PACMAN HIPA

Motivation

The operation of high-intensity proton accelerators like HIPA at the PSI requires precise control of several hundred machine parameters to ensure efficiency, stability, and safety of the beam line. Traditional approaches rely heavily on manual oversight and classical control strategies, which may not fully exploit the rich, continuous stream of sensor data available. Integrating Machine Learning (ML) offers the potential to improve performance by minimizing beam losses, enhancing parameter control,  and preventing unnecessary machine interruptions.

Proposed Approach / Solution

First, we developed a surrogate model based on beam diagnostics data to provide real-time predictions to the machine protection system. The model enables early forecasting of potential interruptions, allowing operators to take preventive actions in advance and maintain stable, uninterrupted accelerator operation. Second, we deployed an automated parameter tuning tool to minimize beam losses. Our approach is based on Safe Bayesian Optimization, a zero-order optimization algorithm that takes safety constraints into account, and in particular prevents interrupts on the accelerator during the optimization.

Impact

This work demonstrated that data-driven prediction and optimization can significantly improve the efficiency and reliability of accelerator operations. By reducing beam losses and unplanned downtimes, the surrogate model contributed to safer, more stable operation at higher beam intensities. As a result, more beam time became available for scientific experiments. The successful integration of predictive modeling marks a step toward smarter, data-driven control of complex accelerator systems.