL) and serves as Group Leader of the Research Software Engineering (RSE) group within LLNL’s Center for Applied Scientific Computing. His work in computer science research spans performance-portable programming models and HPC performance optimization, and he leads the development of the Umpire memory management library for multi-architecture HPC systems. David also serves as Deputy Director of LLNL’s High Performance Computing Innovation Center (HPCIC) and its RADIUSS open-source software initiative, where he drives efforts to make HPC software development more accessible and collaborative.

Emerging innovative AI techniques for sustainable HPC energy use – Fawada Qaiser, Durham University

Sustainability and scientific computing – David McDonagh, STFC Scientific Computing

Overview of Hartree Centre projects on sustainability – Sarah Hanrahan STFC Hartree Centre

Material Science in the 21st Century has high impact on sustainability and efficiency of products. Through innovation in ecology friendly materials and processes, it drives economic growth and reduces the environmental impact of industry.

ElMerFold: Exascale Distillation Workflows for Protein Structure Prediction on El Capitan – Nikoli Dryden, LLNL

Abstract: Understanding the structure of proteins is critical to many biological applications. Recent deep learning approaches, such as AlphaFold, make high-quality structure predictions. However, training state-of-the-art models requires producing distillation datasets, which requires several times more compute than training. We describe ElMerFold, an optimized distillation data generation workflow that scales to 10,800 nodes (43,200 APUs) on El Capitan, by leveraging a combination of inference, workflow, I/O, and resiliency innovations. We demonstrate this for the OpenFold 3 monomer protein distillation dataset, which uses a complex workflow centered on a set of pretrained OpenFold 2 models to predict structures for ~41 million proteins. ElMerFold achieves a 5.7× speedup over the state-of-the-art. As distillation and synthetic data become more common, these workflows will help drive scientific advancement.

Bio: Nikoli Dryden is a research scientist in the Informatics Group of the Center for Applied Scientific Computing (CASC) at Lawrence Livermore National Laboratory (LLNL). Previously, he was an ETH Postdoctoral Fellow at ETH Zurich, working with Professor Torsten Hoefler in the Scalable Parallel Computing Laboratory. He completed his PhD in computer science at the University of Illinois Urbana-Champaign, advised by Professor Marc Snir. His research focuses on the intersection of high-performance computing and machine learning.

Autonomous Multiscale Simulations – Embedded Machine Learning for Smart Simulations – Peer-Timo Bremer, LLNL

Abstract:  Many, if not all, of our mission-critical application codes depend on various subscale models to address issues not covered by the primary simulation. Depending on the fidelity and costs required to derive reliable answers, these subscale models can quickly dominate the runtime of a simulation. A common and effective approach is to replace such models with machine learning surrogates, which are much cheaper to evaluate and, if trained sufficiently, can theoretically provide similar results. In practice, the main challenge is ensuring sufficient training on the available data and, perhaps even more challenging, guaranteeing that the training data adequately covers all future simulations a user might need. The lack of solutions to this problem has so far hindered the use of data-driven surrogates for high-impact applications.

The Autonomous Multiscale Simulation (AMS) project addresses these challenges by reformulating the traditional static workflow of collecting data, training models, and deploying them into a dynamic process with continuously improving models. Each model evaluation begins with an uncertainty quantification of the corresponding query; if the model is deemed insufficiently reliable, the original subscale model is invoked. This process implicitly generates new potential training data, which AMS collects, filters, and ultimately uses to enhance the model as part of independent computing jobs not directly linked to the main simulation. In this manner, AMS establishes the first reliable approach to integrating data-driven modeling into mission codes, potentially leading to transformative changes in both the speed and fidelity of large-scale simulations. AMS is implemented as a modular library using the latest software principles and is already integrated into the ALE3D and MARBL codes. We have successfully demonstrated simulations of shock-driven pore collapse in TATB using ALE3D, comparing it with the in house Cheetah code, and replacing Cheetah with a learned surrogate through AMS.

Bio:  Peer-Timo Bremer received his Diploma in Mathematics and Computer Science at the Leibniz University in Hannover Germany in 2000. Subsequently, he enrolled at UC Davis and was awarded the SEGRF Fellowship at LLNL in 2002. Working closely with researchers in CASC he received a PhD in Computer Science from UC Davis in 2004. After two years as a post-doc at the University of Illinois, Urbana-Champaign he returned to Livermore as full-time staff in 2006. Timo currently serves as the Associate Director of Livermore’s AI Innovation Incubator, as group leader of the Machine Intelligence group in CASC, as the data science POC for DOE ASCR, and on the research council of the Data Science Institute. He also holds a shared appointment as the Associate Director of research at the Center for Extreme Data Management Analysis and Visualization at the University of Utah. His research covers large scale data analysis, visualization, topological techniques, scientific machine learning, and augmented reality.

Multi-scale modelling – An Industrial Perspective – Misbah Sarwar, Johnson Matthey

Multi-scale simulations, ranging from electronic structure to continuum-based approaches have become embedded in product development cycles, innovating the way in which new products are developed. The talk will give an overview of how multi-scale modelling combined with advanced characterization techniques are being used in industry to understand the structure and activity of catalytic materials that are used to accelerate the transition to net zero. The talk will also discuss how newly developed approaches such as MLIPs might fit into such a workflow and be used to accelerate the catalyst discovery process.

AI and Quantum Computing for Microbiome Data – Ruediger Zillmer, Unilever

Abstract: The skin microbiome is vital to skin health, yet its complexity challenges traditional analysis. We present an AI-driven approach using categorical embeddings and neural networks to analyse microbiome responses to skincare treatments, revealing patterns aligned with clinical measures. Further, we illustrate the potential of quantum machine learning to uncover patterns and interactions that are otherwise computationally intractable.

Agentic Frameworks for autonomous discoveries – Michail Smyrnakis, STFC Hartree Centre

This presentation explores agentic frameworks—AI systems capable of autonomous decision-applied across diverse scientific and simulation domains. We highlight how large language models and multi-agent systems can accelerating discovery in materials science and drug development, while enabling strategic reasoning in game environments and adaptive modelling for pandemics.

Discover how organisations are preparing for next-generation digital demands. This session will showcase scalable AI and exascale-ready solutions that can support long-term growth, resilience and agility across a wide range of industries.

Pannel Discussion – moderator Vassil Alexandrov, STFC Hartree Centre

Panel members: Mark Wilkinson, Brian van Essen (LLNL), Dave Brains (IBM), Timo Bremer (LLNL), Rob Akers (UKAEA).

Roadmapping – way forward discussion, moderator Erik Draeger (LLNL)