friday / writing

The Fermilab Copilot

The Fermilab Accelerator Complex must deliver reliable, high-intensity proton beams for the Deep Underground Neutrino Experiment. Operations involve hundreds of diagnostic channels, beam tuning procedures, and subsystem interactions that operators manage manually. AI/ML models could automate routine tasks, but deploying them reliably into an accelerator control system requires infrastructure that most physics labs lack.

Note (2026-11-14; read against the paper's full text): the source is a Fermilab technical memo that describes requirements and infrastructure for using AI and machine learning in the accelerator control system (an MLOps framework, a data quality framework for trustworthy applications, and workflow integration with language models), with use cases in beam diagnostics, beam control and support-system automation; it describes needs and plans, not a deployed system or an evaluation. The essay's central claims, that the bottleneck is operational trust, that an undetected drifting model is worse than none, and that natural-language querying makes operators more likely to notice errors, are the essay's own; I did not find tests of operator behavior in the memo, which presents the data quality framework as what is needed to build trustworthy applications.

The authors (arXiv:2603.19507) identify three required capabilities: an MLOps framework standardizing the AI/ML lifecycle from data management through deployment and monitoring; a data quality framework ensuring that training data meets standards (accelerator data is notoriously heterogeneous and noisy); and LLM integration for assisting operators with information retrieval, code development, and routine analysis.

The through-claim: the bottleneck for AI in accelerator operations isn't algorithm performance — it's operational trust. An ML model that drifts undetected is worse than no model at all, because the operator stops checking what the model handles. The MLOps and data quality frameworks aren't engineering overhead — they're the mechanism by which trust is maintained. The LLM integration is the interface layer: operators who can query the system in natural language are more likely to notice when it's wrong than operators who must read dashboards.