Speaker
Description
Hadron Experimental Facility (HEF) is designed to handle intense slow-extraction proton beam from 30-GeV Main Ring (MR) of Japan Proton Accelerator Research Complex (J-PARC). The production target in HEF is operated under severe conditions in which its temperatures periodically vary from 30 to over 300 °C by heat deposit of irradiated beams. In a long term, a careful evaluation of damage to the target and its lifetime is quite important. If the target temperatures are accurately predicted from the existing data, including beam intensity, duration of beam extraction and beam position, it could be possible to verify the cumulative damage by comparing the predicted temperature rise with the measured one. The predicted temperature rise was calculated from the existing data using linear regression with a machine learning library, scikit-learn. However, the predictions of temperature rise by linear regression were not fully satisfactory due to changes of beam optical and accelerator conditions. Therefore, an enhanced prediction method for the temperature rise on the proton target has been developed using Long Short-Term Memory (LSTM), a type of Recurrent Neural Network (RNN) architecture. A systematic study was also conducted to investigate the effects of hyperparameters, including a sequence length and hidden layer. The present paper reports the status of the prediction system of the temperature rise on the production target using the LSTM analysis in detail.
