Machine-learning plot of parameter space

Pareto-optimal front for the ALS-U storage ring lattice derived by an in-house-developed ML-based multi-objective genetic algorithm. The most desirable solutions are those that offer high brightness (low emittance, shown on the right axis); long beam lifetime (large momentum aperture or MA, shown on the left axis), and good injection efficiency (large negative diffusion rate, shown in arbitrary units on the X axis). These desirable solutions tend to be toward the lower left corner of this plot.

Since modern particle accelerators can be large and incredibly complex instruments, accelerator physicists employ a broad range of tools and techniques to tackle the physics and technology challenges involved in improving the operation and performance of existing accelerators, like the ALS, and in optimizing the design of next-generation accelerators, like ALS-U.

Our group has established its expertise in optimizing future accelerators, focusing on highest-brightness light source designs. As designers push state-of-the-art accelerator technology toward its limits, optimizing the accelerator design (choosing among an ever-increasing number of parameters for multiple and often competing goals) becomes challenging to grasp using traditional first-principles approaches. AI/ML techniques can assist accelerator physicists in developing and understanding design solutions, speeding up and improving the tools we use for their optimization. We have successfully applied such techniques to the ALS-U optimization process and provide expertise and methods for future accelerator design optimization.

Plot shows how machine learning stabilizes vertical beam size.

The quality of the vertical beam size of the ALS electron beam (red) is greatly improved when ML-based stabilization is turned on (yellow area).

Once an accelerator has been commissioned and is operational, physicists are tasked with improving and optimizing the as-built machine to fulfill and, ideally, surpass the theoretical performance promise of the underlying design.

Owing to a modern accelerator’s complex nature, control is usually fully digital, and commissioning and operation tools are mostly software. We develop and apply state-of-the-art codes for these purposes but also recognize that modern approaches, such as AI/ML, present a unique opportunity to tackle problems existing tools have yet to allow us to overcome. We are pioneers in developing and employing AI/ML in synchrotrons for anomaly detection, online nonlinear optimization, and integrating various heterogeneous online data sources toward problem detection and resolution.