Machine-learning-assisted rapid beam energy change at the ATLAS heavy ion linac

Studying nuclear reactions to develop new medical isotopes requires the measurement of production cross sections at varying beam energies. To do this efficiently without wasting beam time available for experimentation, the energy needs to be changed rapidly over a very wide range. We present recent experimental results, employing machine learning methods for rapid tuning of 16O &136Xe beams following manual energy change. For the 16O beam energy is changed from a base value of 106 MeV to 71 MeV and for the 136Xe beam energy is changed from the base energy of 803 MeV to 671 MeV and 525 MeV. Re-tuning from the base energy took ~15 minutes compared to ~30 minutes for operator tuning. The collected data is then used to validate the same procedure in a virtual model of ATLAS running the TRACK simulation code with error ~6% across energies. Preliminary virtual model error is corrected by considering quadrupole misalignments and real values of input beam parameters obtained by Bayesian inference.