Adaptive machine learning with hard physics constraints for 6D phase space diagnostics of intense charged particle beams

Advanced particle accelerators, such as the FACET-II plasma wakefield acceleration test facility are creating intense high energy (10s of GeV) high charge electron beams. For example, a signle electron bunch in FACET-II can have up to 2 nC of charge within an incredibly short bunch length (σ_t) of 1-100 femtoseconds (1 fs = 10^-15 s) resulting in peak currents of up to 200 kA. Such highly energetic, intense, and short beams are difficult to image because they may damage intercepting diagnostics such as scintillating screens and because their lengths are exceeding the resolution of deflecting cavity-based diagnostics, such as TCAVs whose resolution is ~3 fs. Machine learning methods have shown promise towards enabling virtual diagnostics of such beams, but so far such efforts have suffered from an inability to handle the large distribution shift in large accelerator facilities whose many components are time varying, and have suffered from a lack of physics constraints. In this work, we present an adaptive machine learning approach, based on convolutional neural network-based autoencoders coupled with nonlinear model-independent adaptive feedback control algorithms, and with built in hard physics constraints, towards developing robust adaptive 6D phase space diagnostics of intense charged particle beams. We show how incorporating feedback and hard physics constraints makes our approach much more robust to distribution shift as well as much more robust for extrapolation (extrapolation is a well known challenge for ML methods which typically can only handle interpolation without re-training).