Machine learning paves the way for smarter particle accelerators

July 20, 2022 — Scientists have developed a new machine learning platform that makes the algorithms that control particle beams and lasers smarter than ever. Their work could contribute to the development of new and improved particle accelerators that will help scientists unlock the secrets of the subatomic world.

Daniele Filippetto, staff scientist, works on the high repetition rate electron scattering apparatus. Credit: Thor Swift, Berkeley Lab.

Daniele Filippetto and colleagues at the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) developed the setup to automatically compensate for real-time changes in accelerator beams and other components, such as magnets. Their machine learning approach is also better than contemporary beam control systems at understanding why things fail and then using physics to formulate an answer. An article describing the research was published late last year in Nature Scientific Reports.

“We’re trying to teach a chip physics, while giving it the wisdom and experience of an experienced scientist operating the machine,” said Filippetto, a staff scientist in the Accelerator Technology Division and in Applied Physics (ATAP) at Berkeley Lab and Deputy Director of the Berkeley Accelerator Controls and Instrumentation Program (BACI).

Their research also has the potential to impact multiple applied areas of particle accelerators, ranging from autonomous operations in industrial and medical environments to increased precision in scientific applications, such as linear colliders and electron lasers. ultra-fast free.

The new technique was demonstrated at Berkeley Lab’s High Repetition-Rate Electron Scattering Apparatus (HiRES) accelerator in collaboration with researchers from Los Alamos National Laboratory and UCLA. The main application of the HiRES beamline is to perform structural dynamics experiments on new quantum materials. The instrument has contributed to many scientific discoveries such as carrying out the very first ultrafast electron diffraction studies of the optical fusion of tantalum ditelluride, a material with interesting and potentially useful properties. Today, this new machine is showing its usefulness for developing new methods of controlling large classes of accelerators.

Particle accelerators produce and accelerate beams of charged particles, such as electrons, protons and ions, of atomic and subatomic size. As machines become more powerful and complex, controlling and optimizing the particle or laser beam becomes more important to meet the needs of scientific, medical and industrial applications.

Filippetto and his colleagues in the BACI program are leading the global development of machine learning tools. These tools provide a platform to develop intelligent algorithms that react quickly and precisely to unforeseen disturbances, learn from their mistakes, and adopt the best strategy to reach or maintain the target beam setpoint.

The tools they develop have the added benefit of providing an accurate model of the overall behavior of a particle accelerator system, regardless of its complexity. Controllers can use these new and improved features to make more efficient decisions in real time.

Filippetto’s work currently focuses on using the power and prediction of machine learning tools to increase the overall stability of particle beams.

“If you can predict the properties of the beam with an accuracy that exceeds their fluctuations, then you can use the prediction to increase the performance of the accelerator,” he said. “Knowing the key beam parameters in real time would have a huge impact on the final precision of the experiments.”

At first, such an approach might seem unlikely to produce accurate results, similar to the challenges of predicting stock market behavior, but the group’s early results are promising. In fact, the algorithm used, which is based on neural network models, shows a tenfold increase in the accuracy of predicted beam parameters compared to a typical statistical analysis. In related work, a recent Halbach Prize was awarded to Simon Leemann, staff scientist in ATAP’s Accelerator Physics Group, and his collaborators for developing machine learning control methods that improve the performance of the advanced light source by stabilizing the highly relativistic electron beam at the experimental level. source points of about an order of magnitude, an unprecedented level.

Dan Wang, an early career researcher, is working on piezo inertial motor drivers to drive piezo mirrors, for laser alignment in the coherent laser combining system. Credit: Thor Swift, Berkeley Lab.

In related research, Dan Wang, a BACI research scientist who began her career at the Berkeley Lab three years ago as a postdoctoral researcher, is using machine learning tools to advance control technology in complex laser systems. In Wang’s case, the ultimate goal is to be able to precisely combine hundreds of ultra-intense laser pulses into a powerful, coherent beam the size of a human hair. In a coherent beam, the phase of each input laser must be controlled within a few degrees of error, which is very difficult. The laser energy can be combined in different ways but in all cases, it is imperative that the coherence of the beam array is stabilized against environmental disturbances such as thermal drift, air fluctuations or the movement of the supporting table.

To do this, Wang and his colleagues developed a neural network model that corrects system errors in the combined laser network 10 times faster than other conventional methods. The model they developed is also able to teach the system to recognize phase errors and parameter changes in lasers and automatically correct disturbances when they occur.

The researchers’ method works in both simulations and laser experiments, where unprecedented control performance has been achieved. The next step of the research is to implement machine learning models on state-of-the-art computers such as field-programmable gate arrays (FPGAs) for faster response, and also to demonstrate the generalizability of this method. learning-based control in more complex systems where there are many more variables to consider.

“I come from an accelerator background, but during my post-doc my colleagues really helped me embrace the power of machine learning,” Wang said. “What I’ve learned is that machine learning is a powerful tool for solving many different problems, but you still have to use your physics to guide you in how you use and apply it. “

“To meet the needs of new science, this work illustrates the active feedback and machine learning methods that are crucial enablers for the next generation of accelerators and laser performance to power new photon sources and future colliders. particles,” said Cameron Geddes, director of the accelerator. Division of Technology and Applied Physics.

This work was supported by the DOE’s Office of Science, Office of Basic Energy Sciences, and Office of High Energy Physics Sciences, as well as the laboratory-led research and development program.

Source: Will Ferguson, Berkeley Lab

Sherry J. Basler