LLM-driven Accelerator Assistant dramatically speeds ALS particle experiments

Lawrence Berkeley National Laboratory has deployed the Accelerator Assistant, an LLM-driven system at the Advanced Light Source (ALS) to support and autonomously assist with accelerator operations and multistage experiment setup.

• Deployment and capabilities: The Accelerator Assistant runs on an NVIDIA H100 GPU (using CUDA) and routes requests through Gemini, Claude, or ChatGPT via the CBorg gateway; it integrates with the lab-developed Osprey framework, Ollama for local inference, EPICS, and Jupyter Notebook execution environments, and can access a database of >230,000 process variables, supporting 40 beamlines and ~1,700 experiments per year.
• Implementation details and rollout: Operators access the tool via command line or Open WebUI with authenticated sessions and personalized memory; the team reports up to 100x reduction in setup effort (“two orders of magnitude”), has published a research paper (arXiv), and is expanding the framework across the DOE’s Genesys mission and in collaborations with ITER (France) and the ELT (Chile).