Experience integrating online modeling, adaptive digital twin infrastructure, and ML-based tuning for accelerator control at SLAC

SLAC and collaborators are developing infrastructure and algorithms for deploying online physics models, combining them with online machine learning (ML) models, and using both of these in tandem for ML-based optimization and control of accelerators. These system models can predict details of the beam phase space distribution, include nonlinear collective effects, and leverage high performance computing (HPC) and ML-based acceleration of simulations to enable execution in reasonable times for control room use. By (1) enabling accelerator system models to be adapted over time as the machine changes and (2) increasing the speed of model execution over traditional detailed physics simulations, these tools can provide useful information for both human-driven and automated tuning. We have been leveraging these system models to speed up accelerator tuning, both by providing initial guesses of settings (i.e. "warm starts") and by providing physics information to speed up adaptive, on-the-fly learning during tuning. Here, we give and overview of these developments and our experiences with deploying these tools so far. We also discuss ongoing collaborations with other accelerator laboratories in this space and describe the broader DOE context.