Machine learning assisted beam commissioning for the MedAustron accelerator complex

Abstract

MedAustron is a center for ion particle therapy and non-clinical research in Wiener Neustadt, Austria. It uses a synchrotron-based accelerator complex with active scanning as the beam delivery technique. The accelerator can deliver beam to five beam lines for four irradiation rooms: Irradiation room 1 (IR1) with a horizontal beam line, IR2 with a vertical as well as a horizontal beam line, IR3 with a horizontal beam line and IR4 with a proton gantry. Two ion species are currently used at MedAustron: Protons with energies in the range of 60-250 MeV (up to 800 MeV for non-clinical research) and carbon ions with energies in the range of 120-400 MeV/u. The commissioning of the accelerator is a complex and time consuming task, with stringent requirements for the beam parameters in the treatment rooms. Main contributors to the complexity are the the non-ideal behavior of beam line elements such as magnets and beam instrumentation devices as well as the limitations of the simulations used as a basis for the commissioning of the machine. This thesis studies the feasibility of using a machine learning-based approach for beam commissioning by selecting an appropriate machine learning model based on simulated data, and evaluating the performance of the machine learning model as a function of the training data size. The results are then verified by training and evaluating the performance of the model on actual measurement data from the machine.