In Berkeley, California, an AI agent at the ALS particle accelerator supports high-stakes physics experiments. The Accelerator Assistant, powered by an NVIDIA H100 GPU, helps keep X-ray research on track by tapping into institutional knowledge data and solving problems autonomously or with human input. The ALS facility conducts 1,700 scientific experiments per year across 40 beamlines.
ALS beam interruptions can halt experiments for minutes, hours, or days, affecting 40 beamlines conducting X-ray experiments. The ALS control system has over 230,000 process variables, making quick issue identification and troubleshooting crucial to minimize downtime. The Accelerator Assistant has been shown to reduce setup time and effort by 100x in running multistage physics experiments.
Operators at ALS interact with the system through a command line interface or Open WebUI, using Osprey framework for safe AI application in complex control systems. The system integrates with EPICS for secure interaction with accelerator hardware and uses Ollama and CBorg for local or external inference. Conversational input is processed to assist with tasks by utilizing external knowledge and personalized memory.
The Accelerator Assistant aids engineers in defining tasks and generating Python scripts for data analysis, visualization, and safe interaction with the accelerator. By providing relevant context from facility operations, the system can save significant time, enabling engineers to run the facilities autonomously with human oversight. The framework is being deployed across U.S. particle accelerator facilities and is being considered for use in fusion reactor facilities.
Beyond optimizing operations, ALS research has global scientific impact. During the COVID-19 pandemic, ALS researchers characterized an antibody neutralizing SARS-CoV-2. ALS experiments on Metal-Organic Frameworks supported Nobel Prize-winning research in sustainable water harvesting and carbon management. ALS measurements of asteroid samples helped trace the chemical history of asteroid Bennu, deepening our understanding of Earth’s origins.
Read more at NVIDIA: AI Copilot Keeps Berkeley’s X-Ray Particle Accelerator on Track