AI

Strategic Rationale and Evidence Base

Strategic Importance

AI methods are reshaping HEP and RA workflows across simulation, reconstruction, calibration, analysis, and instrument operations: pre-trained models at the LHC trigger classify collision events in real time, AI at telescope edges detects astrophysical transients within milliseconds, and digital twins model detector radiation damage, alignment shifts, and telescope-array calibration drift. ML-based fast simulation has demonstrated two-to-three orders of magnitude computational cost reduction for high-impact HEP and RA use cases whilst maintaining scientific accuracy. The European AI moment is open: the AI Continent Action Plan commits over €20 billion to AI infrastructure, AI Factories enter operations in 2025-2026, AI Gigafactories are proposed at €100 million-plus per site, and the AI Act sets the regulatory baseline. With a Commission-estimated European compute capacity gap of around 20 GW by 2036 and three hyperscalers holding roughly 70% of the European cloud market, sovereign-supply development is a load-bearing dimension of the build-out. HEP and RA must be part of this effort or risk being structurally side-lined; GPU resources are not yet accepted as "pledged" allocations for LHC experiments, exposing a structural gap between AI demand and the resource accounting frameworks that fund scientific computing.

AI agents that plan and execute scientific workflow steps under human oversight (analysis triage, anomaly investigation, calibration response) extend AI integration from training and inference into operational decision-making. This requires P03 service-account access, P06 orchestration, and P09 audit and provenance.

Current State

Europe possesses substantial but fragmented AI infrastructure. EuroHPC has procured supercomputers across Europe, including exascale systems, and 19 AI Factories have been selected across three waves (December 2024, March 2025, October 2025). AI capabilities are not yet integrated as first-class infrastructure across most facilities, which privilege batch-mode access over the elastic processing that production AI workflows require. Access modalities at EuroHPC sites (different AAI stacks, batch-oriented allocation) are not yet aligned with the AI lifecycle. A strategic tension also exists at the hardware layer: industry GPU evolution prioritises low-precision (AI-optimised) computation at the expense of FP64 capability that remains critical for Lattice QCD, parts of HEP detector simulation, and high-precision astrophysical workflows (D6.1 §5.7). Without active engagement at HPC procurement and design stages, future European systems will support either AI workflows or HEP/RA precision needs, not both.

Community Evidence

D6.1 identifies a key gap directly relevant to AI integration: Gap #11: AI/ML integration. AI/ML workloads are increasingly central and are being investigated at every step of data-processing pipelines, from data acquisition to simulation and analysis. Recommendation: enhance services in AI/ML platforms to support distributed training, inference deployment, and access to foundation models, integrated with HPC and domain workflows.

Community surveys reveal that 90% of respondents report software performance improvement as an issue for AI workloads, yet 58% lack the resources to address it, quantifying the software investment gap. D6.1 §5.7 sets out the AI-vs-precision tension and three associated actions: identify analysis and simulation tasks suitable for ML acceleration; communicate precision and performance needs to HPC hardware planners; train physicists and astronomers in ML theory, frameworks and tools to fully exploit AI Factories.

Technical Foundation

D6.1 Section 4.5 (AI/ML and HPC Applications) defines four capabilities for data-intensive research. P07 builds on each as follows:

• Advanced Analytics and Interactive Exploration. Jupyter, RStudio and Dask-based platforms linked to the data federation that allow seamless scale-out from a researcher's laptop to large external resources. Target: hypothesis-generation workflows that start interactive at PB scale and transition to offline batch without per-site re-engineering.
• AI/ML Platforms. Distributed, GPU-rich training environments spanning HPC, cloud, and edge; scalable inference services; FAIR-compliant model registries that work as searchable "model zoos" enabling, for example, an astrophysicist to find a transient-classifier and fine-tune it with local data. Federated learning for privacy-preserving training across distributed observatories where bulk data movement is impractical (e.g. RA observatory-level streams; selected HEP collaboration-internal datasets).
• Simulation Frameworks and Digital Twins. Domain simulation (GEANT4 for HEP detectors, N-body and fluid solvers for astrophysics) combined with AI-assisted prediction. Digital twin platforms such as itwinai allow domain experts to use multi-node training and optimisation without dedicated ML engineering. Concrete targets: digital twin of at least one HL-LHC sub-detector for radiation and alignment modelling; digital twin pilot for an SKA precursor station for calibration prediction.
• Large-Scale Data Processing Pipelines. Reconstruction, Monte Carlo production, RA data reduction. AI components integrated into existing pipelines without per-site customisation, running on heterogeneous compute through standardised access protocols (P04).
• Agentic AI. AI agents that plan and execute scientific workflow steps under human oversight, extending AI integration from training and inference into operational decision-making.

