Machine Learning Platform: Deploying and Managing Models in the CERN Control System

Recent advances make machine learning (ML) a powerful tool to cope with the inherent complexity of accelerators, the large number of degrees of freedom and continuously drifting machine characteristics. A diverse set of ML ecosystems, frameworks and tools are already being used at CERN for a variety of use cases such as optimization, anomaly detection and forecasting. We have adopted a unified approach to model storage, versioning and deployment which accommodates this diversity, and we apply software engineering best practices to achieve the reproducibility needed in the mission-critical context of particle accelerator controls. This paper describes CERN Machine Learning Platform - our central platform for storing, versioning and deploying ML models in the CERN Control Center. We present a unified solution which allows users to create, update and deploy models with minimal effort, without constraining their workflow or restricting their choice of tools. It also provides tooling to automate seamless model updates as the machine characteristics evolve. Moreover, the system allows model developers to focus on domain-specific development by abstracting infrastructural concerns.