Surrogate-based Bayesian inference of transverse beam distribution for non-stationary accelerator systems

Abstract

Constraints on the beam diagnostics available in real-time and time-varying beam source conditions make it difficult to provide users with high-quality beams for long periods without interrupting experiments. Although surrogate model-based inference is useful for inferring the unmeasurable, the system states can be incorrectly inferred due to manufacturing errors and neglected higher-order effects when creating the surrogate model. In this paper, we propose to adaptively assimilate the surrogate model for reconstructing the transverse beam distribution with uncertainty and underspecification using a sequential Monte Carlo from the measurements of quadrant beam loss monitors. The proposed method enables sample-efficient and training-free inference and control of the time-varying transverse beam distribution.