How Is Artificial Intelligence (AI) Helping Physicists Working On Particle Accelerators? Machine learning is a subset of artificial intelligence that has the potential to dramatically improve the efficiency of a wide range of computing operations, from speech and picture recognition to autonomous vehicles to stock market trading to medical diagnostics.
Machine learning algorithms often need to be trained on existing data before they can begin working on a specific task; this allows the algorithms to learn how to quickly and accurately predict future events without any human input. But what if it’s a brand-new position for which there is no precedent from which to draw upon for instruction?
Scientists at the Department of Energy’s SLAC National Accelerator Laboratory have shown that machine learning can be used to improve particle accelerator efficiency by instructing algorithms in the fundamental physics principles underlying accelerator operations without requiring any a priori knowledge of these principles.
According to Adi Hanuka, a former SLAC research associate who led a study published in Physical Review Accelerator and Beams, “injecting physics into machine learning is a really hot topic in many research areas—in materials science, environmental science, battery research, particle physics, and more.” In the field of accelerator physics, this represents one of the earliest applications of machine learning influenced by physics. So, How Is Artificial Intelligence (AI) Helping Physicists Working On Particle Accelerators?
Also Read: How Does The Internet Change Your Brain And Behavior?
Educating AI with physics
Particle accelerators are high-powered equipment used in a variety of fields, from basic physics studies to molecular imaging and cancer radiation therapy. Operators must configure the accelerator for optimal performance in order to obtain the best beam for a certain application.
Because of the complexity of the many moving parts, fine-tuning a large particle accelerator may be a real headache. The fact that some parts are dependent on the settings of others only adds another layer of complexity.
Machine learning has been demonstrated to considerably aid human operators in recent experiments at SLAC, speeding up the optimization process and revealing viable accelerator settings that have never been thought of before. Machine learning can also aid in diagnosing particle beam quality without compromising it, as is the case with most other methods.
In order for these methods to be effective, researchers had to first “train” machine learning algorithms using either historical data from accelerator operations or hypothetical performance data generated by computer simulations. However, they also discovered that the amount of new data necessary might be drastically reduced by utilizing information from physics models in conjunction with existing experimental data.
The new research shows that past data are not required if one is familiar enough with the underlying physics that describes how an accelerator functions.
This method was used to fine-tune the SPEAR3 accelerator that drives SLAC’s Stanford Synchrotron Radiation Lightsource (SSRL). The researchers claim that their findings are on par with, if not better than, those produced by training the algorithm using actual archival data by leveraging information gained directly from physics-based models.
The study’s lead researcher, SLAC staff scientist Joe Duris, said, “Our results are the latest highlight of a continual push at SLAC to build machine learning techniques for adjusting accelerators.”
Trying to foretell the future
That’s not to imply that historical records can’t be instructive. If you’ve got the hang of physics, they’ll still be useful. When researchers combined the physics-informed machine learning model with data from the SPEAR3 accelerator, they saw significant improvements. The approach is also being used to fine-tune one of the world’s most powerful X-ray sources, SLAC’s Linac Coherent Light Source (LCLS), for which historical data from earlier experimental runs is available in the form of archives.
When SLAC staff activate LCLS-II next year, the method’s full potential should become obvious. With this superconducting LCLS improvement, a new accelerator is being installed, and the optimal parameters for it have yet to be identified. The accelerator’s operators might benefit from having AI that is familiar with the fundamentals of the field working alongside them.
