An artificial intelligence algorithm has been developed to more precisely predict how particles are distributed in a particle beam inside an accelerator showing that users can infer very complicated high-dimensional beam shapes from “astonishingly small amounts of data.”
Particle accelerators are among the most important (and biggest) experimental tools in modern physics. Beams of particles are shot through metal piping at near-light speed to study the atomic behaviour of molecules and the smallest subatomic particles.
Read more: “Ghostly” neutrinos used for the first time to peer inside protons at Fermilab accelerator
Knowing how a particle beam will behave in a given experiment is important to maximise the scientifically useful information that can be gleaned. This is especially important as accelerators operate at ever higher energies and produce more complex beam profiles.
But identifying particle behaviour is no easy task.
Because particle beams often involve on the order of billions of particles, it’s not simply a matter of predicting where each one will end up.
Now, researchers at the US Department of Energy’s SLAC (Stanford Linear Accelerator Center) in California, and the University of Chicago, have developed a machine learning algorithm to give a more accurate picture of how particles in an accelerated beam are distributed.
“We have a lot of different ways to manipulate particle beams inside accelerators, but we don’t have a really precise way to describe a beam’s shape and momentum,” says SLAC accelerator scientist Ryan Roussel. “Our algorithm takes into account information about a beam that is normally discarded and uses that information to paint a more detailed picture of the beam.”
Researchers usually use a statistical approach to describe the speed and position of particles to provide a rough shape of the overall beam. But potentially useful information could be ignored in the process.
Alternatively, scientists can assess what a beam would look like under different experimental conditions by taking many measurements of the beam itself. Such methods sometimes already use machine learning, but require huge amounts of data and computational power.
In the latest study, the team built a machine learning model that essentially takes the best elements of both methods.
Their algorithm uses our knowledge of beam dynamics to predict what is collectively known as the “phase space distribution” of particle speeds and positions.
“Most machine learning models don’t directly include any notion of particle beam dynamics to speed up learning and reduce the amount of data required,” says SLAC accelerator scientist Auralee Edelen. “We’ve shown that we can infer very complicated high-dimensional beam shapes from astonishingly small amounts of data.”
The team tested their model at the Argonne Wakefield Accelerator at the DOE’s Argonne National Laboratory near Chicago, Illinois. They were able to interpret experimental data using particle beam physics using only 10 data points – for a machine learning model not trained in particle beam dynamics, the task would have taken up to 10,000 data points.
The model can currently reconstruct a particle beam in a 4D beam phase space – along the up-down and left-right axes. The researchers are working toward a full 6D phase space distribution which includes particle speeds along the direction of the beam itself.
The research is published in Physical Review Letters.