Speaker
Description
We take a step beyond Operator-supervised optimization by introducing an agent that plans and executes fast, sample-efficient injector tuning constrained by competing goals (capture efficiency, energy spread, transmitted charge).
We present an agent built on the Osprey agentic framework that can (i) author a machine-readable configuration for a configurable online optimizer, (ii) invoke that optimizer as a tool, and (iii) autonomously infer next tuning actions by querying and actuating the control system.
The agent reasons over multi-diagnostic signals and operational constraints, selects objective/constraint formulations, proposes bounded setpoint updates, and schedules measurements; safety is enforced via capability-scoped tool use, guardrails, and machine-protection checks.
Control system integration relies on EPICS get/put/monitor primitives for live state estimation and closed-loop execution.
Initial machine studies at the ALS show the agent can compose full tuning plans—configuring objectives tied to known bottlenecks (e.g., beam-loading–driven energy-spread control relevant to booster acceptance) and launching optimization runs—yielding reduced operator intervention and faster convergence, while preserving capture within the booster ring’s tight longitudinal window.
We summarize the architecture (Osprey capabilities, optimizer interface, EPICS I/O), online decision logic, early results, and a path to portable deployment at other accelerator facilities.
Funding Agency
Funded by the US Department of Energy (BES Accelerator and Detector Research Program), and supported by the US Department of Energy, Director of the Office of Science under Contract DEAC02-05CH11231.
| In which format do you inted to submit your paper? | LaTeX |
|---|