Recent developments of machine-learning-based techniques in RIKEN RI Beam Factory The RIKEN RI Beam Factory (RIBF) is a heavy ion accelerator facility consisting of four ECR ion sources, two linacs, five cyclotrons, and a superconducting fragment separator. For the efficient operation of this facility, developments of beam diagnostics and beam tuning methods using machine learning have been initiated. This presentation will focus on the following two topics. The first is the prediction of the beam intensity from the ECR ion sources. By using plasma images as input to the machine learning model, the beam intensity has been successfully estimated for short-term test datasets. The second is the development of a parameter optimization system based on Bayesian Optimization for high intensity heavy ion beams. We successfully optimized primary 36+Kr beam at 30 enA while monitoring 34+Kr by trackers in a spectrometer. Future prospects for the application of machine learning at RIBF, including these methods, will be discussed.

In general, accelerator facilities are controlled by a huge number of parameters. The RIKEN RI Beam Factory (RIBF), a heavy-ion accelerator complex consisting of several cyclotrons and Linacs, is controlled or influenced by more than 600 parameters, including environmental factors. To optimize these parameters more efficiently and accurately, we are attempting to implement Bayesian optimization (BO). Given the importance of space charge effects and beam loading, it is desirable to adjust parameters at high beam intensity, making it crucial to develop an optimization system capable of handling high-intensity heavy ion beams. We have been working on developing indices suitable for high-intensity beams and exploring methods for optimization while maintaining operational safety. So far we developed a technique that enables the simultaneous measurement of beam transmission and spot shape on the target by tracking charge-converted particles after passing through the target. Additionally, we are investigating the use of line BO with a safety function to ensure safe beam optimization. Currently, we are preparing for simulations and tests using beam line.

Long-term beam stability is one of the important issues in supplying heavy ion beams using an Electron Cyclotron Resonance Ion Source (ECRIS). When the beam intensity decreases in long-term operation, we should tune the ECRIS parameters to restore the original beam intensity. Continuous measurement of the beam intensity using a Faraday cup is impractical while the beam is in use. We have had to rely on an unreliable method of monitoring the total drain current to estimate the beam intensity during the beamtime. To resolve this issue, we propose a new method for predicting the beam intensity using machine learning. Our approach uses plasma images, captured through the beam extraction electrode, in addition to the operating parameters as input for the machine learning model. In short-term test datasets, our model has successfully predicted the beam intensity. In addition, ablation studies have shown that plasma images are essential for beam intensity prediction. This presentation will detail the prediction model and its results on the test data.