Modelization of an Injector With Machine Learning

Mathieu Debongnie (1) , Maud Baylac (1) , Frédéric Bouly (1) , Nicolas Chauvin (2) , Angélique Gatera , Tomas Junquera , Didier Uriot (2)

1 LPSC - Laboratoire de Physique Subatomique et de Cosmologie
2 IRFU - Institut de Recherches sur les lois Fondamentales de l'Univers

Modern particle accelerator projects, such as MYRRHA, have very high stability and/or reliability requirements. To meet those, it is necessary to optimize or develop new methods for the control systems. One of the difficulties lies in the relatively long computation time of current beam dynamics codes. In this context, the very low computation time of neural network is of great attraction. However, a neural network has to be trained in order to be of any use. The training of a beam dynamic predictor uses a large dataset (experimental or simulated) that represents the dynamics over the parameter space of interest. Therefore, choosing the right training dataset is crucial for the quality of the neural network predictions. In this work, a study on the sampling choice for the training data is performed to train a neural network to predict the transmission of a beam through a low energy beam transport line and a Radiofrequency Quadrupole. We show and discuss the results obtained on training data set to model the IPHI and MYRRHA injectors.

Mots clés: network; rfq; LEBT; solenoid; proton

Domaines: Physique des accélérateurs [physics.acc-ph]

Mathieu Debongnie, Maud Baylac, Frédéric Bouly, Nicolas Chauvin, Angélique Gatera, et al.. Modelization of an Injector With Machine Learning. 10th International Particle Accelerator Conference, May 2019, Melbourne, Australia. pp.WEPTS006, ⟨10.18429/JACoW-IPAC2019-WEPTS006⟩. ⟨hal-02272677⟩