European Alignment

Engagement with the European AI policy stack is the strategic backbone of this priority:

• Regulatory. AI Act (Regulation (EU) 2024/1689) sets transparency, documentation, dataset-quality and risk-classification obligations that affect AI used in safety-relevant scientific workflows (detector control, telescope operations), as well as the model-registry and provenance work in this priority.
• Coordination. The AI Office (DG CNECT, 2024) is the central EC coordination point for AI strategy and AI Act enforcement, including scientific-advisory channels.
• Funding. The AI Continent Action Plan (2025) commits over €10 billion for AI Factories and over €20 billion for AI infrastructure overall. The Apply AI strategy and the Cloud and AI Development Act (CADA, proposed 3 June 2026 within the European Technological Sovereignty Package)target a sovereign EU AI and cloud stack.
• Infrastructure. AI Factories (EuroHPC, 19 sites across three waves) carry an explicit obligation under their governing regulation to support the official European Data Spaces. AI Gigafactories proposed at scale capable of training 1 billion-parameter-class foundation models. IPCEI-CIS provides a co-funding pathway for sovereign federated cloud.
• Science. The RAISE programme coordinates AI capabilities for science across the EU via SCIANCE, its AI-in-science landscape-mapping project. The SCIANCE working group for Astronomy and Fundamental Physics provides the route for the RA and HEP communities; substantive engagement and use-case contribution through that working group is the actionable EC route in this priority.
• Data. EOSC enables FAIR-compliant model and dataset sharing across communities.
• Strategic frameworks. ESFRI identifies the need for Research Infrastructures to adopt emerging technologies including AI to maintain scientific competitiveness.

Target Outcomes and Impact Assessment

Strategic Outcomes

Outcome 1: AI-Enabled Scientific Workflow

s. AI capabilities are integrated into scientific workflows as first-class infrastructure rather than add-ons. GPU-rich platforms support distributed training, inference deployment, and access to foundation models across HPC, HTC, and cloud resources. AI surrogate models achieve two to three orders of magnitude computational cost reduction for high-impact HEP and RA use cases, opening research questions previously computationally intractable.

Outcome 2: Sustainable AI Capability

. FAIR principles apply not only to training data but also to training processes and models. Federated learning enables model training where data cannot be centralised due to sovereignty, GDPR, or volume constraints. Energy-aware scheduling addresses AI energy consumption whilst building on European Green500 leadership.

Outcome 3: HEP and RA Strategically Positioned in the EU AI Programme

HEP and RA hold formal representation in the RAISE AI for Science pillar. HEP and RA precision requirements are incorporated in AI Factory and AI Gigafactory procurement specifications, preserving FP64 capability alongside AI-optimised acceleration.

Success Indicators

• Resource accounting reform. GPU and AI-Factory allocations to HEP and RA experiments accepted as pledged resources within the WLCG and SRCNet accounting frameworks within the mid-term horizon.
• Operational AI workloads on AI Factories. A representative subset of European AI Factories with operational HEP or RA pilot workloads.
• FAIR AI lifecycle. Training data, processes, and models documented and preserved following FAIR principles; shared model registries operational with measurable cross-community reuse.
• Precision-capable AI infrastructure. Lattice QCD (or equivalent FP64 reference workload) sustained at acceptable performance on at least one AI-Factory-class system.
• Programme participation. HEP and RA representation secured in the AI for Science pillar of the RAISE programme; recurring engagement with the AI Office on scientific AI compliance.
• Workflow integration. AI methods integrated into scientific workflows through common orchestration frameworks (P06).
• Energy efficiency. AI workloads scheduled with energy awareness; reporting integrated with P12 sustainability accounting
• Agentic AI workflows. Agent-mediated scientific workflow execution (analysis triage, anomaly investigation, calibration response) operational across HEP and RA collaborations under a documented oversight regime; agent activity audited within WLCG and SRCNet operational reporting.

European Impact

AI integration is a strategic contribution to European digital sovereignty: indigenous AI capability for scientific research reduces dependence on external providers, and the FAIR-compliant model and dataset infrastructure provides a sovereign alternative to commercial model marketplaces. Engagement with the EU AI policy stack positions HEP and RA as evidence-base contributors rather than peripheral users. Workforce uplift through AI training builds capacity that flows back into European industry. Coupled with sustained FP64 capability in procurement specifications, the priority safeguards the long-term feasibility of precision-dependent science within the AI-led infrastructure transition.

