Speaker
Description
PIP-II project will raise Fermilab Booster intensity and ramp rate. Beam losses will limit maximum power and are hard to simulate. Presently, Booster uses operator-guided empirical tuning - a challenging task due to high dimensionality, multiple objectives, critical safety constraints, and drifts. We developed a synergistic suite of Bayesian optimization (BO) and reinforcement learning (RL) tools to optimize and stabilize beam losses, including novel techniques for fast risk-aware Bayesian optimization and exploration. For initial tune-up, we performed single and multi-objective tuning using scalarized objectives comprised of critical beam loss locations, achieving significant rebalancing of losses as well as an overall improvement in transmission efficiency. To build a data-driven surrogate, active learning was used to collect data while relying on risk-aware constraints to successfully avoid beam trips. A few thousand points were collected, and a GP surrogate validated for uncertainty-aware predictions. Several off-policy RL agent architectures were trained for long term stabilization. In surrogate-based testing, SAC with BPM context and history embedding had best performance with fast and robust convergence when subjected to energy and trajectory perturbations. Experimental testing is ongoing to enable operational use.
| Paper status | Resubmitted proceeding files received and assigned to an editor. Accepted. |
|---|