The 3D model shown above is of course a simplified representation of a particle accelerator and particle beam. Real accelerators are significantly larger and have vastly more components. Where the visualization shows only five magnets, a real accelerator might have hundreds or even thousands of magnets. Combined with other components, it is not uncommon for these facilties to have tens if not hundreds of thousands of control parameters. At the same time the charged particle beam in reality is so small that it is invisible to the naked eye. Using this vast number of control parameters, this tiny particle beam needs to be controlled over multiple kilometers of beam pipe down to precisions of a few tens of micrometers. For reference that is about the same as throwning a pea from Hamburg to Munich, hoping to hit a target the size of another pea. What is more, users of particle accelerators continue to ask for ever more demanding experimental setups, requiring even more precise control of the particle beam.
While today, expert human operators spend over 2000 hours per year tuning these control parameters, AI and machine learning methods are now being explored to help enable the demands of novel experiments, while reducing tuning times and making more time available for actual scientific research.
One approach explored at DESY is the so-called Reinforcement Learning, where the tuning of the accelerator is formulated as a game in which the AI agent is rewarded for achieving the desired beam properties as quickly as possible, while being penalized for entering unsafe operating conditions. This method has so far successfully been tested on a tuning task involving an accelerator section of three quadrupole magnets and two dipole magnets, very similar to the one shown in the visualization. Here, the AI agent was able to learn how to attain a desired position and focus of the beam visible on the downstream diagnostic screen in a fraction of the time it would take a human operator. However, like real human experts (probably even worse than them), the AI agent needs a lot of time to learn how to control the beam in the first place. In the mentioned example, the AI agent actually needed about 3 years of non-stop trial and error on the accelerator to learn how to control the beam better than the human experts. This much time simply is not available in a real accelerator, where beam time is a very precious resource. Instead, the AI agent was able to learn of fourty parallel simulations of the accelerator, where in each simulation was sped up by five orders of magnitude. This way it is possible for the AI agent to learn in just under an hour, allthe while using no time on the real accelerator at all.