Implementation Pathway and Stakeholder Roles

Implementation Phases

• In the short-term:
• Formal engagement with the RAISE programme to secure HEP and RA representation in the AI for Science pillar. Identify research-community interlocutors and a governance route.
• Per-experiment assessment of which analysis and simulation tasks are suitable for ML acceleration, and which require sustained FP64 (per D6.1 §5.7 Action 1).
• Pilot validation of AI platforms, federated storage and orchestration on a small set of AI Factory sites (P04, P05).
• Communicate FP64 precision and performance needs to EuroHPC procurement bodies so future procurements specify balanced hardware mixes (per D6.1 §5.7 Action 2).
• Pilot federated-learning deployment for one cross-observatory RA use case and one HEP collaboration-internal use case.
• Engage the AI Office on AI Act application to scientific AI (model documentation, dataset transparency, risk classification of operational AI, including the agent-class profile).
• Define oversight framework for agentic AI in scientific workflows: autonomy scope, approval requirements, audit requirements, interfaces with P03 service-account access and P09 provenance.
• Pilot one HEP analysis-triage agent and one RA transient-follow-up agent under bounded oversight: P03 service-account access, P06 orchestration, P09 audit and provenance.
• Mid-term:
• Integrating cloud and HPC infrastructures, with the cloud providing a unified access model, can lower the entry barrier for scientists and accelerate the adoption of larger-scale AI workloads across HEP and research infrastructures. This approach leverages cloud interoperability while providing access to HPC-grade compute for the most demanding workloads.
• GPU resources accepted as pledged within WLCG and SRCNet accounting frameworks.
• Agent-mediated workflow execution operational across multiple HEP and RA workflows; oversight framework documented and adopted by participating infrastructures.
• In the long term:
• Sustained, interoperable European continuum with production-grade federated AI/ML environments, automated provenance tracking, formalized co-design and governance models, and embedded AI/ML as strategic capabilities underpinning Europe’s exascale science leadership.
• Agentic AI as a routine capability for HEP and RA; controls, audit trails, and incident reporting integrated with WLCG and SRCNet operational tooling; AI Act compliance baseline applied to scientific AI agents.

These phases rely on standardization, co-design with scientific communities, advanced resource orchestration, sustainable policies, and a skilled workforce.

Stakeholder Roles

Scientific Communities

• Responsibilities: Identify workloads suitable for AI acceleration and those requiring sustained FP64; produce workload-mix specifications as input to HPC and AI Factory procurement; drive FAIR adoption for models and training processes; engage with RAISE governance via designated SPECTRUM interlocutors.
• Benefits: AI-enabled workflows opening intractable research questions; secured AI Factory and Gigafactory access; recognition of GPU contributions in accounting frameworks.

Research Infrastructures

• Responsibilities: contribute HEP and RA use cases and requirements to the SCIANCE Astronomy and Fundamental Physics working group; coordinate federated-learning pilots across member sites; provide foundational datasets and instrument-edge integration points for AI deployment.
• Benefits: improved utilisation; AI-assisted detector calibration, radiation and alignment modelling, and transient detection at lower marginal cost.

Computing & Data Service Providers / e-Infrastructures

• Responsibilities: integrate AI platform capability across the continuum (EuroHPC supercomputers, AI Factories, national HPC, EGI-federated grid and cloud sites, WLCG, SRCNet); reconcile AAI across community AAIs and AI Factory access; provide elastic allocation modes for AI workloads alongside batch; implement model-registry hosting and federation; coordinate energy-aware scheduling.
• Benefits: standardised interfaces, improved utilisation, reduced operating costs through shared FAIR model infrastructure, and positioning of EGI/WLCG/SRCNet alongside AI Factories as the operational substrate for scientific AI.

Policy Makers / Funding Bodies

• Responsibilities: steer coherent EU and national investment across the AI continuum, covering AI Factories, EuroHPC and national facilities, horizontal HTC/Cloud federations like EGI or community federations like WLCG and SRCNet; include HEP and RA precision needs in procurement and provide multi-year compute allocations for science (P02); sustain HEP and RA contribution to SCIANCE; ensure cross-DG coordination (DG RTD, DG CNECT, DG GROW).
• Benefits: coherent return on coordinated EU and national AI investment; visible scientific impact across European AI policy and the SCIANCE programme; sovereign capability for high-value science; alignment of AI Act objectives with operational scientific practice.

Resource Needs

• Compute. Annual GPU-hours across the continuum (AI Factories for training, national HPC and EGI-federated grid sites for inference and analysis, WLCG and SRCNet for community workflows), with sustained FP64 capability allocation alongside AI-optimised allocation.
• Storage. Training-data staging capacity and model-registry storage.
• Workforce. ML engineers and ML-physicist hybrid roles within RIs and infrastructure providers, aligned with the P13 framework.
• Software development. Integration of AI capabilities into existing reconstruction, simulation, and analysis pipelines, aligned with P08 and P09.
• Engagement. Sustained coordination capacity for SCIANCE working-group, AI Office and AI Factory engagement.

Critical Dependencies

• SRIDA priorities: P02 (multi-year AI compute allocations), P03 (cross-AAI and service-account access for agentic workflows), P04 (HPC/HTC/cloud integration), P05 (training-data movement and federated-learning access), P06 (orchestration of AI and agentic execution), P08-P10 (FAIR model lifecycle, reproducibility, audit and provenance), P12 (energy-aware scheduling), P13 (ML and hybrid-role capacity).
• Community and infrastructure: WLCG and SRCNet accounting frameworks accepting GPU and AI-Factory allocations as pledged resources; EGI Federation onboarding AI workload patterns (elastic allocation, model-registry hosting); per-experiment HEP and RA workload assessment for ML suitability versus FP64.
• Policy and timeline: AI Act enforcement (AI Office and Member State authorities, including agent-class profile); AI Factory rollout (EuroHPC, 19 sites scheduled); AI Gigafactory selection (proposed, outcome not yet decided); SCIANCE working-group structure and RAISE governance evolution.
• Risk: if the post-2027 EuroHPC procurement cycle drops FP64 in favour of AI-optimised throughput, the precision-dependent fraction of HEP and RA computing loses its European home.