Automated tuning will be crucial for the fast commissioning of large, complex accelerators in scientific user facilities, such as the Advanced Photon Source Upgrade (APS-U). Such tuning also helps ensure that machines reach and maintain the highest performance during user operation. Machine learning (ML) enables efficient search of global optima in complex parameter spaces and thus can be used to improve tuning methods.

In collaboration with SLAC National Accelerator Laboratory, APS researchers are developing advanced, ML-based tuning methods and applying these methods to important applications on the APS accelerator complex: tuning of the dynamic aperture and the momentum aperture for the APS storage ring, stabilization of injection efficiency and charge injected into the APS, stabilization of the coupling ratio against insertion device (ID) gap moves, and compensation of orbit shifts due to fill pattern changes.

Minimizing unplanned interruption of the X-ray beam to facility users is another important challenge in accelerator operation. Failure of accelerator components can cause hours to days of lost productivity and impact experiments. Automated detection of anomalies and identification of their root causes will be essential to reduce unplanned accelerator down time. Prediction of faults and preventive maintenance are also possible. In a collaborative effort to develop ML-based methods for fault detection, identification, and prediction, we will identify fault types of interest at the APS and develop methodologies to mitigate them.