Operational Accelerator Tuning via Model-Coupled Optics and Bayesian Steering

Abstract

We present an on-line tuning strategy for the ISAC post-accelerator that pre-sets machine optics with a digital twin and then performs Bayesian optimization for steering under online operation with beam. The model computes end-to-end tunes in seconds and interfaces with the control system under device bounds, slew-rate limits, and loss interlocks. We report three experimental case studies demonstrating that decoupling optics from steering yields faster and more reliable convergence than a fully Bayesian optics-plus-steering baseline under identical conditions. Across these cases, iterations to high transmission tunes are reduced by a factor of 4-6, with final average transmissions in the mid- to high-90% range. By factorizing optics from steering, the dimensionality of the parameter space is reduced, convergence becomes more predictable, and operational safeguards are easier to enforce.

Figures (9)

• Figure 1: Overview of the ISAC-I RIB postaccelerator, showing major components.
• Figure 2: All operator input commands sent into the control system over a 24-hour time period, displayed in 5-minute timebins and shown over 5 consecutive days. In this instance, operators were manually tuning the ISAC-I linac in preparation for scheduled beam delivery. Power law normalization applied for clarity.
• Figure 3: Evolution of a converging beam through a drift. a) Starting distribution where beam is converging $\alpha_x<0$. b) Beam reaches a waist where $\alpha_x=0$. c) Beam starts diverging $\alpha_x>0$.
• Figure 4: transoptr simulation from the stable ion source line through the RFQ to the MEBT section. The sequential optimization tune is shown alongside the design tune, in solid and dashed lines respectively. The $x$ and $y$ beam envelopes are 2RMS. The focal strengths are scaled up for the quadrupoles and dipoles and scaled down at the RFQ for visual clarity.
• Figure 5: Optimization history for select sequences. $\|\chi\|$ shown in a red solid line alongside the tunable elements in dotted lines. $\|\chi\|$ was normalized to [0,1]. The tunable elements were normalized using the maximum and minimum of all the parameters.