Machine learning for the surrogate modeling of radiofrequency quadrupole accelerators

The IsoDAR (Isotope Decay-At-Rest) experiment is a novel antielectron-neutrino source proposed to operate with an unprecedented primary proton beam current of 10 mA. Chief among the technical innovations that make this possible is the inclusion of a Radiofrequency Quadrupole (RFQ) that accelerates and pre-bunches the beam during its axial injection into the cyclotron. The accurate start-to-end simulation of through-going beams in this and similar experimental setups is computationally expensive, particularly because the high beam current leads to nonlinear space charge effects which must be accounted for. In this contribution, we demonstrate that machine learning-based surrogate models can approximate, to high predictive accuracy, the output beam parameters when passed through a virtual representation of the IsoDAR RFQ, significantly more computationally efficient than traditional high-fidelity simulations. This is the basis for a toolkit with the potential to transform traditional particle accelerator engineering by incorporating insights from artificial intelligence. We discuss pros and cons of such surrogate models, particularly when it comes to predicting quantities based on the second moments of the beam particle distribution such as transverse beam emittance, and discuss opportunities for the use of such surrogate models as a much faster accelerator design optimization tool.