SLAC and collaborators are developing infrastructure and algorithms for deploying online physics models and combining them with machine learning (ML) models and ML-based feedback from its running accelerators. These models predict details of the beam phase space distribution, include nonlinear collective effects, and leverage high performance computing and ML-based acceleration of simulations to enable execution in reasonable times for control room use. By enabling accelerator system models to be adapted over time and increasing the speed of model execution, these system models can provide useful information for both human-driven and automated tuning. System models such as these are sometimes called "digital twins", which are distinguished from offline models by the bi-directional flow of information with the real system. We have also been leveraging these system models to speed up accelerator tuning, by providing initial guesses of settings (i.e. "warm starts") and physics information to speed up ML-based tuning. For example, we have used these models to provide priors for Bayesian optimization and training platforms for reinforcement learning. Here, we give and overview of these developments (both research and infrastructure), our deployment experience, and applications at LCLS, LCLS-II, and FACET-II, with a focus on emittance tuning, FEL pulse intensity tuning, and phase space shaping. We also discuss ongoing collaborations with LBNL, JLAB, FNAL, and BNL in this space.

Differentiable simulations are in demand in accelerator physics, demonstrating order-of-magnitude improvements for complex tasks such as many-parameter optimization for accelerator working points and reconstruction of hard-to-measure quantities. At its core, a differentiable simulation does not only solve a forward problem, but additionally provides gradients of output parameters (e.g. beam parameters) with respect to input parameters (e.g. beamline or source parameters). How to effectively program large dynamic simulations differentiably is still an open question, but there is general consensus that a “single-source” approach aided by automatic differentiation (AD) is desirable. Addressing this, there are a) emerging domain-specific languages in machine learning that are intrinsically differentiable, and b) highly-performing & scalable, general-purpose languages like ISO C++ of existing codes. The challenge of approach a) is syntax specialization, which can limit ease of implementation & performance for physics algorithms, while b) requires additional work for AD. Performance is important for modeling high-order beam dynamics and collective effects in accelerators. We compare the fast, modern codes ImpactX (C++/Python) and Cheetah (PyTorch) using traditional, gradient-free modeling. We then show progress in introducing single-source differentiability in ImpactX using modern compiler techniques, producing performant executables for gradient-based and gradient-free modeling.

Structured plasma channels are an essential technology for driving high-gradient, plasma-based acceleration and control of electron and positron beams for advanced concepts accelerators. Laser and gas technologies can permit the generation of long plasma columns known as hydrodynamic, optically-field-ionized (HOFI) channels, which feature low on-axis densities and steep walls. By carefully selecting the background gas and laser properties, one can generate narrow, tunable plasma channels for guiding high intensity laser pulses. We present on the development of 1D and 2D simulations of HOFI channels using the FLASH code, a publicly available radiation hydrodynamics code. We explore sensitivities of the channel evolution to laser profile, intensity, and background gas conditions. We examine experimental measurements of plasma channels and their comparison to model predictions. Lastly, we discuss ongoing work to couple these tools to community PIC models to capture variations in initial conditions and channel influence on wakefield accelerator applications.