In 2021, a collaboration of NSD scientists from the Applied Nuclear Physics, Low-Energy Nuclear Physics, and the 88” Cyclotron programs received funding through a DOE Office of Science FOA for “Data Analytics for Autonomous Optimization and Control of Accelerators and Detectors” to explore application of machine learning approaches to the operations of both the VENUS ion source [1] and the GRETA γ-ray detector array [2].
The primary goal of this project is to explore the best applications of existing machine learning approaches and techniques to optimize and automate operations of both devices. Work has primarily been focused so far on the case of VENUS. The team has established a continuous logging system that has been used to populate a time-series database with relevant control and diagnostic parameters from the operation of the source for over a year. The researchers, supported by a group of undergraduate researchers, are exploring the use of predictive neural networks and Bayesian optimization of the parameter space to tune the ion source.
By providing training data, such as the parameter values that VENUS was operating at and the corresponding beam current produced, the team has begun to develop a neural network which is intended to predict the current for other parameter values and use it to optimize the source. Early indications, however, suggest that in order to do this successfully, the range of input parameters must be extended and the researchers are therefore installing environmental sensors at the ion source in order to measure additional parameters such as temperature and air pressure.
In contrast, the use of Gaussian Process Regression (GPR) to map the parameter space and guide optimization of the VENUS tuning procedure has been very successful. Using this approach, a probability map of the full parameter space is generated from just a few measurements at different parameter values. The probabilistic map can then be used to guess parameter values which are likely to provide a better beam current. By performing the procedure a few dozen times an optimal solution is iteratively approached within a few hours, a similar timescale that would be expected from a human operator. The GPR approach has been shown to robustly produce a maximized beam current within a limited parameter space. Fig. 1 shows the result of using GPR to explore a two-dimensional parameter space during a test run with VENUS. The black circles represent measured observables, while the smooth distribution in the background shows the predictions from the GPR. Work is ongoing to extend this effort and potentially identify useful operational settings for VENUS which have not been considered by human operators.
This project is led by Damon Todd from the Low Energy Nuclear Physics program, Marco Salathe from the Applied Nuclear Physics program and Heather Crawford, from the Low Energy Nuclear Physics program.
[1] Recent progress on the superconducting ion source VENUS, J.Y. Benitez et al., Review of Scientific Instruments 83, 02A311 (2012), DOI: 10.1063/1.3662119
[2] The GRETA Final Design Report, 2020. Available online: GRETA_FDR.pdf