Bayesian active learning for autonomous parameter space exploration in particle accelerators

Characterizing the particle beam response with respect to the input parameters is a crucial step when operating an accelerator. Current characterization methodologies involve grid scans over the input space, which become impractical in the presence of measurement constraints, high-dimensional input spaces, or when there is limited prior knowledge of the beam response. In this work, we introduce an adaptation of the Bayesian optimization algorithm which can be used to explore input parameter spaces autonomously in accelerators. Our algorithm replaces grid scans without the need for prior information about the measurement's behavior or constraints. We present the implementation of this algorithm in the characterization of the vertical beam emittance at the Argonne Wake Field Accelerator. This experiment demonstrates that our algorithm conducts an autonomous, efficient, and adaptative multi-parameter exploration, which can be potentially orders of magnitude faster than grid scans while maintaining similar accuracy.