Machine learning based closed orbit correction for steady state microbunching storage ring

Closed orbit correction is the initial step in error correction, and its results significantly influence subsequent correction of higher-order errors. This is particularly important for an steady state microbunching (SSMB) storage ring, where both the transverse dimensions approach the diffraction limit and the longitudinal length is on the order of hundreds of nanometers. The precision and speed of closed orbit correction are essential for achieving high average power and high brightness radiation in the SSMB storage ring. Here, we introduce a new model to address this issue. Our model integrates the attention mechanism with convolutional neural network (CNN). This combination enables the model to capture both global dependencies and local patterns. Furthermore, we considered the effects of various error distributions on the lattice optics function. By incorporating diverse error scenarios into the training dataset, we enhance the model’s robustness. Through simulations of the SSMB storage ring, we compared the correction results of our model with the singular value decomposition method. It achieves the desired residual closed orbit after only two iterations and shows encouraging correction outcomes across various error distributions. We tested the stability of our model by adjusting the strengths of specific correctors and incorporating beam position monitor (BPM) noise. Additionally, to address the issue of BPM failures, we integrated a Gaussian process model to ensure that our correction process remains effective even when BPMs malfunction.