Speaker
Description
The Los Alamos Neutron Science Center (LANSCE) supports five user facilities and operates under a wide range of beam conditions. For decades, beam tuning has relied largely on operator expertise and manual knob adjustments to control losses. While several optimizers have been developed and tested for specific tasks, none have been incorporated into routine operations, in part because accelerator tuning at LANSCE is inherently multi‑objective and must adapt reliably to conditions such as startup, beam recovery, and production.
We present our recent effort to deploy the Badger Optimizer, specifically designed for accelerator control rooms, using an XOpt backend tailored. Our implementation emphasizes operator trust, machine safety, and practical usability. Multi‑objective tuning is reformulated through operator-adjustable weighting, enabling straightforward prioritization of high‑loss locations or balancing between different beamlines. Dynamic boundaries are introduced to ensure responsive but protected optimization behavior, and correlated tuning variables are grouped to improve convergence and operator interpretation. Operational constraints such as nominal setpoints and hardware limits are incorporated directly into the optimization workflow.
This work outlines a practical path toward integrating ML‑based optimization into daily control-room practice, supporting more efficient and reliable tuning across LANSCE’s diverse operating modes.