Operators of electron microscopes, X-ray lasers, medical accelerators and other devices that are powered by beams of accelerated electrons must analyze the quality of the electron beams and make necessary adjustments to optimize the instrument’s performance. Typically, this involves a physical device that requires interaction with the beam in order to measure its shape, intensity and other properties.
Conventional beam diagnostics often destroy or alter the beam, or require its deflection, so they cannot be used along with the actual application. In addition, some situations prevent accurate measurement due to technical limitations. This can be the case when the beam’s electron pulses are fired at a very high rate or are very intense.
To address this problem and remove the limitations of a physical device, researchers at the Department of 
Artistic representation of a neural network superimposed on an electron beam profile. Source: Adi Hanuka/SLAC National Accelerator LaboratoryEnergy’s (DOE's) SLAC National Accelerator Laboratory are advancing a machine learning solution to create a virtual diagnostic approach. Crucial information about beam quality can be obtained in a non-invasive yet efficient way.
"Our method can be used to diagnose virtually any machine that uses electron beams, whether it's an electron microscope for imaging of ultrasmall objects or a medical accelerator used in cancer therapy," said SLAC research associate Adi Hanuka, who led the study.
The new method uses a machine learning algorithm inspired by the brain’s neural network. Once trained with data from the lab’s particle accelerators, the algorithm could accurately predict beam properties for experimental situations.
Predictions generated by the diagnostic tool were compared with experimental and simulated data for the electron beams of three DOE Office of Science user facilities at SLAC: the Linac Coherent Light Source (LCLS) X-ray laser, its future upgrade LCLS-II, and the Facility for Advanced Accelerator Experimental Tests (FACET-II).
The results showed that virtual diagnostics using the machine-learning approach are effective in cases where conventional tools are not. The neural network can provide detailed information about each of the million electron pulses per second LCLS-II produces and can provide accurate information about the high-intensity beam of FACET-II — both of which cannot be analyzed with physical devices.
The research is published in Scientific Reports.
