Combining Physics and Machine Learning to Analyze Particle Beams in Accelerators

The Science

To precisely control the shape of particle beams, scientists need to measure them in fine detail. A beam can be thought of as a cloud of particles. Scientists describe each beam (or cloud) by six numbers: three for its position (x, y, z) and three for its momentum (p_x, p_y, p_z). Traditional measurement methods can usually capture only part of this information and create a flat, two-dimensional snapshot of the beam. Now, researchers have developed a new technique that combines a special type of physics simulation with machine learning. This approach can reconstruct the beam's full details in just minutes. It uses only about 10 to 20 experimental measurements, far fewer than previous methods.

The Impact

Characterizing and controlling detailed particle beam shapes is crucial for advancing experiments and discoveries. However, existing methods are slow. In addition, they require specialized hardware or machine learning models trained on massive simulation datasets or simplified physics that miss key details. In contrast, this new reconstruction technique works across different accelerators and setups. It is easy to adapt to new facilities and requires no prior data. It enables real-time, six-dimensional beam characterization, which improves accelerator monitoring and control. Overall, this approach provides a faster, more flexible tool that complements other techniques. It offers a powerful new way to study and optimize particle beams.

Summary

This new approach uses a "differentiable physics simulation." This is a type of simulation that makes it easy to calculate derivatives for every step of the process. This capability allows scientists to combine the simulation with machine learning components, such as a neural network that models the initial beam distribution. To reconstruct the beam, researchers record 2D projections of the beam on an imaging screen while adjusting a magnet and a radio-frequency cavity. The researchers then feed the measurements into the differentiable simulation of the setup, along with the neural network representation of the beam. This enables the system to reconstruct the beam's full six-dimensional phase space. The differentiable simulation constrains the reconstruction, while the neural network adds flexibility and computational efficiency. The combination of the two enables results in minutes and without any prior data.

After first demonstrating the technique on partial beam reconstruction (the x-y dimensions), researchers extended it to full 6D reconstruction and the characterization of flat beams. Researchers have conducted all of the experiments so far at the Argonne Wakefield Accelerator, a research facility at Department of Energy's Argonne National Laboratory. Future work aims to apply this method to additional types of measurements, which would further increase its versatility. Expanding the use of this method would make high-precision, real-time beam characterization a practical tool for accelerator research and development.

Contact

Ryan Roussel and Auralee Edelen
SLAC National Accelerator Laboratory
[email protected], [email protected]

Funding

Funding for this work is provided by the Office of Science, Basic Energy Sciences' Early Career Research Program (ECRP), and High Energy Physics' General Accelerator Research & Development (GARD) program, as well as the National Science Foundation (NSF) Center for Bright Beams.

Publications

Roussel, R., et al. Efficient six-dimensional phase space reconstructions from experimental measurements using generative machine learning, Physical Review Accelerators and Beams 27, 094601 (2024). [DOI: https://doi.org/10.1103/PhysRevAccelBeams.27.094601 ]

Roussel, R., et al. Phase space reconstruction from accelerator beam measurements using neural networks and differentiable simulations, Physical Review Letters 130, 145001 (2023). [DOI: https://doi.org/10.1103/PhysRevLett.130.145001]

Kim, S., et al. Four-dimensional phase space reconstruction of flat and magnetized beams using neural networks and differentiable simulations, Physical Review Accelerators and Beams 27, 074601 (2024). [DOI: https://doi.org/10.1103/PhysRevAccelBeams.27.074601]