Smart prediction of beam availability at synchrotron facilities

Particle accelerators, particularly synchrotron facilities, are essential instruments in materials, life, environmental, and medical sciences. They accelerate charged particles to near-light speed, producing synchrotron light delivered to beamlines. Maintaining reliable beamtime is crucial but challenging because many interconnected subsystems contribute to beam delivery. This paper presents a machine-learning approach for predicting beam availability to support synchrotron reliability. It compares several Deep Learning (DL) and Machine Learning (ML) models commonly used in related work. The methodology uses operational data from the Synchrotron-light for Experimental Science and Applications in the Middle East (SESAME) facility, with preprocessing across multiple time-window sizes to distinguish trip from no-trip events. The dataset was constructed from 222 trip events and 506 no-trip events. For the best-performing 10-second setting, it contained 7,998 labeled samples with 166 PV-based features, including 30.4% trip and 69.6% no-trip samples. Using grid search to optimize model and time-window selection, the Neural Network (NN) model with a 10-second window achieved the highest recall of 94.6%. Training on 2020–2022 data and testing on held-out 2023 data, the NN achieved 32.8% precision and a 48.7% F1-score on the test set. The evaluation was conducted offline using historical data, and the system has not yet been deployed for online real-time operation. The generated dataset is openly available to support further research.