<header>

AI for Particle Accelerators
Satellite Meeting at IPAC26

</header>

Overview

1. Initiatives

Europe

The ARTIFACT network (artifact-network.org) and its underlying funded projects were presented. The network brings together European AI-for-accelerators initiatives and includes an ~€18M funded program centered on digital twins and federated learning. Agentic workflows were highlighted as a key development direction within this framework.

United States

The MOAT collaboration was presented as a leading US initiative, with a dedicated poster and talk at IPAC26 (Friday). The broader context of the DOE Genesis Mission — a Department of Energy initiative to coordinate AI for science across the national laboratory complex — was described, including multi-facility efforts on agentic AI, digital twins, and shared platforms. Open collaboration frameworks and community mailing lists were highlighted as entry points for international participation.

Korea

Ongoing AI-related developments at Korean accelerator facilities were presented.

Japan

The integration of AI development in particle accelerators within the broader national AI4Science program was discussed, presenting the Japanese approach to coordinating domain-specific and general AI research.

2. Data

The central importance of data for AI-driven efforts was a recurring theme. Key points from the dedicated data discussion:

• Data standards need to be compatible with agentic workflows. A common challenge across facilities is the absence of normalized, AI-ready data formats, and the difficulty of making legacy documentation machine-readable.
• FAIRness principles (Findable, Accessible, Interoperable, Reusable) were discussed as the necessary framework for data publishing. Significant gaps remain in applying FAIR principles to accelerator facility data.
• Ontology descriptions for accelerator facilities were identified as a critical bottleneck for enabling AI and agentic workflows. A dedicated US effort on accelerator facility ontology development (compatible with AI tools) was presented, with integration into the American Science Cloud (ASC) under the Genesis Mission. This architecture allows a central ontologies architecture while individual facilities retain control of their data. UK efforts specifically targeting the harmonization of ontological descriptions were also highlighted.
• European data standardization efforts integrate within the EOSC (European Open Science Cloud), notably through the AI4EOSC initiative, though these are at an earlier stage than the US counterpart.
• A dedicated satellite meeting on data standards and accelerator ontological descriptions was announced for Thursday at 12:30 during IPAC26.

3. Industry

Two companies shared perspectives on AI integration in accelerator systems:

• Cosylab emphasized the importance of linking control systems with simulation tools, positioning this interface as the primary entry point for AI in operational settings.
• IBA highlighted the industrial drivers: cost-efficient design, turnkey systems, and the operational lifecycle — responding to both client requirements and the current economic context.

The broader discussion touched on how industrial actors can connect with the open research community, and the importance of human oversight and system reliability in control room environments.

4. Simulation Tools

A dedicated discussion addressed simulation tools as enablers for AI integration — a topic of particular relevance for agentic and autonomous workflows:

• The Accelerator Middle Layer (AML) toolbox was discussed, including ongoing work to translate these tools into Python with automatic differentiation capabilities.
• XSuite was highlighted for its ongoing work on automatic differentiation features and native integration with surrogate models, making it suitable for AI-in-the-loop workflows.
• The Particle Accelerator Lattice Standard (PALS) was presented as a standardization effort for accelerator lattice descriptions, relevant for interoperability between simulation codes and AI tools.
• Simulation-to-agent interaction was identified as a key capability to develop: enabling AI agents to interact with and reason over simulation outputs in a structured way.
• The broader concept of AI-ready systems — designing simulation environments from the ground up with AI integration in mind — was raised as a guiding principle for future tool development.

5. Infrastructures

A broad discussion addressed the state of AI readiness across facilities and the barriers to AI development and deployment. Participants shared the state of their AI strategies:

• CERN: A new AI roadmap has been released, covering the IT department, accelerator sectors, and broader experimental programs. Open questions remain around cross-sector integration and the translation and interoperability between AI-ready systems.
• ESS: The approach to AI readiness starts from establishing correct data acquisition requirements (synchronous data, metadata, quality standards) and defining internal policy frameworks. The philosophy is to get the data infrastructure right before layering AI on top, with particular attention given to the current commissioning phase.
• GSI: An internal AI strategy does not exist yet but a readiness roadmap is being prepared.
• KIT: Focus on data management and analysis, with integration into the broader Helmholtz Association AI strategy.
• INFN: No formal AI roadmap; solutions are predominantly bottom-up and individual initiative-driven.
• CLARA: Focus on digital twins and their interface with EPICS-based control systems.
• France (national level and facilities): Bottom-up structuration is underway. GANIL is developing a new strategy integrating nuclear physics experiments, accelerator operation and simulation, and IT, within a broader national framework involving CNRS and CEA. ARRONAX emphasized the importance of developing an AI strategy to support current operation and future upgrades, though no formal roadmap is yet in place.
• IFMIF-DONES (under construction): The importance of planning AI-ready systems from the design phase was raised — a facility under construction has the opportunity to embed AI readiness from the start.
• EU level: Bottom-up approaches dominate, characterized as "AI for science / Science for AI" — domain needs are expected to drive AI tool development rather than top-down mandates.

Cross-cutting themes

• Maintainability and explainability of AI systems are key barriers to operational deployment. Operator trust requires interpretable models and well-defined failure modes.
• Top-down vs. bottom-up: A tension between institutionally mandated roadmaps and organic, researcher-driven AI adoption was observed across multiple facilities.
• Cultural shift and training: Significant workforce adaptation is required, particularly in facilities with established automation philosophies and control room cultures.
• Data sanitization and normalization: A recurring challenge — raw operational data is often too heterogeneous or poorly documented to be used directly by AI systems.

Topics Not Addressed

The following topics were identified as important but could not be discussed due to time constraints:

Follow-up and Next Steps

• A satellite meeting on accelerator data standards and ontological descriptions takes place on Thursday at 12:30 during IPAC26. Participants interested in ontology and data publishing standards are encouraged to attend.
• A summary of this discussion will be distributed to all registered attendees via the community mailing list.
• All registered participants will be added to the community newsletter/mailing list for future updates.
• The community is encouraged to review their facility documentation and work toward making it more AI-ready for various workflows (LLM-based ingestion, digital twin workflows, agentic systems, etc.).
• Facilities interested in joining the MOAT/MODE collaboration framework are encouraged to reach out through the collaboration's public channels.

<footer>AI for Particle Accelerators — Satellite Meeting  |  IPAC26, Deauville, France  |  May 19, 2026</footer>