Adaptive Physics-Informed Machine Learning for Robust Beam Control and Diagnostics
Particle accelerators and their beams are uncertain time-varying systems composed out of thousands of coupled components including RF cavities and magnets which are constantly perturbed by disturbances such as temperature drifts which introduce phase shifts in RF cables, vibrations which cause a change in resonance and introduce both phase and amplitude shifts, power source ripples and hysteresis effects which perturb magnetic fields. Furthermore, components are imperfectly aligned and their field profiles are usually only estimated. Finally, accelerated beams are also uncertain and time-varying objects due to fluctuations in beam sources and complex collective effects such as space charge forces and coherent synchrotron radiation. The results of time-variation and uncertainty is that machines differ from existing models and continuously change with time forcing lengthy tuning procedures following shut downs, large variance in beam properties during operations, and lengthy tuning when making large changes in beam characteristics such as bunch lengths, bunch charge, and energy. The problem is further exacerbated by the fact that only limited non-invasive online beam diagnostics are available which typically measure only 1D or 2D projections of the beam’s 6D (x,y,z,px,py,pz) phase space. The fields of adaptive feedback control and machine learning (ML) have the potential to aid in controlling charged particle beams and in developing new non-invasive diagnostics. One of the major challenges faced by ML methods for accelerators is that the accuracy of trained ML tools quickly degrades for time-varying systems and repeatedly collecting massive new data sets for re-training is not always feasible. Combining physics-informed ML with active online feedback has potential for the development of robust adaptive non-invasive diagnostics. This talk presents recently developed adaptive machine learning (AML) methods designed for time-varying systems by combining robust model-independent feedback control with physics-informed ML tools such as generative encoder-decoder convolutional neural networks for automatic control of charged particle beam phase space , for online multi-objective optimization , for providing estimates of unknown input beam distributions  and for learning physics correlations to provide virtual 6D phase space diagnostics based on 2D measurements . It is shown that ML tools can be made to be much more robust by the addition of hard physics constraints and real-time active feedback. The talk starts with a mathematical overview of a family of model-independent adaptive feedback control techniques and then moves on to applications combining the techniques with ML tools.
 A. Scheinker, et al. "Demonstration of model-independent control of the longitudinal phase space of electron beams in the linac-coherent light source with femtosecond resolution." Physical review letters 121.4 (2018): 044801. https://doi.org/10.1103/PhysRevLett.121.044801
 A. Scheinker, et al. "Online multi-objective particle accelerator optimization of the AWAKE electron beam line for simultaneous emittance and orbit control." AIP Advances 10.5 (2020): 055320. https://doi.org/10.1063/5.0003423
 A. Scheinker, et al. "An adaptive approach to machine learning for compact particle accelerators." Scientific reports 11.1 (2021): 1-11. https://doi.org/10.1038/s41598-021-98785-0
 A. Scheinker. "Adaptive machine learning for time-varying systems: low dimensional latent space tuning." Journal of Instrumentation 16.10 (2021): P10008. https://doi.org/10.1088/1748-0221/16/10/P10008