Physics model-informed Gaussian process for online optimization of particle accelerators

High-dimensional optimization is a critical challenge for operating large-scale scientific facilities. We apply physics-informed Gaussian process (GP) optimizer to tune complex system by conducting efficient global search. Typical GP models learn from past observations make predictions, but this reduces their applicability new systems where archive data not available. Instead, here we use fast approximate model physics simulations design the model. The then employed inferences sequential online in order optimize system. Simulation and experimental studies were carried out demonstrate method control of storage ring. show that outperforms current routinely used optimizers terms convergence speed, robustness on task. ability inform machine-learning with may have wide applications science.