Source-linked AI summary
Agentic Artificial Intelligence for Multistage Physics Experiments at a Large-Scale User Facility Particle Accelerator
Thorsten Hellert, Drew Bertwistle, Simon C. Leemann, Antonin Sulc, Marco Venturini
TL;DR
Non-routine accelerator tasks require subsystem expertise and custom scripting, while mistakes can threaten machine availability and concurrent experiments. The Accelerator Assistant turns natural-language requests into safety-gated, multistage procedures on a production synchrotron. In a representative task, it reduced preparation time by two orders of magnitude versus manual scripting while preserving operator-standard safety constraints.
Problem
Non-routine accelerator tasks require custom scripts and deep subsystem knowledge across more than 230,000 process variables, while beam interruptions affect concurrent experiments.
Source: §I · p. 1Method
The Accelerator Assistant translates natural-language requests into inspectable plans that coordinate archive retrieval, PV resolution, script execution, machine interaction, and analysis.
Source: §I · p. 2, §II · p. 3, §III · p. 4Results
Preparation time was reduced by two orders of magnitude relative to manual scripting while operator-standard safety constraints were preserved.
Source: §III · p. 4, §IV · p. 5Takeaways & Limitations
Plan-first orchestration, bounded tool access, and dynamic capability selection provide a blueprint for transparent, portable agentic systems at scientific facilities.
Source: §IV · p. 5
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
Accelerator work demands distributed subsystem expertise, custom scripting, and careful protection of beam availability, yet reports of autonomous multistage operation at production user facilities remain outstanding. The Accelerator Assistant addresses this gap with natural-language-driven, auditable workflows and demonstrated a two-orders-of-magnitude preparation-time reduction while upholding operator-standard safety constraints.
I. INTRODUCTION
- Beam interruptions at the ALS typically cause at least 30 minutes of downtime, affecting dozens of concurrent experiments across more than 40 beamlines.Mistuned RF parameters or incorrect magnet settings can also cause beam loss or hardware damage.Source: §I · p. 1
- ALS troubleshooting spans more than 230,000 process variables and demands custom scripts and subsystem expertise, creating substantial preparation overhead.Unexpected faults require operators to identify relevant channels, retrieve archive data, and assemble ad hoc analysis under time pressure.Source: §I · p. 1
- Prior work covered simulated or low-stakes applications and accelerator prototypes, but autonomous multistage operation at production user facilities remained unreported.Earlier accelerator efforts included a prototype assistant, a single-task beamline optimization study, and a broader integration vision.Source: §I · pp. 1–2
- Preparation time fell by two orders of magnitude versus expert manual scripting while operator-standard safety constraints were upheld.The Accelerator Assistant retrieves archive data, resolves PVs, generates and executes scripts, and analyzes results from a single prompt.Source: §I · p. 2
- Plan-first orchestration, dynamic capability filtering, and reproducible artifacts make the system inspectable and support its portability across scientific facilities.The authors present the ALS deployment as a blueprint for integrating language-model-driven systems into other synchrotrons, accelerators, and infrastructures.Source: §I · p. 2
II. ALS AGENTIC CONTROL FRAMEWORK
The Accelerator Assistant uses modular capabilities to translate conversational input into structured tasks and orchestrate access to accelerator data, execution environments, and hardware. Inspectable plans, human review options, and structured run artifacts support controlled, reproducible operation.
II. ALS AGENTIC CONTROL FRAMEWORK
- The framework normalizes natural-language fragments into structured objects and combines modular capabilities without retraining existing components.Capabilities are classified independently for task relevance, limiting prompt inflation as the inventory grows.Source: §II · pp. 2–3
- Users access the agent through a web interface or command line, while it connects to PV and archive services and Jupyter execution environments.Inference can run locally through Ollama or through external providers via the CBorg gateway; EPICS integration supports safe hardware interaction.Source: §II · p. 2, Figure 2 · p. 3
- Before any tool call, plan-first orchestration creates a complete plan with explicit input-output dependencies that operators or validators can inspect or modify.Plans provide checkpoints for reviewing inputs, outputs, and dependencies before actions begin.Source: §II · p. 3
- Containerized services, bounded retries, checkpointing, and human approval options support controlled execution and produce logs, JSON outputs, and notebooks for provenance.The system can also generate reusable Phoebus monitoring panels from natural-language requests.Source: §II · p. 3
- Configurable PV limits and write blacklists or whitelists let facilities adapt the framework to local practices and operating conditions.The authors describe these configuration mechanisms as a path to facilities with differing maturity, actuator behavior, diagnostic conditions, and metadata quality.Source: §II · p. 3
III. FROM QUERY TO EXPERIMENT
The Accelerator Assistant executed a non-routine insertion-device measurement from a natural-language request, coordinating archived-data retrieval, PV resolution, controlled machine interaction, and analysis. Preparation was two orders of magnitude faster than manual scripting, and the demonstration produced consistent plots while preserving safety gates.
III. FROM QUERY TO EXPERIMENT
- Preparation fell from several hours of manual scripting and debugging to a few minutes, a two-orders-of-magnitude speedup.The beam-based measurement itself still takes about an hour; the preparation savings came from turning the request into a structured plan of modular steps.Source: §III · p. 4
- The execution plan stages time-range normalization, PV resolution, archive retrieval, analysis, controlled machine interaction, and visualization behind safety gates.At the ALS, all write access to the control system requires operator approval.Source: §III · p. 4
- PV Finder resolves atomic user intents into EPICS channels using a normalized MML Accelerator Object model and a strictly bounded API.The model covers approximately 10,000 key PVs; resolved PVs and time ranges are then used for automated archive retrieval.Source: §III · p. 4, Figure 3 · p. 4
- Code generation proceeds through a plan, a structured results schema, and minimal Python code executed in containerized Jupyter kernels with read/write policies.Runs produce notebooks, JSON results, and figures for provenance and reproducibility.Source: Figure 4 · p. 4, §III · p. 4
- The chained capabilities produced consistent, publication-ready plots across devices, with the example showing the expected absence of vertical beam-size hysteresis.The authors state that the same architecture applies to a wide range of machine physics tasks beyond this example.Source: §III · p. 5, Figure 5 · p. 5
- The ALS demonstration measured insertion-device gap versus vertical beam size using a 30-point bidirectional sweep with five repeated measurements per point.The plan combined historical range extraction, automated script generation, and real-time machine control for one device.Source: §III · p. 3, Figure 5 · p. 5
IV. CONCLUSION
The Accelerator Assistant executes multi-stage physics experiments on a production synchrotron, reducing preparation time by two orders of magnitude while preserving operator-standard safety constraints. Its architecture supports transparent, reproducible operation and provides a blueprint for safe integration of agentic AI into scientific operations and other large-scale experimental facilities.
IV. CONCLUSION
- Preparation time fell by two orders of magnitude while operator-standard safety constraints were preserved and reproducible artifacts supported inspection and trust.The system was deployed on a production synchrotron light source storage ring.Source: §IV · p. 5
- Plan-first orchestration, bounded tool access, and dynamic capability selection support scaling, transparency, and portability across accelerator facilities.Source: §IV · p. 5
- The demonstrated results provide a blueprint for safely integrating agentic AI into scientific operations and other large-scale experimental facilities.Source: §IV · p. 5