Agentic AI for Multi-Stage Physics Experiments at a Large-Scale User Facility Particle Accelerator

TL;DR

The paper addresses the bottleneck of manual scripting and domain-specialist load in accelerator experiments. It introduces a language-model-driven agentic AI that operates within the ALS control environment, turning natural-language prompts into structured, auditable execution plans. The approach uses plan-first orchestration, bounded tool access, and dynamic capability selection to ensure safety and reproducibility. An end-to-end pipeline integrates data retrieval, PV resolution, script generation, machine control, and analysis on containerized components. In a production ALS deployment, the system reduces preparation time by two orders of magnitude and maintains operator safety, offering a blueprint for broader adoption.

Abstract

We present the first language-model-driven agentic artificial intelligence (AI) system to autonomously execute multi-stage physics experiments on a production synchrotron light source. Implemented at the Advanced Light Source particle accelerator, the system translates natural language user prompts into structured execution plans that combine archive data retrieval, control-system channel resolution, automated script generation, controlled machine interaction, and analysis. In a representative machine physics task, we show that preparation time was reduced by two orders of magnitude relative to manual scripting even for a system expert, while operator-standard safety constraints were strictly upheld. Core architectural features, plan-first orchestration, bounded tool access, and dynamic capability selection, enable transparent, auditable execution with fully reproducible artifacts. These results establish a blueprint for the safe integration of agentic AI into accelerator experiments and demanding machine physics studies, as well as routine operations, with direct portability across accelerators worldwide and, more broadly, to other large-scale scientific infrastructures.

Figures (5)

• Figure 1: Overview of the agentic workflow. Multi-turn conversational input and external data sources are first processed into a structured task. Relevant capabilities are dynamically classified on each iteration of the interaction, and the description of the selected tools are passed to the execution planner. The planner generates a complete, inspectable execution plan with explicit dependencies, which is then carried out by the agent with context tracking and artifact management.
• Figure 2: System architecture of the Accelerator Assistant. Control room and remote users access the system via a web interface (Open WebUI) or command line, which routes requests to the ALS Agent. The agent orchestrates connections to the PV database, archive data, and execution environments such as Jupyter. Model inference is performed either locally using Ollama or through cloud providers via the CBorg gateway. Integration with EPICS enables safe interaction with accelerator hardware at the ALS.
• Figure 3: Workflow of the PV Finder subsystem. A normalized export of the Matlab Middle Layer Accelerator Object provides the data model, which the agent explores through a strictly bounded API. Natural language queries are split into atomic intents, preprocessed to extract systems and keywords, and resolved into specific EPICS PVs.
• Figure 4: Pipeline for controlled Python code execution in the Accelerator Assistant. Natural language tasks are translated into a plan, results schema, and then Python code, which can dynamically access the agent context, is statically analyzed, and may be reviewed by a human operator. Execution is typically confined to containerized Jupyter kernels with strict read/write policies, and every run produces session artifacts (context, notebooks, JSON) for full reproducibility.
• Figure 5: Example output of the Accelerator Assistant: hysteresis measurement of ID gap versus vertical beam size at the ALS. The execution plan generated by the agent combined historical range extraction, automated script generation, and real-time machine control. The agent performed a 30-point bidirectional gap sweep with 5 repeated measurements per point, producing the plot shown here for one device. This figure illustrates the final output of the agentic workflow, where every step, from parsing natural language to data retrieval, machine control, and plotting, was generated and executed automatically.