AI Meets Particle Accelerators: Engineers Propose a Safety-First Integration Layer

A new review published in EPJ Research Infrastructures argues that the biggest obstacle to bringing artificial intelligence into accelerator control rooms is not the quality of the AI models themselves. It is the engineering gap between research-grade prototypes and the mature, safety-critical control systems that operate real machines. The article, authored by Jure Varlec, Jan Jug and Tilen Žagar of the Slovenian control systems company Cosylab, lays out an architecture designed to close that gap without touching the operational infrastructure facilities have spent decades building.

The authors, writing as an industrial partner within the ARTIFACT network—a European project whose full name is ARTificial Intelligence For Accelerators, user Communities and associated Technologies—frame the problem as one of integration rather than performance. Accelerator control systems rest on established frameworks such as EPICS and TANGO, or on industrial SCADA variants, layered over decades of operational experience, established safety procedures and legacy hardware. Deploying an AI model into that environment demands attention to real-time constraints, safety-interlock compatibility, operator trust and long-term maintainability—considerations that simply do not arise in a research laboratory.

The scale of the fragmentation becomes clear in a 2023 online survey conducted by K. Zenker and colleagues, which gathered responses from 15 facilities. Forty-six percent of participants reported using EPICS, 27 percent used TANGO, and 27 percent relied on systems built on other frameworks, including DOOCS, TINE and LabVIEW. Larger and older facilities often run several frameworks at once—separate systems for the accelerator and its beamlines, for instance, or industrial SCADA for cryogenics. The review presents this diversity as healthy competition among design philosophies, but it carries a practical cost: there is no standardised interface through which external software can connect to a control system, which makes testing new tools on operational accelerators far harder and constrains how quickly they are adopted.

The paper illustrates the difficulty with a concrete case. A recent reinforcement-learning-based beam tuning study at the Paul Scherrer Institut (PSI) required a dedicated Python extraction library to interface with the archival infrastructure, resampled all channels to a fixed 200-millisecond interval, and yielded only 63.2 days of usable data from a much longer archival period after filtering for consistent machine configuration. The same facility provisioned a dedicated Linux server on the control network, running PyEPICS, to act as a mediated access point for the machine-learning agent. And because the agent was coupled to EPICS through PyEPICS inside a custom OpenAI Gym environment, it could not be deployed at a TANGO-based facility without rewriting the interface layer.

Against this backdrop, the authors propose a framework-agnostic abstraction layer that connects AI services to facility control systems without necessarily modifying operational infrastructure. The design rests on a handful of principles. Non-invasiveness: the layer operates as a client of the control system, using the same standard protocols as any other high-level client. Framework-agnostic abstraction: an ML model or agentic workflow should address process variables through a single unified interface regardless of which framework hosts them. Modularity: each connector is an independently deployable component that can be developed and versioned separately. And safety by default: all access to process variables is read-only unless explicitly and narrowly enabled.

The safety architecture is arguably the most striking part. Write access is an opt-in capability, scope-restricted to specific subsets of process variables, and the layer distinguishes between operational writes that actuate the machine and display writes that present AI state to operators. For actuating agents, the architecture supports a human-in-the-loop approval workflow in which the AI proposes an action and an operator reviews and approves or rejects it before execution. A watchdog mechanism returns control to the operator if an autonomous AI service becomes unresponsive—a network partition between the AI service and the control system is described as a realistic failure mode—with configurable responses ranging from freezing the machine at its current setpoints to reverting to pre-AI values. Audit logging records every AI-initiated action, including those passing through human approval, a necessity because frameworks differ sharply in native auditing support: EPICS offers the caPutLog module, OPC-UA mandates audit events in its specification, but TANGO has no native audit logging and DOOCS offers none.

The technical comparison underpinning the design is detailed. EPICS, the most widely deployed framework and the paper’s primary illustrative example, organises hardware state as named process variables served by Input/Output Controllers and accessed over the Channel Access or newer PVAccess protocols. The five frameworks examined—EPICS, TANGO, OPC-UA, DOOCS and FESA—differ most in metadata and command mechanisms. EPICS uses four alarm severity levels with 22 status codes; TANGO defines five quality states including a CHANGING state; OPC-UA provides three severities with hundreds of sub-codes and an Uncertain category; DOOCS exposes a server-dependent integer status flag; and FESA has no per-field quality at all, only a device-level status property. Cycle tagging is similarly uneven: FESA supports first-class multiplexing, DOOCS tags readings with a macropulse or train ID, EPICS offers a 64-bit user tag, while TANGO and OPC-UA have no native cycle tagging at all.

Large language models receive their own treatment, split into two roles with very different risk profiles. As operator copilots, LLMs could interpret alarms, assist with procedures, retrieve knowledge and summarise logbook entries—all on a read-only basis. The authors cite the Osprey framework, developed with contributions from Cosylab and LBNL, which transforms natural language prompts into state-graph workflows that operators can review before execution, and ChatEED, which focuses on operating-log book-keeping. As actuating agents, however, LLMs introduce substantially greater risk. The paper is blunt about their limitations: they are language-processing models that lack the capability to work with exact mathematics, physics and engineering problems, and they frequently commit logical fallacies. Safety-critical systems such as the machine protection system should remain strictly off-limits, and even within permitted scopes, human review is deemed necessary. The authors also note that most accelerator AI applications are unlikely to be classified as high-risk under the EU AI Act, though alignment should still be sought to avoid regulatory uncertainty.

Validation, the authors stress, has not yet happened. The ARTIFACT implementation of the abstraction layer has not commenced, and the TwinRise project—an EU-funded Digital Twin Engine initiative federating AI and simulation components across particle accelerators, oncology centres and fusion research facilities—will be the first context in which the design is deployed and tested against operational demands. The authors anticipate three areas of inquiry: resolving open design decisions through iterative implementation, collecting structured operator feedback on approval gates and interaction patterns, and quantifying latency budgets, subscription throughput and archive-query scaling under realistic load. They further flag certification of AI-integrated control systems under facility safety frameworks such as IEC 61511 as unprecedented territory, alongside software longevity across facilities that operate on decade-long timescales and a shortage of interdisciplinary expertise spanning accelerator physics, control engineering and machine learning. As a review, the article proposes rather than proves; its contribution is a carefully argued architecture awaiting the operational test that TwinRise is expected to provide.

Subject of Research: Integration of artificial intelligence into particle accelerator control systems through a framework-agnostic abstraction layer

Article Title: Bridging the Deployment Gap: Integrating AI into Accelerator Control Systems

Article References: Varlec, J., Jug, J., & Žagar, T. (2026). Bridging the Deployment Gap: Integrating AI into Accelerator Control Systems. EPJ Research Infrastructures, 10(1), Article 18. https://doi.org/10.1007/s41781-026-00173-8

Image Credits: AI Generated

DOI: 10.1007/s41781-026-00173-8

Keywords: particle accelerators, artificial intelligence, control systems, EPICS, TANGO, machine learning, large language models, human-in-the-loop, ARTIFACT network, Cosylab, TwinRise, digital twins