IPAC'25 - the 16th International Particle Accelerator Conferece

Machine learning-enhanced beam trajectory stabilization at Canadian Light Source

THPM105

5 Jun 2025, 15:30

2h

Exhibiton Hall A _Magpie (TWTC)

Poster Presentation MC6.D13 Machine Learning Thursday Poster Session

Speaker

Shervin Saadat (Canadian Light Source (Canada))

Description

This study presents a Neural Network (NN)-based feedback system to improve orbit correction precision at the Canadian Light Source (CLS), replacing the traditional Response Matrix (RM) algorithm. Implemented with TensorFlow-Keras, the NN model features three hidden layers with 96 nodes, processing high-frequency beam position monitor (BPM) data at 1 kHz from 48 vertical and 48 horizontal sensors. The model was trained on a dataset of BPM and orbit corrector (OC) values under varying beam currents, with alignment tolerances set at ±1000 µm and ±500 µm horizontally and vertically. Tests over 11 minutes at 8.0 mA beam current showed the NN model's superior performance, reducing root-mean-square (RMS) horizontal beam position fluctuations by 56.4% in section 11 and 20.3% in section 1. RMS values dropped from 1.28 × 10⁻⁴ (RM) to 1.02 × 10⁻⁴ (NN) in section 11. The NN consistently stabilized beam trajectory without manual intervention, demonstrating the potential of machine learning to revolutionize particle accelerator control systems with adaptive, precise real-time management.