From Notebook to Control Room: Reproducible MLOps for Accelerator Digital Twins

Scientific user facilities like the Linac Coherent Light Source (LCLS), the world’s brightest X-ray free-electron laser at the SLAC National Accelerator Laboratory, require rapid, continuous control adjustments to maximize beam time and experimental throughput. Machine Learning (ML) surrogate models can help operators navigate complex parameter spaces, but only if those models can be deployed reliably and updated without disrupting operations. How do you reliably and sustainably move ML models from various researchers’ training environments into a facility production system? In this talk, we present a reproducible MLOps framework that addresses this challenge in the context of digital twin deployments for online accelerator control. We discuss how we standardized the Kubernetes-based deployment of digital twins and ML inference services by prioritizing reproducibility across environments through container-based packaging, rigorous model versioning and experiment tracking integrated into the deployment pipeline, minimal boilerplate templates so domain scientists can onboard models without DevOps expertise, and flexible configuration to adapt across diverse control scenarios. The framework is built around a software ecosystem that provides abstraction layers to wrap simulator-specific backends (PyTorch-based neural network surrogates and physics simulations) behind a common Python API. This design means the deployment infrastructure generalizes across model types rather than being tied to any single simulation code or modeling approach. The same model instance can be used directly in Python-native workflows for development and testing, or deployed as a process variable server integrated with existing control system infrastructure, without changes to the underlying model code. This approach demonstrates how such software infrastructure can lower the barrier to AI integration in research facilities, allowing scientists to focus on model development while RSEs maintain the reliability, observability, and sustainability of the deployment stack. We discuss lessons learned, remaining challenges around hybrid ML-simulation system models, and paths toward tooling that is shareable across facilities.