Evolution of the high-fidelity objective from first-principles simulations without (grey) and with (blue) the assistance of reduced-model simulations using multi-fidelity Bayesian optimization. This opens a path to cost-effective optimization of laser-plasma accelerators in large parameter spaces, an important step toward fulfilling the high beam quality requirements of future applications.

Particle accelerators often involve many tunable parameters. Tuning these various parameters simultaneously typically requires evaluating each set of parameters, either with simulation or in experiments, which can be very time-consuming.

In the Advanced Modeling Program (AMP), we use AI/ML techniques to drastically reduce the time needed for this type of task. For instance, we use multi-fidelity Bayesian optimization to combine the results of fast, low-fidelity and slower, high-fidelity simulations and arrive at the optimal parameters more quickly. We are also working on ML models that combine simulation and experimental data and can provide real-time feedback on how to tune parameters during experimental campaigns.

AMP’s high fidelity simulations are also a desired source for AI/ML model training, providing a trove of data from gigabytes to petabytes in size. As one example, AMP used high-fidelity wakefield simulations to train ML surrogates that can be used as flexible elements in future, conventional-plasma hybrid accelerators.

Various machine-learning methods are being tested to improve the speed and accuracy of particle beam phase spaces tomography based on limited (Beam Position Monitors) diagnostics. In this example, tomography using a differentiable particle model is tested: at left is the ground truth phase-space density; at right is the mean prediction from the ML-based model.

We also use AI/ML techniques for particle accelerator lattice design. We make use of auto-differentiation to develop differentiable simulation models that include both nonlinear lattice parameters and charged particle collective effects. Bayesian optimization is explored in lattice matching and beam parameter optimization.

We are actively involved in training the next generation of particle accelerator physicists in AI/ML techniques. In particular, AMP members (along with colleagues from SLAC) have developed and taught a course on optimization and ML for accelerators for the U.S. Particle Accelerator School (USPAS).