32a. Machine Invariant Virtual Diagnostics

Processing of particle accelerator diagnostic data can reveal significant information about the accelerator’s operation and be used to improve control of the accelerator. Machine learning offers a powerful new tool for performing better processing of diagnostic data and may help to overcome previous limitations imposed by problems such as noise and nonlinearity. However, machine learning techniques require extensive input data to be useful and are currently customized to work on individual accelerators. We will use variational autoencoders as a diagnostic tool to process data from beam position monitors along with use of self-consistency metric to pre-process input data and assess reliability. Data processed by the autoencoders can then be interpreted by a separate algorithm to extract meaningful physical data about the accelerator. In Phase I the use of variational autoencoders will be studied using simplified models of particle accelerator to generate input data. The impact of more realistic effects and noise from diagnostics will then be assessed and self-consistency metrics will be introduced to combat these problems. Finally, experimental data from existing accelerators will be used to test these new diagnostic tools. Better machine learning techniques for particle accelerators will help to improve operation and reliability of existing particle accelerators. By creating more easily transferable tools it will be possible to reduce time and effort needed to apply these techniques in a general manner. Beyond particle accelerators the use of autoencoders for data analysis and processing has broad application to many fields such as medical imaging and signal processing.