Scientific Achievement
Researchers from the Advanced Light Source Accelerator Physics Program in the Accelerator Technology & Applied Physics (ATAP) Division at Lawrence Berkeley National Laboratory (Berkeley Lab) have successfully integrated a deep-learning-based feed-forward control system to stabilize the vertical electron beam size at Berkeley Lab’s Advanced Light Source (ALS) during regular user operation. This method utilizes a neural network to dynamically compensate for perturbations caused by the user-controlled insertion devices (IDs), ensuring stable beam performance during operations.
Significance and Impact
Integrating deep learning into the control systems at the ALS marks a significant advancement in synchrotron operations. By utilizing neural networks, this research introduces an efficient, real-time solution for beam size stabilization, improving speed and precision compared to traditional methods. The work demonstrates the effectiveness of machine learning in accelerator control, offering substantial improvements in operational performance at scientific facilities.
This pioneering work brings the Lab’s state-of-the-art machine learning R&D to the operational area, improving the ALS’s performance for users and demonstrating how machine-learning techniques for controlling complex accelerators and related systems are important across Basic Energy Sciences, High-Energy Physics, Fusion, and other areas.
Application of deep learning to synchrotron operations
Operational performance of the NN-based ID FF system during a user run. Shown are the vertical ID gaps (top), the elliptical polarized undulator (EPU) phase or longitudinal offsets (center), and the vertical electron beam size (bottom) as measured at ALS diagnostic beamline 3.1 (red) and as inferred (blue) if no correction had been applied. One beam outage occurred at hour 42 during that 5-day window; notably, the beam size control algorithm dis- and re-engaged automatically without human
intervention. (Credit: Berkeley Lab)
As the first synchrotron user facility to implement deep learning-based control algorithms in regular operations, ALS has demonstrated the practical benefits of using machine learning in accelerator control. Furthermore, the deep learning method can acquire training data in an order of magnitude quicker than traditional ID feed-forward table-based correction methods, making it a faster and more efficient solution.
Real-time adaptation through online fine-tuning
A key innovation of this system is its ability to fine-tune the model in real-time during user operations. This online fine-tuning allows the neural network to adapt to new operational conditions not previously encountered during pretraining, providing continuous stability despite changing configurations.
Operational robustness
The system is designed to be highly robust, incorporating an inhibitor chain that prevents incorrect actions during abnormal conditions, such as beam outages. The inhibitor chain automatically disengages and reengages the system as needed, ensuring the system remains reliable and responsive without requiring manual intervention.
Research Details
Illustration of the experimental setup for quantitative evaluation of the beam size correction. The top two plots show the vertical ID gaps and horizontal EPU offsets, respectively. The two bottom plots show the uncorrected beam size and the beam size while running the NN-based FF correction. (Credit: Berkeley Lab)
The research began with a comprehensive analysis of years of archived user operation data from ALS, which laid the groundwork for model development. Dedicated accelerator physics shifts were then used to collect additional data, providing a robust dataset for training the model. A detailed study of various neural network architectures and hyperparameters led to selecting a multilayer perceptron as the optimal choice. The system was integrated into the ALS control framework, communicating through the Experimental Physics and Industrial Control System (EPICS) and operating with a 10 Hz update loop to ensure real-time adjustments and maintain beam size stability.
Contact: Thorsten Hellert
Researchers: Thorsten Hellert, Tynan Ford, Simon C. Leemann, Hiroshi Nishimura, and Marco Venturini (Berkeley Lab); and Andrea Pollastro (University of Naples, Italy, and Instrumentation and Measurement for Particle Accelerator Laboratory, Naples, Italy).
Funding: U.S. Department of Energy’s Office of Science, Office of Basic Energy Sciences.
Publication: Hellert, T., Ford, T., Leemann, S. C., Nishimura, H., Venturini, M., and Pollastro, A. “Application of deep learning methods for beam size control during user operation at the Advanced Light Source,” Phys. Rev. Accel. Beams 27(7), 074602, 2024, https://doi.org/10.1103/PhysRevAccelBeams.27.074602
For more information on ATAP News articles, contact caw@lbl.gov.
