Machine Learning Speeds Up Tuning Operations of Particle Accelerator

Every year, the SLAC National Accelerator Laboratory (SLAC) at the Department of Energy is visited by scientists across the world to carry out scores of experiments in energy research, biology, materials science, and chemistry at the Linac Coherent Light Source (LCLS) X-ray laser.

The LCLS produces extremely bright X-rays from high-energy beams of electrons that are generated in a massive linear particle accelerator.

At LCLS, experiments are conducted day and night, in two 12-hour shifts every day. At the beginning of every shift, operators need to adjust the performance of the accelerator so as to prepare the X-ray beam for the subsequent experiment. At times, more tweaking would also be required during a shift. Previously, this task—known as accelerator tuning—required operators to devote hundreds of hours every year.

Now, with the help of machine learning, SLAC scientists have devised a novel tool that can speed up a part of this tuning process by five times more than that of the earlier techniques. The researchers have explained this technique in the Physical Review Letters journal published on March 25^th, 2020.

Tuning the Beam

A high-quality electron beam had to be initially prepared to create a strong X-ray beam at LCLS. Some of the energy of the electrons subsequently gets changed into X-ray light within exclusive magnets. The characteristics of the electron beam, which has to be tightly focused and dense, are a major factor in establishing the quality of the X-ray beam.

“Even a small difference in the density of the electron beam can have a huge difference in the amount of X-rays you get out at the end,” stated Daniel Ratner, head of SLAC’s machine learning initiative and also a member of the research group that devised the novel method.

A range of 24 exclusive magnets, known as quadrupole magnets, are used by the accelerator to focus the beam of electrons just like the way the glass lenses focus light. Conventionally, human operators will cautiously turn the knobs to tune the individual magnets between their shifts to ensure that the accelerator was creating the required X-ray beam for a specific experiment. But this process consumes plenty of the operators’ time—time that could be spent on other significant tasks to enhance the X-ray beam for experiments.

Some years ago, a computer algorithm was implemented by LCLS operators that expedited and automated the process of magnet tuning. But this algorithm had certain drawbacks. It aimed at enhancing the X-ray beam by making haphazard modifications to the strengths of the magnets and subsequently picking the ones that made the best X-ray beam.

Yet this algorithm was different from human operators. It did not have any previous knowledge of the structure of the accelerator and could not make educated guesses about tuning the accelerator that could have eventually resulted in considerably improved results.

This is the reason why SLAC scientists decided to create a new algorithm that integrates machine learning with knowledge about the accelerator’s physics. Machine learning is a “smart” computer program that slowly learns to become better over time.

The machine learning approach is trying to tie this all together to give operators better tools so that they can focus on other important problems.

Joseph Duris, Study Lead and Scientist, SLAC National Accelerator Laboratory

A Better Beam, Faster

The latest approach employs a method known as a Gaussian process, which predicts the impact of a specific accelerator adjustment on the X-ray beam’s quality. Moreover, it creates uncertainties for its predictions. The computer algorithm subsequently decides the type of modifications that need to be made for the most optimal enhancements.

For instance, the algorithm may decide to make a major alteration, the result of which is extremely vague but may result in a big payoff. In other words, this latest, adventurous computer algorithm has enhanced features than the earlier algorithm that makes the required adjustments to produce the most optimal X-ray beam.

In addition, the SLAC team utilized the data obtained from the earlier LCLS operations to teach the computer algorithm the type of magnet strengths that usually result in brighter X-rays, thus allowing the algorithm a way to make educated guesses about the type of modifications it should try. This method equips the computer algorithm with both expertise and knowledge that are naturally present in human operators, and which was absent in the previous algorithm.

We can rely on that physics knowledge, that institutional knowledge, in order to improve the predictions.

Joseph Duris, Study Lead and Scientist, SLAC National Accelerator Laboratory

The method was also improved by gaining a better understanding of the magnets’ relationships with one another. The quadrupole magnets operate in pairs, and their focusing power can be enhanced by increasing the strength of one magnet in a pair, while decreasing the strength of the other magnet.

Through this latest process, the quadrupole magnets can be adjusted about three to five times faster, estimated the scientists. The process is also likely to generate higher-intensity beams when compared to that of the algorithm used earlier.

“Our ability to increase our tuning efficiency is really, really critical to being able to deliver a beam faster and with better quality to people who are coming from all over the world to run experiments,” stated Jane Shtalenkova, an accelerator operator at SLAC who worked with Duris, Ratner, and others to create the novel tool.

Beyond LCLS

The same technique can also be used to adjust the properties of other electrons or X-ray beams that would be required by other researchers to improve their experiments. For instance, scientists could use the method to increase the signal they receive from their sample after it is struck by LCLS’ X-ray beam. Such a flexibility also makes this novel computer algorithm useful for other kinds of facilities.

The nice thing about this machine learning algorithm is that you can do tech transfer relatively easily.

Adi Hanuka, Scientist, SLAC National Accelerator Laboratory

Hanuka has been testing the new method at three other accelerators—PEGASUS at the University of California, Los Angeles; SPEAR3, the accelerator ring powering SLAC’s Stanford Synchrotron Radiation Lightsource (SSRL); and the Advanced Photon Source (APS) at the Department of Energy’s Argonne National Laboratory.

“This tool now exists in several labs. Hopefully, we’ll be integrating it into even more labs soon,” concluded Hanuka.

APS, SSRL, and LCLS are Department of Energy Office of Science user facilities. The project was mostly funded as part of SLAC’s Laboratory Directed Research and Development Program (LDRD).